Author response:

The following is the authors’ response to the original reviews.

Reviewer #2 (Public Review):

The authors make a compelling case for the biological need to exquisitely control RecB levels, which they suggest is achieved by the pathway they have uncovered and described in this work. However, this conclusion is largely inferred as the authors only investigate the effect on cell survival in response to (high levels of) DNA damage and in response to two perturbations - genetic knock-out or over-expression, both of which are likely more dramatic than the range of expression levels observed in unstimulated and DNA damage conditions.

In the discussion of the updated version of the manuscript, we have clarified the limits of our interpretation of the role of the uncovered regulation.

Lines 411-417: “It is worth noting that the observed decrease in cell viability upon DNA damage was detected for relatively drastic perturbations such as recB deletion and RecBCD overexpression. Verifying these observations in the context of more subtle changes in RecB levels would be important for further investigation of the biological role of the uncovered regulation mechanism. However, the extremely low numbers of RecB proteins make altering its abundance in a refined, controlled, and homogeneous across cells manner extremely challenging and would require the development of novel synthetic biology tools.”

Reviewer #3 (Public Review):

The major weaknesses include a lack of mechanistic depth, and part of the conclusions are not fully supported by the data.

(1) Mechanistically, it is still unclear why upon DNA damage, translation level of recB mRNA increases, which makes the story less complete. The authors mention in the Discussion that a moderate (30%) decrease in Hfq protein was observed in previous study, which may explain the loss of translation repression on recB. However, given that this mRNA exists in very low copy number (a few per cell) and that Hfq copy number is on the order of a few hundred to a few thousand, it's unclear how 30% decrease in the protein level should resides a significant change in its regulation of recB mRNA.

We agree that the entire mechanistic pathway controlling recB expression may be not limited to just Hfq involvement. We have performed additional experiments, proposed by the reviewer, suggesting that a small RNA might be involved (see below, response to comments 3&4). However, we consider that the full characterisation of all players is beyond the scope of this manuscript. In addition to describing the new data (see below), we expanded the discussion to explain more precisely why changes in Hfq abundance upon DNA damage may impact RecB translation. 

Lines 384-391: “A modest decrease (~30%) in Hfq protein abundance has been seen in a proteomic study in E. coli upon DSB induction with ciprofloxacin (DOI: 10.1016/j.jprot.2018.03.002). While Hfq is a highly abundant protein, it has many mRNA and sRNA targets, some of which are also present in large amounts (DOI: 10.1046/j.1365-2958.2003.03734.x). As recently shown, the competition among the targets over Hfq proteins results in unequal (across various targets) outcomes, where the targets with higher Hfq binding affinity have an advantage over the ones with less efficient binding (DOI: 10.1016/j.celrep.2020.02.016). In line with these findings, it is conceivable that even modest changes in Hfq availability could result in significant changes in gene expression, and this could explain the increased translational efficiency of RecB under DNA damage conditions. “

(2) Based on the experiment and the model, Hfq regulates translation of recB gene through binding to the RBS of the upstream ptrA gene through translation coupling. In this case, one would expect that the behavior of ptrA gene expression and its response to Hfq regulation would be quite similar to recB. Performing the same measurement on ptrA gene expression in the presence and absence of Hfq would strengthen the conclusion and model.

Indeed, based on our model, we expect PtrA expression to be regulated by Hfq in a similar manner to RecB. However, the product encoded by the ptrA gene, Protease III, (i) has been poorly characterised; (ii) unlike RecB, is located in the periplasm (DOI: 10.1128/jb.149.3.1027-1033.1982); and (iii) is not involved in any DNA repair pathway. Therefore, analysing PtrA expression would take us away from the key questions of our study.

(3) The authors agree that they cannot exclude the possibility of sRNA being involved in the translation regulation. However, this can be tested by performing the imaging experiments in the presence of Hfq proximal face mutations, which largely disrupt binding of sRNAs.

(4) The data on construct with a long region of Hfq binding site on recB mRNA deleted is less convincing. There is no control to show that removing this sequence region itself has no effect on translation, and the effect is solely due to the lack of Hfq binding. A better experiment would be using a Hfq distal face mutant that is deficient in binding to the ARN motifs.

We performed the requested experiments. We included this data in the manuscript in the supplementary figure (Figure S11), and our interpretation in the discussion.

Lines 354-378: “While a few recent studies have shown evidence for direct gene regulation by Hfq in a sRNA-independent manner (DOI: 10.1101/gad.302547.117; DOI: 10.1111/mmi.14799; DOI: 10.1371/journal.pgen.1004440; DOI: 10.1111/mmi.12961; DOI: 10.1038/emboj.2013.205), we attempted to investigate whether a small RNA could be involved in the Hfq-mediated regulation of RecB expression. We tested Hfq mutants containing point mutations in the proximal and distal sides of the protein, which were shown to disrupt either binding with sRNAs or with ARN motifs of mRNA targets, respectively [DOI: 10.1016/j.jmb.2013.01.006, DOI: 10.3389/fcimb.2023.1282258]. Hfq mutated in either proximal (K56A) or distal (Y25D) faces were expressed from a plasmid in a ∆hfq background. In both cases, Hfq expression was confirmed with qPCR and did not affect recB mRNA levels (Supplementary Figure S11b). When the proximal Hfq binding side (K56A) was disrupted, RecB protein concentration was nearly similar to that obtained in a ∆hfq mutant (Supplementary Figure S11a, top panel). This observation suggests that the repression of RecB translation requires the proximal side of Hfq, and that a small RNA is likely to be involved as small RNAs (Class I and Class II) were shown to predominantly interact with the proximal face of Hfq [DOI: 10.15252/embj.201591569]. When we expressed Hfq mutated in the distal face (Y25D) which is deficient in binding to mRNAs, less efficient repression of RecB translation was detected (Supplementary Figure S11a, bottom panel). This suggests that RecB mRNA interacts with Hfq at this position. We did not observe full de-repression to the ∆hfq level, which might be explained by residual capacity of Hfq to bind its recB mRNA target in the point mutant (Y25D) (either via the distal face with less affinity or via the lateral rim Hfq interface).”

Taken together, these results suggest that Hfq binds to recB mRNA and that a small RNA might contribute to the regulation although this sRNA has not been identified.

(5) Ln 249-251: The authors claim that the stability of recB mRNA is not changed in ∆hfq simply based on the steady-state mRNA level. To claim so, the lifetime needs to be measured in the absence of Hfq.

We measured recB lifetime in the absence of Hfq in a time-course experiment where transcription initiation was inhibited with rifampicin and mRNA abundance was quantified with RT-qPCR. The results confirmed that recB mRNA lifetime in hfq mutants is similar to the one in the wild type (Figure S7d, referred to the line 263 of the manuscript).

(6) What's the labeling efficiency of Halo-tag? If not 100% labeled, is it considered in the protein number quantification? Is the protein copy number quantification through imaging calibrated by an independent method? Does Halo tag affect the protein translation or degradation?

Our previous study (DOI: 10.1038/s41598-019-44278-0) described a detailed characterization of the HaloTag labelling technique for quantifying low-copy proteins in single E. coli cells using RecB as a test case. 

In that study, we showed complete quantitative agreement of RecB quantification between two fully independent methods: HaloTag-based labelling with cell fixation and RecB-sfGFP combined with a microfluidic device that lowers protein diffusion in the bacterial cytoplasm. This second method had previously been validated for protein quantification (DOI: 10.1038/ncomms11641) and provides detection of 80-90% of the labelled protein. Additionally, in our protocol, immediate chemical fixation of cells after the labelling and quick washing steps ensure that new, unlabelled RecB proteins are not produced. We, therefore, conclude that our approach to RecB detection is highly reliable and sufficient for comparing RecB production in different conditions and mutants.

The RecB-HaloTag construct has been designed for minimal impact on RecB production and function. The HaloTag is translationally fused to RecB in a loop positioned after the serine present at position 47 where it is unlikely to interfere with (i) the formation of RecBCD complex (based on RecBCD structure, DOI: 10.1038/nature02988), (ii) the initiation of translation (as it is far away from the 5’UTR and the beginning of the open reading frame) and (iii) conventional C-terminalassociated mechanisms of protein degradation (DOI: 10.15252/msb.20199208). In our manuscript, we showed that the RecB-HaloTag degradation rate is similar to the dilution rate due to bacterial growth. This is in line with a recent study on unlabelled proteins, which shows that RecB’s lifetime is set by the cellular growth rate (DOI: 10.1101/2022.08.01.502339).

Furthermore, we have demonstrated (DOI: 10.1038/s41598-019-44278-0) that (i) bacterial growth is not affected by replacing the native RecB with RecB-HaloTag, (ii) RecB-HaloTag is fully functional upon DNA damage, and (iii) no proteolytic processing of the RecB-HaloTag is detected by Western blot. 

These results suggest that RecB expression and functionality are unlikely to be affected by the translational HaloTag insertion at Ser-47 in RecB.

In the revised version of the manuscript, we have added information about the construct and discuss the reliability of the quantification.

Lines 141-152: “To determine whether the mRNA fluctuations we observed are transmitted to the protein level, we quantified RecB protein abundance with singlemolecule accuracy in fixed individual cells using the Halo self-labelling tag (Fig. 2A&B).

The HaloTag is translationally fused to RecB in a loop after Ser47(DOI: 10.1038/s41598-019-44278-0) where it is unlikely to interfere with the formation of RecBCD complex (DOI: 10.1038/nature02988), the initiation of translation and conventional C-terminal-associated mechanisms of protein degradation (DOI: 10.15252/msb.20199208). Consistent with minimal impact on RecB production and function, bacterial growth was not affected by replacing the native RecB with RecBHaloTag, the fusion was fully functional upon DNA damage and no proteolytic processing of the construct was detected (DOI: 10.1038/s41598-019-44278-0). To ensure reliable quantification in bacteria with HaloTag labelling, the technique was previously verified with an independent imaging method and resulted in > 80% labelling efficiency (DOI: 10.1038/s41598-019-44278-0, DOI: 10.1038/ncomms11641). In order to minimize the number of newly produced unlabelled RecB proteins, labelling and quick washing steps were followed by immediate chemical fixation of cells.”

Lines 164-168: “Comparison to the population growth rate [in these conditions (0.017 1/min)] suggests that RecB protein is stable and effectively removed only as a result of dilution and molecule partitioning between daughter cells. This result is consistent with a recent high-throughput study on protein turnover rates in E. coli, where the lifetime of RecB proteins was shown to be set by the doubling time (DOI: 10.1038/s41467-024-49920-8).”

(7) Upper panel of Fig S8a is redundant as in Fig 5B. Seems that Fig S8d is not described in the text.

We have now stated in the legend of Fig S8a that the data in the upper panel were taken from Fig 5B to visually facilitate the comparison with the results given in the lower panel. We also noticed that we did not specify that in the upper panel in Fig S9a (the data in the upper panel of Fig S9a was taken from Fig 5C for the same reason). We added this clarification to the legend of the Fig S9 as well.

We referred to the Fig S8d in the main text. 

Lines 283-284: “We confirmed the functionality of the Hfq protein expressed from the pQE-Hfq plasmid in our experimental conditions (Fig. S8d).”

Reviewer #1 (Recommendations For The Authors):

(1) Experimental regime to measure protein and mRNA levels.

(a) Authors expose cells to ciprofloxacin for 2 hrs. They provide a justification via a mathematical model. However, in the absence of a measurement of protein and mRNA across time, it is unclear whether this single time point is sufficient to make the conclusion on RecB induction under double-strand break.

In our experiments, we only aimed to compare recB mRNA and RecB protein levels in two steady-state conditions: no DNA damage and DNA damage caused by sublethal levels of ciprofloxacin. We did not aim to look at RecB dynamic regulation from nondamaged to damaged conditions – this would indeed require additional measurements at different time points. We revised this part of the results to ensure that our conclusions are stated as steady-state measurements and not as dynamic changes.

Line 203-205: “We used mathematical modelling to verify that two hours of antibiotic exposure was sufficient to detect changes in mRNA and protein levels and for RecB mRNA and protein levels to reach a new steady state in the presence of DNA damage.”

(b) Authors use cell area to account for the elongation under damage conditions. However, it is unclear whether the number of copies of the recB gene are similar across these elongated cells. Hence, authors should report mRNA and protein levels with respect to the number of gene copies of RecB or chromosome number as well.

Based on the experiments in DNA damaging conditions, our main conclusion is that the average translational efficiency of RecB is increased in perturbed conditions. We believe that this conclusion is well supported by our measurements and that it does not require information about the copy number of the recB gene but only the concentration of mRNA and protein. We did observe lower recB mRNA concentration upon DNA damage in comparison to the untreated conditions, which may be due to a lower concentration of genomic DNA in elongated cells upon DNA damage, as we mention in lines (221-223).

Our calculation of translation efficiency could be affected by variations of mRNA concentration across cells in the dataset. For example, longer cells that are potentially more affected by DNA damage could have lower concentrations of mRNA. We verified that this is not the case, as recB mRNA concentration is constant across cell size distribution (see the figure below or Figure S5a from Supplementary Information).

Therefore, we do not think that the measurements of recB gene copy would change our conclusions. We agree that measuring recB gene copies could help to investigate the reason behind the lower recB mRNA concentration under the perturbed conditions as this could be due to lower DNA content or due to shortage of resources (such as RNA polymerases). However, this is a side observation we made rather than a critical result, whose investigation is beyond the scope of this manuscript.

Author response image 1.

(2) RecB as a proxy for RecBCD. Authors suggest that RecB levels are regulated by hfq. However, how does this regulatory circuit affect the levels of RecC and RecD? Ratio of the three proteins has been shown to be important for the function of the complex.

A full discussion of RecBCD complex formation regulation would require a complete quantitative model based on precise information on the dynamic of the complex formation, which is currently lacking. 

We can however offer the following (speculative) suggestions assuming that all three subunits are present in similar abundance in native conditions (DOI: 10.1038/s41598019-44278-0 for RecB and RecC). As the complex is formed in 1:1:1 ratio (DOI: 10.1038/nature02988), we propose that the regulation mechanism of RecB expression affects complex formation in the following way. If the RecB abundance becomes lower than the level of RecC and RecD subunits, the complex formation would be limited by the number of available RecB subunits and hence the number of functional RecBCDs will be decreased. On the contrary, if the number of RecB is higher than the baseline, then, especially in the context of low numbers, we would expect that the probability of forming a complex RecBC (and then RecBCD) will be increased. Based on this simple explanation, we might speculate that regulation of RecB expression may be sufficient to regulate RecB levels and RecBCD complex formation. However, we feel that this argument is too speculative to be added to the manuscript. 

(3) Role of Hfq in RecB regulation. While authors show the role of hfq in recB translation regulation in non-damage conditions, it is unclear as to how this regulation occurs under damage conditions.

(a) Have the author carried out recB mRNA and protein measurement in hfqdeleted cells under ciprofloxacin treatment?

We attempted to perform experiments in hfq mutants under ciprofloxacin treatment. However, the cells exhibited a very strong and pleiotropic phenotype: they had large size variability and shape changes and were also frequently lysing. Therefore, we did not proceed with mRNA and protein quantification because the data would not have been reliable. 

(b) How do the authors propose that Hfq regulation is alleviated under conditions of DNA damage, when RecB translation efficiency increases?

We propose that Hfq could be involved in a more global response to DNA damage as follows. 

Based on a proteomic study where Hfq protein abundance has been found to decrease (~ 30%) upon DSB induction with ciprofloxacin (DOI: 10.1016/j.jprot.2018.03.002), we suggest that this could explain the increased translational efficiency of RecB. While Hfq is a highly abundant protein, it has many targets (mRNA and sRNA), some of which are also highly abundant. Therefore the competition among the targets over Hfq proteins results in unequal (across various targets) outcomes (DOI: 10.1046/j.13652958.2003.03734.x), where the targets with higher Hfq binding affinity have an advantage over the ones with less efficient binding. We reason that upon DNA damage, a moderate decrease in the Hfq protein abundance (30%) can lead to a similar competition among Hfq targets where high-affinity targets outcompete low-affinity ones as well as low-abundant ones (such as recB mRNAs). Thus, the regulation of lowabundant targets of Hfq by moderate perturbations of Hfq protein level is a potential explanation for the change in RecB translation that we have observed. Potential reasons behind the changes of Hfq levels upon DNA damage would be interesting to explore, however this would require a completely different approach and is beyond the scope of this manuscript.

We have modified the text of the discussion to explain our reasoning:

Lines 384-391: “A modest decrease (~30%) in Hfq protein abundance has been seen in a proteomic study in E. coli upon DSB induction with ciprofloxacin (DOI: 10.1016/j.jprot.2018.03.002). While Hfq is a highly abundant protein, it has many mRNA and sRNA targets, some of which are also present in large amounts (DOI: 10.1046/j.1365-2958.2003.03734.x). As recently shown, the competition among the targets over Hfq proteins results in unequal (across various targets) outcomes, where the targets with higher Hfq binding affinity have an advantage over the ones with less efficient binding (DOI: 10.1016/j.celrep.2020.02.016). In line with these findings, it is conceivable that even modest changes in Hfq availability could result in significant changes in gene expression, and this could explain the increased translational efficiency of RecB under DNA damage conditions.”

(c) Is there any growth phenotype associated with recB mutant where hfq binding is disrupted in damage and non-damage conditions? Does this mutation affect cell viability when over-expressed or under conditions of ciprofloxacin exposure?

We checked the phenotype and did not detect any difference in growth or cell viability affecting the recB-5 UTR* mutants either in normal conditions or upon exposure to ciprofloxacin. However, this is expected because the repair capacity is associated with RecB protein abundance and in this mutant, while translational efficiency of recB mRNA increases, the level of RecB proteins remains similar to the wild-type (Figure 5E).

Minor points:

(1) Introduction - authors should also discuss the role of RecFOR at sites of fork stalling, a likely predominant pathway for break generated at such sites.

The manuscript focuses on the repair of DNA double-strand breaks (DSBs). RecFOR plays a very important role in the repair of stalled forks because of single-strand gaps but is not involved in the repair of DSBs (DOI: 10.1038/35003501). We have modified the beginning of the introduction to mention the role of RecFOR. 

Lines 35-39: “For instance, replication forks often encounter obstacles leading to fork reversal, accumulation of gaps that are repaired by the RecFOR pathway (DOI: 10.1038/35003501) or breakage which has been shown to result in spontaneous DSBs in 18% of wild-type Escherichia coli cells in each generation (DOI: 10.1371/journal.pgen.1007256), underscoring the crucial need to repair these breaks to ensure faithful DNA replication.”

(2) Methods: The authors refer to previous papers for the method used for single RNA molecule detection. More information needs to be provided in the present manuscript to explain how single molecule detection was achieved.

We added additional information in the method section on the fitting procedure allowing quantifying the number of mRNAs per detected focus.

Lines 515-530: “Based on the peak height and spot intensity, computed from the fitting output, the specific signal was separated from false positive spots (Fig. S1a). To identify the number of co-localized mRNAs, the integrated spot intensity profile was analyzed as previously described (DOI: 10.1038/nprot.2013.066). Assuming that (i) probe hybridization is a probabilistic process, (ii) binding each RNA FISH probe happens independently, and (iii) in the majority of cases, due to low-abundance, there is one mRNA per spot, it is expected that the integrated intensities of FISH probes bound to one mRNA are Gaussian distributed. In the case of two co-localized mRNAs, there are two independent binding processes and, therefore, a wider Gaussian distribution with twice higher mean and twice larger variance is expected. In fact, the integrated spot intensity profile had a main mode corresponding to a single mRNA per focus, and a second one representing a population of spots with two co-localized mRNAs (Fig. S1b). Based on this model, the integrated spot intensity histograms were fitted to the sum of two Gaussian distributions (see equation below where a, b, c, and d are the fitting parameters), corresponding to one and two mRNA molecules per focus. An intensity equivalent corresponding to the integrated intensity of FISH probes in average bound to one mRNA was computed as a result of multiple-Gaussian fitting procedure (Fig. S1b), and all identified spots were normalized by the one-mRNA equivalent.

“

Reviewer #2 (Recommendations For The Authors):

Overall the work is carefully executed and highly compelling, providing strong support for the conclusions put forth by the authors.

One point: the potential biological consequences of the post-transcriptional mechanism uncovered in the work would be enhanced if the authors could 1) tune RecB protein levels and 2) directly monitor the role that RecB plays in generating single-standed DNA at DSBs.

We agree that testing viability of cells in case of tunable changes in RecB levels would be important to further investigate the biological role of the uncovered regulation mechanism. However, this is a very challenging experiment as it is technically difficult to alter the low number of RecB proteins in a controlled and homogeneous across-cell manner, and it would require the development of precisely tunable and very lowabundant synthetic designs. 

We did monitor real-time RecB dynamics by tracking single molecules in live E. coli cells in a different study (DOI: 10.1101/2023.12.22.573010) that is currently under revision. There, reduced motility of RecB proteins was observed upon DSB induction indicating that RecB is recruited to DNA to start the repair process.

Author Response:

The following is the authors' response to the original reviews.

Thank you for considering our manuscript “An Unexpected Role of Neutrophils in Clearing Apoptotic Hepatocytes In Vivo". We also thank the referees for their review. We have addressed their comments in detail and added new data to buttress our conclusions.

Reviewer #1 (Public Review):

This study by Cao et al. demonstrates role of Neutrophil in clearing apoptotic hepatocytes by directly burrowing into the apoptotic hepatocytes and ingesting the effete cells from inside without causing inflammation. The authors applied intravital microscopy, Immunostaining and electron microscopy to visualize perforocytosis of neutrophil in hepatocytes. They also found that neutrophil depletion impairs the clearance of apoptotic hepatocytes causing impaired liver function and generation of autoantibodies, implying a role of defective neutrophil- mediated clearance of apoptotic cells in Autoimmune Liver disease. The experiments were well designed and conducted, the results were reasonably interpreted, and the manuscript was clearly written with logical inputs.

Thank you for your comments.

One weak point is that the signals/mechanisms that determine why neutrophil specifically target apoptotic hepatocytes in liver and no other organs or cells is not clearly understood.

We are still studying why neutrophils selectively phagocytose hepatocytes but not HUVEC or 293 cells. We have some intriguing preliminary data so far showing that apoptotic 293 cells have no significant increase of IL-1β production as compared with their nonapoptotic controls; both apoptotic 293 cells and HUVECs do not have increased surface selectin proteins (new Fig. S3C).

Reviewer #2 (Public Review):

[…] By examination of HE-stained, noncancerous liver tissue sections from patients with hepatocellular carcinoma and hepatic hemangioma, the authors observed that cells with neutrophil nuclear morphology were inside apoptotic hepatocytes. The authors also further characterized this observation by staining the sections with neutrophil and apoptosis markers. In addition, the authors observed the same phenomena in mouse livers using intravital microscopy, which also recorded the time course of the disappearance of a neutrophil-associated apoptotic cell. The author went on further characterization of neutrophil-mediated efferocytosis of cultured hepatic cells in vitro and demonstrated the process was specific for apoptotic hepatic cells, but not HEK293 or endothelial cells. The in vitro system was then used to characterize the molecular bases for neutrophil-mediated efferocytosis of apoptotic hepatic cells. The evidence was provided to suggest that IL1b and IL-8 released from and selectins upregulated in apoptotic hepatic cells were important. Importantly, the authors used two methods to deplete the neutrophils and showed that the neutrophil depletion increased apoptotic cells in livers. Finally, the authors showed that neutrophil depletion caused defects in liver function parameters. At the end, the authors presented evidence to suggest that AIL disease may be due to defective neutrophils that fail to perform "perforocytosis."

Thank you for your comments.

Point #1. Although the evidence in its totality indicates that neutrophils burrow into apoptotic hepatocytes, the significance of this "perforocytosis" phenomenon and the circumstances under which it may occur remain to be better defined. In both neutrophil depletion models, the TNUEL-positive cells were not definitively identified rather than assuming they were hepatocytes.

Anatomically, the apoptotic hepatocytes are randomly distributed in the hepatic plate from the central vein to the portal region (please refer to the image below: hematoxylin staining of liver tissues, black arrowhead indicates perforocytosis sites).

Author response image 1.

Histologically, the structure of liver/hepatic lobe are well defined, and the cell types in the livers are easy to histologically identify based on their location, morphology and the relationship to hepatic plate and sinusoid. In addition, the hepatocytes are well known for its rich cytoplasmic components, cellular connection and prominent large round nucleus. Thus, hepatocytes are very easy to identify even without using specific molecular markers such as E-cadherin or albumin. Based on these characteristics, the TUNEL positive cells that we displayed in Fig. 5A are apoptotic hepatocytes.

Point #2. In addition, there are discrepancies in the number of neutrophils and apoptotic cells in mouse liver studies; Fig. 2a WT (many neutrophils; locations unclear) vs Fig. 5A Ctr (a few neutrophils that appear in or near a vessel), and Fig. 2a DTR (a few apoptotic cells) vs Fig. 5A Depletion (many apoptotic cells).

In response, Fig. 2A demonstrates a larger area of the mouse liver (bar, 100 µm), while Fig. 5A exhibits a relatively small area of the liver sample (bars, 20 µm for Ctrl and 15 µm for DTR). Similarly, apoptotic cells in Fig. 2A DTR need to zoom in to quantify. We apologize for the confusion, and we did quantify the apoptotic cells in Fig.2A WT vs DTR (see the bar graph next to the images in Fig. 2A).

Point #3. Importantly, Fig 5a Ctrl, which is presumably a section from a mouse without any surgical treatment or without inflammation, the sole TUNNEL signal does not appear to be associated with neutrophils. Does this mean that "perforocytosis" primarily occurs in inflamed livers (Of note, human liver samples in Fig 1 are from patient with tumors. There should be inflammation in the livers of these patients).

In Fig 5A Ctrl, the TUNEL signal indicates apoptotic hepatocytes. The neutrophils (stained with anti-NE antibody, red) are associated with the apoptotic hepatocyte (Fig. 5A). We observed that perforocytosis primarily occurs in normal noninflamed livers.

Human liver samples in Fig 1 are from patient with tumors, hence it is possible that neutrophil burrowing is somehow associated with cancerous/inflammatory livers as the reviewer pointed out. This possibility was ruled out based on our method of sample preparation and experimental results themselves.

1) Both noncancerous and cancerous liver samples were sliced based on the anatomical appearance of normal and cancer tissues (differences were rather easy to identify, and these samples were prepared by highly experienced pathologists from the Liver Cancer Center of Zhongshan Hospital, Shanghai). Furthermore, the results were confirmed by determining whether the surrounding tissue contained microlesions characteristic of metastatic tumors. We only counted apoptotic hepatocytes in noncancerous regions having normal liver lobes and morphologically normal hepatocytes, plates, sinusoid and Kupffer cells. We also excluded hepatoma, chronic inflammatory regions, and necrotic regions.

2) We did not observe recruitment of neutrophils into apoptotic HCC cells, indicating that the clearance of apoptotic cancer cells was not mediated by neutrophils (unpublished observations).

3) It is hard for us to obtain normal human liver samples; however, we did study samples from patients with liver hemangioma characterized by aberrant vasculature in livers but with normal liver functions and the structure of hemangioma livers that we analyzed are nearly identical to a healthy liver in histology (these liver samples contained no cancerous regions and there was no apparent cirrhosis or inflammation). And here we obtained similar results (these are shown in Fig. 1B; a total of 40 apoptotic hepatocytes were examined).

4) Our data from normal mouse livers, isolated primary cells (hepatocytes and neutrophils) and cell lines (NCTC and HL60) all confirmed the central findings in this paper (Fig. 2, 3).

Point #4. The data on human AIL patient neutrophils raises more questions: how many AIL patients have been examined? Do these AIL neutrophils lack IL1, IL8 receptors, and/or selectin ligands? Are there increases in apoptotic hepatocytes in AIL patients?

In response, we have analyzed 16 AIL patient samples (see table below).

Author response table 1.

We performed microarray assay to screen the differential gene expression of neutrophils from normal and liver autoimmune patients. We have identified that IL-1β receptor, IL1R1 and selectin binding protein, P- selectin glycoprotein ligand 1 (PSGL-1) were all decreased in neutrophils from the AIL patients (new Fig 7D). These findings are consistent with our observations using cells and mouse models.

Point #5. Additionally, the overall numbers of apoptotic cells even in the absence of neutrophils are rare; thus, it is questionable that such rarity of apoptotic cells can cause significant AIL phenotypes.

We quantified apoptotic liver cells in percentages instead of overall numbers (Fig. 5, we were not able to precisely calculate the overall numbers, which could be large since billions of cells undergoing apoptosis daily). Depletion of neutrophils increased the percentage of apoptotic cells about 5-6-fold in livers, and we observed the generation of autoantibodies (Fig. 6).

Reviewer #1 (Recommendations For The Authors):

This study by Cao et al. was well designed and conducted, the results were reasonably interpreted, and the manuscript was clearly written with logical inputs.

It would further gain the significance of this study if authors could address the following questions:

1.  What are the mechanisms/ signals that prevents AIL Liver neutrophils from burrowing into hepatocytes?

We have identified that IL-1β receptor, IL1R1 and selectin binding protein, P-selectin glycoprotein ligand 1 (PSGL-1) were all decreased in neutrophils from the AIL patients (new Fig 7D).

2.  Have authors looked if autoantigens expressed on hepatocytes, which are often found in autoimmune liver disease trigger signaling events that activate neutrophils to burrow?

Thank you for the comment, we have not examined autoantigens expressed in hepatocytes and plan to carry out this research as suggested.

3.  Is perforocytosis observed in apoptotic hepatocytes induced by different agents like LPS, TNF-a , rapamycin, alcohol etc?

We did not observe perforocytosis in LPS or TNF-a treated hepatocytes. One possible reason is that LPS or TNF-a we used induced massive necrosis instead of apoptosis. Howere, we did observe neutrophil perforocytosis in FasL-induced apoptotic hepatocytes (unpublished observations).

Reviewer #2 (Recommendations For The Authors):

In addition to the questions raised in the "Public review" section, the authors are also recommended to address the following issues:

1) Why is CD11b+ not associated with the apoptotic sites as neutrophils express CD11b

We have co-immunostained human liver samples with CD11b antibody (from Abcam: ab133357) and MPO antibody (from R&D: AF3667) and observed that tissue infiltrating neutrophils in livers have low to undetectable levels of CD11b expression (please refer the image below; white arrowheads point to neutrophils). Few CD11b+ cells in liver tissues express MPO (the CD11b+ cells are mostly macrophages, unpublished observations).

Based on these data, we conclude that CD11b is hardly expressed in neutrophils inside livers.

Author response image 2.

2) Can TUNEL signals in Fig. S1C be from apoptotic neutrophils?

In response, the fragmentation of nucleus is a hallmark of apoptosis hence TUNEL staining will uniformly label all fragmented parts of apoptotic nucleus. The nucleus of NE+ neutrophils are not labelled by TUNEL staining in Fig. S1C. The TUNEL+ nuclear fragments seen inside neutrophils are nuclear debris of apoptotic hepatocytes phagocytosed by neutrophils (Fig. S1C).

3) The Fig 2B experiment may be done with induced apoptosis so that neutrophil burrowing steps may be recorded from the very beginning and a better time course for the entire process can be assessed.

Thank you for the suggestions, we had tried many times with various conditions, yet still had no success to capture the very beginning of perforocytosis in vivo. We are continuing to work on this.

4) In "we found thatU937 cells exhibited much lower phagocytosis of apoptotic NCTC cells than did HL60 cells (Fig. S2B, C)," the citation should be only S2C

Thank you for pointing this out, we have corrected this in the manuscript.

5) Both neutrophil depletion models cause neutrophil death, which may complicate the interpretation of the liver function and AIL disease phenotypes. A neutropenic model such as G-CSFR−/− or Cebpe-/- mice may be used to avoid the caveat of antibody/DTR-dependent depletion models.

Thank you for this thoughtful suggestion. We have also induced AIL phenotypes in mice by using α- Galcer. α-Galcer did not cause neutrophil death but impaired neutrophil perforocytosis and futher generated AIL phenotypes in mice (unpublished observations). We plan to perform the simiarl experiments in G-CSFR−/− or Cebpe−/− mice as the reviewer suggested.

6) RNAi silencing experiments need additional controls for off-target effects

These RNAi silencing constructs were purchased from Santa Cruz Biotechnology and the off-target effects have been tested by the company. No significant off-target effects have been detected according to the manufacture report.

Author response:

The following is the authors’ response to the original reviews.

eLife assessment:

This manuscript is a valuable study of the responses of GPi neurons to DBS stimulation in human PD and dystonia patients and it finds evidence for altered short-term and long-term plasticity in response to DBS between the two patient populations. This data set is of interest to both basic and clinical researchers working in the field of DBS and movement disorders. While there was enthusiasm for the potential significance of these findings, support for their conclusions was incomplete. Thir data may be indicative of more interesting and complex interpretations than currently considered in the article. 

The authors would like to express their gratitude to the Editorial Team and Reviewers for their invaluable feedback which helped to improve the manuscript.

Reviewer #1:

Summary:

Sumarac et al investigate differences in globus pallidus internus (GPi) spike activity and short- and long-term plasticity of direct pathway projections in patients with Parkinson's disease (PD) and dystonia. Their main claims are that GPi neurons exhibit distinct characteristics in these two disorders, with PD associated with specific power-frequency oscillations and dystonia showing lower firing rates, increased burstiness, and less regular activity. Additionally, long-term plasticity and synaptic depression appear to differ between the two conditions. The authors suggest that these findings support the concept of hyperfunctional GPi output in PD and hypofunctional output in dystonia, possibly driven by variations in the plasticity of striato-pallidal synapses. Overall enthusiasm is relatively high, but I think the discussion omits discussing findings that don't align well with standard models. 

Strengths: 

These types of studies are valuable as the data arise from patients who have dystonia or PD. This could provide unique insights into disease pathophysiology that might not be recapitulated in animal systems work. 

Thank you for the positive feedback.

Weaknesses: 

- The rate model and indirect/direct pathway ideas lack explanatory power; too much of the hypothesis generation and discussion in this manuscript is set in the context of these old ideas. Their data in my view emphasize this somewhat emphatically. Most patients with the 'hypokinetic' movement disorder PD have dystonia as a part of their motor features. Dystonia is a form of excessive muscle activation that on the one hand is 'hyperkinetic' but on the other usually decreases the speed of motor tasks, even in patients with primary dystonia. Similarly, PD patients display a bewildering variety of hyperkinetic manifestations as well (rest tremor, dystonia, dyskinesia). If these are truly independent classifications, i.e. hyper- versus hypo-kinetic, the authors must acknowledge that there is considerable overlap in the spike activity across groups - numerous dystonia patients display higher discharge rates than the majority of the PD sample. Based on the firing rate alone, it would not be possible to distinguish these groups. 

Thank you for your insightful comments regarding the discussion of the rate model and the distinction between hyperkinetic and hypokinetic movement disorders. We acknowledge that the rate model, primarily derived from limited number of animal subjects [1], may not fully encapsulate the complexities of Parkinson's disease (PD) and dystonia. Our study aimed to validate animal model findings in humans by correlating single-neuron features with disease symptom severity. However, we concur with the Reviewer’s comment regarding the overlapping motor features in hypokinetic and hyperkinetic disorders. We can speculate that the overlap in neuronal properties may be reflected in the overlap of, for example, hyperkinetic features being also present in PD, as suggested by the Reviewer. Per the Reviewer’s request, we have now acknowledged this notion in the manuscript. Interestingly, hypokinetic symptoms have been reported to occur in dystonia in response to GPi-stimulation and have been associated with beta activity in the LFP [2], which reinforces the notion that neural activity may be more related to specific symptoms rather than diseases as a whole. Supplementing our analyses, in addition to total UPDRSIII scores, we have now provided correlations with only hypokinetic (i.e. bradykinesia) subscores of the UPDRSIII to focus on more direct assessment of hypokinetic features in PD versus hyperkinetic features in dystonia. We have updated our methods and results accordingly.

[1] M. R. DeLong, “Primate models of movement disorders of basal ganglia origin.,” Trends Neurosci, vol. 13, no. 7, pp. 281–285, Jul. 1990, doi: 10.1016/0166-2236(90)90110-v.

[2] R. Lofredi et al., “Pallidal Beta Activity Is Linked to Stimulation-Induced Slowness in Dystonia,” Movement Disorders, vol. 38, no. 5, pp. 894–899, 2023, doi: 10.1002/mds.29347.

Amendments to the manuscript:

“Indeed, variability in spike firing rates in PD may be reflected in the considerable overlap in spiking activity between PD and dystonia (Fig. 1A), with many dystonia patients exhibiting higher discharge rates compared to PD patients.”

“Given that UPDRSIII includes both hypokinetic and hyperkinetic symptoms of PD, we further sought to disaggregate the score by only considering items 23-26 in UPDRSIII, which assess hypokinetic symptoms of PD.”

“… with a marginally stronger correlation for PD hypokinetic symptoms only (items 23-26 of UPDRSIII, Spearman's rho=0.32, p=.0330; Supplementary Fig. 3)”

Supplementary Fig. 3: We provided correlations with hypokinetic (i.e., bradykinesia) subscore of the UPDRSIII. There is very little difference between correlation results of UPDRSIII total (Fig. 1) and the hypokinetic-only subscore (Supplementary Fig. 3).

“though our results do not change substantially when only hypokinetic PD features are considered (Supplementary Fig. 3).”

- If beta power is pathognomonic of parkinsonism, the authors found no differences in beta-related spike discharges across the groups. One would have predicted greater beta power in PD than in primary dystonia. This should be discussed explicitly and an interpretation should be provided. 

We agree with the reviewer that considering the previous LFP literature, one might have expected a difference in single-neuron oscillation power between PD and dystonia. However, while prior studies [3], [4] have reported significant differences in oscillatory power between the two diseases, researchers examined local field potential (LFP) activity only. Other work [5] in non-human primates investigated single-neuron oscillations and reported no differences between PD and dystonia at the single-neuron level, in line with our findings. However, despite the lack of difference in overall power presented here, we provide evidence that the strength of the beta-frequency single-neuron oscillations nevertheless correlates with symptom severity in PD but not dystonia; whereas the strength of the theta-frequency single-neuron oscillations correlates with symptom severity in dystonia but not PD.

[3] P. Silberstein et al., “Patterning of globus pallidus local field potentials differs between Parkinson’s disease and dystonia.,” Brain, vol. 126, no. Pt 12, pp. 2597–2608, Dec. 2003, doi: 10.1093/brain/awg267.

[4] D. D. Wang et al., “Pallidal Deep-Brain Stimulation Disrupts Pallidal Beta Oscillations and Coherence with Primary Motor Cortex in Parkinson’s Disease,” J Neurosci, vol. 38, no. 19, pp. 4556–4568, May 2018, doi: 10.1523/JNEUROSCI.0431-18.2018.

[5] P. A. Starr et al., “Spontaneous pallidal neuronal activity in human dystonia: comparison with Parkinson’s disease and normal macaque.,” J Neurophysiol, vol. 93, no. 6, pp. 3165–3176, Jun. 2005, doi: 10.1152/jn.00971.2004.

Amendments to the manuscript:

“Although previous research has reported differences in the LFP power between PD and dystonia [27,28], a study in non-human primates found no such differences in single-neuron oscillatory strength [8], as reflected in our findings. However, despite a lack of difference in overall power across disorders, we were able to derive disease/frequency-specific relationships with respect to clinical scores (Fig. 1C; oscillatory features).”

- The study lacks a healthy control group, making it challenging to differentiate disease-specific findings from normal variations in GPi activity and plasticity. Although this is acknowledged in the discussion, this complicates the interpretation of the results. The sample sizes for PD and dystonia patients are relatively small, and the study combines various forms of dystonia, potentially masking subtype-specific differences. A larger and more homogenous sample could enhance the study's reliability.

Indeed, intraoperative microelectrode recordings cannot be obtained in healthy individuals. We agree with the Reviewer that this limits the interpretation of the data. However, directly comparing clinical correlations with single neuron readouts between two distinct clinical entities may, to some degree, compensate for the lack of healthy control data. This contrast, while not providing a healthy control, is still able to point to disease-specific differences. This approach has previously been used to comparisons at the LFP level [6]. While the sample size is indeed small, it is comparable or even higher to similar studies that have investigated the relation of symptom severity of single neuron readouts [7]. The Reviewer is right in that we do not differentiate between generalized or cervical dystonia. We chose to do so because our subgroup analysis provided in the Supplementary Material did not suggest specific differences; though there is insufficient data from specific dystonia subtypes to make formal statistical comparisons. Indeed, future studies should investigate specific subtypes further.

[6] R. Lofredi et al., “Pallidal beta bursts in Parkinson’s disease and dystonia,” Movement Disorders, vol. 34, no. 3, pp. 420–424, 2019, doi: 10.1002/mds.27524.

[7] A. Gulberti et al., “Subthalamic and nigral neurons are differentially modulated during parkinsonian gait,” Brain, p. awad006, Feb. 2023, doi: 10.1093/brain/awad006.

Amendments to the manuscript:

“While we did not observe differences across dystonia subtypes (Supplementary Fig. 1), future studies in larger patient cohorts would are warranted. Finally, as many findings in Fig. 1 do not survive corrections for multiple comparisons, we suggest interpretation of results with caution. Despite this, many of our findings related to neuronal correlates are generally in line with previous literature, especially related to oscillatory correlates of PD and dystonia.”

- While they mention that data are available on request, sharing data openly would increase transparency and allow for independent validation of the results. It is unclear how sharing deidentified data would compromise patient privacy or present ethical issues of any kind, as claimed by the authors. 

Much of the data in question were collected under an old Research Ethics Board (REB) protocol which did not address data sharing. However, we have consulted with our REB and gained retroactive permission to post de-identified data which are now available in the Supplementary Material.

Amendments to the manuscript:

“The data that support the findings of this study are available in a public repository (see: https://osf.io/nqzd2/)”

- They appropriately acknowledge several limitations, such as the inability to use pharmacological interventions and the need for further research in the chronic setting. 

Thank you for the comment.

- The manuscript highlights differences in GPi activity and plasticity between PD and dystonia but could provide more context on the clinical implications of these findings, particularly regarding what the implications would be novel paradigms for deep brain stimulation. 

Thank you for the comment. Our finding that striato-pallidal plasticity decays more slowly in dystonia compared to PD may relate to the slower time course of symptom relief associated with GPi-DBS in dystonia, as presently outlined in the discussion. On the other hand, symptoms are also suppressed for longer after the cessation of stimulation in dystonia compared to PD, which may reflect long-term plastic changes [8], [9]. In the context of clinical DBS, plasticity modulation may be facilitated by intermittent stimulation algorithms that may achieve the necessary plastic network change by applying stimulation for a defined time but could then be switched off for improved energy consumption and perhaps as a means of mitigating side effects. DBS devices with chronic sensing may enable monitoring of evoked potential amplitudes for future adaptive stimulation applications; however, currently available devices are limited by low sampling rates, but future devices may overcome these technical limitations.

[8] D. Ruge et al., “Deep brain stimulation effects in dystonia: time course of electrophysiological changes in early treatment.,” Mov Disord, vol. 26, no. 10, pp. 1913–1921, Aug. 2011, doi: 10.1002/mds.23731.

[9] D. Ruge et al., “Shaping reversibility? Long-term deep brain stimulation in dystonia: the relationship between effects on electrophysiology and clinical symptoms.,” Brain, vol. 134, no. Pt 7, pp. 2106–2115, Jul. 2011, doi: 10.1093/brain/awr122.

Amendments to the manuscript:

“While further work is certainly required to better understand disease-related differences in plasticity, our findings may nevertheless motivate the development of periodic intermittent (ON/OFF) DBS strategies which periodically modulate synaptic plasticity for therapeutic benefits which outlast stimulation delivery, as have recently been employed in preclinical work [52,53].”

- While statistical tests are mentioned, the manuscript could benefit from a more detailed presentation of statistical methods, including correction for multiple comparisons and effect sizes. Did the authors consider different recording sites within each patient as independent observations? I think this is not appropriate if that was the case. 

Thank you for your constructive feedback. In response to the concerns regarding the statistical methods, we have expanded our analysis to provide a more comprehensive statistical overview. Specifically, we implemented the Bonferroni correction for multiple comparisons across each of the seven tests conducted for the differences in single-neuron features between PD and dystonia. The adjustment revealed that only the burst index and coefficient of variation retain statistical significance after post hoc correction, while the firing rate does not. Results of the Bonferroni corrections are now presented in Supplementary Table 3. Reflecting on the initial comment about firing rates between the two disorders, our updated findings underscore the limitation of using firing rates alone to differentiate between PD and dystonia, and instead, our analysis now points to burstiness and firing irregularity as more reliable discriminators. Regarding the clinical correlations, we refined our statistical analysis by employing nonparametric Monte Carlo permutation tests with 5000 permutations, as used in recent work [10], [11]. This method is chosen for its independence from assumptions regarding data distribution. Specifically, we computed and tested the Spearman rho for significance using the permutation test. Then, to address multiple comparisons, we controlled the false discovery rate (FDR) using the Benjamini-Hochberg procedure. Results of these comparisons are now presented in Supplementary Table 4. Lastly, to address the concern regarding recording site independence within patients, we updated our plasticity analysis methodology. In our study, 6 out of 18 patients had multiple recording sites. Thus, to account for this, we employed linear mixed models (LMM) with patient ID as a random factor to appropriately account for the non-independence of these observations.

[10] v Lofredi et al., “Dopamine-dependent scaling of subthalamic gamma bursts with movement velocity in patients with Parkinson’s disease,” Elife, vol. 7, p. e31895, Feb. 2018, doi: 10.7554/eLife.31895.

[11] R. Lofredi et al., “Subthalamic beta bursts correlate with dopamine-dependent motor symptoms in 106 Parkinson’s patients,” npj Parkinsons Dis., vol. 9, no. 1, Art. no. 1, Jan. 2023, doi: 10.1038/s41531-022-00443-3.

Amendments to the manuscript:

“For comparing differences in single-neuron features between PD and dystonia, significant results were followed up with post hoc multiple comparisons with a Bonferroni correction. For clinical correlations, non-parametric Monte Carlo permutation tests were used, avoiding assumptions about data distribution. The tested values were randomly shuffled 5,000 times to form a probability distribution, with the p-value reflecting the original sample rank. All tests underwent adjustment for multiple comparisons, controlling the false discovery rate (FDR) at an α-level of 0.05.”

“analyzed using a linear mixed model (LMM) with patient ID as a random factor, normalized fEP amplitudes as the response variable, and epoch as a fixed effect”

“using a LMM with patient ID as a random factor”

“However, none of the clinical correlations survived Benjamini-Hochberg FDR-correction for multiple comparisons (Supplementary Table 4).”

“In PD, fEP amplitudes were significantly greater after compared to before HFS (LMM; p = .0075, effect size = 5.42 ± 1.79; Fig. 2C), while in dystonia, the increase approached but did not reach statistical significance (LMM; p = .0708, effect size = 2.82 ± 1.45; Fig. 2C).”

All statistics were updated in the results section and the figures.

“Finally, as many findings in Fig. 1 do not survive corrections for multiple comparisons, we suggest interpretation of results with caution. Despite this, many of our findings related to neuronal correlates are generally in line with previous literature, especially related to oscillatory correlates of PD and dystonia.”

- The manuscript could elaborate on the potential mechanisms underlying the observed differences in GPi activity and plasticity and their relevance to the pathophysiology of PD and dystonia. 

Thank you for your feedback. We have enhanced the manuscript by integrating additional discussions on previous studies related to plasticity in dystonia and PD (e.g., [12], [13]), which highlight excessive plasticity in dystonia. Although these may appear contradictory to our findings of increased plasticity in PD compared to dystonia, we propose (also justified by previous literature) that chronic dopaminergic medication use may lead to synaptic over-sensitization, which has been hypothesized as a biological mechanism underlying levodopa-induced dyskinesias (a hyperkinetic feature) in PD [14].

[12] Y. Tamura et al., “Disordered plasticity in the primary somatosensory cortex in focal hand dystonia.,” Brain, vol. 132, no. Pt 3, pp. 749–755, Mar. 2009, doi: 10.1093/brain/awn348.

[13] D. A. Peterson, T. J. Sejnowski, and H. Poizner, “Convergent evidence for abnormal striatal synaptic plasticity in dystonia.,” Neurobiol Dis, vol. 37, no. 3, pp. 558–573, Mar. 2010, doi: 10.1016/j.nbd.2009.12.003.

[14] P. Calabresi, B. Picconi, A. Tozzi, V. Ghiglieri, and M. Di Filippo, “Direct and indirect pathways of basal ganglia: a critical reappraisal.,” Nat Neurosci, vol. 17, no. 8, pp. 1022–1030, Aug. 2014, doi: 10.1038/nn.3743.

Amendments to the manuscript:

“Converging evidence from past animal and human studies suggests that dystonia is associated with impaired synaptic function and abnormal synaptic plasticity [35–37]. Compared to healthy controls, it has been shown that transcranial magnetic stimulation induced motor evoked potentials (MEPs) are hyperexcitable in dystonia [38,39], and somatosensory and motor cortical plasticity is greater [40]. Likewise, enhanced long-term potentiation at cortico-striatal synapses has been shown in rodent models of dystonia [41,42]. While our finding that long term potentiation effects are greater in PD compared to dystonia (Fig. 2D) is difficult to corroborate with this literature, one potential explanation can be that all of our PD patients are long-term users of levodopa. We have previously shown that the intake of this antiparkinsonian dopaminergic medication leads to potent increases in the magnitude of direct pathway plasticity [15]. Although patients are 12hr withdrawn form antiparkinsonian medications for surgery, it could be that striato-pallidal synapses are nevertheless chronically over-sensitized from prolonged use of dopaminergic medication; which is a well-known hypothesis related to the manifestation of levodopa-induced dyskinesias (a hyperkinetic feature) in PD [43]. Indeed, a lack of depotentiation of striato-pallidal projections has previously been observed in patients with levodopa-induced dyskinesias [44]. As such, excessive plasticity of these projections may corroborate hyperkinetic features of dystonia and levodopa-induced dyskinesias in PD.”

Reviewer #2: 

Summary: 

The authors investigated how neuronal activity and metrics of plasticity using local electrical stimulation in the GPi were different between Parkinson's disease and dystonia patients. 

Strengths: 

The introduction highlights the importance of the work and the fundamental background needed to understand the rest of the paper. It also clearly lays out the novelty (i.e., that the dynamics of plastic effects in GPi between dystonia and PD have not been directly compared). 

The methods are clearly described and the results are well organized in the figures. 

The results are strong with measurements from a large population of patients for each disease group and with distinct findings for each group. 

Thank you for the kind appraisal.

Weaknesses: 

The discussion was hard to follow in several places, making it difficult to fully appreciate how well the authors' claims and conclusions are justified by their data, mostly in relation to the plasticity results. It may help to summarize the relevant findings for each section first and then further expand on the interpretation, comparison with prior work, and broader significance. Currently, it is hard to follow each section without knowing which results are being discussed until the very end of the section. With the current wording in the "Neuronal correlates.." section, it is not always clear which results are from the current manuscript, and where the authors are referring to past work.

Thank you for this feedback. The main findings are now summarized in a paragraph at the beginning of the Discussion section, before being discussed in comparison to other studies in the literature in subsequent sub-sections. Moreover, throughout the Discussion, findings from our study are now always reflected by a reference to the relevant figure to more easily differentiate current findings from previous literature. Additionally, Discussion sub-sections have been expanded to consider additional literature in response to various comments throughout the Review process (including the subsequent Review comment).

Amendments to the manuscript:

Paper findings are referenced to figures which depict the results at hand; discussion sub-sections expanded; and the following text has been added at the start of the Discussion:

“In particular, we found that GPi neurons exhibited lower firing rates, but greater burstiness and variability in dystonia compared to PD (Fig. 1A). While no differences were found in the power of spiketrain oscillations across disorders (Fig. 1B), we found that PD symptom severity positively correlated with the power of low-beta frequency spiketrain oscillations, whereas dystonia symptom severity positively correlated with the power of theta frequency spiketrain oscillations (Fig. 1C). Dystonia symptom severity moreover correlated negatively with firing rate, and positively with neuronal variability. These results are discussed in greater detail with respect to previous literature in the subsequent Discussion section entitled “Neuronal correlates of PD and dystonia.” In response to electrical stimulation (protocol depicted in Fig. 2A), we found significant increases in the amplitudes of positive-going stimulation-evoked field potential amplitudes (considered to reflect striato-pallidal synaptic strength; as exemplified in Fig. 2B) before versus after HFS in both PD and dystonia (Fig. 2C); with recording sites in PD exhibiting significantly greater increases (Fig. 2D). While changes to evoked potential amplitude before versus after stimulation can be considered to be reflective of long-term plasticity [15,18], the dynamics of evoked potentials during HFS (as depicted in Fig. 2E) can be considered as reflective of short-term synaptic plasticity [18,21]. To this end, our findings are suggestive of faster latency synaptic depression in PD compared to dystonia (Fig. 2F/G). Plasticity findings are discussed in greater detail in the Discussion section entitled “Direct pathway plasticity.”

Also, I felt that more discussion could be used to highlight the significance of the current results by comparing and/or contrasting them to prior relevant work and mechanisms. The novelty or impact is not very clear as written. Could this be further substantiated in the Discussion? 

Thank you for the feedback. The discussion has been expanded to include additional literature that is relevant to the findings reported in the manuscript. For example, with regards to the neuronal correlates sub-section, we now highlight the important findings [15] that show changes to the discharge rates and oscillatory tendencies of GPi neurons in non-human primates in response to staged MPTP applications to progressively titrate motor severity; these results substantiate our lack of correlation with firing rates in PD, and presence of a clinical correlation with beta oscillations. We additionally now emphasize human studies that found LFP power difference between PD and dystonia [3], [4]; but simultaneously highlight studies that did not find such differences in spike-train oscillations (in non-human primates) [5], which is reflective of our own findings. With regards to our plasticity sub-section, we have added new content related to previous literature on plasticity in dystonia and PD (also addressed in response to a query from Reviewer #1). For example, we bring to light a variety of previous studies [12], [13] emphasizing excessive plasticity in dystonia. However, while such studies may seem to contradict our findings of greater plasticity in PD compared to dystonia, we additionally provide hypotheses (justified by previous literature) that prolonged used of dopaminergic medication may result in synaptic over-sensitization, thus giving rise to levodopa-induced dyskinesias (a hyperkinetic feature) in PD [14].

[3] P. Silberstein et al., “Patterning of globus pallidus local field potentials differs between Parkinson’s disease and dystonia.,” Brain, vol. 126, no. Pt 12, pp. 2597–2608, Dec. 2003, doi: 10.1093/brain/awg267.

[4] D. D. Wang et al., “Pallidal Deep-Brain Stimulation Disrupts Pallidal Beta Oscillations and Coherence with Primary Motor Cortex in Parkinson’s Disease,” J Neurosci, vol. 38, no. 19, pp. 4556–4568, May 2018, doi: 10.1523/JNEUROSCI.0431-18.2018.

[5] P. A. Starr et al., “Spontaneous pallidal neuronal activity in human dystonia: comparison with Parkinson’s disease and normal macaque.,” J Neurophysiol, vol. 93, no. 6, pp. 3165–3176, Jun. 2005, doi: 10.1152/jn.00971.2004.

[12] Y. Tamura et al., “Disordered plasticity in the primary somatosensory cortex in focal hand dystonia.,” Brain, vol. 132, no. Pt 3, pp. 749–755, Mar. 2009, doi: 10.1093/brain/awn348.

[13] D. A. Peterson, T. J. Sejnowski, and H. Poizner, “Convergent evidence for abnormal striatal synaptic plasticity in dystonia.,” Neurobiol Dis, vol. 37, no. 3, pp. 558–573, Mar. 2010, doi: 10.1016/j.nbd.2009.12.003.

[14] P. Calabresi, B. Picconi, A. Tozzi, V. Ghiglieri, and M. Di Filippo, “Direct and indirect pathways of basal ganglia: a critical reappraisal.,” Nat Neurosci, vol. 17, no. 8, pp. 1022–1030, Aug. 2014, doi: 10.1038/nn.3743.

[15] A. Muralidharan et al., “Physiological changes in the pallidum in a progressive model of Parkinson’s disease: Are oscillations enough?,” Exp Neurol, vol. 279, pp. 187–196, May 2016, doi: 10.1016/j.expneurol.2016.03.002.

Amendments to the manuscript:

“Despite the lack of correlations with firing rate in PD, our findings seem to align with those of Muralidharan and colleagues [25], who showed that GPi neuronal firing rates may not directly correlate with motor severity but exhibit variability across the disease severity continuum in parkinsonian non-human primates (initially increasing, then decreasing, then increasing again at mild, moderate, and severe disease manifestations, respectively). Thus, while GPi discharge rates may change in PD, such changes may not be reflected by linear relationships with motor sign development and progression. Indeed, variability in spike firing rates in PD may be reflected in the considerable overlap in spiking activity between PD and dystonia (Fig. 1A), with many dystonia patients exhibiting higher discharge rates compared to PD patients. While differences in discharge rates were nevertheless observed between PD and dystonia, it may be that the combination of rate and pattern (reflected in the BI and CV) changes best differentiates the two disorders.”

“Converging evidence from past animal and human studies suggests that dystonia is associated with impaired synaptic function and abnormal synaptic plasticity [35–37]. Compared to healthy controls, it has been shown that transcranial magnetic stimulation induced motor evoked potentials (MEPs) are hyperexcitable in dystonia [38,39], and somatosensory and motor cortical plasticity is greater [40]. Likewise, enhanced long-term potentiation (LTP) at cortico-striatal synapses has been shown in rodent models of dystonia [41,42]. While our finding that LTP effects are greater in PD compared to dystonia (Fig. 2D) is difficult to corroborate with this literature, one potential explanation can be that all of our PD patients are long-term users of levodopa. We have previously shown that the intake of this antiparkinsonian dopaminergic medication leads to potent increases in the amount of plasticity elicited in GPi [15]. Although patients are 12hr withdrawn form antiparkinsonian medications for surgery, it could be that striato-pallidal synapses are nevertheless chronically over-sensitized from prolonged use of dopaminergic medication; which is a well-known hypothesis related to the manifestation of levodopa-induced dyskinesias (a hyperkinetic feature) in PD [43]. Indeed, a lack of depotentiation of striato-pallidal projections has previously been observed in patients with levodopa-induced dyskinesias [44]. As such, excessive plasticity of these projections may corroborate hyperkinetic features of dystonia and levodopa-induced dyskinesias in PD.”

Some specific comments and questions about the Discussion: 

Lines 209-211 - This sentence was hard to understand, could it be clarified? 

Lines 211-213 - What do phasic and tonic components mean exactly? Could this be specifically defined? Are there specific timescales (as referred to in Intro)?

Lines 215-217 - It's not clear what was delayed in dystonia, and how the authors are trying to contrast this with the faster time course in PD. I think some of this is explained in the introduction, but could also be re-summarized here as relevant to the results discussed. 

Lines 223-224 - I'm not sure I follow the implication that network reorganization leads to delayed functional benefits. Could this be further elaborated? 

Reply & Amendments to the manuscript: Thank you for your feedback. We've made the following concise revisions to address the comments:

We've clarified lines 209-211 to explain that variations in electrical stimulation effects on pathways in PD and dystonia may reveal the operational mechanisms of DBS, despite a common target:

“The variation in the modulation of these projections / pathways to electrical stimulation may also indicate the mechanism by which DBS operates across PD and dystonia, despite a common stimulation target.”

In response to the second comment on lines 211-213 about phasic and tonic components, we now specify that phasic refers to dynamic muscle contractions, and tonic to continuous muscle contractions, providing clear definitions relevant to our context:

“Clinical studies in dystonia have shown that DBS leads to a more rapid improvement in the transient, dynamic muscle contractions (phasic components) of the disorder when compared to the sustained, continuous muscle contractions (tonic or fixed components) [33]”

For lines 215-217, we've refined our discussion to clearly contrast the delayed response in dystonia with the faster onset in PD:

“This contrast with PD, where the, the maximal clinical response to DBS occurs within a much faster time course [13,36].”

On lines 223-224, we've expanded the explanation of how network reorganization may lead to delayed functional benefits, highlighting adjustments in neural connectivity and synaptic efficacy in response to stimulation:

“which involves adjustments in neural connectivity or synaptic efficacy in response to the stimulation [14,35].”

Could the absence of a relationship between FR and disease in PD be discussed? 

Thank you for raising this point. Despite observing higher firing rates in PD compared to dystonia, it is unexpected that these rates do not correlate with symptom severity according to the rate model of PD [1]. However, despite the lack of correlations with firing rates, our findings align with similar animal work of Muralidharan et al. [15], which reported that neuronal firing rates within the GPi of rhesus monkeys did not increase linearly with respect to varying intensities of parkinsonian motor severity. We did however show that low beta oscillatory strength within the GPi may play a significant role in the manifestation of motor symptoms in PD; which is also in line with findings of Muralidharan and colleagues. As per the Reviewer’s request, we have included this content into our discussion.

[1] M. R. DeLong, “Primate models of movement disorders of basal ganglia origin.,” Trends Neurosci, vol. 13, no. 7, pp. 281–285, Jul. 1990, doi: 10.1016/0166-2236(90)90110-v.

[15] A. Muralidharan et al., “Physiological changes in the pallidum in a progressive model of Parkinson’s disease: Are oscillations enough?,” Exp Neurol, vol. 279, pp. 187–196, May 2016, doi: 10.1016/j.expneurol.2016.03.002.

Amendments to the manuscript:

“Despite the lack of correlations with firing rate in PD, our findings seem to align with those of Muralidharan and colleagues [25], who showed that GPi neuronal firing rates may not directly correlate with motor severity but exhibit variability across the disease severity continuum in parkinsonian non-human primates (initially increasing, then decreasing, then increasing again at mild, moderate, and severe disease manifestations, respectively). Thus, while GPi discharge rates may change in PD, such changes may not be reflected by linear relationships with motor sign development and progression.”

“Indeed, Muralidharan and colleagues [25] also showed linear group-level relationships between low-beta frequency spiketrain oscillations and disease severity in parkinsonian non-human primates, despite the lack of linear relationships with spike discharge rates (as discussed above).”

It wasn't very clear how the direct pathway can be attributed to plasticity changes if the GPi makes up both the direct and indirect pathways. Could this be further clarified? 

The reviewer brings up an important nuanced point. Recent work from our lab [16] shows that inhibitory evoked fields in STN (which receives inhibitory fields from GPe; no other inhibitory sources) are persistent with very minimal depression during HFS. On the other hand, inhibitory fields in the SNr (which receives majority of its inhibitory inputs from striatum; though some come by way of GPe as well per anatomical literature) depress quickly. We have previously also shown these rapidly depressing fields in GPi [17], [18], which also receives the majority of its inhibitory inputs via striatum, though some also from GPe. As such, the disaggregation of striatum-mediated versus GPe-mediated inhibitory fields is achieved based on: lack of rapidly depressing inhibitory evoked field potentials in STN (which receives inhibitory inputs via GPe and not striatum), but a common presence of rapidly depressing evoked field potentials in SNr and GPi (which both receive most of their inhibitory inputs from striatum); differences in the morphology of purportedly GPe- (fast latency) versus striatum-mediated (slow latency) evoked field potentials [16]; and the presence of slow latency caudato-nigral evoked field potentials in slices [19] that are reversed by GABA antagonist application [20]. These points are indeed outlined in the first paragraph of the Discussion sub-section “Direct pathway plasticity.” However, we have now additionally added a point to the Limitations that inhibitory inputs to the GPi also come by way of GPe, though in a lesser abundance.

[16] L. A. Steiner et al., “Persistent synaptic inhibition of the subthalamic nucleus by high frequency stimulation,” Brain Stimul, vol. 15, no. 5, pp. 1223–1232, 2022, doi: 10.1016/j.brs.2022.08.020.

[17] L. D. Liu, I. A. Prescott, J. O. Dostrovsky, M. Hodaie, A. M. Lozano, and W. D. Hutchison, “Frequency-dependent effects of electrical stimulation in the globus pallidus of dystonia patients.,” J Neurophysiol, vol. 108, no. 1, pp. 5–17, Jul. 2012, doi: 10.1152/jn.00527.2011.

[18] L. Milosevic et al., “Modulation of inhibitory plasticity in basal ganglia output nuclei of patients with Parkinson’s disease,” Neurobiology of Disease, vol. 124, pp. 46–56, Apr. 2019, doi: 10.1016/j.nbd.2018.10.020.

[19] M. Yoshida and W. Precht, “Monosynaptic inhibition of neurons of the substantia nigra by caudato-nigral fibers,” Brain Res, vol. 32, no. 1, pp. 225–228, Sep. 1971, doi: 10.1016/0006-8993(71)90170-3.

[20] W. Precht and M. Yoshida, “Blockage of caudate-evoked inhibition of neurons in the substantia nigra by picrotoxin,” Brain Res, vol. 32, no. 1, pp. 229–233, Sep. 1971, doi: 10.1016/0006-8993(71)90171-5.

Amendments to the manuscript:

“Indeed, GPi receives the greatest abundance of inhibitory inputs from striatum (direct pathway), but also it also receives inhibitory inputs by way of GPe (indirect pathway). Although we can functionally disaggregate these pathway-specific responses based on differences in morphology and dynamics of GPe-mediated versus striatum-mediated inhibitory fEPs [21]; the possibility of compounded effects cannot be completely ruled out.”

The mechanism of short- and long-term plasticity as applied in the protocols used in this work are outlined in reference to previous citations [15, 16, 18]. Because this is a central aspect of the current work and interpreting the results, it was difficult to appreciate how these protocols provide distinct metrics of short and long-term plasticity in GPi without some explanation of how it applies to the current work and the specific mechanisms. It would also help to be able to better link how the results fit with the broader conclusions. 

Short-term plasticity is measured as the dynamic change to the fEP during ongoing HFS. For long-term plasticity analyses, the fEP amplitudes during LFS were compared pre- versus post-HFS. To make this analysis more intuitive we have added a protocol illustration to Fig 2. We have moreover greatly expanded the discussion to include more literature related to disease-specific differences in plasticity, and implications of modulating plasticity using DBS.

Amendments to the manuscript:

Added new panel to Fig 2

Author response image 1.

“Converging evidence from past animal and human studies suggests that dystonia is associated with impaired synaptic function and abnormal synaptic plasticity [35–37]. Compared to healthy controls, it has been shown that transcranial magnetic stimulation induced motor evoked potentials (MEPs) are hyperexcitable in dystonia [38,39], and somatosensory and motor cortical plasticity is greater [40]. Likewise, enhanced long-term potentiation at cortico-striatal synapses has been shown in rodent models of dystonia [41,42]. While our finding that long term potentiation effects are greater in PD compared to dystonia (Fig. 2D) is difficult to corroborate with this literature, one potential explanation can be that all of our PD patients are long-term users of levodopa. We have previously shown that the intake of this antiparkinsonian dopaminergic medication leads to potent increases in the amount of plasticity elicited in GPi [15]. Although patients are 12hr withdrawn form antiparkinsonian medications for surgery, it could be that striato-pallidal synapses are nevertheless chronically over-sensitized from prolonged use of dopaminergic medication; which is a well-known hypothesis related to the manifestation of levodopa-induced dyskinesias (a hyperkinetic feature) in PD [43]. Indeed, a lack of depotentiation of striato-pallidal projections has previously been observed in patients with levodopa-induced dyskinesias [44]. As such, excessive plasticity of these projections may corroborate hyperkinetic features of dystonia and levodopa-induced dyskinesias in PD.”

In the Conclusion, it was difficult to understand the sentence about microcircuit interaction (line 232) and how it selectively modulates the efficacy of target synapses. Some further explanation here would be helpful. Also, it was not clear how these investigations (line 237) provide cellular-level support for closed-loop targeting. Could the reference to closed-loop targeting also be further explained? 

We agree with the reviewer that the current wording may be confusing. We have changed the wording to be clearer. We have additionally added content related to closed-loop DBS based on chronic monitoring of evoked potential responses.

Amendments to the manuscript:

“Furthermore, chronic monitoring of evoked fields may allow for tracking of subcortical neuronal projections as indexed by inhibitory fields reported in this study. microcircuit interaction to selectively modulate the efficacy of target synapses.”

future applications of DBS may also benefit from closed loop tuning of basal-ganglia-thalamo-cortical circuit dynamics and plasticity through chronic monitoring of evoked potential responses [56].

How is the burst index calculated (Methods)? 

Thank you for pointing out that the burst index definition was missing from the paper. It has now been added to the manuscript.

Amendments to the manuscript:

“The burst index was computed by taking the ratio of the means from a two-component Gaussian mixture model applied to the log interspike interval distribution, a modification of the previous mode-over-mean ISI method [20]”

Figures and figure captions are missing some details:

Fig. 1 - What does shading represent? 

The shading in Fig. 1 illustrates results that were significant before adjustment for multiple comparisons.

Amendments to the manuscript:

“Depicted scatterplots are results that were significant before correction for multiple comparisons”

Fig. 2 - Can the stimulation artifact be labeled so as not to be confused with the physiological signal? Is A representing the average of all patients or just one example? Are there confidence intervals for this data as it's not clear if the curves are significantly different or not (may not be important to show if just one example)? Same for D. What is being plotted in E? Is this the exponential fitted on data? Can this be stated in the figure citation directly so readers don't have to find it in the text, where it may not be directly obvious which figure the analyses are being applied towards? 

Thank you for your comments regarding Fig. 2. We have made the following revisions to address the concerns:

To clarify the presence of stimulation artifacts and differentiate them from the physiological signal, we have updated Panel B and E in the updated Fig. 2 which highlight the stimulation artifacts accordingly.

Regarding the comment about Panel A (now B in the updated figure), it represents one single example per disease, rather than an average of all patients.

In response to the comment about what is plotted in Panel E, we have revised the figure caption to explicitly state that it includes the exponential fit on the data.

Amendments to the manuscript:

Figure 2 panel B and E now highlight stimulation artifacts.

Author response image 2.

Author response image 3.

The figure captions could use more details, that can be taken from the text, so that readers can understand figures without searching for relevant details across the paper. 

Thank you for your feedback. We have revised the figure captions accordingly to provide more details.

Amendments to the manuscript:

“Fig 1 – GPi spiketrain feature analyses and clinical correlates of PD and dystonia. (A) With respect to (A) rate-based spiketrain features, firing rate was greater in PD while burst index (BI) and coefficient of variation (CV) were greater in dystonia; whereas no differences were found for (B) oscillatory spiketrain features for theta, alpha, low beta, high beta frequencies. MWU statistical results depicted are not corrected for multiple comparisons; after correction using the Bonferroni method, only CV and BI results remain significant (please see Supplementary Table 3). (C) In PD, the power of low beta spiketrain oscillations positively correlated (Spearman correlation) with symptom severity; in dystonia, neuronal firing rate negatively correlated with symptom severity, whereas CV and the power of theta spiketrain oscillations positively correlated with symptom severity. Depicted scatterplots are results that were significant before correction for multiple comparisons; however, none of the results persist after Benjamini-Hochberg correction for false discovery rate (please see Supplementary Table 4).”

“Fig 2 – Long-term and short-term effects of HFS on striato-pallidal plasticity in PD and dystonia. (A) Schematic of the plasticity protocol to assess long-term plasticity via fEP amplitude comparisons pre- versus post-HFS and short-term plasticity via fEP dynamics during HFS. (B) Highlights example fEP traces for measuring long-term plasticity pre- versus post-HFS, with (C) displaying group-level fEP amplitudes pre- versus post-HFS across diseases. (D) Illustrates the amount of plasticity (i.e., percentage change in fEP amplitudes pre- versus post-HFS) in both PD and dystonia, with PD showing higher levels of plasticity. (E) Provides an example of fEP traces during HFS for assessing short-term plasticity, with (F) depicting group-level decay rates of fEP amplitudes using an exponential fit on the fEP amplitudes over the first 5 stimulus pulses across diseases. (G) Shows the half-life of the fitted exponential (i.e., rate of attenuation of fEP amplitudes) between PD and dystonia, with PD demonstrating faster fEP attenuation.”

Author response:

The following is the authors’ response to the original reviews

Reviewer #1:

To gain further insight into the dynamics of microglial aging in the hippocampus, the authors used a bioinformatics method known as "pseudotime" or "trajectory inference" to understand how cells may progress through different functional states, as defined by cellular transcriptome (15,16). These bioinformatics approaches can reveal key patterns in scRNAseq / snRNAseq datasets and, in the present study, the authors conclude that a "stress response" module characterized by expression of TGFb1 represents a key "checkpoint" in microglial aging in midlife, after which the cells can move along distinct transcriptional trajectories as aging progresses. This is an intriguing possibility. However, pseudotime analyses need to be validated via additional bioinformatics as well as follow-up experiments. Indeed, Heumos et al, in their Nature Genetics "Expert Guidelines" Review, emphasize that "inferred trajectories might not necessarily have biological meaning." They recommend that "when the expected topology is unknown, trajectories and downstream hypotheses should be confirmed by multiple trajectory inference methods using different underlying assumptions."(15) Numerous algorithms are available for trajectory inference (e.g. Monocle, PAGA, Slingshot, RaceID/StemID, among many others) and their performance and suitability depends on the individual dataset and nature of the trajectories that are to be inferred. It is recommended to use dynGuidelines(16) for the selection of optimal pseudotime analysis methods. In the present manuscript, the authors do not provide any justification for their use of Monocle 3 over other trajectory inference approaches, nor do they employ a secondary trajectory inference method to confirm observations made with Monocle 3. Finally, follow-up validation experiments that the authors carry out have their own limitations and caveats (see below). Hence, while the microglial aging trajectories identified by this study are intriguing, they remain hypothetical trajectories that need to be proven with additional follow-up experiments.

We thank the reviewer for their suggestion. We have utilized the dynGuidelines kindly provided by the reviewer to utilize an additional trajectory inference tool to analyze our data. We selected Scorpius based on the structure of our data. The tool has provided additional support that microglia progress from a homeostatic state (Cx3cr1, Mef2c) to the induction of stress genes (Hspa1, Atf3) at an intermediate point during aging progression. Furthermore, we observe a concordant increase in ribosomal protein genes at a time point in the pseudotime analysis immediately prior to activation of inflammation-related genes (Il1b, Cst7). These additional analyses support the main findings of our original pseudotime analysis and have been added to the manuscript as Figure S3C,D. Additionally, in the statistical test that uncovers differentially expressed genes along the pseudotime trajectory in this analyses, we find that Tgfb1 is one of the genes that is differentially expressed with peak expression at an intermediate timepoint along the pseudotime trajectory. Furthermore, we have done some preliminary trajectory analysis with slingshot (Street et al, BMC Genomics, PMID: 29914354) that found a similar trajectory with analogous gene expression patterns and dynamic expression of Tgfb1.

To follow up on the idea that TGFb1 signaling in microglia plays a key role in determining microglial aging trajectories, the authors use RNAscope to show that TGFb1 levels in microglia peak in middle age. They also treat primary LPS-activated microglia with TGFb1 and show that this restores expression of microglial homeostatic gene expression and dampens expression of stress response and, potentially, inflammatory genes. Finally, they utilize transgenic approaches to delete TGFb1 from microglia around 8-10mo of age and scRNAseq to show that homeostatic signatures are lost and inflammatory signatures are gained. Hence, findings in this study support the idea that TGFb1 can strongly regulate microglial phenotype. Loss of TGFb1 signaling to microglia in adulthood has already been shown to cause decreased microglial morphological complexity and upregulation of genes typically associated with microglial responses to CNS insults(17-19). TGFb1 signaling to microglia has also been implicated in microglial responses to disease and manipulations to increase this signaling can improve disease progression in some cases(19). In this light, the findings in the present study are largely confirmatory of previous findings in the literature. They also fall short of unequivocally demonstrating that TGFb1 signaling acts as a "checkpoint" for determining subsequent microglial aging trajectory. To show this clearly, one would need to perturb TGFb1 signaling around 12mo of age and carry out sequencing (bulkRNAseq or scRNAseq) of microglia at 18mo and 24mo. Such experiments could directly demonstrate whether the whole microglial population has been diverted to the TGFb1-low aging trajectory (that progresses through a translational burst state to an inflammation state as proposed). Future development of tools to tag TGFb1 high or low microglia could also enable fate tracing type experiments to directly show whether the TGFb1 state in middle age predicts cell state at later phases of aging.

We apologize for the use of the term “checkpoint” when referring to the role of Tgfb1 in microglial aging. Instead, our model posits that Tgfb1 expression increases in response to the early insults of the aging process in an attempt to return microglia to homeostasis. Therefore, this would predict that increasing TGFB1 levels after an insult would decrease activation and age-related progression of microglia, which we demonstrate in vitro (Figure 3). Alternatively, the loss of TGFB1 should prevent microglia from returning to a homeostatic state after an age-related stressor, and thus increase the number of microglia in activated states. We observe this increase in activated microglia in our middle-aged microglia-specific Tgfb1 knockout mouse model. Furthermore, the haploinsufficiency of Tgfb1 at this age indicates that TGFB1 signaling in microglia is sensitive to relative levels of Tgfb1. The transient increase in Tgfb1 expression further suggests that the threshold for TGFB1 signaling is dynamic. Finally, RNA-Seq analysis of both in vitro TGFB1 supplemented microglia and in vivo Tgfb1 depleted microglia highlight that TGFB1 alters the aging microglia transcriptome. Combined, these results provide evidence that Tgfb1 modulates advancement of microglia through an aging continuum.

The present study would also like to draw links between features of microglial aging in the hippocampus and a decline in hippocampal-dependent cognition during aging. To this end, they carry out behavioral testing in 8-10mo old mice that have undergone microglial-specific TGFb1 deletion and find deficits in novel object recognition and contextual fear conditioning. While this provides compelling evidence that TGFb1 signaling in microglia can impact hippocampus-dependent cognition in midlife, it does not demonstrate that this signaling accelerates or modulates cognitive decline (see below). Age-associated cognitive decline refers to cognitive deficits that emerge as a result of the normative brain aging process (20-21). For a cognitive deficit to be considered age-associated cognitive decline, it must be shown that the cognitive operation under study was intact at some point earlier in the adult lifespan. This requires longitudinal study designs that determine whether a manipulation impacts the relationship between brain status and cognition as animals age (22-24). Alternatively, cross-sectional studies with adequate sample sizes can be used to sample the variability in cognitive outcomes at different points of the adult lifespan (22-24) and show that this is altered by a particular manipulation. For this specific study, one would ideally demonstrate that hippocampal-based learning/memory was intact at some point in the lifespan of mice with microglial TGFb1 KO but that this manipulation accelerated or exacerbated the emergence of deficits in hippocampal-dependent learning/memory during aging. In the absence of these types of data, the authors should tone down their claims that they have identified a cellular and molecular mechanism that contributes to cognitive decline.

We agree with the reviewer that to adequately demonstrate an age-dependent effect of microglia-derived TGFB1 on cognition it is necessary to perturb microglial TGFB1 at young and mature ages and assess the age-dependent effect on cognition. To address this, we have now performed a complementary behavioral study utilizing the Tmem119-CreER mouse model to drive the microglia-specific excision of Tgfb1 in two separate cohorts of mice – one young (2-3 months) and one in mature mice (7-8 months) – followed by cognitive testing. Using the novel object recognition test, we find that young mice of all genotypes (WT, Tgfb1 Het and Tgfb1 cKO ) retain the ability to recognize the novel object (as determined by having a significant preference in exploring the novel object). Alternatively, only the WT mature mice demonstrate a preference for the novel object, while the Tgfb1 Het and Tgfb1 cKO show no preference for the novel object. These behavioral data demonstrate an age-dependent necessity for microglia-specific TGFB1 in in maintain proper hippocampal-dependent memory and is now included in the manuscript as revised Figure 4I-J. We have also included additional behavioral tests (Y-Maze and open field) that did not show any difference between the genotypes as Figure S6D-G. Unfortunately, we were unable to perform the fear conditioning testing, as our apparatus broke during this time. Together, these results reveal that there is an age-dependent necessity for microglia-derived TGFB1 for hippocampal-dependent cognitive function.

A final point of clarification for the reader pertains to the mining of previously generated data sets within this study. The language in the results section, methods, and figure legends causes confusion about which experiments were actually carried out in this study versus previous studies. Some of the language makes it sound as though parabiosis experiments and experiments using mouse models of Alzheimer's Disease were carried out in this study. However, parabiosis and AD mouse model experiments were executed in previous studies (25,26), and in the present study, RNAseq datasets were accessed for targeted data mining. It is fantastic to see further mining of datasets that already exist in the field. However, descriptions in the results and methods sections need to make it crystal clear that this is what was done.

The reviewer makes an excellent point. While we referenced the public dataset in the original manuscript, the citation style of superscripted numbers diminishes our ability to adequately reference the datasets. Therefore, we have added the names of the first authors (Palovics for the parabiosis dataset and Sala Frigerio for the Alzheimer’s Disease dataset) to all the instances in the results and figure legends when we refer to these datasets.

Additional recommendations:

Major comments.

(1) There is some ambiguity surrounding how to interpret the microglial TGFb1 knockout that seems incompatible with viewing this molecule as a "checkpoint" in microglial aging. TGFb1 is believed to be primarily produced by microglia. Secreted TGFb1 is then detected by microglial TGFbR2. Are the microglia that have high levels of TGFb1 in middle age signaling to themselves (autocrine signaling)? Or contributing to a local milieu that impacts multiple neighbor microglia (paracrine signaling)? The authors could presumably look in their own dataset to evaluate microglial capacity to detect TGFb1 via its receptors.

We thank the reviewer for this insightful suggestion. We have undertaken analysis of our dataset to assess whether Tgfb1 acts through autocrine or paracrine signaling. To do so, we reanalyzed our microglia aging scRNA-Seq dataset leveraging the variation in microglia Tgfb1 expression to probe the relative activity of TGFB1. Specifically, we partitioned microglia into quartiles based on their Tgfb1 expression, and subsequently investigated the expression of TGFB signaling effectors and targets. High expression of downstream TGFB signaling pathway components in microglia with high Tgfb1 expression would point to autocrine mechanisms while, alternatively, high expression of downstream TGFB signaling pathway components in microglia with low Tgfb1 expression would point to paracrine mechanisms. We observed highest expression of TGFB signaling pathway components and targets in microglia with the highest expression of Tgfb1. These data suggest that Tgfb1 acts through an autocrine mechanism. These results have been added to our manuscript as Figure S4E-G. Additionally, while our manuscript was under review, a paper by Bedolla et al (Nature Communications 2024; PMID: 38906887) was published that investigated the role of Tgfb1 in adult microglia. This paper utilized orthogonal techniques – sparse microglia-specific Tgfb1 knockout and IHC - to also suggest that microglia utilize autocrine Tgfb1 signaling. Together, these complementary data provide strong evidence that Tgfb1 acts through an autocrine mechanism in adult microglia.

(2) Conclusions of the study rest on the assumption that microglial inflammatory responses are a central driver of cognitive decline. They assume that manipulations that increase microglial progression into an inflammatory state will negatively impact cognitive function. Although there are certainly a lot of data in the field that inflammatory factors can impact synaptic function, additional experiments would be required to unequivocally demonstrate that a "TGFb1 dependent" progression of microglia to an inflammatory state underlies any observed changes in cognition. For example, in the context of microglial TGFb1 deletion, can NSAIDs or blockers of soluble TNFa (e.g. XENP345), or blockers of SPP1, etc. rescue behavior? Can microglial depletion in this context rescue behavior? Assuming behavior was carried out in the same microglial TGFb1 KO mice that were used for microglial scRNAseq, they could also carry out linear regression-type analyses to link microglial inflammatory status to the behavioral performance of individual mice. In the absence of additional evidence of this sort, the authors should tone down claims about mechanistic relationships between microglial state and cognitive performance.

We thank the reviewer for realizing that the link between cognition and inflammation in our paper is speculative. Therefore, we have taken the reviewer’s advice and toned down the claims linking inflammation to cognition in our manuscript. Instead, we connect the disruption in cognition to what is observed in our data, a loss of microglia homeostasis and a shift in the microglia aging trajectories.

Additional Recommendations:

Minor comments:

(1) Ideally at some point in the results or discussion, the authors should acknowledge that the hippocampus has highly distinct sub-regions and that microglia show different functions and properties across these sub-regions (e.g. microglia in hilus and subgranular zone vs microglia in stratum radiatum, vs microglia immediately adjacent to or embedded within stratum pyrimidale). Do expression levels of TGFb1 and microglial aging trajectories vary across sub-regions? To what extent can this account for heterogeneity of aging trajectories observed in microglial aging within the hippocampus?

We are interested in how microglia heterogeneity during aging is influenced by the specific functions, and thus microenvironments within the hippocampus. Therefore, we have expanded our IHC analysis of microglia to determine how the microenvironment influences microglia phenotypes by looking at several different regions of the hippocampus. We have included this regional analysis as Figure S2 in the manuscript. This analysis has revealed region-specific effects on microglia activation during aging.

(2) For immunohistochemistry data, it is not particularly convincing to see one example of one cell from each condition. Generally, an accepted approach in the field is to present lower magnification images accompanied by zoom panels for several cells from each field of view. This reassures the reader that specific cells haven't simply been "cherry-picked" to support a particular conclusion.

To allay the concerns of the reviewer that cells haven’t been “cherry-picked”, we have provided low magnification images for the aging CD68 and NF_κB stains in Supplemental Figure S2.

(3) In immunohistochemistry data, have measures been taken to ensure that observed signals are not simply autofluorescence that becomes prominent in tissues with aging? (i.e. use of trueblack or photoquenching of tissue prior to staining) See PMID 37923732

We agree that autofluorescence, at least partially due to the accumulation of lipofuscin, becomes prominent in certain regions and cells of the hippocampus during aging. This most prominently occurs in the microglia of the hilus. This autofluorescence has a particular subcellular distribution, as it is localized to lyso-endosomal bodies. The microglia activation marker CD68 is also localized to lysosomes. A previous publication by Burns et al (eLife; PMID: 32579115) identified autofluorescent microglia (AF+) with unique molecular profiles that accumulate with age. They posited that these AF+ microglia resembled other microglia subsets that have pronounced storage compartments, such as the pro-inflammatory lipid droplet-containing microglia that accumulate with age reported by Marschallinger et al (Nature; PMID: 31959936). As such, autofluorescence present in microglia potentially represents distinctive and functional states of microglia. Our CD68 immunostaining accumulates with age, which could overlap with autofluorescent storage bodies. Thus, we performed a complementary CD68 immunostaining in an independent cohort of young (3 months) and aged (24 months) mice with autofluorescence quencher TrueBlack, and found that the staining pattern and accumulation of CD68 microglia with age persisted as previously observed after use of this quencher (see Authpr response image 1). Images are IBA1 (cyan) and CD68 (yellow) with the molecular layer (ML), granule cell (GC), and hilus illustrated and corresponding quantification provided (Two-way ANOVA with Sidak’s multiple comparisons test; ***P<0.001; ****P<0.0001).

We would like to note that the subcellular localization of the other immunostainings included in the manuscript was distinct from CD68, and not likely to be associated with the autofluorescent storage bodies. Additionally, our RNAScope staining for Tgfb1 did not show an accumulation with age, but rather a transient increase at 12 months of age, which indicates that the interpretation of the RNAScope stain for Tgfb1 was not unduly influenced by autofluorescence.

Author response image 1.

(4) Ideally, more care is needed with the language used to describe microglial state during aging. The terms "dystrophic," "dysfunctional," and "inflammatory" all carry their own implications and assumptions. Many changes exhibited by microglia during aging can initially be adaptive or protective, particularly during middle age. Without additional experiments to show that specific microglial attributes during aging are actively detrimental to the tissue and additional experiments to show that microglia have ceased to be capable of engaging in many of their normal actions to support tissue homeostasis, the authors should exercise caution in using terms like dysfunctional.

We appreciate the reviewers’ suggestion. To allay the concerns of the reviewer about the multiple implications of terms such as “dysfunctional” and “inflammatory”, we have tried to replace them throughout the text with more specific terms.

Reviewer #2:

That said, given what we recently learned about microglia isolation for RNA-seq analysis, there is a danger that some of the observations are a result of not age, but cell stress from sample preparation (enzymatic digestion 10min at 37C; e.g. PMID: 35260865). Changes in cell state distribution along aging were made based on scRNA-seq and were not corroborated by any other method, such as imaging of cluster-specific marker expression in microglia at different ages. This analysis would allow confirming the scRNA-seq data and would also give us an idea of where the subsets are present within the hippocampus, and whether there is any interesting distribution of cell states (e.g. some are present closer to stem cells?). Since TGFb is thought to be crucial to microglia biology, it would be valuable to include more analysis of the mice with microglia-specific Tgfb deletion e.g. what was the efficiency of recombination in microglia? Did their numbers change after induction of Tgfb deletion in Cx3cr1-creERT2::Tgfb-flox mice.

We thank the reviewer for their comment regarding potential ex vivo transcriptional alterations with the approaches used in our study. We performed our aging microglia scRNA-Seq characterization prior to the release of Marsh et al (Nature Neuroscience; PMID: 35260865), which revealed the potential transcriptional artefacts induced by isolation. That being said, we took great care to minimize the amount of time samples were subjected to enzymatic digestion (15 minutes) and kept cells at 4C during the remainder of the isolation. Furthermore, we performed all isolations simultaneously, so that transcriptional changes induced by the isolation would be present across all ages and should not be observed during our analysis unless indicative of a true age-related change. Additionally, we have corroborated changes in cell state distribution across ages using several markers (Tgfb1 and KLF2 for the intermediate stress state, S6 for the translation state, and NFKB and CD68 for activation states). In the revised manuscript, we have added additional hippocampal subregion analysis of several IHC immunostains to provide spatial insights into the microglia aging process (Figure S2). This analysis reveals unique spatial dynamics of microglia aging. For example, as the reviewer foresaw, we found that the granule cell layer (the location of adult hippocampal neurogenesis) had a more pronounced age-associated progression of microglial activation than several other regions. A subset of regions had minimal levels of activation during aging, such as the molecular layer and the stratum radiatum of the CA1 (inner CA1in the manuscript) – regions enriched in synaptic terminals. Furthermore, this analysis highlights the susceptibility of microglia aging to microenvironmental influences.

Regarding the temporally controlled microglia-specific genetic KO mouse model used in our original submission, the Cx3cr1-CreER allele selected (B6.129P2(Cg)-Cx3cr1tm2.1(cre/ERT2)Litt/WganJ) has been reported to have very high recombination efficiency (~94% in Parkhurst et al (Cell; PMID: 24360280)), and we used a tamoxifen induction protocol very similar to Faust et al. (Cell Reports; PMID: 37635351) that achieved ~98% recombination (they injected 100mg/kg for 5 days, while we injected 90mg/kg for 5 days). We analyzed our scRNA-Seq data for the expression of Tgfb1 and found that the knockout mice had a 67% reduction in cells expressing higher levels of Tgfb1 (see panel A in Author response image 2). This is likely a large underestimate of the recombination efficiency, as exon 3 is floxed and residual nonfunctional transcripts could be present, given nonsense-mediated decay is not realized in a number of knockout lines (Lindner et al, Methods, PMID: 33838271). We likely achieved a much higher excision efficiency. We would like to highlight that our data indicating increased microglia activation after tamoxifen treatment (Figure S5A) and the involvement of autonomous signaling (Figure S4E-G) are consistent with recently published work by Bedolla et al, (Nature Communications; PMID: 38906887). Additionally, as part of the revision process, we have now corroborated our behavioral data using and independent temporally controlled microglia-specific KO mouse model - Tmem119-CreER::Tgfb1 knockout mice (Figure 4I-K). We performed qPCR on sorted microglia to determine RNA levels in wildtype and knockout mice. Relative levels of Tgfb1 and exon 3 of Tgfb1 (the floxed exon) on technical replicates of 3 pooled samples indicated overall loss of Tgfb1 expression, as well as undetectable levels of exon 3 as normalized to Actb (see panel B in Author response image 2).

Author response image 2.

With respect to the effects of aging and Tgfb1 on microglia density, we find a slight region-specific increase in microglia density with age (see Author response image 3). The density of Iba1 cells across hippocampal regions was analyzed at 3 and 24 months of age (see panel A in Author response image 3) and along an aging continuum at 3, 6, 12, 18, and 24 months (see panel B in Author response image 3). These data are also included in the revised manuscript (Figure S2D-F).

Author response image 3.

Deletion of Tgfb1 also had region-specific effects on microglia. While there was no difference in microglia density between wildtype and heterozygous microglia, there was a significant increase in microglia density in the hilus and molecular layers in knockout mice (see Author response image 4) and included in the revised manuscript (Figure S5A). These data indicate that there are subtle region-specific increases in microglia density with age, as well as following the deletion of Tgfb1 from microglia of mature mice.

Author response image 4.

Additional Recommendations:

(1) The problem of possible digestion artifacts in scRNA-seq should be at least addressed in the discussion as a caveat in data interpretation. Staining for unique cluster markers in undigested tissue would solve the problem. It can be done with microscopy or using flow cytometry, but for this microglia, isolation should be done with no enzymes or with Actinomycin (PMID: 35260865).

The ex vivo activation signature uncovered by Marsh et al. (Nature Neuroscience; PMID: 35260865) arises from the digestion methods used to isolate microglia. We took the utmost care in processing our microglia identically within experiments, which should minimize the amount of uneven ex vivo activation of microglia. This is borne out by the structures of our single-cell sequencing data. Unlike Marsh et al_. where they observe unique cluster after addition of their inhibitors, we do not see any clusters unique to a single condition, suggesting that any influence of _ex vivo activation was evenly distributed.

Importantly, as suggested by the review, we have we have complemented our scRNA-Seq analysis by corroborating several markers for various stages of microglia aging progression using RNAScope and IHC in intact tissue. Specifically, the transient age-dependent increase in Tgfb1 high microglia was confirmed using RNAScope (Figure 3B), the age-related increase in ribosomal high microglia was confirmed using S6 immunostaining (Figure 3I), and the increase of various markers of age-associated activation (C1q, CD68 and NFkB) was confirmed using immunostaining (Figure 1F and Figure S2D-I). Additionally, we have also performed immunostainings for KLF2 and confirmed peak microglia expression at 18 months of age with lower levels at 24 months of age (Figure 2H).

(2) The figures of GO and violin plots are not easy to follow sometimes... what are the data points in the violin plots, maybe worth showing them as points? For the GO, e.g. in 3D, 3J, including a short description of the figure could help, e.g. in Figure 1. it was clear.

We chose not to include the datapoints in the violin plots for aesthetic purposes. Each violin plot would have had hundreds of points that would have made the plots very busy and hidden the structure of the distribution. In Author response image 5 we show the violin plot in Figure 2M with (panel A) and without (panel B) individual points. In a small format, the points overlap and become jumbled together. Therefore, we chose to present the violin plots without points for clarity on the data structure. As for the gene ontology plots in Figure 3, we have updated the descriptions in both the text and figure legends to provide clarification on what they represent.

Author response image 5.

(3) I'm very curious to see the mechanism of action of "aged" microglia in the TGFb-depletion model. Is it creating hostile conditions for stem cells, or we have increased synapse loss? Something else?

We thank the reviewer for their insightful questions. We would like to note that during the revision process of our manuscript, a complementary study was published reporting that the loss of microglia-derived Tgfb1 leads to an aberrant increase in the density of dendritic spines in the CA1 region of the hippocampus (Bedolla et al, Nature Communications, PMID: 38906887). The data from Bedolla et al, shows sparsely labeled neurons in the CA1 with a mGreenLantern expressing virus in mice the had Tgfb1 deleted from microglia using the Cx3cr1-CreERT driver (Figure 7U,V). Additionally, McNamara et al (Nature; PMID: 36517604) demonstrated that microglia-derived Tgfb1 signaling regulates myelin integrity during development and several studies have revealed links between Tgfb1 signaling and altered neurogenesis (e.g., He et al, Nature, PMID: 24859199 and Dias et al, Neuron, PMID: 25467979). Together, this growing body of work indicates that microglia-derived TGFB1 regulates myelination, neurogenesis and synaptic plasticity, which have all been shown to play a role in cognition.

Author Response

The following is the authors’ response to the original reviews.

eLife assessment

This valuable study introduces an innovative method for measuring interocular suppression depth, which implicates mechanisms underlying subconscious visual processing. The evidence supporting the effectiveness of this method would be solid after successfully addressing concerns raised by the reviewers. The novel method will be of interest not only to cognitive psychologists and neuroscientists who study sensation and perception but also to philosophers who work on theories of consciousness.

Thank you for the recognition and appreciation of our work.

Public Reviews:

Reviewer #1 (Public Review):

Strengths:

The authors introduced a new adapted paradigm from continuous flash suppression (CFS). The new CFS tracking paradigm (tCFS) allowed them to measure suppression depth in addition to breakthrough thresholds. This innovative approach provides a more comprehensive understanding of the mechanisms underlying continuous flash suppression. The observed uniform suppression depth across target types (e.g., faces and gratings) is novel and has new implications for how the visual system works. The experimental manipulation of the target contrast change rate, as well as the modeling, provided strong support for an early interocular suppression mechanism. The authors argue that the breakthrough threshold alone is not sufficient to infer about unconscious processing.

Weaknesses:

A major finding in the current study is the null effect of the image categories on the suppression depth measured in the tCFS paradigm, from which the authors infer an early interocular mechanism underlying CFS suppression. This is not strictly logical as an inference based on the null effect. The authors may consider statistical evaluation of the null results, such as equivalence tests or Bayesian estimation.

We have now included a Bayesian model comparison (implemented in JASP), to assess the strength of evidence in favour of the alternative hypothesis (or null effect). For example in Experiment 1 (comparing discrete to tCFS), we found inconsistent evidence in favour of the null effect of image-category on suppression depth:

Lines 382 – 388: “We quantified the evidence for this null-effect on suppression depth with a subsequent Bayesian model comparison. A Bayesian repeated-measures ANOVA (2 x 2; procedure x image type on suppression depth) found that the best model to explain suppression depth included the main effect of procedure (BF10 = 3231.74), and weak evidence/data insensitivity for image type (BF10 = 0.37). This indicates that the data was insensitive as to whether image-type was better at predicting suppression depth than the null model.”

In Experiment 2, which was specifically designed to investigate the effect of image category on suppression depth, we found strong evidence in favour of the null:

Lines 429 – 431: “A Bayesian repeated-measures ANOVA (1 x 5, effect of image categories on suppression depth), confirmed strong evidence in favour of the null hypothesis (BF01 =20.30).

In Experiment 3, we also had image categories, but the effect of rate of contrast change was our main focus. For completeness, we have also included the Bayes factors for image-category in Experiment 3 in our text.

Lines 487- 490> “This null-effect of image-type was again confirmed with a Bayesian model comparison (3 speed x 4 image categories on suppression depth), demonstrating moderate support for the null effect of image category (BF01= 4.06).”

We have updated our Methods accordingly with a description of this procedure

Lines 297-305: “We performed Bayesian model comparison to quantify evidence for and against the null in JASP, using Bayesian repeated measures ANOVAs (uninformed prior with equal weight to all models). We report Bayes factors (B) for main effects of interest (e.g. effect of image type on suppression depth), as evidence in favour compared to the null model (BF10= B). Following the guidelines recommended in (Dienes 2021), B values greater than 3 indicate moderate evidence for H1 over H0, and B values less than 1/3 indicate moderate evidence in favour of the null. B values residing between 1/3 and 3 are interpreted as weak evidence, or an insensitivity of the data to distinguish between the null and alternative models.”

More importantly, since limited types of image categories have been tested, there may be some exceptional cases. According to "Twofold advantages of face processing with or without visual awareness" by Zhou et al. (2021), pareidolia faces (face-like non-face objects) are likely to be an exceptional case. They measured bidirectional binocular rivalry in a blocked design, similar to the discrete condition used in the current study. They reported that the face-like non-face object could enter visual awareness in a similar fashion to genuine faces but remain in awareness in a similar fashion to common non-face objects. We could infer from their results that: when compared to genuine faces, the pareidolia faces would have a similar breakthrough threshold but a higher suppression threshold; when compared to common objects, the pareidolia faces would have a similar suppression threshold but a low breakthrough threshold. In this case, the difference between these two thresholds for pareidolia faces would be larger than either for genuine faces or common objects. Thus, it would be important for the authors to discuss the boundary between the findings and the inferences.

This is correct. We acknowledge that our sampling of image-categories is limited, and have added a treatment of this limitation in our discussion. We have expanded on the particular case of Zhou et al (2021), and the possibility of the asymmetries suggested:

Lines 669 – 691: “As a reminder, we explicitly tested image types that in other studies have shown differential susceptibility to CFS attributed to some form of expedited unconscious processing. Nevertheless, one could argue that our failure to obtain evidence for category specific suppression depth is based on the limited range of image categories sampled in this study. We agree it would be informative to broaden the range of image types tested using tCFS to include images varying in familiarity, congruence and affect. We can also foresee value in deploying tCFS to compare bCFS and reCFS thresholds for visual targets comprising physically meaningless ‘tokens’ whose global configurations can synthesise recognizable perceptual impressions. To give a few examples, dynamic configurations of small dots varying in location over time can create the compelling impression of rotational motion of a rigid, 3D object (structure from motion) or of a human engaged in given activity (biological motion) (Grossmann & Dobbins, 2006; Watson et al., 2004). These kinds of visual stimuli are associated with neural processing in higher-tier visual areas of the human brain, including the superior occipital lateral region (e.g., Vanduffel et al., 2002) and the posterior portion of the superior temporal sulcus (e.g., Grossman et al., 2000). These kinds of perceptually meaningful impressions of objects from rudimentary stimulus tokens are capable of engaging binocular rivalry. Such stimuli would be particularly useful in assessing high-level processing in CFS because they can be easily manipulated using phase-scrambling to remove the global percept without altering low-level stimulus properties. In a similar vein, small geometric shapes can be configured so as to resemble human or human-like faces, such as those used by (Zhou et al., 2021)[1]. These kinds of faux faces could be used in concert with tCFS to compare suppression depth with that associated with actual faces.

[1] Zhou et al. (2021) derived dominance and suppression durations with fixed-contrast images. In their study, genuine face images and faux faces remained suppressed for equivalent durations whereas genuine faces remained dominant significantly longer than did faux faces. The technique used by those investigators - interocular flash suppression (Wolfe, 1994) - is quite different from CFS in that it involves abrupt, asynchronous presentation of dissimilar stimuli to the two eyes. It would be informative to repeat their experiment using the tCFS procedure.

Reviewer #2 (Public Review):

Summary

The paper introduces a valuable method, tCFS, for measuring suppression depth in continuous flash suppression (CFS) experiments. tCFS uses a continuous-trial design instead of the discrete trials standard in the literature, resulting in faster, better controlled, and lower-variance estimates. The authors measured suppression depth during CFS for the first time and found similar suppression depths for different image categories. This finding provides an interesting contrast to previous results that breakthrough thresholds differ for different image categories and refine inferences of subconscious processing based solely on breakthrough thresholds. However, the paper overreaches by claiming breakthrough thresholds are insufficient for drawing certain conclusions about subconscious processing.

We agree that breakthrough thresholds can provide useful information to draw conclusions about unconscious processing – as our procedure is predicated on breakthrough thresholds. Our key point is that breakthrough provides only half of the needed information.

We have amended our manuscript thoroughly (detailed below) to accommodate this nuance and avoid this overreaching claim.

Strengths

(1) The tCFS method, by using a continuous-trial design, quickly estimates breakthrough and re-suppression thresholds. Continuous trials better control for slowly varying factors such as adaptation and attention. Indeed, tCFS produces estimates with lower across-subject variance than the standard discrete-trial method (Fig. 2). The tCFS method is straightforward to adopt in future research on CFS and binocular rivalry.

(2) The CFS literature has lacked re-suppression threshold measurements. By measuring both breakthrough and re-suppression thresholds, this work calculated suppression depth (i.e., the difference between the two thresholds), which warrants different interpretations from the breakthrough threshold alone.

(3) The work found that different image categories show similar suppression depths, suggesting some aspects of CFS are not category-specific. This result enriches previous findings that breakthrough thresholds vary with image categories. Re-suppression thresholds vary symmetrically, such that their differences are constant.

Thank you for this positive and succinct summary of our contribution. We have adopted your 3rd point “... suggesting that some aspects...” in our revised manuscript to more appropriately treat the ways that bCFS and reCFS thresholds may interact with suppression depths. For example:

Lines 850 – 852: “These [low level] factors could be parametrically varied to examine specifically whether they modulate bCFS thresholds alone, or whether they also cause a change in suppression depth by asymmetrically affecting reCFS thresholds”.

Weaknesses

(1) The results and arguments in the paper do not support the claim that 'variations in breakthrough thresholds alone are insufficient for inferring unconscious or preferential processing of given image categories,' to take one example phrasing from the abstract. The same leap in reasoning recurs on lines 28, 39, 125, 566, 666, 686, 759, etc.

We have thoroughly updated our manuscript with respect to mentions of preferential processing, to avoid this leap in reasoning throughout. For example, this phrase in the abstract now reads:

Lines 27-30: “More fundamentally, it shows that variations in bCFS thresholds alone are insufficient for inferring whether the barrier to achieving awareness exerted by interocular suppression is weaker for some categories of visual stimuli compared to others”.

Take, for example, the arguments on lines 81-83. Grant that images are inequivalent, and this explains different breakthrough times. This is still no argument against differential subconscious processing. Why are images non-equivalent? Whatever the answer, does it qualify as 'residual processing outside of awareness'? Even detecting salience requires some processing. The authors appear to argue otherwise on lines 694-696, for example, by invoking the concept of effective contrasts, but why is effective contrast incompatible with partial processing? Again, does detecting (effective) contrast not involve some processing? The phrases 'residual processing outside of awareness' and 'unconscious processing' are broad enough to encompass bottom-up salience and effective contrast. Salience and (effective) contrast are arguably uninteresting, but that is a different discussion. The authors contrast 'image categories' or semantics with 'low-level factors.' In my opinion, this is a clearer contrast worth emphasizing more. However, semantic processing is not equal to subconscious processing writ large.

We are in agreement with your analysis that differential subconscious processing may contribute to differences between images, and have updated our manuscript to clarify this possibility. In particular, we have now included a section in our Discussion which offers a suggestion for future research, linking sensitivity to different low-level image features with differences in gain of the respective contrast-response functions.

From Lines 692 – 722: “Next we turn to another question raised about our conclusion concerning invariant depth of suppression: If certain image types have overall lower bCFS and reCFS contrast thresholds relative to other image types, does that imply that images in the former category enjoy “preferential processing” relative to those in the latter? Given the fixed suppression depth, what might determine the differences in bCFS and reCFS thresholds? Figure 3 shows that polar patterns tend to emerge from suppression at slightly lower contrasts than do gratings and that polar patterns, once dominant, tend to maintain dominance to lower contrasts than do gratings and this happens even though the rate of contrast change is identical for both types of stimuli. But while rate of contrast change is identical, the neural responses to those contrast changes may not be the same: neural responses to changing contrast will depend on the neural contrast response functions (CRFs) of the cells responding to each of those two types of stimuli, where the CRF defines the relationship between neural response and stimulus contrast. CRFs rise monotonically with contrast and typically exhibit a steeply rising initial response as stimulus contrast rises from low to moderate values, followed by a reduced growth rate for higher contrasts. CRFs can vary in how steeply they rise and at what contrast they achieve half-max response. CRFs for neurons in mid-level vision areas such as V4 and FFA (which respond well to polar stimuli and faces, respectively) are generally steeper and shifted towards lower contrasts than CRFs for neurons in primary visual cortex (which responds well to gratings). Therefore, the effective strength of the contrast changes in our tCFS procedure will depend on the shape and position of the underlying CRF, an idea we develop in more detail in Supplementary Appendix 1, comparing the case of V1 and V4 CRFs. Interestingly, the comparison of V1 and V4 CRFs shows two interesting points: (i) that V4 CRFs should produce much lower bCFS and reCFS thresholds than V1 CRFs, and (ii) that V4 CRFs should produce more suppression than V1 CRFs. Our data do not support either prediction: Figure 3 shows that bCFS and reCFS thresholds are very similar for all image categories and suppression depth is uniform. There is no room in these results to support the claim that certain images receive “preferential processing” or processing outside of awareness, although there are many other kinds of images still to be tested and exceptions may potentially be found. As a first step in exploring this idea, one could use standard psychophysical techniques (e.g., (Ling & Carrasco, 2006)) to derive CRFs for different categories of patterns and then measure suppression depth associated with those patterns using tCFS.”

We have also expanded on this nuanced line of reasoning in a new Supplementary Appendix for the interested reader.

The preceding does not detract from the interest in finding uniform suppression depth. Suppression depth and absolute bCFS can conceivably be due to orthogonal mechanisms warranting their own interpretations. In fact, the authors briefly take this position in the Discussion (lines 696-704, 'A hybrid model ...'). The involvement of different mechanisms would defeat the argument on lines 668-670.

We agree with this analysis, and note our response to Reviewer 1 and the possibility of exceptional cases that may affect absolute bCFS or reCFS thresholds independently.

Similarly, we agree with the notion that some aspects of CFS may not be category specific. The symmetric relationship of thresholds for a given category of stimuli should be assessed in the context of other categories, such as with pontillist images and by incorporating semantic features of images into the mask as in Che et al. (2019) and Han et al. (2021). This line of reasoning and suggestions for future research is provided in the revised discussion, beginning:

Lines 67: “Nevertheless, one could argue that our failure to obtain evidence for category specific suppression depth is based on a limited range of image categories….”

(2) These two hypotheses are confusing and should be more clearly distinguished: a) varying breakthrough times may be due to low-level factors (lines 76-79); b) uniform suppression depth may also arise from early visual mechanisms (e.g., lines 25-27).

Thank you for highlighting this opportunity for clarification. We have updated our text:

Lines 25 – 27: “This uniform suppression depth points to a single mechanism of CFS suppression, one that likely occurs early in visual processing, because suppression depth was not modulated by target salience or complexity”

Lines 78 – 79: “Sceptics argue, however, that differences in breakthrough times can be attributed to low-level factors such as spatial frequency, orientation and contrast that vary between images”

Neutral remarks

The depth between bCFS and reCFS depended on measurement details such as contrast change speed and continuous vs. discrete trials. With discrete trials, the two thresholds showed inverse relations (i.e., reCFS > bCFS) in some participants. The authors discuss possible reasons at some length (adaptation, attention, etc. ). Still, a variable measure does not clearly indicate a uniform mechanism.

We have ensured our revised manuscript makes no mention of a uniform mechanism, although we frequently mention our result of uniform suppression depth.

Reviewer #3 (Public Review):

Summary:

In the 'bCFS' paradigm, a monocular target gradually increases in contrast until it breaks interocular suppression by a rich monocular suppressor in the other eye. The present authors extend the bCFS paradigm by allowing the target to reduce back down in contrast until it becomes suppressed again. The main variable of interest is the contrast difference between breaking suppression and (re) entering suppression. The authors find this difference to be constant across a range of target types, even ones that differ substantially in the contrast at which they break interocular suppression (the variable conventionally measured in bCFS). They also measure how the difference changes as a function of other manipulations. Interpretation in terms of the processing of unconscious visual content, as well as in terms of the mechanism of interocular suppression.

Thank you for your positive assessment of our methodology.

Strengths:

Interpretation of bCFS findings is mired in controversy, and this is an ingenuous effort to move beyond the paradigm's exclusive focus on breaking suppression. The notion of using the contrast difference between breaking and entering suppression as an index of suppression depth is interesting, but I also feel like it can be misleading at times, as detailed below.

Weaknesses:

Here's one doubt about the 'contrast difference' measure used by the authors. The authors seem confident that a simple subtraction is meaningful after the logarithmic transformation of contrast values, but doesn't this depend on exactly what shape the contrast-response function of the relevant neural process has? Does a logarithmic transformation linearize this function irrespective of, say, the level of processing or the aspect of processing that we're talking about?

Given that stimuli differ in terms of the absolute levels at which they break (and re-enter) suppression, the linearity assumption needs to be well supported for the contrast difference measure to be comparable across stimuli.

Our motivation to quantify suppression depth after log-transform to decibel scale was two-fold. First, we recognised that the traditional use of a linear contrast ramp in bCFS is at odds with the well-characterised profile of contrast discrimination thresholds which obey a power law (Legge, 1981) and the observations that neural contrast response functions show the same compressive non-linearity in many different cortical processing areas (e.g.: V1, V2, V3, V4, MT, MST, FST, TEO. See (Ekstrom et al., 2009)). Increasing contrast in linear steps could thus lead to a rapid saturation of the response function, which may account for the overshoot that has been reported in many canonical bCFS studies. For example, in (Jiang et al., 2007), target contrast reached 100% after 1 second, yet average suppression times for faces and inverted faces were 1.36 and 1.76 seconds respectively. As contrast response functions in visual neurons saturate at high contrast, the upper levels of a linear contrast ramp have less and less effect on the target's strength. This approach to response asymptote may have exaggerated small differences between stimulus conditions and may have inflated some previously reported differences. In sum, the use of a log-transformed contrast ramp allows finer increments in contrast to be explored before saturation, a simple manipulation which we hope will be adopted by our field.

Second, by quantifying suppression depth as a decibel change we enable the comparison of suppression depth between experiments and laboratories, which inevitably differ in presentation environments. As a comparison, a reaction-time for bCFS of 1.36 s can not easily be compared without access to near-identical stimulation and testing environments. In addition once ramp contrast is log transformed it effectively linearises the neural contrast response function. This means that comparing different studies that use different contrast levels for masker or target can be directly compared because a given suppression depth (for example, 15 dB) is the same proportionate difference between bCFS and reCFS regardless of the contrasts used in the particular study.

We also acknowledge that different stimulus categories may engage neural and visual processing associated with different contrast gain values (e.g., magno- vs parvo-mediated processing). But the breaks and returns to suppression of a given stimulus category would be dependent on the same contrast gain function appropriate for that stimulus which thus permits their direct comparison. Indeed, this is why our novel approach offers a promising technique for comparing suppression depth associated with various stimulus categories (a point mentioned above). Viewed in this way, differences in actual durations of break times (such as we report in our paper) may tell us more about differences in gain control within neural mechanisms responsible for processing of those categories.

We have now included a summary of these arguments in a new paragraph of our discussion (from lines 696- cf Reviewer 2 above), as well as a new Supplementary Appendix.

Here's a more conceptual doubt. The authors introduce their work by discussing ambiguities in the interpretation of bCFS findings with regard to preferential processing, unconscious processing, etc. A large part of the manuscript doesn't really interpret the present 'suppression depth' findings in those terms, but at the start of the discussion section (lines 560-567) the authors do draw fairly strong conclusions along those lines: they seem to argue that the constant 'suppression depth' value observed across different stimuli argues against preferential processing of any of the stimuli, let alone under suppression. I'm not sure I understand this reasoning. Consider the scenario that the visual system does preferentially process, say, emotional face images, and that it does so under suppression as well as outside of suppression. In that scenario, one might expect the contrast at which such a face breaks suppression to be low (because the face is preferentially processed under suppression) and one might also expect the contrast at which the face enters suppression to be low (because the face is preferentially processed outside of suppression). So the difference between the two contrasts might not stand out: it might be the same as for a stimulus that is not preferentially processed at all. In sum, even though the author's label of 'suppression depth' on the contrast difference measure is reasonable from some perspectives, it also seems to be misleading when it comes to what the difference measure can actually tell us that bCFS cannot.

We have addressed this point with respect to the differences between suppression depth and overall value of contrast thresholds in our revised discussion (reproduced above), and supplementary appendix.

The authors acknowledge that non-zero reaction time inflates their 'suppression depth' measure, and acknowledge that this inflation is worse when contrast ramps more quickly. But they argue that these effects are too small to explain either the difference between breaking contrast and re-entering contrast to begin with, or the increase in this difference with the contrast ramping rate. I agree with the former: I have no doubt that stimuli break suppression (ramping up) at a higher contrast than the one at which they enter suppression (ramping down). But about the latter, I worry that the RT estimate of 200 ms may be on the low side. 200 ms may be reasonable for a prepared observer to give a speeded response to a clearly supra-threshold target, but that is not the type of task observers are performing here. One estimate of RT in a somewhat traditional perceptual bistability task is closer to 500 ms (Van Dam & Van Ee, Vis Res 45 2005), but I am uncertain what a good guess is here. Bottom line: can the effect of contrast ramping rate on 'suppression depth' be explained by RT if we use a longer but still reasonable estimated RT than 200 ms?

A 500 ms reaction time estimate would not account for the magnitude of the changes observed in Experiment 3. Suppression depths in our slow, medium, and fast contrast ramps were 9.64 dB, 14.64 dB and 18.97 dB, respectively (produced by step sizes of .035, .07 and .105 dB per video frame at 60 fps). At each rate, assuming a 500 ms reaction time for both thresholds would capture a change of 2.1 dB, 4.2 dB, 6.3 dB. This difference cannot account for the size of the effects observed between our different ramp speeds. Note that any critique using the RT argument also applies to all other bCFS studies which inevitably will have inflated breakthrough points for the same reason.

We’ve updated our discussion with this more conservative estimate:

Lines 744 – 747: “For example, if we assume an average reaction time of 500 ms for appearance and disappearance events, then suppression depth will be inflated by ~4.2 dB at the rate of contrast change used in Experiments 1 and 2 (.07 dB per frame at 60 fps). This cannot account for suppression depth in its entirety, which was many times larger at approximately 14 dB across image categories.”

Lines 755 – 760: [In Experiment 3] “Using the same assumptions of a 500 ms response time delay, this would predict a suppression depth of 2.1 dB, 4.2 dB and 6.3 dB for the slow, medium and fast ramp speeds respectively. However, this difference cannot account for the size of the effects (Slow 9.64 dB, Medium 14.6 dB, Fast 18.97 dB). The difference in suppression depth based on reaction-time delays (± 2.1 dB) also does not match with our empirical data (Medium - Slow = 4.96 dB; Fast - Medium = 4.37 dB)”

A second remark about the 'ramping rate' experiment: if we assume that perceptual switches occur with a certain non-zero probability per unit time (stochastically) at various contrasts along the ramp, then giving the percept more time to switch during the ramping process will lead to more switches happening at an earlier stage along the ramp. So: ramping contrast upward more slowly would lead to more switches at relatively low contrast, and ramping contrast downward more slowly would lead to more switches at relatively high contrasts. This assumption (that the probability of switching is non-zero at various contrasts along the ramp) seems entirely warranted. To what extent can that type of consideration explain the result of the 'ramping rate' experiment?

We agree that for a given ramp speed there is a variable probability of a switch in perceptual state for both bCFS and reCFS portions of the trial. To put it in other words, for a given ramp speed and a given observer the distribution of durations at which transitions occur will exhibit variance. We see that variance in our data (just as it’s present in conventional binocular rivalry duration histograms), as a non-zero probability of switches at very short durations (for example). One might surmise that slower ramp speeds would afford more opportunity for stochastic transitions to occur and that the measured suppression depths for slow ramps are underestimates of the suppression depth produced by contrast adaptation. Yet by the same token, the same underestimation would occur during fast ramp speeds, indicating that that difference may be even larger than we reported. In our revision we will spell this out in more detail, and indicate that a non-zero probability of switches at any time may lead to an underestimation of all recorded suppression depths.

In our data, we believe the contribution of these stochastic switches are minimal. Our current Supplementary Figure 1(d) indicates that there is a non-zero probability of responses early in each ramp (e.g. durations < 2 seconds), yet these are a small proportion of all percept durations. This small proportion is clear in the empirical cumulative density function of percept durations, which we include below. Notably, during slow-ramp conditions, average percept durations actually increased, implying a resistance to any effect of early stochastic switching.

Author response image 1.

The data from Supplementary FIgure 1D. (right) Same data reproduced as a cumulative density function. The non-zero probability of a switch occurring (for example at very short percept durations) is clear, but a small proportion of all switches. Notably, In slow ramp trials, there is more time for this stochastic switching to occur, which should underestimate the overall suppression depth. Yet during slow-ramp conditions, average percept durations increased (vertical arrows), implying a resistance to any effect of early stochastic switching.

When tying the 'dampened harmonic oscillator' finding to dynamic systems, one potential concern is that the authors are seeing the dampened oscillating pattern when plotting a very specific thing: the amount of contrast change that happened between two consecutive perceptual switches, in a procedure where contrast change direction reversed after each switch. The pattern is not observed, for instance, in a plot of neural activity over time, threshold settings over time, etcetera. I find it hard to assess what the observation of this pattern when representing a rather unique aspect of the data in such a specific way, has to do with prior observations of such patterns in plots with completely different axes.

We acknowledge that fitting the DHO model to response order (rather than time) is a departure from previous investigations modelling oscillations over time. Our alignment to response order was a necessary step to avoid the smearing which occurs due to variation in individual participant threshold durations.

Our Supplementary Figure 1 shows the variation in participant durations for the three rates of contrast change. From this pattern we can expect that fitting the DHO to perceptual changes over time would result in the poorest fit for slow rates of change (with the largest variation in durations), and best fit for fast rates of change (with least variation in durations).

That is indeed what we see, reproduced in the review figure below. We include this to show the DHO is still applicable to perceptual changes over time when perceptual durations have relatively low variance (in the fast example), but not the alternate cases. Thus the DHO is not only produced by our alignment to response number - but this step is crucial to avoid the confound of temporal smearing when comparing between conditions.

Author response image 2.

DHO fit to perceptual thresholds over time. As a comparison to manuscript Figure 5 (aligning to response order), here we display the raw detrended changes in threshold over time per participant, and their average. Individual traces are shown in thin lines, the average is thick. Notably, in the slow and medium conditions, when perceptual durations had relatively high variance, the DHO is a poor fit to the average (shown in pink). The DHO is still an excellent fit in fast conditions, when modelling changes in threshold over time, owing to the reduced variance in perceptual durations (cf. Supplementary Figure 1). As a consequence, to remove the confound of individual participant durations, we have fitted the DHO when aligned to response order in our manuscript.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

The terminology used: "suppression depth". The depth of interocular suppression indexed by detection threshold has long been used in the literature, such as in Tsuchiya et al., 2006. I notice that this manuscript has created a totally different manipulative definition of the depth of suppression, the authors should make this point clear to the readers to avoid confusion.

We believe that our procedure does not create a new definition for suppression depth, but rather utilises the standard definition used for many years in the binocular rivalry literature: the ratio between a threshold measured for a target while it is in the state of suppression and for that same target when in the dominance state.

We have now revised our introduction to make the explicit continuation from past methods to our present methodology clear:

Lines 94 – 105: “One method for measuring interocular suppression is to compare the threshold for change-detection in a target when it is monocularly suppressed and when it is dominant, an established strategy in binocular rivalry research (Alais, 2012; Alais et al., 2010; Alais & Melcher, 2007; Nguyen et al., 2003). Probe studies using contrast as the dependent variable for thresholds measured during dominance and during suppression can advantageously standardise suppression depth in units of contrast within the same stimulus (e.g., Alais & Melcher, 2007; Ling et al., 2010). Ideally, the change should be a temporally smoothed contrast increment to the rival image being measured (Alais, 2012), a tactic that preludes abrupt onset transients and, moreover, provides a natural complement to the linear contrast ramps that are standard in bCFS research. In this study, we measure bCFS thresholds as the analogue of change-detection during suppression, and as their complement, record thresholds for returns to suppression (reCFS).”

The paper provides a new method to measure CFS bidirectionally. Given the possible exceptional case of pareidolia faces, it would be important to discuss how the bidirectional measurement offers more information, e.g., how the bottom-up and top-down factors would be involved in the breakthrough phase and the re-suppression phase.

In our discussion, we have now included the possibility of exceptional cases (such as pareidolia faces), and how an asymmetry may arise with respect to separate image categories affecting either bCFS or reCFS thresholds orthogonally.

Lines 688 - 691: “...In a similar vein, small geometric shapes can be configured so as to resemble human faces, such as those used by Zhou et al. (2021)[footnote]. These kinds of faux faces could be used in concert with tCFS to compare suppression depth with that associated with actual faces.

[footnote] Zhou et al. (2021) derived dominance and suppression durations with fixed-contrast images. In their study, genuine face images and faux faces remained suppressed for equivalent durations whereas genuine faces remained dominant significantly longer than did faux faces. The technique used by those investigators - interocular flash suppression (Wolfe, 1994) - is quite different from CFS in that it involves abrupt, asynchronous presentation of dissimilar stimuli to the two eyes. It would be informative to repeat their experiment using the tCFS procedure.”

What makes the individual results in the discrete condition much less consistent than the tCFS (in Figure 2c)? The authors discussed that motivation or attention to the task would change between bCFS and reCFS blocks (Line 589). But this point is not clear. Does not the attention to task also fluctuate in the tCFS paradigm, as the target continuously comes and goes?

We believe the discrete conditions have greater variance owing to the blocked design of the discrete conditions. A sequence of bCFS thresholds was collected in order (over ~15 mins), before switching to a sequence of back-to-back discrete reCFS thresholds (another ~15 mins), or a sequence of the tCFS condition. As the order of these blocks was randomized, thresholds collected in the discrete bCFS vs reCFS blocks could be separated by many minutes. In contrast, during tCFS, every bCFS threshold used to calculate the average is accompanied by a corresponding reCFS threshold collected within the same trial, separated by seconds. Thus the tCFS procedure naturally controls for waxing and waning attention, as within every change in attention, both thresholds are recorded for comparison.

A second advantage is that because the tCFS design changes contrast based on visibility, targets spend more time close to the threshold governing awareness. This reduced distance to thresholds remove the opportunity for other influences (such as oculomotor influences, blinks, etc), from introducing variance into the collected thresholds.

Experiment 3 reported greater suppression depth with faster contrast change. Because the participant's response was always delayed (e.g., they report after they become aware that the target has disappeared), is it possible that the measured breakthrough threshold gets lower, the re-suppression threshold gets higher, just because the measuring contrast is changing faster?

We have included an extended discussion of the contribution of reaction-times to the differences in suppression depth we report. Importantly, even a conservative reaction time of 500 ms, for both bCFS and reCFS events, cannot account for the difference in suppression depth between conditions.

Lines 755 – 760> “Using the same assumptions of a 500 ms response time delay, this would predict a suppression depth of 2.1 dB, 4.2 dB and 6.3 dB for the slow, medium and fast ramp speeds respectively. However, this difference cannot account for the size of the effects (Slow 9.64 dB, Medium 14.6 dB, Fast 18.97 dB). The difference in suppression depth based on reaction-time delays (± 2.1 dB) also does not match with our empirical data (Medium - Slow = 4.96 dB; Fast - Medium = 4.37 dB).”

In the current manuscript, some symbols are not shown properly (lines 145, 148, 150, 303).

Thank you for pointing this out, we will arrange with the editors to fix the typos.

Reviewer #2 (Recommendations For The Authors):

Line 13: 'time needed'-> contrast needed?

This sentence was referring to previous experiments which predominantly focus on the time of breakthrough.

Line 57: Only this sentence uses saliency; everywhere else in the paper uses salience.

We have updated to salience throughout.

Fig. 1c: The higher variance in discrete measurement results may be due to more variation in discrete trials, e.g., trial duration and inter-trial intervals (ITIs). Tighter control is indeed one advantage of the continuous tCFS design. For the discrete condition, it would help to report more information about variation across trials. How long and variable are the trials? The ITIs? This information is also relevant to the hypothesis about adaptation in Experiment 3.

In the discrete condition, each trial ended after the collection of a single response. Thus the variability of the trials is the same as the variability of the contrast thresholds reported in Figure 2. The distribution of these ‘trials’ (aka percept durations), is also shown in Supplementary Figure 1.

The ITI between discrete trials was self-paced, and not recorded during the experiment.

Line 598: 'equivalently' is a strong word. The benefit is perhaps best stated relatively: bCFS and reCFS are measured under closer conditions (e.g., adaptation, attention) with continuous experiments compared to discrete ones.

We agree - and have amended our manuscript:

Lines 629 – 632: “Alternating between bCFS/reCFS tasks also means that any adaptation occurring over the trial will occur in close proximity to each threshold, as will any waning of attention. The benefit being that bCFS and reCFS thresholds are measured under closer conditions in continuous trials, compared to discrete ones.”

Reviewer #3 (Recommendations For The Authors):

Figure 1 includes fairly elaborate hypothetical results and how they would be interpreted by the authors, but I didn't really see any mention of this content in the main text. It wasn't until I started reading the caption that I figured it out. A more elaborate reference to the figure would prevent readers from overlooking (part of) the figure's message.

We have now made it clearer in the text that those details are contained in the caption to Figure 1.

Lines 113 – 115: “Figure 1 outlines hypothetical results that can be obtained when recording reCFS thresholds as a complement to bCFS thresholds in order to measure suppression depth.”

A piece of text seems to have been accidentally removed on line 267.

Thank you, this has now been amended

Author Response

The following is the authors’ response to the original reviews.

eLife assessment

The authors have developed a compelling coarse-grained simulation approach for nucleosome-nucleosome interactions within a chromatin array. The data presented are solid and provide new insights that allow for predictions of how chromatin interactions might occur in vivo, but some of the claims should be tempered. The tools will be valuable for the chromosome biology field.

Response: We want to thank the editors and all the reviewers for their insightful comments. We have made substantial changes to the manuscript to improve its clarity and temper necessary claims, as detailed in the responses, and we performed additional analyses to address the reviewers’ concerns. We believe that we have successfully addressed all the comments, and the quality of our paper has improved significantly.

In the following, we provide point-to-point responses to all the reviewer comments. 

RESPONSE TO REFEREE 1:

Comment 0: This study develops and applies a coarse-grained model for nucleosomes with explicit ions. The authors perform several measurements to explore the utility of a coarse-grained simulation method to model nucleosomes and nucleosome arrays with explicit ions and implicit water. ’Explicit ions’ means that the charged ions are modeled as particles in simulation, allowing the distributions and dynamics of ions to be measured. Since nucleosomes are highly charged and modulated by charge modifications, this innovation is particularly relevant for chromatin simulation.

Response: We thank the reviewer’s excellent summary of the work.

Comment 1: Strengths: This simulation method produces accurate predictions when compared to experiments for the binding affinity of histones to DNA, counterion interactions, nucleosome DNA unwinding, nucleosome binding free energies, and sedimentation coefficients of arrays. The variety of measured quantities makes both this work and the impact of this coarse-grained methodology compelling. The comparison between the contributions of sodium and magnesium ions to nucleosome array compaction, presented in Figure 3, was exciting and a novel result that this simulation methodology can assess.

Response: We appreciate the reviewer’s strong assessment of the paper’s significance, novelty, and broad interest, and we thank him/her for the detailed suggestions and comments.

Comment 2: Weaknesses: The presentation of experimental data as representing in vivo systems is a simplification that may misrepresent the results of the simulation work. In vivo, in this context, typically means experimental data from whole cells. What one could expect for in vivo experimental data is measurements on nucleosomes from cell lysates where various and numerous chemical modifications are present. On the contrary, some of the experimental data used as a comparison are from in vitro studies. In vitro in this context means nucleosomes were formed ’in a test tube’ or under controlled conditions that do not represent the complexity of an in vivo system. The simulations performed here are more directly compared to in vitro conditions. This distinction likely impacts to what extent these simulation results are biologically relevant. In vivo and in vitro differences could be clarified throughout and discussed.

Response: As detailed in Response to Comment 3, we have made numerous modifications in the Introduction, Results, and Discussion Section to emphasize the differences between reconstituted and native nucleosomes. The newly added texts also delve into the utilization of the interaction strength measured for reconstituted nucleosomes as a reference point for conceptualizing the interactions among native nucleosomes.

Comment 3: In the introduction (pg. 3), the authors discuss the uncertainty of nucleosome-tonucleosome interaction strengths in vivo. For example, the authors discuss works such as Funke et al. However, Funke et al. used reconstituted nucleosomes from recombinant histones with one controlled modification (H4 acetylation). Therefore, this study that the authors discuss is measuring nucleosome’s in vitro affinity, and there could be significant differences in vivo due to various posttranslational modifications. Please revise the introduction, results section ”Close contacts drive nucleosome binding free energy,” and discussion to reflect and clarify the difference between in vitro and in vivo measurements. Please also discuss how biological variability could impact your findings in vivo. The works of Alexey Onufriev’s lab on the sensitivity of nucleosomes to charge changes (10.1016/j.bpj.2010.06.046, 10.1186/s13072-018-0181-5), such as some PTMs, are one potential starting place to consider how modifications alter nucleosome stability in vivo.

Response: We thank the reviewer for the insightful comments and agree that native nucleosomes can differ from reconstituted nucleosomes due to the presence of histone modifications.

We have revised the introduction to emphasize the differences between in vitro and in vivo nucleosomes. The new text now reads

"The relevance of physicochemical interactions between nucleosomes to chromatin organization in vivo has been constantly debated, partly due to the uncertainty in their strength [cite]. Examining the interactions between native nucleosomes poses challenges due to the intricate chemical modifications that histone proteins undergo within the nucleus and the variations in their underlying DNA sequences [cite]. Many in vitro experiments have opted for reconstituted nucleosomes that lack histone modifications and feature wellpositioned 601-sequence DNA to simplify the chemical complexity. These experiments aim to establish a fundamental reference point for understanding the strength of interactions within native nucleosomes. Nevertheless, even with reconstituted nucleosomes, a consensus regarding the significance of their interactions remains elusive. For example, using force-measuring magnetic tweezers, Kruithof et al. estimated the inter-nucleosome binding energy to be ∼ 14 kBT [cite]. On the other hand, Funke et al. introduced a DNA origamibased force spectrometer to directly probe the interaction between a pair of nucleosomes [cite], circumventing any potential complications from interpretations of single molecule traces of nucleosome arrays. Their measurement reported a much weaker binding free energy of approximately 2 kBT. This large discrepancy in the reported reference values complicates a further assessment of the interactions between native nucleosomes and their contribution to chromatin organization in vivo."

We modified the first paragraph of the results section to read

"Encouraged by the explicit ion model’s accuracy in reproducing experimental measurements of single nucleosomes and nucleosome arrays, we moved to directly quantify the strength of inter-nucleosomes interactions. We once again focus on reconstituted nucleosomes for a direct comparison with in vitro experiments. These experiments have yielded a wide range of values, ranging from 2 to 14 kBT [cite]. Accurate quantification will offer a reference value for conceptualizing the significance of physicochemical interactions among native nucleosomes in chromatin organization in vivo."

New text was added to the Discussion Section to emphasize the implications of simulation results for interactions among native nucleosomes.

"One significant finding from our study is the predicted strong inter-nucleosome interactions under the physiological salt environment, reaching approximately 9 kBT. We showed that the much lower value reported in a previous DNA origami experiment is due to the restricted nucleosomal orientation inherent to the device design. Unrestricted nucleosomes allow more close contacts to stabilize binding. A significant nucleosome binding free energy also agrees with the high forces found in single-molecule pulling experiments that are needed for chromatin unfolding [cite]. We also demonstrate that this strong inter-nucleosomal interaction is largely preserved at longer nucleosome repeat lengths (NRL) in the presence of linker histone proteins. While posttranslational modifications of histone proteins may influence inter-nucleosomal interactions, their effects are limited, as indicated by Ding et al. [cite], and are unlikely to completely abolish the significant interactions reported here. Therefore, we anticipate that, in addition to molecular motors, chromatin regulators, and other molecules inside the nucleus, intrinsic inter-nucleosome interactions are important players in chromatin organization in vivo."

The suggested references (10.1016/j.bpj.2010.06.046, 10.1186/s13072-018-0181-5) are now included as citations # 44 and 45.

Comment 4: Due to the implicit water model, do you know if ions can penetrate the nucleosome more? For example, does the lack of explicit water potentially cause sodium to cluster in the DNA grooves more than is biologically relevant, as shown in Figure 1?

Response: We thank the reviewer for the insightful comments. The parameters of the explicit-ion model were deduced from all-atom simulations and fine-tuned to replicate crucial aspects of the local ion arrangements around DNA (1). The model’s efficacy was demonstrated in reproducing the radial distribution function of Na+ and Mg2+ ion distributions in the proximity of DNA (see Author response image 1). Consequently, the number of ions near DNA in the coarse-grained models aligns with that observed in all-atom simulations, and we do not anticipate any significant, unphysical clustering. It is worth noting that previous atomistic simulations have also reported the presence of a substantial quantity of Na+ ions in close proximity to nucleosomal DNA (refer to Author response image 2).

Author response image 1.

Comparison between the radial distribution functions of Na+ (left) and Mg2+ (right) ions around the DNA phosphate groups computed from all-atom (black) and coarse-grained (red) simulations. Figure adapted from Figure 4 of Ref. 1. The coarse-grained explicit ion model used in producing the red curves is identical to the one presented in the current manuscript. (© 2011, AIP Publishing. This figure is reproduced with permission from Figure 4 in Freeman GS, Hinckley DM, de Pablo JJ (2011) A coarse-grain three-site-pernucleotide model for DNA with explicit ions. The Journal of Chemical Physics 135:165104. It is not covered by the CC-BY 4.0 license and further reproduction of this figure would need permission from the copyright holder.)

Author response image 2.

Three-dimensional distribution of sodium ions around the nucleosome determined from all-atom explicit solvent simulations. Darker blue colors indicate higher sodium density and high density of sodium ions around the DNA is clearly visible. The crystallographically identified acidic patch has been highlighted as spheres on the surface of the histone core and a high level of sodium condensation is observed around these residues. Figure adapted from Ref. 2. (© 2009, American Chemical Society. This figure is reproduced with permission from Figure 7 in Materese CK, Savelyev A, Papoian GA (2009) Counterion Atmosphere and Hydration Patterns near a Nucleosome Core Particle. J. Am. Chem. Soc. 131:15005–15013.. It is not covered by the CC-BY 4.0 license and further reproduction of this figure would need permission from the copyright holder.)

Comment 5: Histone side chain to DNA interactions, such as histone arginines to DNA, are essential for nucleosome stability. Therefore, can the authors provide validation or references supporting your model of the nucleosome with one bead per amino acid? I would like to see if the nucleosomes are stable in an extended simulation or if similar dynamic motions to all-atom simulations are observed.

Response: The nucleosome model, which employs one bead per amino acid and lacks explicit ions, has undergone extensive calibration and has found application in numerous prior studies. For instance, the de Pablo group utilized a similar model to showcase its ability to accurately replicate the experimentally measured nucleosome unwinding free energy penalty (3), sequence-dependent nucleosome sliding (4), and the interaction between two nucleosomes (5). Similarly, the Takada group employed a comparable model to investigate acetylation-modulated tri-nucleosome structures (6), chromatin structures influenced by chromatin factors (7), and nucleosome sliding (8). Our group also employed this model to study the structural rearrangement of a tetranucleosome (9) and the folding of larger chromatin systems (10). In cases where data were available, simulations frequently achieved quantitative reproduction of experimental results.

We added the following text to the manuscript to emphasize previous studies that validate the model accuracy.

"We observe that residue-level coarse-grained models have been extensively utilized in prior studies to examine the free energy penalty associated with nucleosomal DNA unwinding [cite], sequence-dependent nucleosome sliding [cite], binding free energy between two nucleosomes [cite], chromatin folding [cite], the impact of histone modifications on tri-nucleosome structures [cite], and protein-chromatin interactions [cite]. The frequent quantitative agreement between simulation and experimental results supports the utility of such models in chromatin studies. Our introduction of explicit ions, as detailed below, further extends the applicability of these models to explore the dependence of chromatin conformations on salt concentrations."

We agree that arginines are important for nucleosome stability. Since we assign positive charges to these residues, their contribution to DNA binding can be effectively captured. The model’s ability in reproducing nucleosome stability is supported by the good agreement between the simulated free energy penalty associated with nucleosomal DNA unwinding and experimental value estimated from single molecule experiments (Figure 1).

To further evaluate nucleosome stability in our simulations, we conducted a 200-ns-long simulation of a nucleosome featuring the 601-sequence under physiological salt conditions– 100 mM NaCl and 0.5 mM MgCl2, consistent with the conditions in Figure 1 of the main text. We found that the nucleosome maintains its overall structure during this simulation. The nucleosome’s radius of gyration (Rg) remained proximate to the value corresponding to the PDB structure (3.95 nm) throughout the entire simulation period (see Author response image 3).

Author response image 3.

Time trace of the radius of gyration (Rg) of a nucleosome with the 601-sequence along an unbiased, equilibrium trajectory. It is evident the Rg fluctuates around the value found in the PDB structure (3.95 nm), supporting the stability of the nucleosome in our simulation.

Occasional fluctuations in Rg corresponded to momentary, partial unwrapping of the nucleosomal DNA, a phenomenon observed in single-molecule experiments. However, we advise caution due to the coarse-grained nature of our simulations, which prevents a direct mapping of simulation timescale to real time. Importantly, the rate of DNA unwrapping in our simulations is notably overestimated.

It’s plausible that coarse-grained models, lacking side chains, might underestimate the barrier for DNA sliding along the nucleosome. Specifically, our model, without differentiation between interactions among various amino acids and nucleotides, accurately reproduces the average nucleosomal DNA binding affinity but may not capture the energetic variations among binding interfaces. Since sliding’s contribution to chromatin organization is minimal due to the use of strongly positioning 601 sequences, we imposed rigidity on the two nucleotides situated at the dyad axis to prevent nucleosomal DNA sliding. In future studies, enhancing the calibration of protein-DNA interactions to achieve improved sequence specificity would be an intriguing avenue. To underscore this limitation of the model, we have included the following text in the discussion section of the main text.

"Several aspects of the coarse-grained model presented here can be further improved. For instance, the introduction of specific protein-DNA interactions could help address the differences in non-bonded interactions between amino acids and nucleotides beyond electrostatics [cite]. Such a modification would enhance the model’s accuracy in predicting interactions between chromatin and chromatin-proteins. Additionally, the single-bead-per-amino-acid representation used in this study encounters challenges when attempting to capture the influence of histone modifications, which are known to be prevalent in native nucleosomes. Multiscale simulation approaches may be necessary [cite]. One could first assess the impact of these modifications on the conformation of disordered histone tails using atomistic simulations. By incorporating these conformational changes into the coarse-grained model, systematic investigations of histone modifications on nucleosome interactions and chromatin organization can be conducted. Such a strategy may eventually enable the direct quantification of interactions among native nucleosomes and even the prediction of chromatin organization in vivo."

Comment 6: The solvent salt conditions vary in the experimental reference data for internucleosomal interaction energies. The authors note, for example, that the in vitro data from Funke et al. differs the most from other measurements, but the solvent conditions are 35 mM NaCl and 11 mM MgCl2. Since this simulation method allows for this investigation, could the authors speak to or investigate if solvent conditions are responsible for the variability in experimental reference data? The authors conclude on pg. 8-9 and Figure 4 that orientational restraints in the DNA origami methodology are responsible for differences in interaction energy. Can the authors rule out ion concentration contributions?

Response: We thank the reviewer for the insightful comment. We would like to clarify that the black curve presented in Figure 4B of the main text was computed using the salt concentration specified by Funke et al. (35 mM NaCl and 11 mM MgCl2). Furthermore, there were no restraints placed on nucleosome orientations during these calculations. Consequently, the results in Figure 4B can be directly compared with the black curve in Figure 5C. The data in Figure 5C were calculated under physiological salt conditions (150 mM NaCl and 2 mM MgCl2), which are the standard solvent salt conditions used in most studies. It is worth noting that the free energy of nucleosome binding is significantly higher at the salt concentration employed by Funke et al. (14 kBT) than the value at the physiological salt condition (9 kBT). Therefore, comparing the results in Figure 4B and 5C eliminates ion concentration conditions as a potential cause for the the almost negligible result reported by Funke et al.

Comment 7: In the discussion on pg. 12 residual-level should be residue-level.

Response: We apologize for the oversight and have corrected the grammatical error in our manuscript.

RESPONSE TO REFEREE 2:

Comment 0: In this manuscript, the authors introduced an explicit ion model using the coarse-grained modelling approach to model the interactions between nucleosomes and evaluate their effects on chromatin organization. The strength of this method lies in the explicit representation of counterions, especially divalent ions, which are notoriously difficult to model. To achieve their aims and validate the accuracy of the model, the authors conducted coarse-grained molecular dynamics simulations and compared predicted values to the experimental values of the binding energies of protein-DNA complexes and the free energy profile of nucleosomal DNA unwinding and inter-nucleosome binding. Additionally, the authors employed umbrella sampling simulations to further validate their model, reproducing experimentally measured sedimentation coefficients of chromatin under varying salt concentrations of monovalent and divalent ions.

Response: We thank the reviewer’s excellent summary of the work.

Comment 1: The significance of this study lies in the authors’ coarse-grained model which can efficiently capture the conformational sampling of molecules while maintaining a low computational cost. The model reproduces the scale and, in some cases, the shape of the experimental free energy profile for specific molecule interactions, particularly inter-nucleosome interactions. Additionally, the authors’ method resolves certain experimental discrepancies related to determining the strength of inter-nucleosomal interactions. Furthermore, the results from this study support the crucial role of intrinsic physicochemical interactions in governing chromatin organization within the nucleus.

Response: We appreciate the reviewer’s strong assessment of the paper’s significance, novelty, and broad interest, and we thank him/her for the detailed suggestions and comments.

Comment 2: The method is simple but can be useful, given the authors can provide more details on their ion parameterization. The paper says that parameters in their ”potentials were tuned to reproduce the radial distribution functions and the potential of mean force between ion pairs determined from all-atom simulations.” However, no details on their all-atom simulations were provided; at some point, the authors refer to Reference 67 which uses all-atom simulations but does not employ the divalent ions. Also, no explanation is given for their modelling of protein-DNA complexes.

Response: We appreciate the reviewer’s suggestion on clarifying the parameterization of the explicition model. The parameterization was not carried out in reference 67 nor by us, but by the de Pablo group in citation 53. Specifically, ion potentials were parameterized to fit the potential of mean force between both monovalent and divalent ion pairs, calculated either from all-atom simulations or from the literature. The authors carried out extensive validations of the model parameters by comparing the radial distribution functions of ions computed using the coarse-grained model with those from all-atom simulations. Good agreements between coarse-grained and all-atom results ensure that the parameters’ accuracy in reproducing the local structures of ion interactions.

To avoid confusion, we have revised the text from:

"Parameters in these potentials were tuned to reproduce the radial distribution functions and the potential of mean force between ion pairs determined from all-atom simulations."

to

"Parameters in these potentials were tuned by Freeman et al. [cite] to reproduce the radial distribution functions and the potential of mean force between ion pairs determined from all-atom simulations."

We modified the Supporting Information at several places to clarify the setup and interpretation of protein-DNA complex simulations.

For example, we clarified the force fields used in these simulation with the following text

"All simulations were carried out using the software Lammps [cite] with the force fields defined in the previous two sections."

We added details on the preparation of these simulations as follows

"We carried out a series of umbrella-sampling simulations to compute the binding free energies of a set of nine protein-DNA complexes with experimentally documented binding dissociation constants [cite]. Initial configurations of these simulations were prepared using the crystal structures with the corresponding PDB IDs listed in Fig. S1."

We further revised the caption of Figure S1 (included as Author response image 4) to facilitate the interpretation of simulation results.

Author response image 4.

The explicit-ion model predicts the binding affinities of protein-DNA complexes well, related to Fig. 1 of the main text. Experimental and simulated binding free energies are compared for nine protein-DNA complexes [cite], with a Pearson Correlation coefficient of 0.6. The PDB ID for each complex is indicated in red, and the diagonal line is drawn in blue. The significant correlation between simulated and experimental values supports the accuracy of the model. To further enhance the agreement between the two, it will be necessary to implement specific non-bonded interactions that can resolve differences among amino acids and nucleotides beyond simple electrostatics. Such modifications will be interesting avenues for future research. See text Section: Binding free energy of protein-DNA complexes for simulation details.

Comment 3: Overall, the paper is well-written, concise and easy to follow but some statements are rather blunt. For example, the linker histone contribution (Figure 5D) is not clear and could be potentially removed. The result on inter-nucleosomal interactions and comparison to experimental values from Ref#44 is the most compelling. It would be nice to see if the detailed shape of the profile for restrained inter-nucleosomal interactions in Figure 4B corresponds to the experimental profile. Including the dependence of free energy on a vertex angle would also be beneficial.

Response: We thank the reviewer for the comments and agree that the discussion on linker histone results was brief. However, we believe the results are important and demonstrate our model’s advantage over mesoscopic approaches in capturing the impact of chromatin regulators on chromatin organization.

Therefore, instead of removing the result, we expanded the text to better highlight its significance, to help its comprehension, and to emphasize its biological implications. The image in Figure 5D was also redesigned to better visualize the cross contacts between nucleosomes mediated by histone H1. The added texts are quoted as below, and the new Figure 5 is included.

Author response image 5.

Revised main text Figure 5, with Figure 5D modified for improved visual clarity.

"Importantly, we found that the weakened interactions upon extending linker DNA can be more than compensated for by the presence of histone H1 proteins. This is demonstrated in Fig. 5C and Fig. S8, where the free energy cost for tearing part two nucleosomes with 167 bp DNA in the presence of linker histones (blue) is significantly higher than the curve for bare nucleosomes (red). Notably, at larger inter-nucleosome distances, the values even exceed those for 147 bp nucleosomes (black). A closer examination of the simulation configurations suggests that the disordered C-terminal tail of linker histones can extend and bind the DNA from the second nucleosome, thereby stabilizing the internucleosomal contacts (as shown in Fig. 5D). Our results are consistent with prior studies that underscore the importance of linker histones in chromatin compaction [cite], particularly in eukaryotic cells with longer linker DNA [cite]."

We further compared the simulated free energy profile, depicting the center of mass distance between nucleosomes, with the experimental profile, as depicted in Author response image 6. The agreement between the simulated and experimental results is evident. The nuanced features observed between 60 to 80 Ain the simulated profile stem from DNA unwinding˚ to accommodate the incoming nucleosome, creating a small energy barrier. It’s worth noting that such unwinding is unlikely to occur in the experimental setup due to the hybridization method used to anchor nucleosomes onto the DNA origami. Moreover, our simulation did not encompass configurations below 60 A, resulting in a lack of data in˚ that region within the simulated profile.

We projected the free energy profile onto the vertex angle of the DNA origami device, utilizing the angle between two nucleosome faces as a proxy. Once more, the simulated profile demonstrates reasonable agreement with the experimental data (Author response image 6). Author response image 6 has been incorporated as Figure S4 in the Supporting Information.

Author response image 6.

Explicit ion modeling reproduces the experimental free energy profiles of nucleosome binding. (A) Comparison between the simulated (black) and experimental (red) free energy profile as a function of the inter-nucleosome distance. Error bars were computed as the standard deviation of three independent estimates. The barrier observed between 60A and 80˚ A arises from the unwinding of nucleosomal DNA when the two nu-˚ cleosomes are in close proximity, as highlighted in the orange circle. (B) Comparison between the simulated (black) and experimental (red) free energy profile as a function of the vertex angle. Error bars were computed as the standard deviation of three independent estimates. (C) Illustration of the vertex angle Φ used in panel (B).

Comment 4: Another limitation of this study is that the authors’ model sacrifices certain atomic details and thermodynamic properties of the modelled systems. The potential parameters of the counter ions were derived solely by reproducing the radial distribution functions (RDFs) and potential of mean force (PMF) based on all-atom simulations (see Methods), without considering other biophysical and thermodynamic properties from experiments. Lastly, the authors did not provide any examples or tutorials for other researchers to utilize their model, thus limiting its application.

Response: We agree that residue-level coarse-grained modeling indeed sacrifices certain atomistic details. This sacrifice can be potentially limiting when studying the impact of chemical modifications, especially on histone and DNA methylations. We added a new paragraph in the Discussion Section to point out such limitations and the relevant text is quoted below.

"Several aspects of the coarse-grained model presented here can be further improved. For instance, the introduction of specific protein-DNA interactions could help address the differences in non-bonded interactions between amino acids and nucleotides beyond electrostatics [cite]. Such a modification would enhance the model’s accuracy in predicting interactions between chromatin and chromatin-proteins. Additionally, the single-bead-per-amino-acid representation used in this study encounters challenges when attempting to capture the influence of histone modifications, which are known to be prevalent in native nucleosomes. Multiscale simulation approaches may be necessary [cite]. One could first assess the impact of these modifications on the conformation of disordered histone tails using atomistic simulations. By incorporating these conformational changes into the coarse-grained model, systematic investigations of histone modifications on nucleosome interactions and chromatin organization can be conducted. Such a strategy may eventually enable the direct quantification of interactions among native nucleosomes and even the prediction of chromatin organization in vivo."

Nevertheless, it’s important to note that while the model sacrifices accuracy, it compensates with superior efficiency. Atomistic simulations face significant challenges in conducting extensive free energy calculations required for a quantitative evaluation of ion impacts on chromatin structures.

The explicit ion model, introduced by the de Pablo group, follows a standard approach adopted by other research groups, such as the parameterization of ion models using the potential of mean force from atomistic simulations (11; 12). According to multiscale coarse-graining theory, reproducing potential mean force (PMF) enables the coarsegrained model to achieve thermodynamic consistency with the atomistic model, ensuring identical statistical properties derived from them. However, it’s crucial to recognize that an inherent limitation of such approaches is their dependence on the accuracy of atomistic force fields in reproducing thermodynamic properties from experiments, as any inaccuracies in the atomistic force fields will similarly affect the resulting coarse-grained (CG) model.

We have provided the implementation of CG model and detailed instructions on setting up and performing simulations GitHub repository. Examples include simulation setup for a protein-DNA complex and for a nucleosome with the 601-sequence.

References [1] Freeman GS, Hinckley DM, de Pablo JJ (2011) A coarse-grain three-site-pernucleotide model for DNA with explicit ions. The Journal of Chemical Physics 135:165104.

[2] Materese CK, Savelyev A, Papoian GA (2009) Counterion Atmosphere and Hydration Patterns near a Nucleosome Core Particle. J. Am. Chem. Soc. 131:15005–15013.

[3] Lequieu J, Cordoba A, Schwartz DC, de Pablo JJ´ (2016) Tension-Dependent Free Energies of Nucleosome Unwrapping. ACS Cent. Sci. 2:660–666.

[4] Lequieu J, Schwartz DC, De Pablo JJ (2017) In silico evidence for sequence-dependent nucleosome sliding. Proc. Natl. Acad. Sci. U.S.A. 114.

[5] Moller J, Lequieu J, de Pablo JJ (2019) The Free Energy Landscape of Internucleosome Interactions and Its Relation to Chromatin Fiber Structure. ACS Cent. Sci. 5:341–348.

[6] Chang L, Takada S (2016) Histone acetylation dependent energy landscapes in trinucleosome revealed by residue-resolved molecular simulations. Sci Rep 6:34441.

[7] Watanabe S, Mishima Y, Shimizu M, Suetake I, Takada S (2018) Interactions of HP1 Bound to H3K9me3 Dinucleosome by Molecular Simulations and Biochemical Assays. Biophysical Journal 114:2336–2351.

[8] Brandani GB, Niina T, Tan C, Takada S (2018) DNA sliding in nucleosomes via twist defect propagation revealed by molecular simulations. Nucleic Acids Research 46:2788–2801.

[9] Ding X, Lin X, Zhang B (2021) Stability and folding pathways of tetra-nucleosome from six-dimensional free energy surface. Nat Commun 12:1091.

[10] Liu S, Lin X, Zhang B (2022) Chromatin fiber breaks into clutches under tension and crowding. Nucleic Acids Research 50:9738–9747.

[11] Savelyev A, Papoian GA (2010) Chemically accurate coarse graining of doublestranded DNA. Proc. Natl. Acad. Sci. U.S.A. 107:20340–20345.

[12] Noid WG (2013) Perspective: Coarse-grained models for biomolecular systems. The Journal of Chemical Physics 139:090901.

Author response:

The following is the authors’ response to the original reviews.

Response to Reviewer 1

Summary:

The authors introduce a denoising-style model that incorporates both structure and primary-sequence embeddings to generate richer embeddings of peptides. My understanding is that the authors use ESM for the primary sequence embeddings, take resolved structures (or use structural predictions from AlphaFold when they're not available), and then develop an architecture to combine these two with a loss that seems reminiscent of diffusion models or masked language model approaches. The embeddings can be viewed as ensemble-style embedding of the two levels of sequence information, or with AlphaFold, an ensemble of two methods (ESM+AlphaFold). The authors also gather external datasets to evaluate their approach and compare it to previous approaches. The approach seems promising and appears to out-compete previous methods at several tasks. Nonetheless, I have strong concerns about a lack of verbosity as well as the exclusion of relevant methods and references.

Thank you for the comprehensive summary. Regarding the concerns listed in the review below, we have made point-to-point response. We also modified our manuscript in accordance. 

Advances:

I appreciate the breadth of the analysis and comparisons to other methods. The authors separate tasks, models, and sizes of models in an intuitive, easy-to-read fashion that I find valuable for selecting a method for embedding peptides. Moreover, the authors gather two datasets for evaluating embeddings' utility for predicting thermostability. Overall, the work should be helpful for the field as more groups choose methods/pretraining strategies amenable to their goals, and can do so in an evidence-guided manner.

Thank you for recognizing the strength of our work in terms of the notable contributions, the solid analysis, and the clear presentation.

Considerations:

(1) Primarily, a majority of the results and conclusions (e.g., Table 3) are reached using data and methods from ProteinGym, yet the best-performing methods on ProteinGym are excluded from the paper (e.g., EVEbased models and GEMME). In the ProteinGym database, these methods outperform ProtSSN models. Moreover, these models were published over a year---or even 4 years in the case of GEMME---before ProtSSN, and I do not see justification for their exclusion in the text.

We decided to exclude the listed methods from the primary table as they are all MSA-based methods, which are considered few-shot methods in deep learning (Rao et al., ICML, 2021). In contrast, the proposed ProtSSN is a zero-shot method that makes inferences based on less information than few-shot methods. Moreover, it is possible for MSA-based methods to query aligned sequences based on predictions. For instance, Tranception (Notin et al., ICML, 2022) selects the model with the optimal proportions of logits and retrieval results according to the average correlation score on ProteinGym (Table 10, Notin et al., 2022).

With this in mind, we only included zero-shot deep learning methods in Table 3, which require no more than the sequence and structure of the underlying wild-type protein when scoring the mutants. In the revision, we have added the performance of SaProt to Table 3, and the performance of GEMME, TranceptEVE, and SaProt to Table 5. Furthermore, we have released the model's performance on the public leaderboard of ProteinGym v1 at proteingym.org.

(2) Secondly, related to the comparison of other models, there is no section in the methods about how other models were used, or how their scores were computed. When comparing these models, I think it's crucial that there are explicit derivations or explanations for the exact task used for scoring each method. In other words, if the pre-training is indeed an important advance of the paper, the paper needs to show this more explicitly by explaining exactly which components of the model (and previous models) are used for evaluation. Are the authors extracting the final hidden layer representations of the model, treating these as features, and then using these features in a regression task to predict fitness/thermostability/DDG etc.? How are the model embeddings of other methods being used, since, for example, many of these methods output a k-dimensional embedding of a given sequence, rather than one single score that can be correlated with some fitness/functional metric? Summarily, I think the text lacks an explicit mention of how these embeddings are being summarized or used, as well as how this compares to the model presented.

Thank you for the suggestion. Below we address the questions in three points. 

(1) The task and the scoring for each method. We followed your suggestion and added a new paragraph titled “Scoring Function” on page 9 to provide a detailed explanation of the scoring functions used by other deep learning zero-shot methods.

(2) The importance of individual pre-training modules. The complete architecture of the proposed ProtSSN model has been introduced on page 7-8. Empirically, the influence of each pre-training module on the overall performance has been examined through ablation studies on page 12. In summary, the optimal performance is achieved by combining all the individual modules and designs.

(3) The input of fitness scoring. For a zero-shot prediction task, the final score for a mutant will be calculated by wildly-used functions named log-odds ratio (for encoder models, including ours) or loglikelihood (for autoregressive models or inverse folding models. In the revision, we explicitly define these functions in sections “Inferencing” (page 7) and “Scoring Function” (page 9). 

(3) I think the above issues can mainly be addressed by considering and incorporating points from Li et al. 2024[1] and potentially Tang & Koo 2024[2]. Li et al.[1] make extremely explicit the use of pretraining for downstream prediction tasks. Moreover, they benchmark pretraining strategies explicitly on thermostability (one of the main considerations in the submitted manuscript), yet there is no mention of this work nor the dataset used (FLIP (Dallago et al., 2021)) in this current work. I think a reference and discussion of [1] is critical, and I would also like to see comparisons in line with [1], as [1] is very clear about what features from pretraining are used, and how. If the comparisons with previous methods were done in this fashion, this level of detail needs to be included in the text.

The initial version did not include an explicit comparison with the mentioned reference due to the difference in the learning task. In particular, [1] formulates a supervised learning task on predicting the continuous scores of mutants of specific proteins. In comparison, we make zero-shot predictions, where the model is trained in a self-supervised learning manner that requires no labels from experiments. In the revision, we added discussions in “Discussion and Conclusion” (lines 476-484):

Recommendations For The Authors:

Comment 1

I found the methods lacking in the sense that there is never a simple, explicit statement about what is the exact input and output of the model. What are the components of the input that are required by the user (to generate) or supply to the model? Are these inputs different at training vs inference time? The loss function seems like it's trying to de-noise a modified sequence, can you make this more explicit, i.e. exactly what values/objects are being compared in the loss?

We have added a more detailed description in the "Model Pipeline" section (page 7), which explains the distinct input requirements for training and inference, as well as the formulation of the employed loss function. To summarize:

(1) Both sequence and structure information are used in training and inference. Specifically, structure information is represented as a 3D graph with coordinates, while sequence information consists of AA-wise hidden representations encoded by ESM2-650M. During inference, instead of encoding each mutant individually, the model encodes the WT protein and uses the output probability scores relevant to the mutant to calculate the fitness score. This is a standard operation in many zero-shot fitness prediction models, commonly referred to as the log-odds-ratio.

(2) The loss function compares the differences between the noisy input sequence and the output (recovered) AA sequence. Noise is added to the input sequences, and the model is trained to denoise them (see “Ablation Study” for the different types of noise we tested). This approach is similar to a one-step diffusion process or BERT-style token permutation. The model learns to recover the probability of each node (AA) being one of 33 tokens. A cross-entropy loss is then applied to compare this distribution with the ground-truth (unpermuted) AA sequence, aiming to minimize the difference.

To better present the workflow, we revised the manuscript accordingly.

Comment 2

Related to the above, I'm not exactly sure where the structural/tertiary structure information comes from. In the methods, they don't state exactly whether the 3D coordinates are given in the CATH repository or where exactly they come from. In the results section they mention using AlphaFold to obtain coordinates for a specific task---is the use of AlphaFold limited only to these tasks/this is to show robustness whether using AlphaFold or realized coordinates?

The 3D coordinates of all proteins in the training set are derived from the crystal structures in CATH v4.3.0 to ensure a high-quality input dataset (see "Training Setup," Page 8). However, during the inference phase, we used predicted structures from AlphaFold2 and ESMFold as substitutes. This approach enhances the generalizability of our method, as in real-world scenarios, the crystal structure of the template protein to be engineered is not always available. The associated descriptions can be found in “Training Setup” (lines 271-272) and “Folding Methods” (lines 429-435).

Comment 3

Lines 142+144 missing reference "Section establishes", "provided in Section ."

199 "see Section " missing reference

214 missing "Section"

Thank you for pointing this out. We have fixed all missing references in the revision.

Comment 4

Table 2 - seems inconsistent to mention the number of parameters in the first 2 methods, then not in the others (though I see in Table 3 this is included, so maybe should just be omitted in Table 2).

In Table 2, we present the zero-shot methods used as baselines. Since many methods have different versions due to varying hyperparameter settings, we decided to list the number of parameters in the following tables.

We have double-checked both Table 3 and Table 5 and confirm that there is no inconsistency in the reported number of parameters. One potential explanation for the observed difference in the comment could be due to the differences in the number of parameters between single and ensemble methods. The ensemble method averages the predictions of multiple models, and we sum the total number of parameters across all models involved. For example, RITA-ensemble has 2210M parameters, derived from the sum of four individual models with 30M, 300M, 680M, and 1200M parameters.

Comment 5

In general, I found using the word "type" instead of "residue" a bit unnatural. As far as I can tell, the norm in the field is to say "amino acid" or "residue" rather than "type". This somewhat confused me when trying to understand the methods section, especially when talking about injecting noise (I figured "type" may refer to evolutionarily-close, or physicochemically-close residues). Maybe it's not necessary to change this in every instance, but something to consider in terms of ease of reading.

Thank you for your suggestion. The term "type" we used is a common expression similar to "class" in the NLP field. To avoid further confusion to the biologists, we have revised the manuscript accordingly. 

Comment 6

197 should this read "based on the kNN "algorithm"" (word missing) or maybe "based on "its" kNN"?

We have corrected the typo accordingly. It now reads “the 𝑘-nearest neighbor algorithm (𝑘NN)” (line 198).

Comment 7

200 weights of dimension 93, where does this number come from?

The edge features are derived by Zhou et al., 2024. We have updated the reference in the manuscript for clarity (lines 201-202).

Comment 8

210-212 "representations of the noisy AA sequence are encoded from the noisy input" what is the "noisy AA sequence?" might be helpful to exactly defined what is "noisy input" or "noisy AA sequence". This sentence could potentially be worded to make it clearer, e.g. "we take the modified input sequence and embed it using [xyz]."

We have revised the text accordingly. In the revised see lines 211-212:

Comment 9

In Table 3

Formatting, DTm (million), (million) should be under "# Params" likely?

Also for DDG this is reported on only a few hundred mutations, it might be worth plotting the confidence intervals over the Spearman correlation (e.g. by bootstrapping the correlation coefficient).

We followed the suggestion and added “million” under the "# Params". We have added the bootstrapped results for DDG and DTm to Table 6. For each dataset, we randomly sampled 50% of the data for ten independent runs. ProtSSN achieves the top performance with a considerably small variance.

Comment 10

The paragraph in lines 319 to lines 328 I feel may lack sufficient evidence.

"While sequence-based analysis cannot entirely replace the role of structure-based analysis, compared to a fully structure-based deep learning method, a protein language model is more likely to capture sufficient information from sequences by increasing the model scale, i.e., the number of trainable parameters."

This claim is made without a citation, such as [1]. Increasing the scale of the model doesn't always align with improving out-of-sample/generalization performance. I don't feel fully convinced by the claim that worse prediction is ameliorated by increasing the number of parameters. In Table 3 the performance is not monotonic with (nor scales with) the number of parameters, even within a model. See ProGen2 Expression scores, or ESM-2 Stability scores, as a function of their model sizes. In [1], the authors discuss whether pretraining strategies are aligned with specific tasks. I think rewording this paragraph and mentioning this paper is important. Figure 3 shows that maybe there's some evidence for this but I don't feel entirely convinced by the plot.

We agree that increasing the number of learnable parameters does not always result in better performance in downstream tasks. However, what we intended to convey is that language models typically need to scale up in size to capture the interactions among residues, while structure-based models can achieve this more efficiently with lower computational costs. We have rephrased this paragraph in the paper to clarify our point in lines 340-342.

Comment 11

Line 327 related to my major comment, " a comprehensive framework, such as ProtSSN, exhibits the best performance." Refers to performance on ProteinGym, yet the best-performing methods on ProteinGym are excluded from the comparison.

The primary comparisons were conducted using zero-shot models for fairness, meaning that the baseline models were not trained on MSA and did not use test performance to tune their hyperparameters. It's also worth noting that SaProt (the current SOTA model) had not been updated on the leaderboard at the time of submitting this paper. In the revised manuscript, we have included GEMME and TranceptEVE in Table 5 and SaProt in Tables 3, 5, and 6. While ProtSSN does not achieve SOTA performance in every individual task, our key argument in the analysis is to highlight the overall advantage of hybrid encoders compared to single sequence-based or structure-based models. We made clearer statement in the revised manuscript (line 349):

Comment 12

Line 347, line abruptly ends "equivariance when embedding protein geometry significantly." (?).

We have fixed the typo, (lines 372-373): 

Comment 13

Figure 3 I think can be made clearer. Instead of using True/false maybe be more explicit. For example in 3b, say something like "One-hot encoded" or "ESM-2 embedded".

The labels were set to True/False with the title of the subfigures so that they can be colored consistently.

Following the suggestion, we have updated the captions in the revised manuscript for clarity.

Comment 14

Lines 381-382 "average sequential embedding of all other Glycines" is to say that the score is taken as the average score in which Glycine is substituted at every other position in the peptide? Somewhat confused by the language "average sequential embedding" and think rephrasing could be done to make things clearer.

We have revised the related text accordingly a for clearer presentation (lines 406-413). 

Comment 15

Table 5, and in mentions to VEP, if ProtSSN is leveraging AlphaFold for its structural information, I disagree that ProtSSN is not an MSA method, and I find it unfair to place ProtSSN in the "non-MSA" categories. If this isn't the case, then maybe making clearer the inputs etc. in the Methods will help.

Your response is well-articulated and clear, but here is a slight revision for improved clarity and flow:

We respectfully disagree with classifying a protein encoding method based solely on its input structure. While AF2 leverages MSA sequences to predict protein structures, this information is not used in our model, and our model is not exclusive to AF2-predicted structures. When applicable, the model can encode structures derived from experimental data or other folding methods. For example, in the manuscript, we compared the performance of ProtSSN using proteins folded by both AF2 and ESMFold.

However, we would like to emphasize that comparing the sensitivity of an encoding method across different structures or conformations is not the primary focus of our work. In contrast, some methods explicitly use MSA during model training. For instance, MSA-Transformer encodes MSA information directly into the protein embedding, and Tranception-retrieval utilizes different sets of MSA hyperparameters depending on the validation set's performance.

To avoid further confusion, we have revised the terms "MSA methods" and "non-MSA methods" in the manuscript to "zero-shot methods" and "few-shot methods."

Comment 16

Table 3 they're highlighted as the best, yet on ProteinGym there's several EVE models that do better as well as GEMMA, which are not referenced.

The comparison in Table 3 focuses on zero-shot methods, whereas GEMME and EVE are few-shot models. Since these methods have different input requirements, directly comparing them could lead to

unfair conclusions. For this reason, we reserved the comparisons with these few-shot models for Table 5, where we aim to provide a more comprehensive evaluation of all available methods.            

Response to Reviewer 2

Summary:

To design proteins and predict disease, we want to predict the effects of mutations on the function of a protein. To make these predictions, biologists have long turned to statistical models that learn patterns that are conserved across evolution. There is potential to improve our predictions however by incorporating structure. In this paper, the authors build a denoising auto-encoder model that incorporates sequence and structure to predict mutation effects. The model is trained to predict the sequence of a protein given its perturbed sequence and structure. The authors demonstrate that this model is able to predict the effects of mutations better than sequence-only models.

Thank you for your thorough review and clear summary of our work. Below, we provide a detailed, pointby-point response to each of your questions and concerns. 

Strengths:

The authors describe a method that makes accurate mutation effect predictions by informing its predictions with structure.

Thank you for your clear summary of our highlights.

Weaknesses:

Comment 1

It is unclear how this model compares to other methods of incorporating structure into models of biological sequences, most notably SaProt.

(https://www.biorxiv.org/content/10.1101/2023.10.01.560349v1.full.pdf).

In the revision, we have updated the performance of SaProt single models (with both masked and unmasked versions with the pLDDT score) and ensemble models in the Tables 3, 5, and 6.

In the revised manuscript, we have updated the performance results for SaProt's single models (both masked and unmasked versions with the pLDDT score) as well as the ensemble models. These updates are reflected in Tables 3, 5, and 6.

Comment 2

ProteinGym is largely made of deep mutational scans, which measure the effect of every mutation on a protein. These new benchmarks contain on average measurements of less than a percent of all possible point mutations of their respective proteins. It is unclear what sorts of protein regions these mutations are more likely to lie in; therefore it is challenging to make conclusions about what a model has necessarily learned based on its score on this benchmark. For example, several assays in this new benchmark seem to be similar to each other, such as four assays on ubiquitin performed at pH 2.25 to pH 3.0.

We agree that both DTm and DDG are smaller datasets, making them less comprehensive than ProteinGym. However, we believe DTm and DDG provide valuable supplementary insights for the following reasons:

(1) These two datasets are low-throughput and manually curated. Compared to datasets from highthroughput experiments like ProteinGym, they contain fewer errors from experimental sources and data processing, offering cleaner and more reliable data.

(2) Environmental factors are crucial for the function and properties of enzymes, which is a significant concern for many biologists when discussing enzymatic functions. Existing benchmarks like ProteinGym tend to simplify these factors and focus more on global protein characteristics (e.g., AA sequence), overlooking the influence of environmental conditions.

(3) While low-throughput datasets like DTm and DDG do not cover all AA positions or perform extensive saturation mutagenesis, these experiments often target mutations at sites with higher potential for positive outcomes, guided by prior knowledge. As a result, the positive-to-negative ratio is more meaningful than random mutagenesis datasets, making these benchmarks more relevant for evaluating model performance.

We would like to emphasize that DTm and DDG are designed to complement existing benchmarks rather than replace ProteinGym. They address different scales and levels of detail in fitness prediction, and their inclusion allows for a more comprehensive evaluation of deep learning models.

Recommendations For The Authors:

Comment 1

I recommend including SaProt in your benchmarks.

In the revision, we added comparisons with SaProt in all the Tables (3, 5 and 6). 

Comment 2

I also recommend investigating and giving a description of the bias in these new datasets.

The bias of the new benchmarks could be found in Table 1, where the mutants are distributed evenly at different level of pH values.

In the revision, we added a discussion regarding the new datasets in “Discussion and Conclusion” (lines 496-504 of the revised version).

Comment 3

I also recommend reporting the model's ability to predict disease using ClinVar -- this experiment is conspicuously absent.

Following the suggestion, we retrieved 2,525 samples from the ClinVar dataset available on ProteinGym’s website. Since the official source did not provide corresponding structure files, we performed the following three steps:

(1) We retrieved the UniProt IDs for the sequences from the UniProt website and downloaded the corresponding AlphaFold2 structures for 2,302 samples.

(2) For the remaining proteins, we used ColabFold 1.5.5 to perform structure prediction.

(3) Among these, 12 proteins were too long to be folded by ColabFold, for which we used the AlphaFold3 server for prediction.

All processed structural data can be found at https://huggingface.co/datasets/tyang816/ClinVar_PDB. Our test results are provided in the following table. ProtSSN achieves the top performance over baseline methods.

Author response table 1.

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public Review): 

This manuscript examines the individual and dual effects of CHIP and LOY in MI employing a cohort of ~460 individuals. CHIP is assessed by NGS and LOY is assessed by PCR. The threshold for CHIP is set at 2% (an arbitrary cutoff that is often used) and LOY at 9% (according to the Discussion text - this reviewer may have missed the section that describes why this threshold was employed). The investigation assessed whether LOY could modulate inflammation, atherosclerotic burden, or MI risk associated with CHIP. Neither CHIP nor LOY independently affected hsCRP, atherosclerotic burden, or MI incidence, nor did LOY presence diminish these outcomes in CHIP+ male subjects.

This study represents the first dual analysis of CHIP and LOY on CVD outcomes. The results are largely negative, contradictory to other studies (many with much larger sample sizes). I would attribute the limitation of sample size as a major contributor to the negative data. While the negative data are suspect, the "positive" finding that LOY abolishes the prognostic significance of CHIP on MI is of interest (and consistent with what is understood from mechanistic studies).

Overall, I enjoyed reading the paper, and it is of interest to the research community.

However, I disagree with some of the authors' interpretations of the data.

Generally, many conclusions on CHIP interpretation are based on the comparison of findings from very large datasets that have been evaluated by shallow NGS DNA sequencing. These studies lack sensitivity and accuracy, but this is counterbalanced by their very large sample sizes. Thus, they draw conclusions from the sickest individuals (ICD codes) with the largest clones (explaining the 10% VAF threshold). Here, the study has a well-phenotyped cohort, but as far as this reviewer can tell, the DNA sequencing is "shallow" NGS. Typically, to assess smaller datasets, investigators employ an error-correction method (DNA barcodes, duplex sequencing, etc.) for the sensitivity and accuracy of calling variants. Thus, the current study appears to suffer from this limitation (small sample sizes combined with NGS).

We thank the reviewer for his/her positive and open comment. We acknowledge that we did not use error-corrected sequencing method for our study. However, we do not fully agree with the statement that our NGS sequencing technique is “shallow”.

Considering our entire sequencing panel, we achieve a sequencing depth ≥100X and ≥300X for 100% [99%;100%] and 99% [99%;100%] of the targeted regions respectively. This corresponds to a median depth of 2111X [1578;2574] for all regions sequenced. When considering “CHIP genes”, the median depth is 2694X [1875;3785] for patients from the CHAth study and 3455X [2266;4885] for patients from the 3C study. More specifically, for DNMT3A and TET2 genes, the median depths of sequencing are 2531X [1818;3313] and 3710X [2444;4901] for patients from the CHAth and 3C study respectively. These values are far much higher than the 300X recommended for NGS sequencing by capture technology by the French National Institute of Cancer. Coupling this high depth of sequencing with our bioinformatic pipeline that uses 3 different variant callers, a manual curing for all variants by trained hematobiologists and a bioinformatic tool to estimate the background noise allow us to detect somatic mutation with a VAF of 1% with a high accuracy. Noteworthy, our accuracy in detecting mutations in leukemia-associated genes is tested twice a year as part of our quality control program organized by the French Group of Molecular Biologists in Hematology (GBMHM). We added the information about the depth of sequencing in the Supplementary Methods section.

While the "negative" data from this study are inconclusive, the positive data (i.e. CHIP being prognostic for MI in the absence but not presence of MI) is of interest. Thus, the investigators may want to consider a shorter report that largely focuses on this finding.

We thank the reviewer for his/her interest in this result. We also agree that it would be interesting to focus specifically on demonstrating the impact of mLOY in countering the cardiovascular risk associated with CHIP. We performed additional analysis to demonstrate that this effect was independent of age and cardiovascular risk factors and included this information in the results section.

However, we believe that it is also of interest to show negative results that, although probably due to limitation in sample size, suggest that the cardiovascular risk associated with CHIP is not as strong and clinically pertinent as initially suggested. Of note, if CHIP really increase the risk of Myocardial Infarction in a significant manner, they would be more frequently detected in subjects who suffered from a MI compared to those who did not, which was not observed in our cohort. Moreover, we were able to determine that if CHIP increases the risk of MI, they do it to a much lesser extent (HR = 1.03 for CHIP) -than other established cardiovascular risk factors such as hypercholesterolemia or tobacco use HR = 1.47 and HR = 1.86 respectively in our cohort), which questions the pertinence of considering for CHIP in the management of patients with atherothrombosis. These data have been added in the Results and Discussion sections.

We also believe that our study has the merit to assess directly the impact of CHIP on atheroma burden, which has been performed in only a limited number of studies in the context of coronary artery disease. This could not be possible by analyzing only male subjects in our cohort because it would further decrease the statistical power of our analyses.

Reviewer #2 (Public Review):

Summary: 

The preprint by Fawaz et al. presents the findings of a study that aimed to assess the relationship between somatic mutations associated with clonal hematopoiesis (CHIP) and the prevalence of myocardial infarction (MI). The authors conducted targeted DNA sequencing analyses on samples from 149 MI patients and 297 non-MI controls from a separate cohort. Additionally, they investigated the impact of the loss of the Y chromosome (LOY), another somatic mutation frequently observed in clonally expanded blood cells. The results of the study primarily demonstrate no significant associations, as neither CHIP nor LOY were found to be correlated with an increased prevalence of MI. Of note, the null findings regarding CHIP are in conflict with several larger studies in the literature.

Strengths:

Overall, this is a useful research work on an emerging risk factor for cardiovascular disease (CVD). The use of a targeted sequencing approach is a strength, as it offers higher sensitivity than the whole exome sequencing approaches used in many previous studies.

Weaknesses:

Reporting null findings is definitely relevant in an emerging field such as the role of somatic mutations in cardiovascular disease. Nevertheless, the study suffers from severe limitations, which casts doubts on the authors' conclusions, as detailed below:

(1) The small sample size of the study population is a critical limitation, particularly when reporting null findings that conflict (partly) with positive findings in much larger studies, totaling hundreds of thousands of individuals (e.g. Zekavat et al, Nature CVR 2023, Vlasschaert et al, Circulation 2023; Zhao et al, JAMA Cardio 2024). The authors claim that they have 90% power to detect an effect size of CHIP on MI comparable to that in a previous report (Jaiswal et al, NEJM 2017). However, the methodology used to estimate statistical power is not described.

We thank the reviewer for his/her pertinent and constructive comments. We totally agree that our study presents a substantially smaller sample size as compared to the studies of Zekavat et al, Vlasschaert et al or Zhao et al.

The CHAth study was designed as a prospective study (which is not frequent in CHIP reports) to demonstrate that, if CHIP increase the risk of MI, they would be detected more frequently in patients who suffered from a MI compared to those who did not. To achieve this, we defined eligibility criteria to have a rather high prevalence of CHIP and optimize the statistical power of a study based on a limited number of patients. We thus enrolled patients who suffered from a first MI after the age of 75 years. These patients had to be compared with subjects from the Three-City study who had 65 years or more at inclusion and did not present any cardiovascular event before inclusion.

To determine the number of patients necessary to achieve our objective, we considered a CHIP prevalence of 20% in the general population after the age of 75 years, as estimated when we set up our study (Genovese et al, NEJM 2014, Jaiswal et al, NEJM 2014, Jaiswal et al, NEJM 2017). At this time the relative risk of MI associated with CHIP was shown to be 1.7, leading to an expected prevalence of CHIP of 37% in subjects who presented a MI. Based on these hypotheses, the recruitment of 112 patients in the CHAth would have been sufficient to detect a significant higher prevalence of CHIP in MI(+) patients compared to MI(-) subjects with a power of 0.90 at a type I error rate of 5%. These calculations were performed by the Research Methodology Support Unit of the University Hospital of Bordeaux. These data were added in the Supplementary Methods section to expose more clearly the design and objectives of the CHAth study.

Finally, we recruited 149 patients in the CHAth study and compared them to 297 control subjects. Although recruiting more patients than initially needed, we observed a similar prevalence of CHIP between our 2 cohorts, suggesting that the cardiovascular risk associated with CHIP is lower than the 1.7 increased risk claimed in most publications related to CHIP in the cardiovascular field. We have to notice that our study was not designed to demonstrate the impact of CHIP on the occurrence of MI during follow-up, which could explain our negative results due to a limited number of patients as stated by the reviewers. This statement has been added in the Supplementary Methods section. However, performing such analysis allowed us to confirm that the risk of MI associated with CHIP was lower than 1.7 and lower than the one associated with hypercholesterolemia or smoking.

We would like also to notice that the eligibility criteria for both CHAth and the Three-City study can have led to a selection bias, possibly contributing to the contradiction of our results with other studies. As stated before, in the CHAth study, only patients who experience a first MI after the age of 75 were enrolled. In the Three-City study, all subjects had 65 years or more at inclusion. On the contrary, most of the cohorts showing an association between CHIP and cardiovascular events were composed of younger subjects:

-          Bioimage : median age 70 years (55-80 years)

-          MDC : median age 60 years

-          ATVB : subjects with a MI before 45 years

-          PROMIS : subjects between 30 and 80 years

-          UK Biobank : between 40 and 70 years at inclusion, median age of 58 years in the study of Vlasschaert et al.

-          Zhao et al : median age of 53.83 years (45.35-62.39 years).

This last information was added in the Discussion section (lines 452-454).

Furthermore, the work by Jaiswal et al (NEJM 2017) showed a hazard ratio of approx. 2.0, but more recent work in much larger populations suggests that the overall effect of CHIP on atherosclerotic CVD is smaller, most likely due to the heterogeneity of effects of different mutated genes (e.g. Zekavat et al, Nature CVR 2023, Vlasschaert et al, Circulation 2023; Zhao et al, JAMA Cardio 2024).

We thank the reviewer for insisting on the fact that the initial HR of 2.0 observed by Jaiswal et al was shown to be smaller in more recent studies. This corresponds to what we wrote in the introduction (lines 103-109) and discussion (lines 365-370, 465-471).

In addition, several analyses in the current manuscript are conducted separately in MI(+) (n= 149) and MI(-) (N=297) individuals, further limiting statistical power. Power is still lower in the investigation of the effects of LOY and its interaction with CHIP, as only men are included in these analyses. Overall, I believe the study is severely underpowered, which calls into question the validity of the reported null findings.

We agree with the reviewer that the statistical power of our study is lower than the one of other studies, in particular those based on several hundred thousand patients. Whenever possible, we analyzed our data by combining MI(+) and MI(-) subjects. However, for some aspects such as atherosclerosis, we did not have the same parameters available for these 2 groups and had to analyze them separately, leading to a more limited statistical power. We also have to acknowledge that our study was not designed to demonstrate an effect of CHIP on incident MI (as stated before), limiting our statistical power to demonstrate an effect of CHIP +/- mLOY on the incident risk of coronary artery disease.

However, when designing our prospective study (CHAth study), we aimed to address the limitations of a small cohort and obtain rapid, significant results regarding the impact of CHIP. We hypothesized that if CHIP really increases the risk of myocardial infarction (MI), it would be detected more frequently in patients who have experienced a MI compared to those who have not. This study design would demonstrate the importance of CHIP in MI pathophysiology without requiring thousands of patients. However, we did not observe such an association questioning the relevance of detecting CHIP for the management of patients in the field of Cardiology. This was confirmed by the fact that in our cohort, the cardiovascular risk associated with CHIP appears to be low (HR = 1.03 [0.657;1.625] after adjustment on sex, age and cardiovascular risk factors) compared to hypercholesterolemia (HR = 1.474 [0.758;2.866]) or smoking (HR = 1.865 [0.943;3.690]). These data have been added in the Results and Discussion sections.

In addition, we would like to mention that despite the limited number of subjects studied, we do not have only negative results. When studying only men subjects, we were able to show that CHIP accelerate the occurrence of MI, particularly in the absence of mLOY (Figure 2D). This effect was independent of age and cardiovascular risk factors (diabetes, cholesterol and high blood pressure). We added this last information in the results section of the manuscript, although we acknowledge that this has to be confirmed in future work.

(2) Related to the above, it is widely accepted that the effects of CHIP on CVD are highly heterogeneous, as some mutated genes appear to have a strong impact on atherosclerosis, whereas the effect of others is negligible (e.g. Zekavat et al, Nature CVR 2023, Vlasschaert et al, Circulation 2023, among others). TET2 mutations are frequently considered a "positive control", given the multiple lines of evidence suggesting that these mutations confer a higher risk of atherosclerotic disease.

However, no association with MI or related variables was found for TET2 mutations in the current work. Reporting the statistical power specifically for assessing the effect of TET2 mutations would enhance the interpretation of these results.

We thank the reviewer for this pertinent remark. It has indeed been shown that depending on the somatic mutation, the impact of CHIP on inflammation, atherosclerosis and cardiovascular risk is different. The studies cited by the reviewer suggest that DNMT3A mutations have a low impact on atherosclerosis/atherothrombosis while other “non-DNMT3A” mutations, including TET2 mutations, have a greater impact. In particular, Zekavat et al suggested that TP53, PPM1D, ASXL1 and spliceosome mutations have a similar impact on atherosclerosis/atherothrombosis to TET2.

To answer to the reviewer in our cohort, we did not find a clear association between the detection of TET2 mutation with a VAF≥2% and:

-          A history of MI at inclusion (p=0.5339)

-          Inflammation (p=0.440)

-          Atherosclerosis burden :

-   In the CHAth study:

-  p=0.031 for stenosis≥50%

-  p=0.442 fir multitruncular lesions

-  p=0.241 for atheroma volume

-   in the 3C study :

-  p=0.792 for the presence of atheroma

-  p=0.3966 for the number of plaques

-  p=0.876 for intima-media thickness

-          Incidence of MI (p=0.5993)

Similarly we did not find any association between the detection of TET2 mutations with a VAF≥1% and:

-          A history of MI at inclusion (p=0.5339)

-          Inflammation (p=0.802)

-          Atherosclerosis burden :

-   In the CHAth study :

-  p=0.104 for stenosis≥50%

-  p=0.617 fir multitruncular lesions

-  p=0.391 for atheroma volume

-   in the 3c study:

-  p=0.3291 for the presence of atheroma

-  p=0.2060 for the number of plaques

-  p=0.2300 for intima-media thickness

-          Incidence of MI (p=0.195)

However, analyzing the specific effect of TET2 mutations reduces the cohort of CHIP(+) subjects to 61 individuals. In these conditions, considering a prevalence of “TET2-CHIP” of 13.5% (in our cohort) and a hazard ratio of 1.3 (Vlasschaert et al), the statistical power to show an increased risk of MI is only 16%.

(3) One of the most essential features of CHIP is the tight correlation with age. In this study, the effect of age on CHIP (Supplementary Tables S5, S6) seems substantially milder than in previous studies. Given the relatively weak association with age here, it is not surprising that no association with MI or atherosclerotic disease was found, considering that this association would have a much smaller effect size.

We thank the reviewer for highlighting this point. Although the difference of median age between subjects with or without a CHIP is not very important in our cohort, we did observe a significant association of CHIP with age:

-          The differences in age were statistically significant both in the CHAth and 3C study (Supplementary Tables S5 and S6)

-          We observed a significant association between age and CHIP prevalence (p<0.001 for the total cohort, p=0.0197 for the CHAth study, and p=0.0394 for the 3C cohort after adjustment on sex). This association was already shown in the figure 1. We added the significant association between age and CHIP prevalence in the Results section (line 279).

As stated before, we have to remind the reviewer that we enrolled only subjects of ≥75 years and ≥65 years in the CHAth and 3C studies respectively. This led to a median age in our cohort that was substantially higher than in other cohorts (in particular the UK Biobank and the different cohorts studied by Jaiswal et al). This could have contributed to an apparent milder effect of age on CHIP, even if this association was still observed.

In addition, there are previous reports of sex-related differences in the prevalence of CHIP, is there an association between CHIP and age after adjusting for sex? 

The reviewer correctly pointed out that sex has been associated with various aspects of CHIP. While Zekavat et al reported that CHIP carriers were more frequently males, Kar et al (Nature Genetics 2022), and Kamphuis et al (Hemasphere 2023) did not observe a difference in the prevalence of CHIP between males and females, but rather a difference in the mutational spectrum. Male presented more frequently SRSF2, ASXL1, SF3B1, U2AF1, JAK2, TP53 and PPM1D mutations while females had more frequently DNMT3A, CBL and GNB1 mutations.

In our study, the association between CHIP prevalence and age was indeed significant even after adjustment on sex (p<0.001 for the total cohort, p=0.0197 for the CHAth study and p=0.0394 for the 3C).

(4) The mutated genes included in the definition of "CHIP" here are markedly different than those in most previous studies, particularly when considering specifically the studies that demonstrated an association between CHIP and atherosclerotic CVD. For instance, the definition of CHIP in this manuscript includes genes such as ANKRD26, CALR, CCND2, and DDX41... that are not prototypical CHIP genes. This is unlikely to have a major impact on the main results, as the vast majority of mutations detected are indeed in bona fide CHIP genes, but it should be at least acknowledged.

We agree with the reviewer that our gene panel includes genes that are not considered prototypical CHIP genes. This acknowledgment has been added in the Supplementary Methods section. To perform this study, we did not design a specific targeted sequencing panel. We used the one that is used for the diagnosis of myeloid malignancies at the University Hospital of Bordeaux. ANKRD26 and DDX41 are genes that, when mutated, predispose to the development of hematological malignancies. CALR mutations are frequently detected in Myeloproliferative Neoplasms while CCND2 mutation can be detected in acute myeloid leukemia among other diseases. As usually performed in our routine practice, we analyzed all the genes in the panel. However, as stated by the reviewer, most of the mutations we detected involved bona fide CHIP genes.

Furthermore, the strategy used here for the CHIP variant calling and curation seems substantially different than that used in previous studies, which precludes a direct comparison. This is important because such differences in the definition of CHIP and the curation of variants are the basis of most conflicting findings in the literature regarding the effects of this condition. Ideally, the authors should conduct sensitivity analyses restricted to prototypical CHIP genes, using the criteria that have been previously established in the field (e.g. Vlasschaert et al, Blood 2023).

We agree with the reviewer, our strategy for CHIP variant calling and curation was substantially different from what has been used in other studies. We decided to apply the criteria we used in previous studies for the analysis of somatic mutation in myeloid malignancies. Because CHIP are defined by the detection of “somatic mutations in leukemia driver genes”, this appeared to follow the definition of CHIP.

We also acknowledge that this discrepancy with the criteria defined by Vlasschaert et al could contribute to our findings that differ from those of other studies. We thus checked whether the variants detected were in accordance or not with the criteria defined by Vlasschaert et al. Pooling the 2 cohorts, we detected 439 variants, 381 of which were in accordance with the criteria established by Vlasschaert et al, representing a concordance rate of 86.8%. Moreover, the variants “wrongly” retained according to these criteria had an impact on the conclusion on the detection of CHIP in only 15 patients (because these variants were associated with a mutation in a bona fide CHIP gene and/or because its VAF was below 2%). Thus, the impact of CHIP variant calling and curation had only a limited impact on our results. This has been added in the discussion (lines 455-459).

However, we would like to discuss the criteria that have been defined by Vlasschaert et al which are probably too restrictive. For some genes, such as ZRSR2, in addition to frameshift and non-sens mutations that are expected to be associated with a loss of function, only some single nucleotide variations were retained (probably those detected by this group). In our patient 20785, we detected a c.524A>G, p.(Tyr175Cys) mutation that was not reported in the list published by Vlasscheart et al. However, this variant presents a VAF presumptive of a somatic origin (3%), affects the Zn finger domain of the protein and is observed in a male subject. Thus, it presents several criteria to consider it as associated with a loss of function. Similarly, the CBL variant c.1139T>C, p.(Leu380Pro) observed in our patient 21536, although not affecting the residues 381-421 of the protein (the criteria defined by Vlasschaert et al), has been reported in 29 cases of hematological malignancies. It is thus likely to have a significant impact on the behavior of hematopoietic cells. Moreover, in the same patient, a TET2 c.4534G>A, p.(Ala1512Thr) variant was detected. Although not affecting directly the CD1 domain, it has been reported in a case of AML with a VAF suggestive of a somatic origin (Papaemmanuil et al, NEJM 2016). The SH2B3 gene is not considered by Vlasschaert et al as a bona fide CHIP gene, contrary to other genes involved in cell signaling such as JAK2, GNAS, GNB1, CBL. However, inactivating mutations in SH2B3 can be detected in myeloid malignancies and were recently shown to drive the phenotype in some patients with a MPN (Zhang et al, American Journal of Hematology 2024). We could thus expect that this also happens in our patients 22591 and 21998 who harbor mutations of SH2B3 (a SNV in the PH domain and a frameshift mutation respectively).

Regarding BCOR, STAG2, SMC3 and RAD21 genes, although frameshift mutations are the most prevalent, there are several reports on the existence of SNV in the context of hematological malignancies (COSMIC, Blood (2021) 138 (24): 2455–2468, Blood Cancer Journal (2023)13:18 ; https://doi.org/10.1038/s41408-023-00790-1).

We can also add that although Vlasschaert et al did not consider CSF3R and CALR as CHIP-genes, Kessler et al did. Because CHIP are an emerging field, it should be considered that the concepts that define it are expected to evolve, as demonstrated by the recent study of the Jyoti Nangalia’s group (Bernstein et al, Nature Genetics 2024) who showed that 17 additional genes (including SH2B3) should be considered as driver of clonal hematopoiesis.

(5) An important limitation of the current study is the cross-sectional design of most of the analyses. For instance, it is not surprising that no association is found between CHIP and prevalent atherosclerosis burden by ultrasound imaging, considering that many individuals may have developed atherosclerosis years or decades before the expansion of the mutant clones, limiting the possible effect of CHIP on atherosclerosis burden. Similarly, the analysis of the relationship between CHIP and a history of MI may be confounded by the potential effects of MI on the expansion of mutant clones. In this context, it is noteworthy that the only positive results here are found in the analysis of the relationship between CHIP at baseline and incident MI development over follow-up. Increasing the sample size for these longitudinal analyses would provide deeper insights into the relationship between CHIP and MI. 

We agree with the reviewer that increasing the sample size for longitudinal analyses would provide deeper insights into the relationship between CHIP and MI. Unfortunately, for the moment, we do not have access to additional samples of the 3C study and are not able to perform these additional analyses.

(6) The description of some analyses lacks detail, but it seems that statistical analyses were exclusively adjusted for age or age and sex. The lack of adjustment for conventional cardiovascular risk factors in statistical analyses may confound results, particularly given the marked differences in several variables observed between groups.

The reviewer is right when saying that we adjusted our analyses on age and/or sex. This was done because as stated before, our results did not show a lot of significant differences. However, we reanalyzed our data, adjusting further the tests for conventional cardiovascular risk factors, and observed similar results. These data have been added in the results section (lines 286-287, 303, 319, 331-332, 341).

(7) The variant allele fraction (VAF) threshold for identifying clinically relevant clonal hematopoiesis is still a subject of debate. The authors state that subjects without any detectable mutation or with mutations with a VAF below 2% were considered non-CHIP carriers. While this approach is frequent in the field, it likely misses many impactful mutations with lower VAFs. Such false negatives could contribute to the null findings reported here. Ideally, the authors should determine the lower detection limit of their sequencing approach (either computationally or through serial dilution experiments) and identify the threshold of VAF that can be detected reliably with their sequencing assay. The association between CHIP and MI should then be evaluated considering all mutations above this VAF threshold, in addition to sensitivity analyses with other thresholds frequent in the literature, such as 1% VAF, 2% VAF, and 10% VAF.

We agree with the reviewer that the VAF threshold for identifying clinically relevant CH is still debated. As stated in the manuscript and by the reviewer, we used the conventional threshold of 2%. Considering that different studies have shown that the cardiovascular risk is increased in a more important manner for CHIP with a high VAF (Jaiswal et al, NEJM 2017, Kessler et al Nature 2022, Vlasschaert et al, Circulation 2023), it is not sure that considering variant with a very low VAF (below 2%) would help us in finding an impact of CHIP on inflammation, atherosclerosis or atherothrombotic risk.

However, as mentioned by the reviewer, variants with a low VAF could have a clinical impact as recently reported by Zhao et al. In France, the use of biological analysis for medical purposes imposes to demonstrate that all its aspects are mastered, including their performances. In that context, we determined that our NGS strategy allowed us to reliably detect mutation with a VAF down to 1% (data not shown). As stated in the discussion, we also analyzed our results considering variants with a VAF of 1% and found similar results (lines 394-395). The sensitivity analyses were already mentioned in the manuscript, as we also searched for an effect of CHIP with a high VAF (≥5%) and found no effect neither. We did not have a sufficient number of subjects carrying variants with a VAF≥10% to perform analysis with this threshold.

(8) The authors should justify the use of 3D vascular ultrasound imaging exclusively in the supra-aortic trunk. I am not familiar with this technique, but it seems to be most typically used to evaluate atherosclerosis burden in superficial vascular beds such as carotids or femorals. I am concerned about the potential impact of tissue depth on the accurate quantification of atherosclerosis burden in the current study (e.g. https://doi.org/10.1016/j.atherosclerosis.2016.03.002). It is unclear whether the carotids or femorals were imaged in the study population. 

We apologize for the lack of precision in the Methods section. As stated by the reviewer, we evaluated the atherosclerosis burden in superficial vascular beds. We measured atheroma volume at the site of the common carotid (as described by B Lopez-Melgar, in Atheroslerosis, 2016). We did not analyze femoral arteries in this study. The sentence is now corrected in the Methods (lines 176-179).

(9) The specific criteria used to define LOY need to be justified. LOY is stated to be defined based on a "A cut off of 9% of cells with mLOY defined the detection of a mLOY based on the study of 30 men of less than 40 years who had a normal karyotype as assessed by conventional cytogenetic study." As acknowledged by the authors, this definition of LOY is substantially different than that used in recent studies employing the same technique to detect LOY (Mas-Peiro et al, EHJ 2023). In addition, it seems essential to provide more detailed information on the ddPCR assay used to determine LOY, including the operating range and, more importantly, the lower limit of detection (%LOY) of the assay. A dilution series of a control DNA with no LOY would be helpful in this context. 

We apologize if the definition of the threshold for detecting mLOY was unclear. To test the performance of our ddPCR technique, we first determined the background noise by testing DNA obtained from total leukocytes in 30 men of ≤40 years who presented a normal karyotype as assessed by conventional cytogenetic technics. In this control population supposed not to carry mLOY, we detected of proportion of cells with mLOY of 2,34+/-1,98 (see Author response image 1, panel A). We thus considered a threshold above 9% as being different from background noise (mean + 3 times the standard deviation).

We then compared the proportion of cells with mLOY measured by ddPCR and conventional karyotype and observed a rather good correlation between the 2 technics (R2\=0.6430, p=0.0053, see Author response image 1, panel B). Finally, we tested the reliability of our ddPCR assay in detecting different levels of mLOY using a dilution series of control DNA (from an equivalent of 2% of cell with mLOY to 98% of cells with mLOY). We observed a very nice correlation between the theoretical and measured proportions of cells with mLOY (R2\=0.9989, p<0.001, see Author response image 1, panel C). Of note, the proportion of mLOY measured for values ≤10% were concordant with theoretical values. However, considering the background noise determined with control DNA, we were unable to confirm that this “signal” was different from the background noise. Therefore, we set a threshold of 9% to define the detection of mLOY by ddPCR. It is also noteworthy that the 10% cell population with mLOY was consistently detected by the ddPCR technique. This has been added in the Methods section (lines 228-235).

Author response image 1.

(10) Our understanding of the relationship between CHIP and CVD is evolving fast, and the manuscript should be considered in the context of recent literature in the field. For instance, the recent work by Zhao et al (JAMA Cardio 2024, doi:10.1001/jamacardio.2023.5095) should be considered, as it used a similar targeted DNA sequencing approach as the one used here, but found a clear association between CHIP and coronary heart disease (in a population of 6181 individuals). 

We thank the reviewer for this pertinent reference. We did not include it in the first version of our manuscript because it was not published yet when we submitted our work. We included this reference in the discussion (lines 451, 455, 464). We also included the recent study of Heimlich et al (Circ Gen Pre Med 2024, lines 464-468) who studied the association of CHIP with atherosclerosis burden.

(11) The use of subjective terms like "comprehensive" or "thorough" in the title of the manuscript does not align with the objective nature of scientific reporting. 

We removed the terms “comprehensive” and “thorough” from the title and the text.

Recommendations for the authors:

Reviewing Editor:

The Editors believe that in light of the small study the word Comprehensive has to be removed (including from the title and abstract).

We agree and removed the term comprehensive from the title and the text.

Reviewer #1 (Recommendations For The Authors):

Other comments:

It has long been recognized that hsCRP does not adequately address the inflammation associated with CHIP. For example, see Bick et al Nature 2020; 586:763. Through an assessment of a large dataset, the regulation of multiple inflammatory mediators was associated with CHIP but not with CRP. 

We agree that hsCRP is probably not the most sensitive marker for inflammatory state associated with CHIP. However, it is the most commonly used one in medical practise. However, as indicated in the discussion (lines 418-420), we did not observe any association between CHIP and the plasmatic level of different cytokines (IL1ß, IL6, IL18 and TNFα) in patients enrolled in the CHAth study.

Many of the citations lack journal names, volumes, page numbers, etc. 

We apologize for this and corrected the citations.

Please provide more details on the methodology (i.e. is CHIP assessed only through NGS with no error correction?). Specify the rationale for why the 9% LOY threshold was employed. Provide this information in the Methods section.

We added more details on the methodology as demanded in the results section (lines 212-214 and 228-235).

Supplementary Table S3 lacks headings. What are the designations for columns 6-8? 

We apologize for this and corrected the Table. Columns 6-8 correspond to the VAF, coverage of the variants and depth of sequencing, as for Table S4.

Author response:

The following is the authors’ response to the current reviews. 

eLife assessment:

This useful modeling study explores how the biophysical properties of interneuron subtypes in the basolateral amygdala enable them to produce nested oscillations whose interactions facilitate functions such as spike-timing-dependent plasticity. The strength of evidence is currently viewed as incomplete because of insufficient grounding in prior experimental results and insufficient consideration of alternative explanations. This work will be of interest to investigators studying circuit mechanisms of fear conditioning as well as rhythms in the basolateral amygdala.

We disagree with the overall assessment of our paper. The current reviews published below focus on two kinds of perceived inadequacies. Reviewer 1 (R1) was concerned that the fear conditioning paradigm used in the model is not compatible with some of the experiments we are modeling. The reviewer helpfully suggested in the Recommendations for the Authors some papers, which R1 believed exposed this incompatibility. In our reading, those data are indeed compatible with our hypotheses, as we will explain in our reply. Furthermore, the point raised by R1 is an issue for the entire field. We will suggest a solution to that issue based on published data.

Reviewer 2 (R2) said that there is no evidence that the BLA is capable of producing, by itself, the rhythms that have been observed during fear conditioning in BLA and, furthermore, that the paper we cited to support such evidence, in fact, refutes our argument. We believe that the reasoning used by reviewer 2 is wrong and that the framework of R2 for what counts as evidence is inadequate. We spell out our arguments below in the reply to the reviewers.

Finally, we believe this work is of interest far beyond investigators studying fear conditioning. The work shows how rhythms can create the timing necessary for spike-timing-dependent plasticity using multiple time scales that come from multiple different kinds of interneurons found both in BLA and, more broadly, in cortex. Thus, the work is relevant for all kinds of associative learning, not just fear conditioning. Furthermore, it is one of the first papers to show how rhythms can be central in mechanisms of higher-order cognition.

Reviewer #1

We thank Reviewer 1 for his kind remarks about our first set of responses and their understanding of the importance of the work. There was only one remaining point to be addressed:

Deficient in this study is the construction of the afferent drive to the network, which does elicit activities that are consistent with those observed to similar stimuli. It still remains to be demonstrated that their mechanism promotes plasticity for training protocols that emulate the kinds of activities observed in the BLA during fear conditioning.

It is true that some fear conditioning protocols involve non-overlapping US and CS, raising the question of how plasticity happens or whether behavioral effects may happen without plasticity. This is an issue for the entire field (Sun et al., F1000Research, 2020). Several papers (Quirk, Repa and LeDoux, 1995; Herry et al, 2007; Bordi and Ledoux 1992) show that the pips in auditory fear conditioning increase the activity of some BLA neurons: after an initial transient, the overall spike rate is still higher than baseline activity. The question remains as to whether the spiking is sustained long enough and at a high enough rate for STDP to take place when US is presented sometime after the stop of the CS.

Experimental recordings cannot speak to the rate of spiking of BLA neurons during US due to recording interference from the shock. However, evidence seems to suggest that ECS activity should increase during the US due to the release of acetylcholine (ACh) from neurons in the basal forebrain (BF) (Rajebhosale et al., 2024). Pyramidal cells of the BLA robustly express M1 muscarinic ACh receptors (Muller et al., 2013; McDonald and Mott, 2021) and M1 receptors target spines receiving glutamatergic input (McDonald et al., 2019). Thus, ACh from BF should elicit a long-lasting depolarization in pyramidal cells. Indeed, the pairing of ACh with even low levels of spiking of BLA neurons results in a membrane depolarization that can last 7 – 10 s (Unal et al., 2015). This implies that the release of ACh can affect the consequences of the CS in successive trials. This should include higher spiking rates and more sustained activity in the ECS neurons after the first presentation of US, thus ensuring a concomitant activation of ECS and fear (F) neurons necessary for STDP to take place. Hence, we suggest that a solution to the problem raised by R1 may be solved by considering the role of ACh release by BF. To the best of our knowledge, there is nothing in the literature that contradicts this potential solution. The model we have may be considered a “minimal” model that puts in by hand the higher frequency due to the cholinergic drive without explicitly modeling it. As R1 says, it is important for us to give the motivation of that higher frequency; in the next revision, we will be explicit about how the needed adequate firing rate can come about without an overlap of CS and US in any given trial.

Reviewer #2

The authors of this study have investigated how oscillations may promote fear learning using a network model. They distinguished three types of rhythmic activities and implemented an STDP rule to the network aiming to understand the mechanisms underlying fear learning in the BLA.

After the revision, the fundamental question, namely, whether the BLA networks can or cannot intrinsically generate any theta rhythms, is still unanswered. The author added this sentence to the revised version: "A recent experimental paper, (Antonoudiou et al., 2022), suggests that the BLA can intrinsically generate theta oscillations (3-12 Hz) detectable by LFP recordings under certain conditions, such as reduced inhibitory tone." In the cited paper, the authors studied gamma oscillations, and when they applied 10 uM Gabazine to the BLA slices observed rhythmic oscillations at theta frequencies. 10 uM Gabazine does not reduce the GABA-A receptor-mediated inhibition but eliminates it, resulting in rhythmic populations burst driven solely by excitatory cells. Thus, the results by Antonoudiou et al., 2022 contrast with, and do not support, the present study, which claims that rhythmic oscillations in the BLA depend on the function of interneurons. Thus, there is still no convincing evidence that BLA circuits can intrinsically generate theta oscillations in intact brain or acute slices. If one extrapolates from the hippocampal studies, then this is not surprising, as the hippocampal theta depends on extra-hippocampal inputs, including, but not limited to the entorhinal afferents and medial septal projections (see Buzsaki, 2002). Similarly, respiratory related 4 Hz oscillations are also driven by extrinsic inputs. Therefore, at present, it is unclear which kind of physiologically relevant theta rhythm in the BLA networks has been modelled.

Reviewer 2 (R2) says “the fundamental question, namely, whether the BLA networks can or cannot intrinsically generate any theta rhythms, is still unanswered.” In our revision, we cited (Antonoudiou et al., 2022), who showed that BLA can intrinsically generate theta oscillations (3-12 Hz) detectable by LFP recordings. R2 pointed out that this paper produces such theta under conditions in which the inhibition is totally removed. R2 then states that the resulting rhythmic populations burst at theta “are driven solely by excitatory cells. Thus, the results by (Antonoudiou et al., 2022) contrast with, and do not support, the present study, which claims that rhythmic oscillations in the BLA depend on the function of interneurons. Thus, there is still no convincing evidence that BLA circuits can intrinsically generate theta oscillations in intact brain or acute slices.”

This reasoning of R2 is faulty. With all GABAergic currents omitted, the LFP is composed of excitatory currents and intrinsic currents. Our model of the LFP includes all synaptic and membrane currents. In our model, the high theta comes from the spiking activity of the SOM cells, which increase their activity if the inhibition from VIP cells is removed. We are including a new simulation, which models the activity of the slice in the presence of kainate (as done in Antonoudiou et al., 2022), providing additional excitation to the network. If the BLA starts at high excitation, our model produces an ongoing gamma in the VIP cells that suppress SOM cells and allows a PING gamma to form between PV and F cells; with Gabazine (modeled as the removal of all the GABAergic synapses), this PING is no longer possible and so the gamma rhythm disappears. As expected, the simulation shows that the model produces theta with Gabazine; the model also shows that a PING rhythm is produced without Gabazine, and that this rhythm goes away with Gabazine because PING requires feedback inhibition (see Author response image 1). Thus, the theta increase with Gabazine in the (Antonoudiou et al., 2022) paper can be reproduced in our model, so that paper does support the model.

Author response image 1.

Spectral properties of the BLA network without (black) versus with Gabazine (magenta). Power spectra of the LFP proxy, which is the linear sum of AMPA, GABA (only present in the absence of Gabazine, D-, NaP-, and H-currents. Both power spectra are represented as mean and standard deviation across 10 network realizations. Bottom: inset between 35 and 50 Hz.

Nevertheless, we agree that this paper alone is not sufficient evidence that the BLA can produce a low theta. We have recently learned of a new paper (Bratsch-Prince et al., 2024) that is directly related to the issue of whether the BLA by itself can produce low theta, and in what circumstances. In this study, intrinsic BLA theta is produced in slices with ACh stimulation (without needing external glutamate input) which, in vivo, would be produced by the basal forebrain (Rajebhosale et al., eLife, 2024) in response to salient stimuli. The low-theta depends on muscarinic activation of CCK interneurons, a group of interneurons that overlaps with the VIP neurons in our model (Krabbe 2017; Mascagni and McDonald, 2003).

We suspect that the low theta produced in (Bratsch-Prince et al., 2024) is the same as the low theta in our model. We do not explicitly include ACh modulation of BLA in our paper, but in current work with experimentalists, we aim to show that ACh is essential to the theta by activating the BLA VIP cells. In our re-revised version, we will discuss Bratsch-Prince et al., 2024 and its connection to our hypothesis that the theta oscillations can be produced within the BLA.

Note that we have already included a paragraph stating explicitly that our hypothesis in no way contradicts the idea that inputs to the BLA may include theta oscillations. Indeed, the following paragraphs in the revised paper describe the complexity of trying to understand the origin of brain rhythms in vivo. R2 did not appear to take this complexity, and the possible involvement of neuromodulation, into account in their current position that the theta rhythms cannot be produced intrinsically in the BLA.

From revised paper: “Where the rhythms originate, and by what mechanisms. A recent experimental paper, (Antonoudiou et al. 2022), suggests that the BLA can intrinsically generate theta oscillations (3-12 Hz) detectable by LFP recordings under certain conditions, such as reduced inhibitory tone. They draw this conclusion in mice by removing the hippocampus, which can volume conduct to BLA, and noticing that other nearby brain structures did not display any oscillatory activity. Our model also supports the idea that intrinsic mechanisms in the BLA can support the generation of the low theta, high theta, and gamma rhythms.

Although the BLA can produce these rhythms, this does not rule out that other brain structures also produce the same rhythms through different mechanisms, and these can be transmitted to the BLA. Specifically, it is known that the olfactory bulb produces and transmits the respiratory-related low theta (4 Hz) oscillations to the dorsomedial prefrontal cortex, where it organizes neural activity (Bagur et al., 2021). Thus, the respiratory-related low theta may be captured by BLA LFP because of volume conduction or through BLA extensive communications with the prefrontal cortex. Furthermore, high theta oscillations are known to be produced by the hippocampus during various brain functions and behavioral states, including during spatial exploration (Vanderwolf, 1969) and memory formation/retrieval (Raghavachari et al., 2001), which are both involved in fear conditioning. Similarly to the low theta rhythm, the hippocampal high theta can manifest in the BLA. It remains to understand how these other rhythms may interact with the ones described in our paper.”

We believe our current paper is important to show how detailed biophysical modeling can unearth the functional implications of physiological details (such as the biophysical bases of rhythms), which are often (indeed, usually) ignored in models, and why rhythms may be essential to some cognitive processes (including STDP). Indeed, for evaluating our paper it is necessary to go back to the purpose of a model, especially one such as ours, which is “hypothesis/data driven”. The hypotheses of the model serve to illuminate the functional roles of the physiological details, giving meaning to the data. Of course, the hypotheses must be plausible, and we think that the discussion above easily clears that bar. Hypotheses should also be checked experimentally, and a model that explains the implications of a hypothesis, such as ours, provides motivation for doing the hard work of experimental testing. We think that R1 understands this and has been very helpful.

—————

The following is the authors’ response to the original reviews.

eLife assessment

This useful modeling study explores how the biophysical properties of interneuron subtypes in the basolateral amygdala enable them to produce nested oscillations whose interactions facilitate functions such as spike-timing-dependent plasticity. The strength of evidence is currently viewed as incomplete because the relevance to plasticity induced by fear conditioning is viewed as insufficiently grounded in existing training protocols and prior experimental results, and alternative explanations are not sufficiently considered. This work will be of interest to investigators studying circuit mechanisms of fear conditioning as well as rhythms in the basolateral amygdala. 

Most of our comments below are intended to rebut the sentence: “The strength of evidence is currently viewed as incomplete because the relevance to plasticity induced by fear conditioning is viewed as insufficiently grounded in existing training protocols and prior experimental results, and alternative explanations are not sufficiently considered”. 

We believe this work will be interesting to investigators interested in dynamics associated with plasticity, which goes beyond fear learning. It will also be of interest because of its emphasis on the interactions of multiple kinds of interneurons that produce dynamics used in plasticity, in the cortex (which has similar interneurons) as well as BLA. We note that the model has sufficiently detailed physiology to make many predictions that can be tested experimentally. Details are below in the answer to reviewers.

Reviewer #1 (Public Comments):  

(1) … the weakness is that their attempt to align with the experimental literature (specifically Krabbe et al. 2019) is performed inconsistently. Some connections between cell types were excluded without adequate justification (e.g. SOM+ to PV+). 

In order to constrain our model, we focused on what is reported in (Krabbe et al., 2019) in terms of functional connectivity instead of structural connectivity. Thus, we included only those connections for which there was strong functional connectivity. For example, the SOM to PV connection is shown to be small (Krabbe et al., 2019, Supp. Fig. 4, panel t). We also omitted PV to SOM, PV to VIP, SOM to VIP, VIP to excitatory projection neurons; all of these are shown in (Krabbe et al. 2019, Fig. 3 (panel l), and Supp. Fig. 4 (panels m,t)) to have weak functional connectivity, at least in the context of fear conditioning. 

We reply with more details below to the Recommendations for the Authors, including new text.

(2) The construction of the afferent drive to the network does not reflect the stimulus presentations that are given in fear conditioning tasks. For instance, the authors only used a single training trial, the conditioning stimulus was tonic instead of pulsed, the unconditioned stimulus duration was artificially extended in time, and its delivery overlapped with the neutral stimulus, instead of following its offset. These deviations undercut the applicability of their findings.  

Regarding the use of a single long presentation of US rather than multiple presentations (i.e., multiple trials): in early versions of this paper, we did indeed use multiple presentations. We were told by experimental colleagues that the learning could be achieved in a single trial. We note that, if there are multiple presentations in our modeling, nothing changes; once the association between CS and US is learned, the conductance of the synapse is stable. Also, our model does not need a long period of US if there are multiple presentations.  

We agree that, in order to implement the fear conditioning paradigm in our in-silico network, we made several assumptions about the nature of the CS and US inputs affecting the neurons in the BLA and the duration of these inputs. A Poisson spike train to the BLA is a signal that contains no structure that could influence the timing of the BLA output; hence, we used this as our CS input signal. We also note that the CS input can be of many forms in general fear conditioning (e.g., tone, light, odor), and we wished to de-emphasize the specific nature of the CS. The reference mentioned in the Recommendations for authors, (Quirk, Armony, and LeDoux 1997), uses pulses 2 seconds long. At the end of fear conditioning, the response to those pulses is brief. However, in the early stages of conditioning, the response goes on for as long as the figure shows. The authors do show the number of cells responding decreases from early to late training, which perhaps reflects increasing specificity over training. This feature is not currently in our model, but we look forward to thinking about how it might be incorporated. Regarding the CS pulsed protocol used in (Krabbe et al., 2019), it has been shown that intense inputs (6kHz and 12 kHz inputs) can lead to metabotropic effects that last much longer than the actual input (200 ms duration) (Whittington et al., Nature, 1995). Thus, the effective input to the BLA may indeed be more like Poisson.

Our model requires the effect of the CS and US inputs on the BLA neuron activity to overlap in time in order to instantiate fear learning. Despite paradigms involving both overlapping (delay conditioning, where US coterminates with CS (Lindquist et al., 2004), or immediately follows CS (e.g., Krabbe et al., 2019)) and non-overlapping (trace conditioning) CS/US inputs existing in the literature, we hypothesized that concomitant activity in CS- and US-encoding neuron activity should be crucial in both cases. This may be mediated by the memory effect, as suggested in the Discussion of our paper, or by metabotropic effects as suggested above, or by the contribution from other brain regions. We will emphasize in our revision that the overlap in time, however instantiated, is a hypothesis of our model. It is hard to see how plasticity can occur without some memory trace of US. This is a consequence of our larger hypothesis that fear learning uses spiketiming-dependent plasticity; such a hypothesis about plasticity is common in the modeling literature. 

We reply with more details below to the Recommendations for the Authors, including new text.

Reviewer #1 (Recommendations For The Authors): 

Major points: 

(1) This paper draws extensively from Krabbe et al. 2019, but it does not do so consistently. The paper would be strengthened if it tried to better match the circuit properties and activations.

Specifically: 

a. Krabbe found that PV interneurons were comparably activated by the US (see Supp Fig 1). Your model does not include that. The basis for the Krabbe 2019 claim that PV US responses are weaker is that they have a slightly larger proportion of cells inhibited by the US, but this is not especially compelling. In addition, their Fig 2 showed that VIP and SOM cells receive afferents from the same set of upstream regions. 

b. The model excluded PV-SOM connections, but this does not agree with Krabbe et al. 2019, Table 2. PV cells % connectivity and IPSC amplitudes were comparable to those from VIP interneurons. 

c. ECS to PV synapses are not included. This seems unlikely given the dense connectivity between PV interneurons and principal neurons in cortical circuits and the BLA (Woodruff and Sah 2007 give 38% connection probability in BLA). 

We thank the Reviewer for raising these points, which allow us to clarify how we constrained our model and to do more simulations. Specifically: 

a. (Wolff et al., Nature, 2014), cited by (Krabbe et al. 2018), reported that PV and SOM interneurons are on average inhibited by the US during the fear conditioning. However, we agree that (Krabbe et al., 2019) added to this by specifying that PV interneurons respond to both CS+ and US, although the fraction of US-inhibited PV interneurons is larger. As noted by the Reviewer, in the model we initially considered the PV interneurons responding only to CS+ (identified as “CS” in our manuscript). For the current revision, we ran new simulations in which the PV interneuron receives the US input, instead of CS+. It turned out that this did not affect the results, as shown in the figure below: all the network realizations learn the association between CS and fear. In the model, the PING rhythm between PV and F is the crucial component for establishing fine timing between ECS and F, which is necessary for learning. Having PV responding to the same input as F, i.e., US, facilitates their entrainment in PING and, thus, successful learning. 

As for afferents of VIP and SOM from upstream regions, in (Krabbe et al., 2019) is reported that “[…] BLA SOM interneurons receive a different array of afferent innervation compared to that of VIP and PV interneurons, which might contribute to the differential activity patterns observed during fear learning.” Thus, in the model, we are agnostic about inputs to SOM interneurons; we modeled them to fire spontaneously at high theta.

To address these points in the manuscript, we added some new text in what follows:

(1) New Section “An alternative network configuration characterized by US input to PV, instead of CS, also learns the association between CS and fear” in the Supplementary information:

“We constrained the BLA network in Fig. 2 with CS input to the PV interneuron, as reported in (Krabbe et al., 2018). However, (Krabbe et al., 2019) notes that a class of PV interneurons may be responding to US rather than CS. Fig. S3 presents the results obtained with this variation in the model (see Fig. 3 A,B for comparison) and shows that all the network realizations learn the association between CS and fear. In the model, the PING rhythm between PV and F is the crucial component for establishing fine timing between ECS and F, which is necessary for learning. Having PV responding to the same input as F, i.e., US, facilitates their entrainment in PING and, thus, successful fear learning.

We model the VIP interneuron as affected by US; in addition, (Krabbe et al. 2019) reports that a substantial proportion of them is mildly activated by CS. Replacing the US by CS does not change the input to VIP cells, which is modeled by the same constant applied current. Thus, the VIP CS-induced activity is a bursting activity at low theta, similar to the one elicited by US in Fig. 2.”

(2) Section “With the depression-dominated plasticity rule, all interneuron types are needed to provide potentiation during fear learning” in Results: “Finally, since (Krabbe et al., 2019) reported that a fraction of PV interneurons are affected by US, we have also run the simulations for single neuron network with the PV interneuron affected by US instead of CS. In this case as well, all the network realizations are learners (see Fig. S3). ”

(3) Section “Conditioned and unconditioned stimuli” in Materials and Methods: “To make Fig. S3, we also considered a variation of the model with PV interneurons affected by US, instead of CS, as reported in (Krabbe et al. 2019).”

b. Re the SOM to PV connection: As reported in the reply to the public reviews, we considered the prominent functional connections reported in (Krabbe et al., 2019), instead of structural connections. That is, we included only those connections for which there was strong functional connectivity. For example, the SOM to PV connection is shown to be small (Supp. Fig. 4, panel t, in (Krabbe et al., 2019)). We also omitted PV to SOM, PV to VIP, SOM to VIP, and VIP to excitatory projection neurons; all of these are shown in (Krabbe et al. 2019, Fig. 3 (panel l), and Supp. Fig. 4 (panels m,t)) to have weak functional connectivity, at least in the context of fear conditioning.

In order to clarify this point, in Section “Network connectivity and synaptic currents” in Materials and Methods, we now say:

“We modeled the network connectivity as presented in Fig. 2B, derived from the prominent functional, instead of structural, connections reported in (Krabbe et al., 2019).”

c. Re the ECS to PV synapses: We thank the Reviewer for the reference provided; as the Reviewer says, the ECS to PV synapses are not included. Upon adding this connection in our network, we found that, unlike the connection suggested in part a above, introducing these synapses would, in fact, change the outcome. Thus, the omission of this connection must be considered an implied hypothesis. Including those synapses with a significant strength would alter the PING rhythm created by the interactions between F and PV, which is crucial for ECS and F fine timing. Thanks very much for showing us that this needs to be said. Our hypothesis does not contradict the dense connections mentioned by the Reviewer; such dense connectivity does not mean that all pyramidal cells connect to all interneurons. This hypothesis may be taken as a prediction of the model.

The absence of this connection is now discussed at the end of a new Section of the Discussion entitled “Assumptions and predictions of the model”, which reads as follows:

“Finally, the model assumes the absence of significantly strong connections from the excitatory projection cells ECS to PV interneurons, unlike the ones from F to PV. Including those synapses would alter the PING rhythm created by the interactions between F and PV, which is crucial for ECS and F fine timing. We note that in (Woodruff and Sah, 2007) only 38% of the pyramidal cells are connected to PV cells. The functional identity of the connected pyramidal cells is unknown. Our model suggests that successful fear conditioning requires F to PV connections and that ECS to PV must be weak or absent.”

(2) Krabbe et al. 2019 and Davis et al. 2017 were referenced for the construction of the conditioned and unconditioned stimulus pairing protocol. The Davis citation is not applicable here because that study was a contextual, not cued, fear conditioning paradigm. Regarding Krabbe, the pairing protocol was radically different from what the authors used. Their conditioned stimulus was a train of tone pips presented at 0.9 Hz, which lasted 30 s, after which the unconditioned stimulus was presented after tone offset. The authors should determine how their network behaves when this protocol is used. Also, note that basolateral amygdala responses to tone stimuli are primarily brief onset responses (e.g. Quirk, Armony, and LeDoux 1997), and not the tonic activation used in the model.  

We replied to this point in our responses to the Reviewer’s Public Comments as follows:

“We agree that, in order to implement the fear conditioning paradigm in our in-silico network, we made several assumptions about the nature of the CS and US inputs affecting the neurons in the BLA and the duration of these inputs. A Poisson spike train to the BLA is a signal that contains no structure that could influence the timing of the BLA output; hence, we used this as our CS input signal. We also note that the CS input can be of many forms in general fear conditioning (e.g., tone, light, odor), and we wished to de-emphasize the specific nature of the CS. The reference mentioned in the Recommendations for authors, (Quirk, Armony, and LeDoux 1997), uses pulses 2 seconds long. At the end of fear conditioning, the response to those pulses is brief. However, in the early stages of conditioning, the response goes on for as long as the figure shows. The authors do show the number of cells responding decreases from early to late training, which perhaps reflects increasing specificity over training. This feature is not currently in our model, but we look forward to thinking about how it might be incorporated. Regarding the CS pulsed protocol used in (Krabbe et al., 2019), it has been shown that intense inputs (6kHz and 12 kHz inputs) can lead to metabotropic effects that last much longer than the actual input (200 ms duration) (Whittington et al., Nature, 1995). Thus, the effective input to the BLA may indeed be more like

Poisson.”

Current answer to the Reviewer:

There are several distinct issues raised by the Reviewer in the more detailed critique. We respectfully disagree that the model is not applicable to context-dependent fear learning where the context acts as a CS, though we should have been more explicit. Specifically, our CS input can describe both the cue and the context. We included the following text in the Results section “Interneuron rhythms provide the fine timing needed for depression-dominated STDP to make the association between CS and fear”:

“In our simulations, the CS input describes either the context or the cue in contextual and cued fear conditioning, respectively. For the context, the input may come from the hippocampus or other non-sensory regions, but this does not affect its role as input in the model.”

The second major issue is whether the specific training protocols used in the cited papers need to be exactly reproduced in the signals received by the elements of our model; we note that there are many transformations that can occur between the sensory input and the signals received by the BLA. In the case of auditory fear conditioning, a series of pips, rather than individual pips, are considered the CS (e.g., (Stujenske et al., 2014; Krabbe et al. 2019)). Our understanding is that a single pip does not elicit a fear response; a series of pips is required for fear learning. This indicates that it is not the neural code of a single pip that matters, but rather the signal entering the amygdala that incorporates any history-dependent signaling that could lead to spiking throughout the sequence of pips.  Also, as mentioned above, intense inputs at frequencies about 6kHz and 12kHz can lead to metabotropic effects that last much longer than each brief pip (~200 ms), thus possibly producing continuous activity in neurons encoding the input. Thus, we believe that our use of the Poisson spike train is reasonable. 

However, we are aware that the activity of neurons encoding CS can be modulated by the pips: neurons encoding auditory CS display a higher firing rate when each pip is presented and a Poisson-like spike train between pips (Herry et al., Journal of Neuroscience, 2007). Here we confirm that potentiation is present even in the presence of the fast transient response elicited by the pips. We said in the original manuscript that there is learning for a Poisson spike train CS input at ~50 Hz; this describes the neuronal activity in between pips. For the revision, we asked whether learning is preserved when CS is characterized by higher frequencies, which would describe the CS during and right after each pip. We show in the new Fig. S4 that potentiation is ensured for a range of CS frequencies. The figure shows the learning speed as a function of CS and US frequencies. For all the CS frequencies considered, i) there is learning, ii) learning speed increases with CS frequency. Thus, potentiation is present even when pips elicit a faster transient response.

To better specify this in the manuscript, 

We added the following sentences in the Results section “With the depressiondominated plasticity rule, all interneuron types are needed to provide potentiation during fear learning”: 

“We note that the CS and US inputs modeled as independent Poisson spike trains represent stimuli with no structure. Although we have not explicitly modeled pulsating pips, as common in auditory fear conditioning (e.g., (Stujenske 2014; Krabbe 2019)), we show in Fig. S4 that potentiation can be achieved over a relatively wide range of gamma frequencies. This indicates that overall potentiation is ensured if the gamma frequency transiently increases after the pip.”

We added the section “The full network potentiates for a range of CS frequencies“ and figure S4 in the Supplementary Information:

We included in Materials and Methods “Conditioned and unconditioned stimuli” the following sentences:

“Finally, for Fig.S4, we considered a range of frequencies for the CS stimulus. To generate the three Poisson spike trains with average frequencies from 48 to 64 Hz in Fig. S4, we set 𝜆 = 800, 1000, 1200.”

Finally, to address the comment about the need for CS and US overlapping in time to instantiate fear association, we added the following text in the Results section “Assumptions and predictions of the model”:

“Finally, our model requires the effect of the CS and US inputs on the BLA neuron activity to overlap in time in order to instantiate fear learning. Despite paradigms involving both overlapping (delay conditioning, where US co-terminates with CS (e.g., (Lindquist et al., 2004)), or immediately follows CS (e.g., Krabbe et al., 2019)) and non-overlapping (trace conditioning) CS/US inputs exist, we hypothesized that concomitant activity in CS- and US-encoding neuron activity should be crucial in both cases. This may be mediated by the memory effect due to metabotropic effects (Whittington et al., Nature, 1995) as suggested above, or by the contribution from other brain regions (see section “Involvement of other brain structures” in the Discussion). The fact that plasticity occurs with US memory trace is a consequence of our larger hypothesis that fear learning uses spike-timing-dependent plasticity; such a hypothesis about plasticity is common in the modeling literature.”

(3) As best as I could tell, only a single training trial was used in this study. Fair enough, especially given that fear learning can occur with a single trial. However, most studies of amygdala fear conditioning have multiple trials (~5 or more). How does the model perform when multiple trials are given?  

The association between CS and fear acquired after one trial, i.e., through a potentiated ECS to F connection, is preserved in the presence of multiple trials.  Indeed, the association would be weakened or erased (through depression of the ECS to F connection) only if ECS and F did not display good fine timing, i.e., F does not fire right after ECS most of the time. However, the implemented circuit supports the role of interneurons in providing the correct fine timing, thus preventing the association acquired from being erased.  

In the second paragraph of the Results section “With the depression-dominated plasticity rule, all interneuron types are needed to provide potentiation during fear learning”, we made the above point by adding the following text:

“We note that once the association between CS and fear is acquired, subsequent presentations of CS and US do not weaken or erase it: the interneurons ensure the correct timing and pauses in ECS and F activity, which are conducive for potentiation.”

(4) The LFP calculations are problematic. First, it is unclear how they were done. Did the authors just take the transmembrane currents they included and sum them, or were they scaled by distance from the 'electrode' and extracellular conductivity (as one would derive from the Laplace equation)? Presumably, the spatial arrangement of model neurons was neglected so distance was not a factor. 

Second, if this is the case, then the argument for excluding GABAergic conductances seems flawed. If the spatial arrangement of neurons is relevant to whether to include or exclude GABAergic conductances, then wouldn't a simulation without any spatial structure not be subject to the concern of laminar vs. nuclear arrangement? 

Moreover, to the best I can tell, the literature the authors use to justify the exclusion of

GABAergic currents does not make the case for a lack of GABAergic contribution in non-laminar structures. Instead, those studies only argue that in a non-laminar structure, AMPA currents are detectable, not that GABA cannot be detected. Thus, the authors should either include the GABAergic currents when calculating their simulated LFP, or provide a substantially better argument or citation for their exclusion. 

We thank the Reviewer for pointing this out; this comment helped us rethink how to model the LFP. The origin of the LFP signal in BLA has not been fully determined, but factors thought to be important include differences in the spatial extension of the arborization in excitatory and inhibitory neurons, in the number of synaptic boutons, and spatial distributions of somata and synapses (Lindén et al 2011; Łęski 2013; Mazzoni et al. 2015). In the first version of the manuscript, we excluded the GABAergic currents because it is typically assumed that they add very little to the extracellular field as the inhibitory reversal potential is close to the resting membrane potential. For the revision, we re-ran the simulations during pre and post fear conditioning and we modeled the LFP as the sum of the AMPA, GABA and NaP-/H-/D- currents. With this new version of the LFP, we added a new Fig. 6 showing that there is a significant increase in the low theta power, but not in the high theta power, with fear learning (Fig. 6 C, D, E). This increase in the low theta power was mainly due to the AMPA currents created by the newly established connection from ECS to F, which allowed F to be active after fear conditioning in response to CS. 

However, as the Reviewer mentioned, our network has no spatial extent: neurons are modeled as point cells. Thus, our current model does not include the features necessary to model some central aspects of the LFP. Despite that, our model does clearly demonstrate how rhythmic activity in the spike timing of neurons within the network changes due to fear learning (Fig. 6B). The spiking outputs of the network are key components of the inputs to the LFP, and thus we expect the rhythms in the spiking to be reflected in more complex descriptions of the LFP. But we also discovered that different LFP proxies provide different changes in rhythmic activity comparing pre- and post-fear learning; although we have no principled way to choose a LFP proxy, we believe that the rhythmic firing is the essential finding of the model.

We have added the following to the manuscript:

(1) In the new version of Fig. 6, we present the power spectra of the network spiking activity (panel B), along with the power spectra of the LFP proxy that includes the GABA, AMPA, and NaP-/H-/D- currents (panels C, D, E). 

(2) We modified the conclusion of the Results section entitled “Increased low-theta frequency is a biomarker of fear learning” by saying:

“In this section, we explore how plasticity in the fear circuit affects the network dynamics, comparing after fear conditioning to before. We first show that fear conditioning leads to an increase in low theta frequency power of the network spiking activity compared to the pre-conditioned level (Fig. 6 A,B); there is no change in the high theta power. We also show that the LFP, modeled as the linear sum of all the AMPA, GABA, NaP-, D-, and H- currents in the network, similarly reveals a low theta power increase and no significant variation in the high theta power (Fig. 6 C,D,E). These results reproduce the experimental findings in (Davis et al., 2017), and (Davis et al., 2017), and Fig 6 F,G show that the low theta increase is due to added excitation provided by the new learned pathway. The additional unresponsive ECS and F cells in the network were included to ensure we had not biased the LFP towards excitation. Nevertheless, although both the AMPA and GABA currents contribute to the power increase in the low theta frequency range (Fig. 6F), the AMPA currents show a dramatic power increase relative to the baseline (the average power ratio of AMPA and GABA post- vs pre-conditioning across 20 network realizations is 3*103 and 4.6, respectively). This points to the AMPA currents as the major contributor to the low theta power increase. Specifically, the newly potentiated AMPA synapse from ECS to F ensures F is active after fear conditioning, thus generating strong currents in the PV cells to which it has strong connections (Fig. 6G). Finally, the increase in power is in the low theta range because ECS and F are allowed to spike only during the active phase of the low theta spiking VIP neurons. We have also explored another proxy for the LFP (see Supplementary Information and Fig. S6).”

In the Supplementary Information, we included a figure and some text in the new section entitled “A higher low theta power increase emerges in LFP approximated with the sum of the absolute values of the currents compared to their linear sum”:

“Given that our BLA network comprises a few neurons described as single-compartment cells with no spatial extension and location, the LFP cannot be computed directly from our model’s read-outs. In the main text, we choose as an LFP proxy the linear sum of the AMPA, GABA, and P-/H-/D-currents. We note that if the LFP is modeled as the sum of the absolute value of the currents, as suggested by (Mazzoni et al. 2008; Mazzoni et al. 2015), an even higher low theta power increase arises after fear conditioning compared to the linear sum. Differences in the power spectra also arise if other LFP proxies (e.g., only AMPA currents, only GABA currents) are considered. A principled description of an LFP proxy would require modeling the three-dimensional BLA anatomy, including that of the interneurons VIP and SOM; this is outside the scope of the current paper. (See (Feng et al. 2019) for a related project in the BLA.)”

(3) We updated the Materials and Methods section “Local field potentials and spectral analysis” to explain how we compute the LFP in the revised manuscript: 

“We considered as an LFP proxy as the linear sum of all the AMPA, GABA, NaP, D, and H currents in the network. The D-current is in the VIP interneurons, and NaP-current and H-current are in SOM interneurons.”

Although it is beyond the scope of the current work, an exploration of the most accurate proxy of the LFP in the amygdala is warranted. Such a study could be accomplished by adopting a similar approach as in (Mazzoni et al., 2015), where several LFP proxies based on point-neuron leaky-integrate and fire neuronal network were compared with a “groundtruth” LFP obtained in an analogous realistic three-dimensional network model. 

To explicitly mention this issue in the paper, we add a paragraph in the “Limitations and caveats” section in the Discussion, which reads as follows:

“LFPs recorded in the experiments are thought to be mainly created by transmembrane currents in neurons located around the electrode and depend on several factors, including the morphology of the arborization of contributing neurons and the location of AMPA and GABA boutons (Katzner et al. 2009; Lindén et al 2011; Łęski 2013; Mazzoni et al. 2015). Since our model has no spatial extension, we used an LFP proxy; this proxy was shown to reflect the rhythmic output of the network, which we believe to be the essential result (for more details see Results “Increased low-theta frequency is a biomarker of fear learning”, and Supplementary Information “A higher low theta power increase emerges in LFP approximated with the sum of the absolute values of the currents compared to their linear sum”).”

(4)     We have removed the section “Plasticity between fear neuron and VIP slows down overall potentiation” in Results and sections “Plasticity between the fear neuron (F) and VIP slows down overall potentiation” and “Plastic F to VIP connections further increase lowtheta frequency power after fear conditioning” in the Supplementary Information. This material is extraneous since we are using a new proxy for LFP.

Minor points: 

(1) In Figure 3C, the y-axis tick label for 0.037 is written as "0.37."

We thank the reviewer for finding this typo; we fixed it.

(2) Figure 5B is unclear. It seems to suggest that the added ECS and F neurons did not respond to either the CS or UCS. Is this true? If so, why include them in the model? How would their inclusion change the model behavior? 

It is correct that the added ECS and F neurons did not respond to the CS or US (UCS); they are constructed to be firing at 11 Hz in the absence of any connections from other cells.  These cells were included to be part of our computation of the LFP.  Specifically, adding in those cells would make the LFP take inhibition into account more, and we wanted to make sure that were not biasing our computation away from the effects of inhibition.  As shown in the paper (Fig. 6B), even with inhibition onto these non-responsive cells, the LFP has the properties claimed in the paper concerning the changes in the low theta and high-theta power, because the LFP is dominated by new excitation rather than the inhibition. 

First, in the Results section “Network with multiple heterogeneous neurons can establish the association between CS and fear”, we commented on the added ECS and F neurons that do not respond to either CS or US by saying the following:

“The ECS cells not receiving CS are inhibited by ongoing PV activity during the disinhibition window (Fig. 5B); they are constructed to be firing at 11 Hz in the absence of any connections from other cells. The lack of activity in those cells during fear conditioning implies that there is no plasticity from those ECS cells to the active F. Those cells are included for the calculation of the LFP (see below in “Increased low-theta frequency is a biomarker of fear learning”.)”

Furthermore, we add the following sentence in the Results section “Increased low-theta frequency is a biomarker of fear learning”: 

“The additional unresponsive ECS and F cells in the network were included to ensure we had not biased the LFP towards excitation.”

(3) Applied currents are given as current densities, but these are difficult to compare with current levels observed from whole-cell patch clamp recordings. Can the currents be given as absolute levels, in pA/nA. 

In principle, it is possible to connect current densities with absolute levels, as requested. However, we note that the number of cells in models is orders of magnitude smaller than the number being modeled. It is common in modeling to adjust physiological parameters to achieve the qualitative properties that are important to the model, rather than trying to exactly match particular recordings.

We added to the Methods description why we choose units per unit area, rather than absolute units. 

“All the currents are expressed in units per area, rather than absolute units, to avoid making assumptions about the size of the neuron surface.”

(4) Regarding: "We note that the presence of SOM cells is crucial for plasticity in our model since they help to produce the necessary pauses in the excitatory projection cell activity. However, the high theta rhythm they produce is not crucial to the plasticity: in our model, high theta or higher frequency rhythms in SOM cells are all conducive to associative fear learning. This opens the possibility that the high theta rhythm in the BLA mostly originates in the prefrontal cortex and/or the hippocampus (Stujenske et al., 2014, 2022)." The chain of reasoning in the above statement is unclear. The second sentence seems to be saying contradictory things. 

We agree that the sentence was confusing; thank you for pointing it out. We have revised the paragraph to make our point clearer. The central points are: 1) having the SOM cells in the BLA is critical to the plasticity in the model, and 2) these cells may or may not be the source of the high theta observed in the BLA during fear learning.

We deleted from the discussion the text reported by the Reviewer, and we added the following one to make this point clearer:

“We note that the presence of SOM cells is crucial for plasticity in our model since they help to produce the necessary pauses in the excitatory projection cell activity. The BLA SOM cells do not necessarily have to be the only source of the high theta observed in the BLA during fear learning; the high theta detected in the LFP of the BLA also originates from the prefrontal cortex and/or the hippocampus (Stujenske et al., 2014, 2022).”

(5) Regarding: "This suggests low theta power change is not just an epiphenomenon but rather a biomarker of successful fear conditioning." Not sure this is the right framing for the above statement. The power of the theta signal in the LFP reflects the strengthening of connections, but it itself does not have an impact on network activity. Moreover, whether something is epiphenomenal is not relevant to the question of whether it can serve as a successful biomarker. A biomarker just needs to be indicative, not causal. 

We intended to say why the low theta power change is a biomarker in the sense of the Reviewer. That is: experiments have shown that, with learning, the low theta power increases. The modeling shows in addition that, when learning does not take place, the low power does not increase. That means that the low theta power increases if and only if there is learning, i.e., the change in low theta power is a biomarker. To make our meaning clearer, we have changed the quoted sentences to read: 

“This suggests that the low theta power change is a biomarker of successful fear conditioning: it occurs when there is learning and does not occur when there is no learning.”

Reviewer #2 (Public Comments): 

We thank the Reviewer for raising these interesting points. Below are our public replies and the changes we made to the manuscript to address the Reviewer’s objections.

(1) Gamma oscillations are generated locally; thus, it is appropriate to model in any cortical structure. However, the generation of theta rhythms is based on the interplay of many brain areas therefore local circuits may not be sufficient to model these oscillations.

Moreover, to generate the classical theta, a laminal structure arrangement is needed (where neurons form layers like in the hippocampus and cortex)(Buzsaki, 2002), which is clearly not present in the BLA. To date, I am not aware of any study which has demonstrated that theta is generated in the BLA. All studies that recorded theta in the BLA performed the recordings referenced to a ground electrode far away from the BLA, an approach that can easily pick up volume conducted theta rhythm generated e.g., in the hippocampus or other layered cortical structure. To clarify whether theta rhythm can be generated locally, one should have conducted recordings referenced to a local channel (see Lalla et al., 2017 eNeuro). In summary, at present, there is no evidence that theta can be generated locally within the BLA. Though, there can be BLA neurons, firing of which shows theta rhythmicity, e.g., driven by hippocampal afferents at theta rhythm, this does not mean that theta rhythm per se can be generated within the BLA as the structure of the BLA does not support generation of rhythmic current dipoles. This questions the rationale of using theta as a proxy for BLA network function which does not necessarily reflect the population activity of local principal neurons in contrast to that seen in the hippocampus.

In both modeling and experiments, a laminar structure does not seem to be needed to produce a theta rhythm. A recent experimental paper, (Antonoudiou et al. 2022), suggests that the BLA can intrinsically generate theta oscillations (3-12 Hz) detectable by LFP recordings under certain conditions, such as reduced inhibitory tone. The authors draw this conclusion by looking at mice ex vivo slices. The currents that generate these rhythms are in the BLA, since the hippocampus was removed to eliminate hippocampal volume conduction and other nearby brain structures did not display any oscillatory activity. Also, in the modeling literature, there are multiple examples of the production of theta rhythms in small networks not involving layers; these papers explain the mechanisms producing theta from non-laminated structures (Dudman et al., 2009, Kispersky et al., 2010, Chartove et al. 2020).  We are not aware of any model description of the mechanisms of theta that do require layers.

We added the following text in the introduction of the manuscript to make this point clearer:  “A recent rodent experimental study (Antonoudiou et al. 2022) suggests that BLA can intrinsically generate theta oscillations (3-12 Hz).”

(2) The authors distinguished low and high theta. This may be misleading, as the low theta they refer to is basically a respiratory-driven rhythm typically present during an attentive state (Karalis and Sirota, 2022; Bagur et al., 2021, etc.). Thus, it would be more appropriate to use breathing-driven oscillations instead of low theta. Again, this rhythm is not generated by the BLA circuits, but by volume conducted into this region. Yet, the firing of BLA neurons can still be entrained by this oscillation. I think it is important to emphasize the difference.

Many rhythms of the nervous system can be generated in multiple parts of the brain by multiple mechanisms. We do not dispute that low theta appears in the context of respiration; however, this does not mean that other rhythms with the same frequencies are driven by respiration. Indeed, in the response to question 1 above, we showed that theta can appear in the BLA without inputs from other regions. In our paper, the low theta is generated in the BLA by VIP neurons. Using intrinsic currents known to exist in VIP neurons (Porter et al., 1998), modeling has shown that such neurons can intrinsically produce a low theta rhythm. This is also shown in the current paper. This example is part of a substantial literature showing that there are multiple mechanisms for any given frequency band. 

To elaborate more on this in the manuscript, we added the following new section in the discussion:

“Where the rhythms originate, and by what mechanisms. A recent experimental paper, (Antonoudiou et al. 2022), suggests that the BLA can intrinsically generate theta oscillations (3-12 Hz) detectable by LFP recordings under certain conditions, such as reduced inhibitory tone. They draw this conclusion in mice by removing the hippocampus, which can volume conduct to BLA, and noticing that other nearby brain structures did not display any oscillatory activity. Our model also supports the idea that intrinsic mechanisms in the BLA can support the generation of the low theta, high theta, and gamma rhythms. 

Although the BLA can produce these rhythms, this does not rule out that other brain structures also produce the same rhythms through different mechanisms, and these can be transmitted to the BLA. Specifically, it is known that the olfactory bulb produces and transmits the respiratory-related low theta (4 Hz) oscillations to the dorsomedial prefrontal cortex, where it organizes neural activity (Bagur et al., 2021). Thus, the respiratory-related low theta may be captured by BLA LFP because of volume conduction or through BLA extensive communications with the prefrontal cortex. Furthermore, high theta oscillations are known to be produced by the hippocampus during various brain functions and behavioral states, including during spatial exploration (Vanderwolf, 1969) and memory formation/retrieval (Raghavachari et al., 2001), which are both involved in fear conditioning. Similarly to the low theta rhythm, the hippocampal high theta can manifest in the BLA. It remains to understand how these other rhythms may interact with the ones described in our paper.”

We also note that the presence of D-currents in the BLA VIP interneurons should be confirmed experimentally, and that the ability of VIP interneurons to generate the BLA low theta rhythm constitutes a prediction of our computational model. These points are specified in the first paragraph in the Discussion entitled “Assumptions and predictions of the model”:

“The interneuron descriptions in the model were constrained by the electrophysiological properties reported in response to hyperpolarizing currents (Sosulina et al., 2010). Specifically, we modeled the three subtypes of VIP, SOM, and PV interneurons displaying bursting behavior, regular spiking with early spike-frequency adaptation, and regular spiking without spike-frequency adaptation, respectively. Focusing on VIP interneurons, we were able to model the bursting behavior by including the D-type potassium current. This current is thought to exist in the VIP interneurons in the cortex (Porter et al., 1998), but whether this current is also found in the VIP interneurons the BLA is still unknown. Similarly, we endowed the SOM interneurons with NaP- and H-currents, as the OLM cells in the hippocampus. Due to these currents, the VIP and SOM cells are able to show  low- and high-theta oscillations, respectively. The presence of these currents and the neurons’ ability to exhibit oscillations in the theta range during fear conditioning and at baseline in BLA, which are assumptions of our model, should be tested experimentally.”

(3) The authors implemented three interneuron types in their model, ignoring a large fraction of GABAergic cells present in the BLA (Vereczki et al., 2021). Recently, the microcircuit organization of the BLA has been more thoroughly uncovered, including connectivity details for PV+ interneurons, firing features of neurochemically identified interneurons (instead of mRNA expression-based identification, Sosulina et al., 2010), synaptic properties between distinct interneuron types as well as principal cells and interneurons using paired recordings. These recent findings would be vital to incorporate into the model instead of using results obtained in the hippocampus and neocortex. I am not sure that a realistic model can be achieved by excluding many interneuron types.

The interneurons and connectivity that we used were inspired by the functional connectivity reported in (Krabbe et al., 2019) (see above answer to Reviewer #1). As reported in (Vereczki et al., 2021), there are multiple categories and subcategories of interneurons; that paper does not report on which ones are essential for fear conditioning. We did use all the highly represented categories of the interneurons, except NPYcontaining neurogliaform cells.

The Reviewer says “I am not sure that a realistic model can be achieved by excluding many interneuron types”. We agree with the Reviewer that discarding the introduction of other interneurons subtypes and the description of more specific connectivity (soma-, dendrite-, and axon-targeting connections) may limit the ability of our model to describe all the details in the BLA. However, this work represents a first effort towards a biophysically detailed description of the BLA rhythms and their function. As in any modeling approach, assumptions about what to describe and test are determined by the scientific question; details postulated to be less relevant are omitted to obtain clarity. The interneuron subtypes we modeled, especially VIP+ and PV+, have been reported to have a crucial role in fear conditioning (Krabbe et al., 2019). Other interneurons, e.g. cholecystokinin and SOM+, have been suggested as essential in fear extinction. Thus, in the follow-up of this work to explain fear extinction, we will introduce other cell types and connectivity. In the current work, we have achieved our goals of explaining the origin of the experimentally found rhythms and their roles in the production of plasticity underlying fear learning. Of course, a more detailed model may reveal flaws in this explanation, but this is science that has not yet been done.

We elaborate more on this in a new section in the Discussion entitled “Assumptions and predictions of the model”. The paragraph related to this point reads as follows:

“Our model, which is a first effort towards a biophysically detailed description of the BLA rhythms and their functions, does not include the neuron morphology, many other cell types, conductances, and connections that are known to exist in the BLA; models such as ours are often called “minimal models” and constitute the majority of biologically detailed models. Such minimal models are used to maximize the insight that can be gained by omitting details whose influence on the answers to the questions addressed in the model are believed not to be qualitatively important. We note that the absence of these omitted features constitutes hypotheses of the model: we hypothesize that the absence of these features does not materially affect the conclusions of the model about the questions we are investigating. Of course, such hypotheses can be refuted by further work showing the importance of some omitted features for these questions and may be critical for other questions. Our results hold when there is some degree of heterogeneity of cells of the same type, showing that homogeneity is not a necessary condition.”

(4) The authors set the reversal potential of GABA-A receptor-mediated currents to -80 mV. What was the rationale for choosing this value? The reversal potential of IPSCs has been found to be -54 mV in fast-spiking (i.e., parvalbumin) interneurons and around -72 mV in principal cells (Martina et al., 2001, Veres et al., 2017).

A GABA-A reversal potential around -80 mV is common in the modeling literature (Jensen et al., 2005; Traub et al., 2005; Kumar et al., 2011; Chartove et al., 2020). Other computational works of the amygdala, e.g. (Kim et al., 2016), consider GABA-A reversal potential at -75 mV based on the cortex (Durstewitz et al., 2000). The papers cited by the reviewer have a GABA-A reversal potential of -72 mV for synapses onto pyramidal cells; this is sufficiently close to our model that it is not likely to make a difference. For synapses onto PV+ cells, the papers cited by the reviewer suggest that the GABA-A reversal potential is -54 mV; such a reversal potential would lead these synapses to be excitatory instead of inhibitory. However, it is known (Krabbe et al., 2019; Supp. Fig. 4b) that such synapses are in fact inhibitory. Thus, we wonder if the measurements of Martina and Veres were made in a condition very different from that of Krabbe. For all these reasons, we consider a GABA-A reversal potential around -80 mV in amygdala to be a reasonable assumption.

In section “Network connectivity and synaptic currents” in “Materials and Methods” we provided references to motivate our choice of considering a GABA-A reversal potential around -80 mV:

“The GABAa current reversal potential (𝐸!) is set to −80        𝑚𝑉, as common in the modeling literature (Jensen et al., 2005; Traub et al., 2005; Kumar et al., 2011; Chartove et al., 2020).”

(5) Proposing neuropeptide VIP as a key factor for learning is interesting. Though, it is not clear why this peptide is more important in fear learning in comparison to SST and CCK, which are also abundant in the BLA and can effectively regulate the circuit operation in cortical areas.

Other peptides seem to be important in overall modulation of fear, but VIP is especially important in the first part of fear learning, the subject of our paper. Re SST: we hypothesize that SST interneurons are critical in fear extinction and preventing fear generalization, but not to initial fear learning. The peptide of the CCK neurons, which overlap with VIP cells, has been proposed to promote the switch between fear and safety states after fear extinction (Krabbe al. 2018). Thus, these other peptides are likely more important for other aspects of fear learning.  

In the Discussion, we have added:

“We hypothesize that SST peptide is critical in fear extinction and preventing fear generalization, but not to initial fear learning. Also, the CCK peptide has been proposed to promote the switch between fear and safety states after fear extinction (Krabbe al. 2018).”

Reviewer #2 (Recommendations For The Authors): 

We note that Reviewer #2’s Recommendations For The Authors have the same content as the Public Comments. Thus, the changes to the manuscript we implemented above address also the private critiques listed below.

(1) As the breathing-driven rhythm is a global phenomenon accompanying fear state, one might restrict the analysis to this oscillation. The rationale beyond this restriction is that the 'high' theta in the BLA has an unknown origin (since it can originate from the ventral hippocampus, piriform cortex etc.). 

In response to point 4 made by Reviewer 1 (Recommendations for the Authors) (p. 13), referring to high theta in the BLA, we previously wrote: 1) having the SOM cells in the BLA is critical to the plasticity in the model, and 2) these cells may or may not be the source of the high theta observed in the BLA during fear learning.

In the Public Critiques, Reviewer 2 relates the respiratory rhythm to the low theta. We answered this point in point 2 of the Reviewer’s Public Comments (at p. 15).

(2) I would include more interneurons in the network model incorporating recent findings. 

This point was answered in our response to point 3 of the Reviewer’s Public Comments.

(3) The reversal potential for GABA-A receptor-mediated currents would be good to set to measured values. In addition, I would use AMPA conductance values that have been measured in the BLA. 

We addressed this objection in our response to point 4 of the Reviewer’s Public Comments.

Reviewer #3 (Public comments):

Weaknesses: 

(1) The main weakness of the approach is the lack of experimental data from the BLA to constrain the biophysical models. This forces the authors to use models based on other brain regions and leaves open the question of whether the model really faithfully represents the basolateral amygdala circuitry. 

(2) Furthermore, the authors chose to use model neurons without a representation of the morphology. However, given that PV+ and SOM+ cells are known to preferentially target different parts of pyramidal cells and given that the model relies on a strong inhibition form SOM to silence pyramidal cells, the question arises whether SOM inhibition at the apical dendrite in a model representing pyramidal cell morphology would still be sufficient to provide enough inhibition to silence pyramidal firing.

3) Lastly, the fear learning relies on the presentation of the unconditioned stimulus over a long period of time (40 seconds). The authors justify this long-lasting input as reflecting not only the stimulus itself but as a memory of the US that is present over this extended time period. However, the experimental evidence for this presented in the paper is only very weak.

We are repeating here the answers we gave in response to the public comments, adding further relevant points.

(1) Our neurons were constrained by electrophysiology properties in response to hyperpolarizing currents in the BLA (Sosulina et al., 2010). We can reproduce these electrophysiological properties by using specific membrane currents known to be present in similar neurons in other brain regions (D-current in VIP interneurons in the cortex, and NaP- and H-currents in OLM/SOM cells in the hippocampus). Also, though a much more detailed description of BLA interneurons was given in (Vereczki et al., 2021), it is not clear that this level of detail is relevant to the questions that we were asking, especially since the experiments described were not done in the context of fear learning.

(2) It is true that we did not include the morphology, which undoubtedly makes a difference to some aspects of the circuit dynamics. Furthermore, it is correct that the model relies on a strong inhibition from SOM and PV to silence the excitatory projection neurons. We agree that the placement of the SOM inhibition on the pyramidal neurons can make a difference on some aspects of the circuit behavior. We are assuming that the inhibition from the SOM cells can inhibit the pyramidal cells firing, which can be seen as a hypothesis of our model. It is well known that VIP cells disinhibit pyramidal cells through inhibition of SOM and PV cells (Krabbe et al. 2019); hence, this hypothesis is generally believed. This choice of parameters comes from using simplified models: it is standard in modeling to adjust parameters to compensate for simplifications.

Re points 1) and 2), in a new paragraph (“Assumptions and predictions of the model”) in the Discussion reported in response to Reviewer #2 (public comments)’s point 3, we stated that modeling requires the omission of many details to bring out the significance of other details.

(3) 40 seconds is the temporal interval we decided to use to present the results. In the Results, we also showed that there is learning over a shorter interval of time (15 seconds) where CS and US/memory of US should both be present. Thus, our model requires 15 seconds over a single or multiple trials for associative learning to be established. We included references to additional experimental papers to support our reasoning in the last paragraph of section “Assumptions and predictions of the model” in the Discussion, also reported in response to Reviewer #1 point 2 (Recommendations for the Authors). We said there that some form of memory or overlap in the activity of the excitatory projection neurons is necessary for spike-timing-dependent plasticity.

The authors achieved the aim of constructing a biophysically detailed model of the BLA not only capable of fear learning but also showing spectral signatures seen in vivo. The presented results support the conclusions with the exception of a potential alternative circuit mechanism demonstrating fear learning based on a classical Hebbian (i.e. non-depression-dominated) plasticity rule, which would not require the intricate interplay between the inhibitory interneurons. This alternative circuit is mentioned but a more detailed comparison between it and the proposed circuitry is warranted.

Our model accounts for the multiple rhythms observed in the context of fear learning, as well as the known involvement of multiple kinds of interneurons. We did not say explicitly enough why our complicated model may be functionally important in ways that cannot be fulfilled with a simpler model with the non depression-dominated Hebbian rule. To explain this, we have added the following in the manuscript discussion: 

“Although fear learning can occur without the depression-dominated rule, we hypothesize that it is necessary for other aspects of fear learning and regulation. That is, in pathological cases, there can be overgeneralization of learning. We hypothesize that the modulation created by the involvement of these interneurons is normally used to prevent such overgeneralization. However, this is beyond the scope of the present paper.”

We have also written an extra paragraph about generalization in the Discussion “Synaptic plasticity in our model”:

“With the classical Hebbian plasticity rule, we show that learning can occur without the involvement of the VIP and SOM cells. Although fear learning can occur without the depressiondominated rule, we hypothesize that the latter is necessary for other aspects of fear learning and regulation. Generalization of learning can be pathological, and we hypothesize that the modulation created by the involvement of VIP and SOM interneurons is normally used to prevent such overgeneralization. However, in some circumstances, it may be desirable to account for many possible threats, and then a classical Hebbian plasticity rule could be useful. We note that the involvement or not of the VIP-SOM circuit has been implicated when there are multiple strategies for solving a task (Piet et al., 2024). In our situation, the nature of the task (including reward structure) may determine whether the learning rule is depression-dominated and therefore whether the VIP-SOM circuit plays an important role.”

Reviewer #3 (Recommendations For The Authors): 

We thank the Reviewer for all the recommendations. We replied to each of them below.

In general, there are some inconsistencies in the naming (e.g. sometimes you write PV sometimes PV+,...), please use consistent abbreviations throughout the manuscript. You also introduce some of the abbreviations multiple times. 

We modified the manuscript to remove all the inconsistencies in the naming. 

Introduction: 

- In the last section you speak about one recent study but actually cite two articles. 

We removed the reference to (Perrenoud and Cardin, 2023), which is a commentary on the Veit et al. article.

Results: 

- 'Brain rhythms are thought to be encoded and propagated largely by interneurons' What do you mean by encoded here? 

We agree with the Reviewer that the verb “to encode” is not accurate. We modified the sentence as follows:

“Brain rhythms are thought to be generated and propagated largely by interneurons”.

- The section 'Interneurons interact to modulate fear neuron output' could be clearer. Start with describing the elements of the circuit, then the rhythms in the baseline. 

We reorganized the section as follows:

“Interneurons interact to modulate fear neuron output. Our BLA network consists of interneurons, detailed in the previous section, and excitatory projection neurons (Fig. 2A). Both the fear-encoding neuron (F), an excitatory projection neuron, and the VIP interneuron are activated by the noxious stimulus US (Krabbe et al., 2019). As shown in Fig. 2A (top, right), VIP disinhibits F by inhibiting both SOM and PV, as suggested in (Krabbe et al., 2019). We do not include connections from PV to SOM and VIP, nor connections from SOM to PV and VIP, since those connections have been shown to be significantly weaker than the ones included (Krabbe et al., 2019). The simplest network we consider is made of one neuron for each cell type. We introduce a larger network with some heterogeneity in the last two sections of the Results.

Fig. 2A (bottom) shows a typical dynamic of the network before and after the US input onset, with US modeled as a Poisson spike train at ~50 Hz; the network produces all the rhythms originating from the interneurons alone or through their interactions with the excitatory projection neurons (shown in Fig. 1). Specifically, since VIP is active at low theta during both rest and upon the injection of US, it then modulates F at low theta cycles via SOM and PV. In the baseline condition, the VIP interneuron has short gamma bursts nested in low theta rhythm. With US onset, VIP increases its burst duration and the frequency of low theta rhythm. These longer bursts make the SOM cell silent for long periods of each low theta cycle, providing F with windows of disinhibition and contributing to the abrupt increase in activity right after the US onset. Finally, in Fig. 2A, PV lacks any external input and fires only when excited by F. Thanks to their reciprocal interactions, PV forms a PING rhythm with F, as depicted in Fig.1C.”

- Figure 3C: The lower dashed line has the tick label '0.37' which should read '0.037'. 

We fixed it.

- The section describing the network with multiple neurons could be clearer, especially, it is not really clear how these different ECS and F neurons receive their input. 

We answered the same objection in the reply to Reviewer #1 in point 2 under “minor issues.”

Discussion: 

- The paragraph 'It has also been suggested that ventral tegmental area has a role in fear expression (Lesas et al.,2023). Furthermore, it has been reported that the prelimbic cortex (PL) modulates the BLA SOM cells during fear retrieval, and the latter cells are crucial to discriminate non-threatening cues when desynchronized by the PL inputs (Stujenske et al., 2022).' is merely stating facts but I don't see how they relate to the presented work. 

We thank the Reviewer for pointing out that this was confusing. What we meant to emphasize was that later stages of fear conditioning and extinction appear to require more than the BLA. We specifically mention the discrimination of non-threatening cues at the end of the paragraph, which now reads as follows:

“Other brain structures may be involved in later stages of fear responsiveness, such as fear extinction and prevention of generalization. It has been reported that the prelimbic cortex (PL) modulates the BLA SOM cells during fear retrieval, and the latter cells are crucial to discriminate non-threatening cues when desynchronized by the PL inputs (Stujenske et al., 2022). Brain structures such as the prefrontal cortex and hippocampus have been documented to play a crucial role also in fear extinction, the paradigm following fear conditioning aimed at decrementing the conditioned fearful response through repeated presentations of the CS alone. As reported by several studies, fear extinction suppresses the fear memory through the acquisition of a distinct memory, instead of through the erasure of the fear memory itself (Harris et al., 2000; Bouton, 2002; Trouche et al., 2013; Thompson et al., 2018). Davis et al., 2017 found a high theta rhythm following fear extinction that was associated with the suppression of threat in rodents. Our model can be extended to include structures in the prefrontal cortex and the hippocampus to further investigate the role of rhythms in the context of discrimination of non-threatening cues and extinction. We hypothesize that a different population of PV interneurons plays a crucial role in mediating competition between fearful memories, associated with a low theta rhythm, and safety memories, associated with a high theta rhythm; supporting experimental evidence is in (Lucas et al., 2016; Davis et al., 2017; Chen et al., 2022).”

- The comparison to other models BLA is quite short and seems a bit superficial. A more indepth comparison seems warranted. 

We thank the reviewer for suggesting that a more in-depth comparison between our and other models in the literature would improve the manuscript. We rewrote entirely the first paragraph of that section. The new content reads as follows:

“Comparison with other models. Many computational models that study fear conditioning have been proposed in the last years; the list includes biophysically detailed models (e.g., (Li 2009; Kim et al., 2013a)), firing rate models (e.g., Krasne 2011; Ball 2012; Vlachos 2011), and connectionist models (e.g., Moustafa 2013; Armony 1997; Edeline 1992) (for a review see (Nair et al., 2016)). Both firing rate models and connectionist models use an abstract description of the interacting neurons or regions. The omission of biophysical details prevents such models from addressing questions concerning the roles of dynamics and biophysical details in fear conditioning, which is the aim of our model.  There are also biophysically detailed models (Li 2009; Kim 2013; Kim 2016; Feng 2019), which differ from ours in both the physiology included in the model and the description of how plastic changes take place.  One main difference in the physiology is that we differentiated among types of interneurons, since the fine timing produced for the latter was key to our use of rhythms to produce spike-time dependent plasticity. The origin of the gamma rhythm (but not the other rhythms) was investigated in Feng et al 2019, but none of these papers connected the rhythms to plasticity.

The most interesting difference between our work and that in (Li 2009; Kim 2013; Kim 2016) is the modeling of plasticity.  We use spike-time dependent plasticity rules.  The models in (Li 2009; Kim 2013; Kim 2016) were more mechanistic about how the plasticity takes place, starting with the known involvement of calcium with plasticity.  Using a hypothesis about back propagation of spikes, the set of papers together come up with a theory that is consistent with STDP and other instantiations of plasticity (Shouval 2002a; Shouval 2002b).  For the purposes of our paper, this level of detail, though very interesting, was not necessary for our conclusions.  By contrast, in order for the rhythms and the interneurons to have the dynamic roles they play in the model, we needed to restrict our STDP rule to ones that are depression-dominated.  Our reading of (Shouval 2002) suggests to us that such subrules are possible outcomes of the general theory.  Thus, there is no contradiction between the models, just a difference in focus; our focus was on the importance of the much-documented rhythms (Seidenbecher et al., 2003; Courtin et al., 2014b; Stujenske et al., 2014; Davis et al., 2017) in providing the correct spike timing.  We showed in the Supplementary Information (“Classical Hebbian plasticity rule, unlike the depression-dominated one, shows potentiation even with no strict pre and postsynaptic spike timing”) that if the STDP rule was not depression dominated, the rhythms need not be necessary.  We hypothesize that the necessity of strict timing enforced by the depression-dominated rule may foster the most appropriate association with fear at the expense of less relevant associations.”

- The paragraph 'This could happen among some cells responding to weaker sensory inputs that do not lead to pre-post timing with fear neurons. This timing could be modified by the "triconditional rule", as suggested in (Grewe et al., 2017).' is not very clear. What exactly is 'this' in the first sentence referring to? If you mention the 'tri-conditional rule' here, please briefly explain it and how it would solve the issue at hand here.  

We apologize that the sentence reported was not sufficiently clear. “This” refers to “depression”. We meant that, in our model, depression during fear conditioning happens every time there is no pre-post timing between neurons encoding the neutral stimuli and fear cells; poor pre-post timing can characterize the activity of neurons responding to weaker sensory inputs and does not lead to associative learning. We modified that paragraph as follows:

“The study in (Grewe et al., 2017) suggests that associative learning resulting from fear conditioning induces both potentiation and depression among coactive excitatory neurons; coactivity was determined by calcium signaling and thus did not allow measurements of fine timing between spikes. In our model, we show how potentiation between coactive cells occurs when strict pre-post spike timing and appropriate pauses in the spiking activity arise. Depression happens when one or both of these components are not present. Thus, in our model, depression represents the absence of successful fear association and does not take part in the reshaping of the ensemble encoding the association, as instead suggested in (Grewe et al., 2017). A possible follow-up of our work involves investigating how fear ensembles form and modify through fear conditioning and later stages. This follow-up work may involve using a tri-conditional rule, as suggested in (Grewe et al. 2017), in which the potential role of neuromodulators is taken into account in addition to the pre- and postsynaptic neuron activity; this may lead to both potentiation and depression in establishing an associative memory.”

- In the limitations and caveats section you mention that the small size of the network implies that they represent a synchronous population. What are the potential implications for the proposed rhythm-dependent mechanism? What are your expectations for larger networks? 

We apologize if we were not adequately clear. We are guessing that the Reviewer thought we meant the entire population was synchronous, which it is not. We meant that, when we use a single cell to represent a subpopulation of cells of that type, that subpopulation is effectively synchronous. For larger networks in which each subtype is represented by many cells, there can be heterogeneity within each subtype. We have shown in the paper that the basic results still hold under some heterogeneity; however, they may fail if the heterogeneity is too large.

We mentioned in a new section named “Assumptions and predictions of the model” in response to point 3 made by Reviewer #2.

- The discussion is also missing a section on predictions/new experiments that can be derived from the model. How can the model be confirmed, what experiments/results would break the model? 

To answer this question, we put in a new section in the Discussion entitled “Assumptions and predictions of the model”. The first paragraph of this section is in the reply to Reviewer #2 point 2; the second paragraph is in the reply to Reviewer #2 point 3; the last paragraph is in the Reply to Reviewer #1 point c; the rest of the section reads as follows:

“Our study suggests that all the interneurons are necessary for associative learning provided that the STDP rule is depression-dominated. This prediction could be tested experimentally by selectively silencing each interneuron subtype in the BLA: if the associative learning is hampered by silencing any of the interneuron subtypes, this validates our study. Finally, the model prediction could be tested indirectly by acquiring more information about the plasticity rule involved in the BLA during associative learning. We found that all the interneurons are necessary to establish fear learning only in the case of a depression-dominated rule. This rule ensures that fine timing and pauses are always required for potentiation: interneurons provide both fine timing and pauses to pyramidal cells, making them crucial components of the fear circuit. 

The modeling of the interneurons assumes the involvement of various intrinsic currents; the inclusion of those currents can be considered hypotheses of the model. Our model predicts that blockade of D-current in VIP interneurons (or silencing VIP interneurons) will both diminish low theta and prevent fear learning. Finally, the model assumes the absence of significantly strong connections from the excitatory projection cells ECS to PV interneurons, unlike the ones from F to PV. Including those synapses would alter the PING rhythm created by the interactions between F and PV, which is crucial for fine timing between ECS and F needed for LTP.”

Author Response

The following is the authors’ response to the original reviews.

Reviewer #1 (Public Review):

• A summary of what the authors were trying to achieve.

The authors cultured pre- and Post-vaccine PBMCs with overlapping peptides encoding S protein in the presence of IL-2, IL-7, and IL-15 for 10 days, and extensively analyzed the T cells expanded during the culture; by including scRNAseq, scTCRseq, and examination of reporter cell lines expressing the dominant TCRs. They were able to identify 78 S epitopes with HLA restrictions (by itself represents a major achievement) together with their subset, based on their transcriptional profiling. By comparing T cell clonotypes between pre- and post-vaccination samples, they showed that a majority of pre-existing S-reactive CD4+ T cell clones did not expand by vaccinations. Thus, the authors concluded that highly-responding S-reactive T cells were established by vaccination from rare clonotypes.

• An account of the major strengths and weaknesses of the methods and results.

Strengths

• Selection of 4 "Ab sustainers" and 4 "Ab decliners" from 43 subjects who received two shots of mRNA vaccinations.

• Identification of S epitopes of T cells together with their transcriptional profiling. This allowed the authors to compare the dominant subsets between sustainers and decliners.

Weaknesses

• Fig. 3 provides the epitopes, and the type of T cells, yet the composition of subsets per subject was not provided. It is possible that only one subject out of 4 sustainers expressed many Tfh clonotypes and explained the majority of Tfh clonotypes in the sustainer group. To exclude this possibility, the data on the composition of the T cell subset per subject (all 8 subjects) should be provided.

In accordance with the reviewer’s suggestion, we provided the composition of the T cell subset per subject (all 8 subjects) in the revised manuscript (shown below).

Author response image 1.

• S-specific T cells were obtained after a 10-day culture with peptides in the presence of multiple cytokines. This strategy tends to increase a background unrelated to S protein. Another shortcoming of this strategy is the selection of only T cells amenable to cell proliferation. This strategy will miss anergic or less-responsive T cells and thus create a bias in the assessment of S-reactive T cell subsets. This limitation should be described in the Discussion.

We thank the reviewer for raising the question related to our experimental strategy. We chose this method because a background unrelated to S protein was lower than widely used AIM methods, which is verified by reconstituting many TCRs and testing the responses in vitro. One more reason is this method can identify S-reactive functional (proliferative) T cell clonotypes than anergic or less-responsive T cells as the reviewer mentioned, which is our objective in this study. In accordance with the reviewer’s suggestion, we have carefully described our limitation and rationale of our experimental strategy in the revised manuscript.

• Fig. 5 shows the epitopes and the type of T cells present at baseline. Do they react to HCoV-derived peptides? I guess not, as it is not clearly described. If the authors have the data, it should be provided.

As the reviewer mentioned, the pre-existing highly expanded clonotypes that we analyzed did not react to HCoV-derived peptides. After we determined the epitopes of the clonotypes, the S peptide sequences were analyzed for homology in HCoVs. The only two clonotypes whose epitope sequences were relatively conserved in HCoV strains (clonotypes #8-pre_9 and #8-pre_10) were tested for their reactivity to the similar HCoV epitope counterparts, but no activation was observed (shown below). We added these data in the revised manuscript.

Author response image 2.

• As the authors discussed (L172), pre-existing S-reactive T cells were of low affinity. The raw flow data, as shown in Fig. S3, for pre-existing T cells may help discuss this aspect.

As the reviewer mentioned, some pre-existing S-reactive T cells might appear to react with S peptides judging from the NFAT-GFP expression of their reporter cell lines. However, the percentage of GFP-expressing cells is affected by many factors such as TCR expression level and HLA molecule expression level. Thus, the affinity of pre-existing S-reactive T cells was not fully deduced from the activation of reporter cell lines as shown in Fig. S3 in the present manuscript. We thank the reviewer for this constructive suggestion, but we therefore decided not to use these data quantitatively to evaluate affinity in this manuscript.

Reviewer #2 (Public Review):

Summary:

A short-term comparison of durability of S antibody levels after 2-dose vaccination, showing that better or more poorly sustained responses correlate with the presence of Tfh cells.

Strengths:

Novelty of approach in expanding, sequencing and expressing TCRs for functional studies from the implicated populations.

Weaknesses:

Somewhat outdated question, short timeline, small numbers, over-interpretation of sequence homology data

Reviewer #2 (Recommendations For The Authors):

In line with my above comments, it might be useful for the authors to look at moderating some of the assertions in what is a rather small-scale descriptive account of correlates of some quite nuanced, short-term, S antibody response differences

We clearly described that some homologous microbe-derived peptides were indeed recognized by S-reactive T cells. Also, we have removed our overstatement from the revised manuscript.

Reviewer #3 (Public Review):

Summary:

The paper aims to investigate the relationship between anti-S protein antibody titers with the phenotypes&clonotypes of S-protein-specific T cells, in people who receive SARS-CoV2 mRNA vaccines. To do this, the paper recruited a cohort of Covid-19 naive individuals who received the SARS-CoV2 mRNA vaccines and collected sera and PBMCs samples at different timepoints. Then they mainly generate three sets of data: 1). Anti-S protein antibody titers on all timepoints. 2) Single-cell RNAseq/TCRseq dataset for divided T cells after stimulation by S-protein for 10 days. 3) Corresponding epitopes for each expanded TCR clones. After analyzing these results, the paper reports two major findings & claims: A) Individuals having sustained anti-S protein antibody response also have more so-called Tfh cells in their single-cell dataset, which suggests Tfh-polarization of S-specific T cells can be a marker to predict the longevity of anti-S antibody. B). S-reactive T cells do exist before the vaccination, but they seem to be unable to respond to Covid-19 vaccination properly.

The paper's strength is it uses a very systemic and thorough strategy trying to dissect the relationship between antibody titers, T cell phenotypes, TCR clonotypes and corresponding epitopes, and indeed it reports several interesting findings about the relationship of Tfh/sustained antibody and about the S-reactive clones that exist before the vaccination. However, the main weakness is these interesting claims are not sufficiently supported by the evidence presented in this paper. I have the following major concerns:

(1) The biggest claim of the paper, which is the acquisition of S-specific Tfh clonotypes is associated with the longevity of anti-S antibodies, should be based on proper statistical analysis rather than just a UMAP as in Fig2 C, E, F. The paper only shows the pooled result, but it looks like most of the so-called Tfh cells come from a single donor #27. If separating each of the 4 decliners and sustainers and presenting their Tfh% in total CD4+ T cells respectively, will it statistically have a significant difference between those decliners and sustainers? I want to emphasize that solid scientific conclusions need to be drawn based on proper sample size and statistical analysis.

In accordance with the reviewer’s request, we have also analyzed the T cells separately (shown below). We observed the average frequency was much lower in decliners than sustainers, while the difference did not reach statistical significance partly because of the large deviation due to one sustainer (#27) who possessed quite a high Tfh%. We modified our description in the revised manuscript.

Author response image 3.

(2) The paper does not provide any information to justify its cell annotation as presented in Fig 2B, 4A. Moreover, in my opinion, it is strange to see that there are two clusters of cells sit on both the left and right side of UMAP in Fig2B but both are annotated as CD4 Tcm and Tem. Also Tfh and Treg belong to a same cluster in Fig 2B but they should have very distinct transcriptomes and should be separated nicely. Therefore I believe the paper can be more convincing if it can present more information and discussion about the basis for its cell annotation.

We agree with the reviewer’s concern. Since antigen stimulation only induced the proliferation of antigen-specific T cells, the multiple clusters were mostly due to the fluctuation of cell cyclerelated genes. We therefore carefully and manually annotated these clusters by selecting the cell type-related genes (Kaech et al, Nat. Rev. Immunol., 2002; Sallusto et al, Annu Rev Immunol., 2004) and determined their subsets regardless of the automatic clustering based on the whole transcriptome. Indeed, antigen-responded Tfh and Treg are close, as ICOS and PDCD1 are expressed. We mainly used IL21 and FOXP3 to distinguish the Tfh and Treg populations, respectively. We thank the reviewer for pointing out this important process that we carefully addressed. We added the description of annotation methods to the revised manuscript.

(3) Line 103-104, the paper claims that the Tfh cluster likely comes from cTfh cells. However considering the cells have been cultured/stimulated for 10 days, cTfh cells might lose all Tfh features after such culture. To my best knowledge there is no literature to support the notion that cTfh cells after stimulated in vitro for 10 days (also in the presence of IL2, IL7 and IL15), can still retain a Tfh phenotype after 10 days. It is possible that what actually happens is, instead of having more S-specific cTfh cells before the cell culture, the sustainers' PBMC can create an environment that favors the Tfh cell differentiation (such as express more pro-Tfh cytokines/co-stimulations). Thus after 10-days culture, there are more Tfh-like cells detected in the sustainers. The paper may need to include more evidence to support cTfh cells can retain Tfh features after 10-days' culture.

We thank the reviewer for raising this important issue. As the reviewer pointed out, culturing T cells for 10 days indeed changed the repertoire and features, so the Tfh clonotypes we detected after the expansion may not correspond to the cTfh clonotypes in vivo. Because our observation and analysis were mostly based on the dominant T cell clonotypes expanded in vitro, we modified our description and conclusion accordingly in the revised manuscript.

(4) It is in my opinion inaccurate to use cell number in Fig4B to determine whether such clone expands or not, given that the cell number can be affected by many factors like the input number, the stimulation quality and the PBMC sample quality. A more proper analysis should be considered by calculating the relative abundance of each TCR clone in total CD4 T cells in each timepoint.

We thank the reviewer for pointing out our inaccuracy. As the reviewer suggested, we used percentages to demonstrate the relative abundance of each clonotype in Fig. 4B of the revised manuscript.

(5) It is well-appreciated to express each TCR in cell line and to determine the epitopes. However, the author needs to make very sure that this analysis is performed correctly because a large body of conclusions of the paper are based on such epitope analysis. However, I notice something strange (maybe I am wrong) but for example, Table 4 donor #8 clonotype post_6 and _7, these two clonotypes have exactly the same TRAV5 and TRAJ5 usage. Because alpha chain don't have a D region, in theory these clonotypes, if have the same VJ usage, they should have the same alpha chain CDR3 sequences, however, in the table they have very different CDR3α aa sequences. I wish the author could double check their analysis and I apologize in advance if I raise such questions based on wrong knowledge.

We thank the reviewer for carefully reading our manuscript. Although the two clonotypes, donor #8 clonotype post_6 and _7, have the exactly same TRAV5 and TRAJ5 usage, they have different CDR3a aa sequences due to random nucleotide addition in the rearrangement. Likewise, donor #27 clonotype post_1 and donor #13 clonotype post_15 had the same TRAV9-2 and TRAJ17 usage but different CDR3a.

Reviewer #3 (Recommendations For The Authors):

(1) Related to my public review 1. To make a solid conclusion, I think the author can include more sustainers and decliners if possible, can just stimulate their PBMCs for 10 days and check the Tfh features in proliferated CD4 T cells (e.g. IL21 secretion, PD-1 expression etc). And then compare these values in sustainers vs decliners

We thank the reviewer for the suggestion. Unfortunately, additional PBMCs from more sustainers and decliners are not available to us. Instead, we carefully described the current observation in the revised manuscript.

(2) Related to my public review 3. The author can attempt to sort CXCR5+ cTfh and CXCR5- non cTfh, stimulate in vitro for 10 days and compare whether the stimulated cTfh still have more Tfh-related features such as increased IL- 21 secretion.

As the reviewer recommended, sorting and culturing the cTfh and non cTfh separately will clarify this issue. Due to the limitation of the samples, we could not perform these experiments.

(3) I couldn't find information about the availability of data and code to analyze the single cell RNA-seq dataset in the manuscript

We clarified the availability of data and added the codes for the single cell RNA-seq dataset in the revised manuscript.

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Recommendations For The Authors):

Specific comments to improve the quality of the work:

(1) The choice of subunits to tag are really not ideal. In the available structures of the human proteasome, The C-terminus of Rpn3/PSMD3 points directly toward the ATPase pore and is likely to disrupt the structure and/or dynamics of the proteasome during proteolysis (see comments regarding controls for functionality below). Similarly, the C-terminal tail of Rpt1/PSMC2 has a key role in the opening of the 20S core particle gate for substrate translocation and processing (see 2018 Nature Communications, 9:1360 and 2018 Cell Reports 24:1301-1315), and Alpha3/PSMA4 can be substituted by a second copy of Alpha4/PSMA7 in some conditions (although tagging Alpha3/PSMA4 would admittedly provide a picture of the canonical proteasome interactome while actively excluding the interactome of the non-canonical proteasomes that form via replacement of Alpha3/PSMA4). Comparison of these cell lines with lines harboring tags on subunits that are commonly used for tagging in the field because of a lack of impacts, such as the N-terminus of Rpn1/PSMD2, the C-terminus of Rpn11/PSMD14, and the C-terminus of Beta4/PSMB2 would help instill confidence that the interactome reported largely arises from mature, functional proteasomes rather than subcomplexes, defective proteasomes, or other species that may occur due to tagging at these positions.

We thank the reviewer for pointing this out. The original purpose of our strategy was to establish proximity labeling of proteasomes to enable applications both in cell culture and in vivo. The choice of PSMA4 and PSMC2 was dictated by previous successful tagging with GFP in mammalian cells (Salomons et al., Exp Cell Res 2010)(Bingol and Schuman, Nature 2006). However, the choice of C-terminal PSMC2 might have been not optimal. HEK293 cells overexpressing PSMC2-BirA show slower growth and the BioID data retrieve higher enrichment of assembly factors suggesting slower assembly of this fusion protein in proteasome. Although we did not observe a negative impact on overall proteasome activity and PSMC2-BirA was (at least in part) incorporated into fully assembled proteasomes as indicated by enrichment of 20S proteins.We apologize for not making it clear that we labeled the N-terminus of PSMD3/Rpn3 and not the C-terminus (Figure 1a and S1a). Therefore, we included in Figure S1a of the revised manuscript structures of the proteasome where the tagged subunit termini are highlighted: C-terminus for PSMA4 and PSMC2 and N-terminus for PSMD3. Additionally, we would like to point out that, differently from PSMC2-BirA, cells expressing BirA-PSMD3 did not show slower growth, and BioID data showed a more homogenous enrichment of both 19S and 20S proteins, as compared to PSMC2-BirA (Figure 1D and 1E). However, the overall level of enrichment of proteasome subunits was not comparable to PSMA4-BirA and, therefore, we opted for focusing the rest of the manuscript on this construct.

In support of this point, the data provided in Figure 1E in which the change in the abundances of each proteasome subunit in the tagged line vs. the BirA control line demonstrates substantial enrichment of the subcomplexes of the proteasome that are tagged in each case; this effect may represent the known feedback-mediated upregulation of new proteasome subunit synthesis that occurs when proteasomal proteolysis is impaired, or alternatively, the accumulation of subcomplexes containing the tagged subunit that cannot readily incorporate into mature proteasomes. Acknowledging this limitation in the text would be valuable to readers who are less familiar with the proteasome.

We would like to clarify that the data shown in Figure 1E do not represent whole proteome data, but rather log2 fold changes vs. BirA* control calculated on streptavidin enrichment samples. The differences in the enrichment of the various subcomplexes between cell lines derives from the fact that the effect size of the enrichment depends on both protein abundance in the isolated complexes, but also on the efficiency of biotinylation. The latter will be higher for proteins located in closer proximity to the bait. A similar observation was pointed out in a recent publication (PMID:36410438) that compared BioID and Co-IP for the same bait. When a component of the nuclear pore complex (Nup158) was analyzed by BioID only the more proximal proteins were enriched as compared to the whole complex in Co-IP data (Author response image 1):

Author response image 1.

Proteins identified in the NUP158 BioID or pulldown experiments are filled in red or light red for significance intervals A or B, respectively. The bait protein NUP158 is filled in yellow. Proteins enriched in the pulldown falling outside the SigA/B cutoff are filled in gray. NPC, nuclear pore complex. SigA, significant class A; SigB, significant class B. Reproduced from Figure 6 of PMID: 36410438.

However, we would like to point out that despite quantitative differences between different proteasome subunits, both 19S and 20S proteins were found to be strongly enriched (typically >2 fold) in all the constructs compared to BirA* control line (Figure 1E). This indicates that at least a fraction of all the tagged subunits are incorporated into fully assembled proteasomes.

Regarding the upregulation of proteasome subunits as a consequence of proteasome dysfunction, we did not find evidence of this, at least in the case of PSMA4. The immunoblot shown in Figure 2A and its quantification in S3A indicate no increased abundance of endogenous PSMA4 upon tetracycline induction of PSMA4-BirA*.

(2) The use of myc as a substrate of the proteasome for demonstration that proteolysis is unaffected is perhaps not ideal. Myc is known to be degraded via both ubiquitin-dependent and ubiquitin-independent mechanisms, such that disruption of one means of degradation (e.g., ubiquitin-dependent degradation) via a given tag could potentially be compensated by another. A good example of this is that the C-terminal tagging of PSMC2/Rpt1 is likely to disrupt interaction between the core particle and the regulatory particle (as suggested in Fig. 1D); this may free up the core particle for ubiquitin-independent degradation of myc.

Aside from using specific reporters for ubiquitin-dependent vs. independent degradation or a larger panel of known substrates, analysis of the abundance of K48-ubiquitinated proteins in the control vs. tag lines would provide additional evidence as to whether or not proteolysis is generally perturbed in the tag lines.

We thank the reviewer for this suggestion. We have included an immunoblot analysis showing that the levels of K48 ubiquitylation (Figure S3d) are not affected by the expression of tagged PSMA4.

(3) On pg. 8 near the bottom, the authors accidentally refer to ARMC6 as ARMC1 in one instance.

We have corrected the mistake.

(4) On pg. 10, the authors explain that they analyzed the interactome for all major mouse organs except the brain; although they explain in the discussion section why the brain was excluded, including this explanation on pg. 10 here instead of in the discussion might be a better place to discuss this.

We moved the explanation from the discussion to the results part.

Reviewer #2 (Recommendations For The Authors):

(1) Perhaps the authors can quantify the fraction of unassembled PSMA4-BirA* from the SEC experiment (Fig. 2b) to give the readers a feeling for how large a problem this could be.

The percentages based on Area Under the Curve calculations have been added to Figure S3b.

(2) Do the authors observe any difference in the enrichment scores between proteins that are known to interact with the proteasome vs proteins that the authors can justify as "interactors of interactors" vs the completely new potential interactors? This could be an interesting way to show that the potential new interactors are not simply because of poor false positive rate calibration, but that they behave in the same way as the other populations.

We thank the reviewer for this suggestion. We analyzed the enrichment scores for 20S proteasome subunits, known PIPs, first neighbors and the remaining enriched proteins. The remaining proteins (potential new interactors) have very similar scores as the first neighbors of known interactors. This plot has been added to Figure S3g.

(3) Did the authors try to train a logistic model for the miniTurbo experiments, like it was done for the BirA* experiments? Perhaps combining the results of both experiments would yield higher confidence on the proteasome interactors.

Following the reviewers suggestion, we applied the classifier on the dataset of the comparison between miniTurbo and PSMA-miniTurbo. We found a clear separation between the FPR and the TPR with 136 protein groups enriched in PSMA-miniTurbo. We have added the classifier and corresponding ROC curve to Figure S4f and S4g.

75 protein groups were found to be enriched for both PSMA4-BirA* and PSMA4-miniTurbo (Author response image 2), including the proteasome core particles, regulatory particles, known interactors and potential new interactors. As we focused more on the identification of substrates with PSMA4-miniTurbo, we did not pursue these overlapping protein groups further, but rather used the comparison to the mouse model to identify potential new interactors.

Author response image 2.

Overlap between ProteasomeID enriched proteins (fpr<0.05) between PSMA4-BirA* and PSMA4-miniTurbo.

(4) Perhaps this is already known, but did the authors check if MG132 affect proteasome assembly? The authors could for example repeat their SEC experiments in the presence of MG132.

We thank the reviewer for the suggestion, however to our knowledge there are no reports that MG132 has an effect on the assembly of the proteasome. MG132 is one of the most used proteasome inhibitors in basic research and as such has been extensively characterized in the last 3 decades. The small peptide aldehyde acts as a substrate analogue and binds directly to the active site of the protease PSMB5/β5. We therefore think it is unlikely that MG132 is interfering with the assembly of the proteasome.

(5) Minor comment: at the bottom of page 8, the authors probably mean ARMC6 and not ARMC1.

We have corrected the mistake.

(6) It would be interesting to expand the analysis of the already acquired in vivo data to try to identify tissue-specific proteasome interactors. Can the authors draw a four-way Venn diagram with the interactors of each tissue?

We thank the reviewer for this suggestion. We have generated an UpSet plot showing the overlap of ProteasomeID enriched proteins in the four tissues that gave us meaningful results (Author response image 3). In order to investigate whether the observed differences in ProteasomeID enriched proteins could be meaningful in terms of proteasome biology, we have highlighted proteins belonging to the UPS that show tissue specific enrichments. We found proteasome activators such as PSME1/PA28alpha and PSME2/PA28beta to enrich preferentially in kidney and liver, respectively, as well as multiple deubiquitinases to enrich preferentially in the heart. These differences might be related to the specific cellular composition of the different tissues, e.g., number of immune cells present, or the tissue-specific interaction of proteasomes with enzymes involved in the ubiquitin cycle. Given the rather preliminary nature of these findings, we have opted for not including this figure in the main manuscript, but rather include it only in this rebuttal letter.

Author response image 3.

Upset plot showing overlap between ProteasomeID enriched proteins in different mouse organs.

Reviewer #3 (Recommendations For The Authors):

(1) In the first paragraph of the Introduction, the authors link cellular senescence caused by partial proteasome inhibition with the efficacy of proteasome inhibitors in cancer therapy. Although this is an interesting hypothesis, I am not aware of any direct evidence for this; rather, I believe the efficacy of bortezomib/carfilzomib in haematological malignancies is most commonly attributed to these cells having adapted to high levels of proteotoxic stress (e.g., chronic unfolded protein response activation). I would suggest rephrasing this sentence.

We thank the reviewer for the comment and have amended the introduction.

(2) For the initial validation experiments (e.g., Fig. 1B), have the authors checked what level of Streptavidin signal is obtained with "+ bio, - tet" ? Although I accept that the induction of PSMA4-BirA* upon doxycycline addition is clear from the anti-Flag blots, it would still be informative to ascertain what level of background labelling is obtained without induction (but in the presence of exogenous biotin).

We tested four different conditions +/- tet and +/- biotin (24h) in PSMA4-BirA* cell lines (Author response image 4). As expected, biotinylation was most pronounced when tet and biotin were added. When biotin was omitted, streptavidin signal was the lowest regardless of the addition of tet. Compared to the -biotin conditions, a slight increase of streptavidin signal could be observed when biotin was added but tet was not added. This could be either due to the promoter leaking (PMID: 12869186) or traces of tetracycline in the FBS we used, as we did not specifically use tet-free FBS for our experiments.

Author response image 4.

Streptavidin-HRP immunoblot following induction of BirA fusion proteins with tetracycline (+tet) and supplementation of biotin (+bio). For the sample used as expression control tetracycline was omitted (-tet). To test background biotinylation, biotin supplementation was omitted (-bio). Immunoblot against BirA and PSMA was used to verify induction of fusion proteins, while GAPDH was used as loading control.

(3) For the proteasome structure models in Fig. 1D, a scale bar would be useful to inform the reader of the expected 10 nm labelling radius (as the authors have done later, in Fig. 2D).

We have added 10 nm scale bars to Figure 1d.

(4) In the "Identification of proteasome substrates by ProteasomeID" Results subsection, I believe there is a typo where the authors refer to ARMC1 instead of ARMC6.

We have corrected the mistake.

(5) I think Fig. S5 was one of the most compelling in the manuscript. Given the interest in confirming on-target efficacy of targeted degradation modalities, as well as identifying potential off-target effects early-on in development, I would consider promoting this out of the supplement.

We thank the reviewer for the comment and share the excitement about using ProteasomeID for targeted degradation screening. We have moved the data on PROTACs (Figure S5) into a new main Figure 5.

In addition, in relation to the comment of this reviewer regarding the detection of endogenous substrates, we have now included validation for one more hit emerging from our analysis (TIGD5) and included the results in Figure 4f, 4g and S4j.

Author response:

The following is the authors’ response to the original reviews.

Summary of reviewers’ comments and our revisions: 

We thank the reviewers for their thoughtful feedback. This feedback has motivated multiple revisions and additions that, in our view, have greatly improved the manuscript. This is especially true with regard to a major goal of this study: clearly defining existing scientific perspectives and delineating their decoding implications. In addition to building on this conceptual goal, we have expanded existing analyses and have added a new analysis of generalization using a newly collected dataset. We expect the manuscript will be of very broad interest, both to those interested in BCI development and to those interested in fundamental properties of neural population activity and its relationship with behavior.

Importantly, all reviewers were convinced that MINT provided excellent performance, when benchmarked against existing methods, across a broad range of standard tasks:

“their method shows impressive performance compared to more traditional decoding approaches” (R1) 

“The paper was thorough in considering multiple datasets across a variety of behaviors, as well as existing decoding methods, to benchmark the MINT approach. This provided a valuable comparison to validate the method.” (R2) 

“The fact that performance on stereotyped tasks is high is interesting and informative…” (R3)

This is important. It is challenging to design a decoder that performs consistently across multiple domains and across multiple situations (including both decoding and neural state estimation). MINT does so. MINT consistently outperformed existing lightweight ‘interpretable’ decoders, despite being a lightweight interpretable decoder itself. MINT was very competitive with expressive machine-learning methods, yet has advantages in flexibility and simplicity that more ‘brute force’ methods do not. We made a great many comparisons, and MINT was consistently a strong performer. Of the many comparisons we made, there was only one where MINT was at a modest disadvantage, and it was for a dataset where all methods performed poorly. No other method we tested was as consistent. For example, although the GRU and the feedforward network were often competitive with MINT (and better than MINT in the one case mentioned above), there were multiple other situations where they performed less well and a few situations where they performed poorly. Moreover, no other existing decoder naturally estimates the neural state while also readily decoding, without retraining, a broad range of behavioral variables.

R1 and R2 were very positive about the broader impacts of the study. They stressed its impact both on decoder design, and on how our field thinks, scientifically, about the population response in motor areas: 

“This paper presents an innovative decoding approach for brain-computer interfaces” (R1)

“presents a substantial shift in methodology, potentially revolutionizing the way BCIs interpret and predict neural behaviour” (R1)

“the paper's strengths, particularly its emphasis on a trajectory-centric approach and the simplicity of MINT, provide a compelling contribution to the field” (R1)

“The authors made strong arguments, supported by evidence and literature, for potentially high-dimensional neural states and thus the need for approaches that do not rely on an assumption of low dimensionality” (R2)

“This work is motivated by brain-computer interfaces applications, which it will surely impact in terms of neural decoder design.” (R2)

“this work is also broadly impactful for neuroscientific analysis... Thus, MINT will likely impact neuroscience research generally.” (R2)

We agree with these assessments, and have made multiple revisions to further play into these strengths. As one example, the addition of Figure 1b (and 6b) makes this the first study, to our knowledge, to fully and concretely illustrate this emerging scientific perspective and its decoding implications. This is important, because multiple observations convince us that the field is likely to move away from the traditional perspective in Figure 1a, and towards that in Figure 1b. We also agree with the handful of weaknesses R1 and R2 noted. The manuscript has been revised accordingly. The major weakness noted by R1 was the need to be explicit regarding when we suspect MINT would (and wouldn’t) work well in other brain areas. In non-motor areas, the structure of the data may be poorly matched with MINT’s assumptions. We agree that this is likely to be true, and thus agree with the importance of clarifying this topic for the reader. The revision now does so. R1 also wished to know whether existing methods might benefit from including trial-averaged data during training, something we now explore and document (see detailed responses below). R2 noted two weaknesses: 1) The need to better support (with expanded analysis) the statement that neural and behavioral trajectories are non-isometric, and 2) The need to more rigorously define the ‘mesh’. We agree entirely with both suggestions, and the revision has been strengthened by following them (see detailed responses below).

R3 also saw strengths to the work, stating that:

“This paper is well-structured and its main idea is clear.” 

“The fact that performance on stereotyped tasks is high is interesting and informative, showing that these stereotyped tasks create stereotyped neural trajectories.” 

“The task-specific comparisons include various measures and a variety of common decoding approaches, which is a strength.”

However, R3 also expressed two sizable concerns. The first is that MINT might have onerous memory requirements. The manuscript now clarifies that MINT has modest memory requirements. These do not scale unfavorably as the reviewer was concerned they might. The second concern is that MINT is: 

“essentially a table-lookup rather than a model.”

Although we don’t agree, the concern makes sense and may be shared by many readers, especially those who take a particular scientific perspective. Pondering this concern thus gave us the opportunity to modify the manuscript in ways that support its broader impact. Our revisions had two goals: 1) clarify the ways in which MINT is far more flexible than a lookup-table, and 2) better describe the dominant scientific perspectives and their decoding implications.

The heart of R3’s concern is the opinion that MINT is an effective but unprincipled hack suitable for situations where movements are reasonably stereotyped. Of course, many tasks involve stereotyped movements (e.g. handwriting characters), so MINT would still be useful. Nevertheless, if MINT is not principled, other decode methods would often be preferable because they could (unlike MINT in R3’s opinion) gain flexibility by leveraging an accurate model. Most of R3’s comments flow from this fundamental concern: 

“This is again due to MINT being a lookup table with a library of stereotyped trajectories rather than a model.”

“MINT models task-dependent neural trajectories, so the trained decoder is very task-dependent and cannot generalize to other tasks.”

“Unlike MINT, these works can achieve generalization because they model the neural subspace and its association to movement.”

“given that MINT tabulates task-specific trajectories, it will not generalize to tasks that are not seen in the training data even when these tasks cover the exact same space (e.g., the same 2D computer screen and associated neural space).”

“For proper training, the training data should explore the whole movement space and the associated neural space, but this does not mean all kinds of tasks performed in that space must be included in the training set (something MINT likely needs while modeling-based approaches do not).”

The manuscript has been revised to clarify that MINT is considerably more flexible than a lookup table, even though a lookup table is used as a first step. Yet, on its own, this does not fully address R3’s concern. The quotes above highlight that R3 is making a standard assumption in our field: that there exists a “movement space and associated neural space”. Under this perspective, one should, as R3 argues fully explore the movement space. This would perforce fully explore the associated neural subspace. One can then “model the neural subspace and its association to movement”. MINT does not use a model of this type, and thus (from R3’s perspective) does not appear to use a model at all. A major goal of our study is to question this traditional perspective. We have thus added a new figure to highlight the contrast between the traditional (Figure 1a) and new (Figure 1b) scientific perspectives, and to clarify their decoding implications.

While we favor the new perspective (Figure 1b), we concede that R3 may not share our view. This is fine. Part of the reason we believe this study is timely, and will be broadly read, is that it raises a topic of emerging interest where there is definitely room for debate. If we are misguided – i.e. if Figure 1a is the correct perspective – then many of R3’s concerns would be on target: MINT could still be useful, but traditional methods that make the traditional assumptions in Figure 1a would often be preferable. However, if the emerging perspective in Figure 1b is more accurate, then MINT’s assumptions would be better aligned with the data than those of traditional methods, making it a more (not less) principled choice.

Our study provides new evidence in support of Figure 1b, while also synthesizing existing evidence from other recent studies. In addition to Figure 2, the new analysis of generalization further supports Figure 1b. Also supporting Figure 1b is the analysis in which MINT’s decoding advantage, over a traditional decoder, disappears when simulated data approximate the traditional perspective in Figure 1a.

That said, we agree that the present study cannot fully resolve whether Figure 1a or 1b is more accurate. Doing so will take multiple studies with different approaches (indeed we are currently preparing other manuscripts on this topic). Yet we still have an informed scientific opinion, derived from past, present and yet-to-be-published observations. Our opinion is that Figure 1b is the more accurate perspective. This possibility makes it reasonable to explore the potential virtues of a decoding method whose assumptions are well-aligned with that perspective. MINT is such a method. As expected under Figure 1b, MINT outperforms traditional interpretable decoders in every single case we studied. 

As noted above, we have added a new generalization-focused analysis (Figure 6) based on a newly collected dataset. We did so because R3’s comments highlight a deep point: which scientific perspective one takes has strong implications regarding decoder generalization. These implications are now illustrated in the new Figure 6a and 6b. Under Figure 6a, it is possible, as R3 suggests, to explore “the whole movement space and associated neural space” during training. However, under Figure 6b, expectations are very different. Generalization will be ‘easy’ when new trajectories are near the training-set trajectories. In this case, MINT should generalize well as should other methods. In contrast, generalization will be ‘hard’ when new neural trajectories have novel shapes and occupy previously unseen regions / dimensions. In this case, all current methods, including MINT, are likely to fail. R3 points out that traditional decoders have sometimes generalized well to new tasks (e.g. from center-out to ‘pinball’) when cursor movements occur in the same physical workspace. These findings could be taken to support Figure 6a, but are equally consistent with ‘easy’ generalization in Figure 6b. To explore this topic, the new analysis in Figure 6c-g considers conditions that are intended to span the range from easy to hard. Results are consistent with the predictions of Figure 6b. 

We believe the manuscript has been significantly improved by these additions. The revisions help the manuscript achieve its twin goals: 1) introduce a novel class of decoder that performs very well despite being very simple, and 2) describe properties of motor-cortex activity that will matter for decoders of all varieties.

Reviewer #1: 

Summary: 

This paper presents an innovative decoding approach for brain-computer interfaces (BCIs), introducing a new method named MINT. The authors develop a trajectory-centric approach to decode behaviors across several different datasets, including eight empirical datasets from the Neural Latents Benchmark. Overall, the paper is well written and their method shows impressive performance compared to more traditional decoding approaches that use a simpler approach. While there are some concerns (see below), the paper's strengths, particularly its emphasis on a trajectory-centric approach and the simplicity of MINT, provide a compelling contribution to the field. 

We thank the reviewer for these comments. We share their enthusiasm for the trajectory-centric approach, and we are in complete agreement that this perspective has both scientific and decoding implications. The revision expands upon these strengths.

Strengths: 

The adoption of a trajectory-centric approach that utilizes statistical constraints presents a substantial shift in methodology, potentially revolutionizing the way BCIs interpret and predict neural behaviour. This is one of the strongest aspects of the paper. 

Again, thank you. We also expect the trajectory-centric perspective to have a broad impact, given its relevance to both decoding and to thinking about manifolds.

The thorough evaluation of the method across various datasets serves as an assurance that the superior performance of MINT is not a result of overfitting. The comparative simplicity of the method in contrast to many neural network approaches is refreshing and should facilitate broader applicability. 

Thank you. We were similarly pleased to see such a simple method perform so well. We also agree that, while neural-network approaches will always be important, it is desirable to also possess simple ‘interpretable’ alternatives.

Weaknesses:  

Comment 1) Scope: Despite the impressive performance of MINT across multiple datasets, it seems predominantly applicable to M1/S1 data. Only one of the eight empirical datasets comes from an area outside the motor/somatosensory cortex. It would be beneficial if the authors could expand further on how the method might perform with other brain regions that do not exhibit low tangling or do not have a clear trial structure (e.g. decoding of position or head direction from hippocampus) 

We agree entirely. Population activity in many brain areas (especially outside the motor system) presumably will often not have the properties upon which MINT’s assumptions are built. This doesn’t necessarily mean that MINT would perform badly. Using simulated data, we have found that MINT can perform surprisingly well even when some of its assumptions are violated. Yet at the same time, when MINT’s assumptions don’t apply, one would likely prefer to use other methods. This is, after all, one of the broader themes of the present study: it is beneficial to match decoding assumptions to empirical properties. We have thus added a section on this topic early in the Discussion: 

“In contrast, MINT and the Kalman filter performed comparably on simulated data that better approximated the assumptions in Figure 1a. Thus, MINT is not a ‘better’ algorithm – simply better aligned with the empirical properties of motor cortex data. This highlights an important caveat. Although MINT performs well when decoding from motor areas, its assumptions may be a poor match in other areas (e.g. the hippocampus). MINT performed well on two non-motor-cortex datasets – Area2_Bump (S1) and DMFC_RSG (dorsomedial frontal cortex) – yet there will presumably be other brain areas and/or contexts where one would prefer a different method that makes assumptions appropriate for that area.”

Comment 2) When comparing methods, the neural trajectories of MINT are based on averaged trials, while the comparison methods are trained on single trials. An additional analysis might help in disentangling the effect of the trial averaging. For this, the authors could average the input across trials for all decoders, establishing a baseline for averaged trials. Note that inference should still be done on single trials. Performance can then be visualized across different values of N, which denotes the number of averaged trials used for training. 

We explored this question and found that the non-MINT decoders are harmed, not helped, by the inclusion of trial-averaged responses in the training set. This is presumably because the statistics of trialaveraged responses don’t resemble what will be observed during decoding. This statistical mismatch, between training and decoding, hurts most methods. It doesn’t hurt MINT, because MINT doesn’t ‘train’ in the normal way. It simply needs to know rates, and trial-averaging is a natural way to obtain them. To describe the new analysis, we have added the following to the text.

“We also investigated the possibility that MINT gained its performance advantage simply by having access to trial-averaged neural trajectories during training, while all other methods were trained on single-trial data. This difference arises from the fundamental requirements of the decoder architectures: MINT needs to estimate typical trajectories while other methods don’t. Yet it might still be the case that other methods would benefit from including trial-averaged data in the training set, in addition to single-trial data. Alternatively, this might harm performance by creating a mismatch, between training and decoding, in the statistics of decoder inputs. We found that the latter was indeed the case: all non-MINT methods performed better when trained purely on single-trial data.”

Reviewer #2:

Summary: 

The goal of this paper is to present a new method, termed MINT, for decoding behavioral states from neural spiking data. MINT is a statistical method which, in addition to outputting a decoded behavioral state, also provides soft information regarding the likelihood of that behavioral state based on the neural data. The innovation in this approach is neural states are assumed to come from sparsely distributed neural trajectories with low tangling, meaning that neural trajectories (time sequences of neural states) are sparse in the high-dimensional space of neural spiking activity and that two dissimilar neural trajectories tend to correspond to dissimilar behavioral trajectories. The authors support these assumptions through analysis of previously collected data, and then validate the performance of their method by comparing it to a suite of alternative approaches. The authors attribute the typically improved decoding performance by MINT to its assumptions being more faithfully aligned to the properties of neural spiking data relative to assumptions made by the alternatives. 

We thank the reviewer for this accurate summary, and for highlighting the subtle but important fact that MINT provides information regarding likelihoods. The revision includes a new analysis (Figure 6e) illustrating one potential way to leverage knowledge of likelihoods.

Strengths:  

The paper did an excellent job critically evaluating common assumptions made by neural analytical methods, such as neural state being low-dimensional relative to the number of recorded neurons. The authors made strong arguments, supported by evidence and literature, for potentially high-dimensional neural states and thus the need for approaches that do not rely on an assumption of low dimensionality. 

Thank you. We also hope that the shift in perspective is the most important contribution of the study. This shift matters both scientifically and for decoder design. The revision expands on this strength. The scientific alternatives are now more clearly and concretely illustrated (especially see Figure 1a,b and Figure 6a,b). We also further explore their decoding implications with new data (Figure 6c-g).

The paper was thorough in considering multiple datasets across a variety of behaviors, as well as existing decoding methods, to benchmark the MINT approach. This provided a valuable comparison to validate the method. The authors also provided nice intuition regarding why MINT may offer performance improvement in some cases and in which instances MINT may not perform as well. 

Thank you. We were pleased to be able to provide comparisons across so many datasets (we are grateful to the Neural Latents Benchmark for making this possible).

In addition to providing a philosophical discussion as to the advantages of MINT and benchmarking against alternatives, the authors also provided a detailed description of practical considerations. This included training time, amount of training data, robustness to data loss or changes in the data, and interpretability. These considerations not only provided objective evaluation of practical aspects but also provided insights to the flexibility and robustness of the method as they relate back to the underlying assumptions and construction of the approach. 

Thank you. We are glad that these sections were appreciated. MINT’s simplicity and interpretability are indeed helpful in multiple ways, and afford opportunities for interesting future extensions. One potential benefit of interpretability is now explored in the newly added Figure 6e. 

Impact: 

This work is motivated by brain-computer interfaces applications, which it will surely impact in terms of neural decoder design. However, this work is also broadly impactful for neuroscientific analysis to relate neural spiking activity to observable behavioral features. Thus, MINT will likely impact neuroscience research generally. The methods are made publicly available, and the datasets used are all in public repositories, which facilitates adoption and validation of this method within the greater scientific community. 

Again, thank you. We have similar hopes for this study.

Weaknesses (1 & 2 are related, and we have switched their order in addressing them): 

Comment 2) With regards to the idea of neural and behavioral trajectories having different geometries, this is dependent on what behavioral variables are selected. In the example for Fig 2a, the behavior is reach position. The geometry of the behavioral trajectory of interest would look different if instead the behavior of interest was reach velocity. The paper would be strengthened by acknowledgement that geometries of trajectories are shaped by extrinsic choices rather than (or as much as they are) intrinsic properties of the data. 

We agree. Indeed, we almost added a section to the original manuscript on this exact topic. We have now done so:

“A potential concern regarding the analyses in Figure 2c,d is that they require explicit choices of behavioral variables: muscle population activity in Figure 2c and angular phase and velocity in Figure 2d. Perhaps these choices were misguided. Might neural and behavioral geometries become similar if one chooses ‘the right’ set of behavioral variables? This concern relates to the venerable search for movement parameters that are reliably encoded by motor cortex activity [69, 92–95]. If one chooses the wrong set of parameters (e.g. chooses muscle activity when one should have chosen joint angles) then of course neural and behavioral geometries will appear non-isometric. There are two reasons why this ‘wrong parameter choice’ explanation is unlikely to account for the results in Figure 2c,d. First, consider the implications of the left-hand side of Figure 2d. A small kinematic distance implies that angular position and velocity are nearly identical for the two moments being compared. Yet the corresponding pair of neural states can be quite distant. Under the concern above, this distance would be due to other encoded behavioral variables – perhaps joint angle and joint velocity – differing between those two moments. However, there are not enough degrees of freedom in this task to make this plausible. The shoulder remains at a fixed position (because the head is fixed) and the wrist has limited mobility due to the pedal design [60]. Thus, shoulder and elbow angles are almost completely determined by cycle phase. More generally, ‘external variables’ (positions, angles, and their derivatives) are unlikely to differ more than slightly when phase and angular velocity are matched. Muscle activity could be different because many muscles act on each joint, creating redundancy. However, as illustrated in Figure 2c, the key effect is just as clear when analyzing muscle activity. Thus, the above concern seems unlikely even if it can’t be ruled out entirely. A broader reason to doubt the ‘wrong parameter choice’ proposition is that it provides a vague explanation for a phenomenon that already has a straightforward explanation. A lack of isometry between the neural population response and behavior is expected when neural-trajectory tangling is low and output-null factors are plentiful [55, 60]. For example, in networks that generate muscle activity, neural and muscle-activity trajectories are far from isometric [52, 58, 60]. Given this straightforward explanation, and given repeated failures over decades to find the ‘correct’ parameters (muscle activity, movement direction, etc.) that create neural-behavior isometry, it seems reasonable to conclude that no such isometry exists.”

Comment 1) The authors posit that neural and behavioral trajectories are non-isometric. To support this point, they look at distances between neural states and distances between the corresponding behavioral states, in order to demonstrate that there are differences in these distances in each respective space. This supports the idea that neural states and behavioral states are non-isometric but does not directly address their point. In order to say the trajectories are non-isometric, it would be better to look at pairs of distances between corresponding trajectories in each space. 

We like this idea and have added such an analysis. To be clear, we like the original analysis too: isometry predicts that neural and behavioral distances (for corresponding pairs of points) should be strongly correlated, and that small behavioral distances should not be associated with large neural distances. These predictions are not true, providing a strong argument against isometry. However, we also like the reviewer’s suggestion, and have added such an analysis. It makes the same larger point, and also reveals some additional facts (e.g. it reveals that muscle-geometry is more related to neural-geometry than is kinematic-geometry). The new analysis is described in the following section:

“We further explored the topic of isometry by considering pairs of distances. To do so, we chose two random neural states and computed their distance, yielding dneural1. We repeated this process, yielding dneural2. We then computed the corresponding pair of distances in muscle space (dmuscle1 and dmuscle2) and kinematic space (dkin1 and dkin2). We considered cases where dneural1 was meaningfully larger than (or smaller than) dneural2, and asked whether the behavioral variables had the same relationship; e.g. was dmuscle1 also larger than dmuscle2? For kinematics, this relationship was weak: across 100,000 comparisons, the sign of dkin1 − dkin2 agreed with dneural1 − dneural2 only 67.3% of the time (with 50% being chance). The relationship was much stronger for muscles: the sign of dmuscle1 − dmuscle2 agreed with dneural1 − dneural2 79.2% of the time, which is far more than expected by chance yet also far from what is expected given isometry (e.g. the sign agrees 99.7% of the time for the truly isometric control data in Figure 2e). Indeed there were multiple moments during this task when dneural1 was much larger than dneural2, yet dmuscle1 was smaller than dmuscle2. These observations are consistent with the proposal that neural trajectories resemble muscle trajectories in some dimensions, but with additional output-null dimensions that break the isometry [60].”

Comment 3) The approach is built up on the idea of creating a "mesh" structure of possible states. In the body of the paper the definition of the mesh was not entirely clear and I could not find in the methods a more rigorous explicit definition. Since the mesh is integral to the approach, the paper would be improved with more description of this component. 

This is a fair criticism. Although MINTs actual operations were well-documented, how those operations mapped onto the term ‘mesh’ was, we agree, a bit vague. The definition of the mesh is a bit subtle because it only emerges during decoding rather than being precomputed. This is part of what gives MINT much more flexibility than a lookup table. We have added the following to the manuscript.

“We use the term ‘mesh’ to describe the scaffolding created by the training-set trajectories and the interpolated states that arise at runtime. The term mesh is apt because, if MINT’s assumptions are correct, interpolation will almost always be local. If so, the set of decodable states will resemble a mesh, created by line segments connecting nearby training-set trajectories. However, this mesh-like structure is not enforced by MINT’s operations.

Interpolation could, in principle, create state-distributions that depart from the assumption of a sparse manifold. For example, interpolation could fill in the center of the green tube in Figure 1b, resulting in a solid manifold rather than a mesh around its outer surface. However, this would occur only if spiking observations argued for it. As will be documented below, we find that essentially all interpolation is local”

We have also added Figure 4d. This new analysis documents the fact that decoded states are near trainingset trajectories, which is why the term ‘mesh’ is appropriate.

Reviewer #3:

Summary:  

This manuscript develops a new method termed MINT for decoding of behavior. The method is essentially a table-lookup rather than a model. Within a given stereotyped task, MINT tabulates averaged firing rate trajectories of neurons (neural states) and corresponding averaged behavioral trajectories as stereotypes to construct a library. For a test trial with a realized neural trajectory, it then finds the closest neural trajectory to it in the table and declares the associated behavior trajectory in the table as the decoded behavior. The method can also interpolate between these tabulated trajectories. The authors mention that the method is based on three key assumptions: (1) Neural states may not be embedded in a lowdimensional subspace, but rather in a high-dimensional space. (2) Neural trajectories are sparsely distributed under different behavioral conditions. (3) These neural states traverse trajectories in a stereotyped order.  

The authors conducted multiple analyses to validate MINT, demonstrating its decoding of behavioral trajectories in simulations and datasets (Figures 3, 4). The main behavior decoding comparison is shown in Figure 4. In stereotyped tasks, decoding performance is comparable (M_Cycle, MC_Maze) or better (Area 2_Bump) than other linear/nonlinear algorithms

(Figure 4). However, MINT underperforms for the MC_RTT task, which is less stereotyped (Figure 4).  

This paper is well-structured and its main idea is clear. The fact that performance on stereotyped tasks is high is interesting and informative, showing that these stereotyped tasks create stereotyped neural trajectories. The task-specific comparisons include various measures and a variety of common decoding approaches, which is a strength. However, I have several major concerns. I believe several of the conclusions in the paper, which are also emphasized in the abstract, are not accurate or supported, especially about generalization, computational scalability, and utility for BCIs. MINT is essentially a table-lookup algorithm based on stereotyped task-dependent trajectories and involves the tabulation of extensive data to build a vast library without modeling. These aspects will limit MINT's utility for real-world BCIs and tasks. These properties will also limit MINT's generalizability from task to task, which is important for BCIs and thus is commonly demonstrated in BCI experiments with other decoders without any retraining. Furthermore, MINT's computational and memory requirements can be prohibitive it seems. Finally, as MINT is based on tabulating data without learning models of data, I am unclear how it will be useful in basic investigations of neural computations. I expand on these concerns below.  

We thank the reviewer for pointing out weaknesses in our framing and presentation. The comments above made us realize that we needed to 1) better document the ways in which MINT is far more flexible than a lookup-table, and 2) better explain the competing scientific perspectives at play. R3’s comments also motivated us to add an additional analysis of generalization. In our view the manuscript is greatly improved by these additions. Specifically, these additions directly support the broader impact that we hope the study will have.

For simplicity and readability, we first group and summarize R3’s main concerns in order to better address them. (These main concerns are all raised above, in addition to recurring in the specific comments below. Responses to each individual specific comment are provided after these summaries.)

(1) R3 raises concerns about ‘computational scalability.’ The concern is that “MINT's computational and memory requirements can be prohibitive.” This point was expanded upon in a specific comment, reproduced below:

I also find the statement in the abstract and paper that "computations are simple, scalable" to be inaccurate. The authors state that MINT's computational cost is O(NC) only, but it seems this is achieved at a high memory cost as well as computational cost in training. The process is described in section "Lookup table of log-likelihoods" on line [978-990]. The idea is to precompute the log-likelihoods for any combination of all neurons with discretization x all delay/history segments x all conditions and to build a large lookup table for decoding. Basically, the computational cost of precomputing this table is O(V^{Nτ} x TC) and the table requires a memory of O(V^{Nτ}), where V is the number of discretization points for the neural firing rates, N is the number of neurons, τ is the history length, T is the trial length, and C is the number of conditions. This is a very large burden, especially the V^{Nτ} term. This cost is currently not mentioned in the manuscript and should be clarified in the main text. Accordingly, computation claims should be modified including in the abstract.

The revised manuscript clarifies that our statement (that computations are simple and scalable) is absolutely accurate. There is no need to compute, or store, a massive lookup table. There are three tables: two of modest size and one that is tiny. This is now better explained:

“Thus, the log-likelihood of , for a particular current neural state, is simply the sum of many individual log-likelihoods (one per neuron and time-bin). Each individual log-likelihood depends on only two numbers: the firing rate at that moment and the spike count in that bin. To simplify online computation, one can precompute the log-likelihood, under a Poisson model, for every plausible combination of rate and spike-count. For example, a lookup table of size 2001 × 21 is sufficient when considering rates that span 0-200 spikes/s in increments of 0.1 spikes/s, and considering 20 ms bins that contain at most 20 spikes (only one lookup table is ever needed, so long as its firing-rate range exceeds that of the most-active neuron at the most active moment in Ω). Now suppose we are observing a population of 200 neurons, with a 200 ms history divided into ten 20 ms bins. For each library state, the log-likelihood of the observed spike-counts is simply the sum of 200 × 10 = 2000 individual loglikelihoods, each retrieved from the lookup table. In practice, computation is even simpler because many terms can be reused from the last time bin using a recursive solution (Methods). This procedure is lightweight and amenable to real-time applications.”

In summary, the first table simply needs to contain the firing rate of each neuron, for each condition, and each time in that condition. This table consumes relatively little memory. Assuming 100 one-second-long conditions (rates sampled every 20 ms) and 200 neurons, the table would contain 100 x 50 x 200 = 1,000,000 numbers. These numbers are typically stored as 16-bit integers (because rates are quantized), which amounts to about 2 MB. This is modest, given that most computers have (at least) tens of GB of RAM. A second table would contain the values for each behavioral variable, for each condition, and each time in that condition. This table might contain behavioral variables at a finer resolution (e.g. every millisecond) to enable decoding to update in between 20 ms bins (1 ms granularity is not needed for most BCI applications, but is the resolution used in this study). The number of behavioral variables of interest for a particular BCI application is likely to be small, often 1-2, but let’s assume for this example it is 10 (e.g. x-, y-, and z-position, velocity, and acceleration of a limb, plus one other variable). This table would thus contain 100 x 1000 x 10 = 1,000,000 floating point numbers, i.e. an 8 MB table. The third table is used to store the probability of s spikes being observed given a particular quantized firing rate (e.g. it may contain probabilities associated with firing rates ranging from 0 – 200 spikes/s in 0.1 spikes/s increments). This table is not necessary, but saves some computation time by precomputing numbers that will be used repeatedly. This is a very small table (typically ~2000 x 20, i.e. 320 KB). It does not need to be repeated for different neurons or conditions, because Poisson probabilities depend on only rate and count.

(2) R3 raises a concern that MINT “is essentially a table-lookup rather than a model.’ R3 states that MINT 

“is essentially a table-lookup algorithm based on stereotyped task-dependent trajectories and involves the tabulation of extensive data to build a vast library without modeling.”

and that,

“as MINT is based on tabulating data without learning models of data, I am unclear how it will be useful in basic investigations of neural computations.”

This concern is central to most subsequent concerns. The manuscript has been heavily revised to address it. The revisions clarify that MINT is much more flexible than a lookup table, even though MINT uses a lookup table as its first step. Because R3’s concern is intertwined with one’s scientific assumptions, we have also added the new Figure 1 to explicitly illustrate the two key scientific perspectives and their decoding implications. 

Under the perspective in Figure 1a, R3 would be correct in saying that there exist traditional interpretable decoders (e.g. a Kalman filter) whose assumptions better model the data. Under this perspective, MINT might still be an excellent choice in many cases, but other methods would be expected to gain the advantage when situations demand more flexibility. This is R3’s central concern, and essentially all other concerns flow from it. It makes sense that R3 has this concern, because their comments repeatedly stress a foundational assumption of the perspective in Figure 1a: the assumption of a fixed lowdimensional neural subspace where activity has a reliable relationship to behavior that can be modeled and leveraged during decoding. The phrases below accord with that view:

“Unlike MINT, these works can achieve generalization because they model the neural subspace and its association to movement.”

“it will not generalize… even when these tasks cover the exact same space (e.g., the same 2D computer screen and associated neural space).”

“For proper training, the training data should explore the whole movement space and the associated neural space”

“I also believe the authors should clarify the logic behind developing MINT better. From a scientific standpoint, we seek to gain insights into neural computations by making various assumptions and building models that parsimoniously describe the vast amount of neural data rather than simply tabulating the data. For instance, low-dimensional assumptions have led to the development of numerous dimensionality reduction algorithms and these models have led to important interpretations about the underlying dynamics”

Thus, R3 prefers a model that 1) assumes a low-dimensional subspace that is fixed across tasks and 2) assumes a consistent ‘association’ between neural activity and kinematics. Because R3 believes this is the correct model of the data, they believe that decoders should leverage it. Traditional interpretable method do, and MINT doesn’t, which is why they find MINT to be unprincipled. This is a reasonable view, but it is not our view. We have heavily revised the manuscript to clarify that a major goal of our study is to explore the implications of a different, less-traditional scientific perspective.

The new Figure 1a illustrates the traditional perspective. Under this perspective, one would agree with R3’s claim that other methods have the opportunity to model the data better. For example, suppose there exists a consistent neural subspace – conserved across tasks – where three neural dimensions encode 3D hand position and three additional neural dimensions encode 3D hand velocity. A traditional method such as a Kalman filter would be a very appropriate choice to model these aspects of the data.

Figure 1b illustrates the alternative scientific perspective. This perspective arises from recent, present, and to-be-published observations. MINT’s assumptions are well-aligned with this perspective. In contrast, the assumptions of traditional methods (e.g. the Kalman filter) are not well-aligned with the properties of the data under this perspective. This does not mean traditional methods are not useful. Yet under Figure 1b, it is traditional methods, such as the Kalman filter, that lack an accurate model of the data. Of course, the reviewer may disagree with our scientific perspective. We would certainly concede that there is room for debate. However, we find the evidence for Figure 1b to be sufficiently strong that it is worth exploring the utility of methods that align with this scientific perspective. MINT is such a method. As we document, it performs very well.

Thus, in our view, MINT is quite principled because its assumptions are well aligned with the data. It is true that the features of the data that MINT models are a bit different from those that are traditionally modeled. For example, R3 is quite correct that MINT does not attempt to use a biomimetic model of the true transformation from neural activity, to muscle activity, and thence to kinematics. We see this as a strength, and the manuscript has been revised accordingly (see paragraph beginning with “We leveraged this simulated data to compare MINT with a biomimetic decoder”).

(3) R3 raises concerns that MINT cannot generalize. This was a major concern of R3 and is intimately related to concern #2 above. The concern is that, if MINT is “essentially a lookup table” that simply selects pre-defined trajectories, then MINT will not be able to generalize. R3 is quite correct that MINT generalizes rather differently than existing methods. Whether this is good or bad depends on one’s scientific perspective. Under Figure 1a, MINT’s generalization would indeed be limiting because other methods could achieve greater flexibility. Under Figure 1b, all methods will have serious limits regarding generalization. Thus, MINT’s method for generalizing may approximate the best one can presently do. To address this concern, we have made three major changes, numbered i-iii below:

i) Large sections of the manuscript have been restructured to underscore the ways in which MINT can generalize. A major goal was to counter the impression, stated by R3 above, that: 

“for a test trial with a realized neural trajectory, [MINT] then finds the closest neural trajectory to it in the table and declares the associated behavior trajectory in the table as the decoded behavior”.

This description is a reasonable way to initially understand how MINT works, and we concede that we may have over-used this intuition. Unfortunately, it can leave the misimpression that MINT decodes by selecting whole trajectories, each corresponding to ‘a behavior’. This can happen, but it needn’t and typically doesn’t. As an example, consider the cycling task. Suppose that the library consists of stereotyped trajectories, each four cycles long, at five fixed speeds from 0.5-2.5 Hz. If the spiking observations argued for it, MINT could decode something close to one of these five stereotyped trajectories. Yet it needn’t. Decoded trajectories will typically resemble library trajectories locally, but may be very different globally. For example, a decoded trajectory could be thirty cycles long (or two, or five hundred) perhaps speeding up and slowing down multiple times across those cycles.

Thus, the library of trajectories shouldn’t be thought of as specifying a limited set of whole movements that can be ‘selected from’. Rather, trajectories define a scaffolding that outlines where the neural state is likely to live and how it is likely to be changing over time. When we introduce the idea of library trajectories, we are now careful to stress that they don’t function as a set from which one trajectory is ‘declared’ to be the right one:

“We thus designed MINT to approximate that manifold using the trajectories themselves, rather than their covariance matrix or corresponding subspace. Unlike a covariance matrix, neural trajectories indicate not only which states are likely, but also which state-derivatives are likely. If a neural state is near previously observed states, it should be moving in a similar direction. MINT leverages this directionality.

Training-set trajectories can take various forms, depending on what is convenient to collect. Most simply, training data might include one trajectory per condition, with each condition corresponding to a discrete movement. Alternatively, one might instead employ one long trajectory spanning many movements. Another option is to employ many sub-trajectories, each briefer than a whole movement. The goal is simply for training-set trajectories to act as a scaffolding, outlining the manifold that might be occupied during decoding and the directions in which decoded trajectories are likely to be traveling.”

Later in that same section we stress that decoded trajectories can move along the ‘mesh’ in nonstereotyped ways:

“Although the mesh is formed of stereotyped trajectories, decoded trajectories can move along the mesh in non-stereotyped ways as long as they generally obey the flow-field implied by the training data. This flexibility supports many types of generalization, including generalization that is compositional in nature. Other types of generalization – e.g. from the green trajectories to the orange trajectories in Figure 1b – are unavailable when using MINT and are expected to be challenging for any method (as will be documented in a later section).”

The section “Training and decoding using MINT” has been revised to clarify the ways in which interpolation is flexible, allowing decoded movements to be globally very different from any library trajectory.

“To decode stereotyped trajectories, one could simply obtain the maximum-likelihood neural state from the library, then render a behavioral decode based on the behavioral state with the same values of c and k. This would be appropriate for applications in which conditions are categorical, such as typing or handwriting. Yet in most cases we wish for the trajectory library to serve not as an exhaustive set of possible states, but as a scaffolding for the mesh of possible states. MINT’s operations are thus designed to estimate any neural trajectory – and any corresponding behavioral trajectory – that moves along the mesh in a manner generally consistent with the trajectories in Ω.”

“…interpolation allows considerable flexibility. Not only is one not ‘stuck’ on a trajectory from Φ, one is also not stuck on trajectories created by weighted averaging of trajectories in Φ. For example, if cycling speed increases, the decoded neural state could move steadily up a scaffolding like that illustrated in Figure 1b (green). In such cases, the decoded trajectory might be very different in duration from any of the library trajectories. Thus, one should not think of the library as a set of possible trajectories that are selected from, but rather as providing a mesh-like scaffolding that defines where future neural states are likely to live and the likely direction of their local motion. The decoded trajectory may differ considerably from any trajectory within Ω.”

This flexibility is indeed used during movement. One empirical example is described in detail:

“During movement… angular phase was decoded with effectively no net drift over time. This is noteworthy because angular velocity on test trials never perfectly matched any of the trajectories in Φ. Thus, if decoding were restricted to a library trajectory, one would expect growing phase discrepancies. Yet decoded trajectories only need to locally (and approximately) follow the flow-field defined by the library trajectories. Based on incoming spiking observations, decoded trajectories speed up or slow down (within limits).

This decoding flexibility presumably relates to the fact that the decoded neural state is allowed to differ from the nearest state in Ω. To explore… [the text goes on to describe the new analysis in Figure 4d, which shows that the decoded state is typically not on any trajectory, though it is typically close to a trajectory].”

Thus, MINT’s operations allow considerable flexibility, including generalization that is compositional in nature. Yet R3 is still correct that there are other forms of generalization that are unavailable to MINT. This is now stressed at multiple points in the revision. However, under the perspective in Figure 1b, these forms of generalization are unavailable to any current method. Hence we made a second major change in response to this concern…  ii) We explicitly illustrate how the structure of the data determines when generalization is or isn’t possible. The new Figure 1a,b introduces the two perspectives, and the new Figure 6a,b lays out their implications for generalization. Under the perspective in Figure 6a, the reviewer is quite right: other methods can generalize in ways that MINT cannot. Under the perspective in Figure 6b, expectations are very different. Those expectations make testable predictions. Hence the third major change… iii) We have added an analysis of generalization, using a newly collected dataset. This dataset was collected using Neuropixels Probes during our Pac-Man force-tracking task. This dataset was chosen because it is unusually well-suited to distinguishing the predictions in Figure 6a versus Figure 6b. Finding a dataset that can do so is not simple. Consider R3’s point that training data should “explore the whole movement space and the associated neural space”. The physical simplicity of the Pac-Man task makes it unusually easy to confirm that the behavioral workspace has been fully explored. Importantly, under Figure 6b, this does not mean that the neural workspace has been fully explored, which is exactly what we wish to test when testing generalization. We do so, and compare MINT with a Wiener filter. A Wiener filter is an ideal comparison because it is simple, performs very well on this task, and should be able to generalize well under Figure 1a. Additionally, the Wiener filter (unlike the Kalman Filter) doesn’t leverage the assumption that neural activity reflects the derivative of force. This matters because we find that neural activity does not reflect dforce/dt in this task. The Wiener filter is thus the most natural choice of the interpretable methods whose assumptions match Figure 1a.

The new analysis is described in Figure 6c-g and accompanying text. Results are consistent with the predictions of Figure 6b. We are pleased to have been motivated to add this analysis for two reasons. First, it provides an additional way of evaluating the predictions of the two competing scientific perspectives that are at the heart of our study. Second, this analysis illustrates an underappreciated way in which generalization is likely to be challenging for any decode method. It can be tempting to think that the main challenge regarding generalization is to fully explore the relevant behavioral space. This makes sense if a behavioral space has “an associated neural space”. However, we are increasingly of the opinion that it doesn’t. Different tasks often involve different neural subspaces, even when behavioral subspaces overlap. We have even seen situations where motor output is identical but neural subspaces are quite different. These facts are relevant to any decoder, something highlighted in the revised Introduction:

“MINT’s performance confirms that there are gains to be made by building decoders whose assumptions match a different, possibly more accurate view of population activity. At the same time, our results suggest fundamental limits on decoder generalization. Under the assumptions in Figure 1b, it will sometimes be difficult or impossible for decoders to generalize to not-yet-seen tasks. We found that this was true regardless of whether one uses MINT or a more traditional method. This finding has implications regarding when and how generalization should be attempted.”

We have also added an analysis (Figure 6e) illustrating how MINT’s ability to compute likelihoods can be useful in detecting situations that may strain generalization (for any method). MINT is unusual in being able to compute and use likelihoods in this way.

Detailed responses to R3: we reproduce each of R3’s specific concerns below, but concentrate our responses on issues not already covered above.

Main comments: 

Comment 1. MINT does not generalize to different tasks, which is a main limitation for BCI utility compared with prior BCI decoders that have shown this generalizability as I review below. Specifically, given that MINT tabulates task-specific trajectories, it will not generalize to tasks that are not seen in the training data even when these tasks cover the exact same space (e.g., the same 2D computer screen and associated neural space). 

First, the authors provide a section on generalization, which is inaccurate because it mixes up two fundamentally different concepts: 1) collecting informative training data and 2) generalizing from task to task. The former is critical for any algorithm, but it does not imply the latter. For example, removing one direction of cycling from the training set as the authors do here is an example of generating poor training data because the two behavioral (and neural) directions are non-overlapping and/or orthogonal while being in the same space. As such, it is fully expected that all methods will fail. For proper training, the training data should explore the whole movement space and the associated neural space, but this does not mean all kinds of tasks performed in that space must be included in the training set (something MINT likely needs while modeling-based approaches do not). Many BCI studies have indeed shown this generalization ability using a model. For example, in Weiss et al. 2019, center-out reaching tasks are used for training and then the same trained decoder is used for typing on a keyboard or drawing on the 2D screen. In Gilja et al. 2012, training is on a center-out task but the same trained decoder generalizes to a completely different pinball task (hit four consecutive targets) and tasks requiring the avoidance of obstacles and curved movements. There are many more BCI studies, such as Jarosiewicz et al. 2015 that also show generalization to complex realworld tasks not included in the training set. Unlike MINT, these works can achieve generalization because they model the neural subspace and its association to movement. On the contrary, MINT models task-dependent neural trajectories, so the trained decoder is very task-dependent and cannot generalize to other tasks. So, unlike these prior BCIs methods, MINT will likely actually need to include every task in its library, which is not practical. 

I suggest the authors remove claims of generalization and modify their arguments throughout the text and abstract. The generalization section needs to be substantially edited to clarify the above points. Please also provide the BCI citations and discuss the above limitation of MINT for BCIs. 

As discussed above, R3’s concerns are accurate under the view in Figure 1a (and the corresponding Figure 6a). Under this view, a method such as that in Gilja et al. or Jarosiewicz et al. can find the correct subspace, model the correct neuron-behavior correlations, and generalize to any task that uses “the same 2D computer screen and associated neural space”, just as the reviewer argues. Under Figure 1b things are quite different.

This topic – and the changes we have made to address it – is covered at length above. Here we simply want to highlight an empirical finding: sometimes two tasks use the same neural subspace and sometimes they don’t. We have seen both in recent data, and it is can be very non-obvious which will occur based just on behavior. It does not simply relate to whether one is using the same physical workspace. We have even seen situations where the patterns of muscle activity in two tasks are nearly identical, but the neural subspaces are fairly different. When a new task uses a new subspace, neither of the methods noted above (Gilja nor Jarosiewicz) will generalize (nor will MINT). Generalizing to a new subspace is basically impossible without some yet-to-be-invented approach. On the other hand, there are many other pairs of tasks (center-out-reaching versus some other 2D cursor control) where subspaces are likely to be similar, especially if the frequency content of the behavior is similar (in our recent experience this is often critical). When subspaces are shared, most methods will generalize, and that is presumably why generalization worked well in the studies noted above.

Although MINT can also generalize in such circumstances, R3 is correct that, under the perspective in Figure 1a, MINT will be more limited than other methods. This is now carefully illustrated in Figure 6a. In this traditional perspective, MINT will fail to generalize in cases where new trajectories are near previously observed states, yet move in very different ways from library trajectories. The reason we don’t view this is a shortcoming is that we expect it to occur rarely (else tangling would be high). We thus anticipate the scenario in Figure 6b.

This is worth stressing because R3 states that our discussion of generalization “is inaccurate because it mixes up two fundamentally different concepts: 1) collecting informative training data and 2) generalizing from task to task.” We have heavily revised this section and improved it. However, it was never inaccurate. Under Figure 6b, these two concepts absolutely are mixed up. If different tasks use different neural subspaces, then this requires collecting different “informative training data” for each. One cannot simply count on having explored the physical workspace.

Comment 2. MINT is shown to achieve competitive/high performance in highly stereotyped datasets with structured trials, but worse performance on MC_RTT, which is not based on repeated trials and is less stereotyped. This shows that MINT is valuable for decoding in repetitive stereotyped use-cases. However, it also highlights a limitation of MINT for BCIs, which is that MINT may not work well for real-world and/or less-constrained setups such as typing, moving a robotic arm in 3D space, etc. This is again due to MINT being a lookup table with a library of stereotyped trajectories rather than a model. Indeed, the authors acknowledge that the lower performance on MC_RTT (Figure 4) may be caused by the lack of repeated trials of the same type. However, real-world BCI decoding scenarios will also not have such stereotyped trial structure and will be less/un-constrained, in which MINT underperforms. Thus, the claim in the abstract or lines 480-481 that MINT is an "excellent" candidate for clinical BCI applications is not accurate and needs to be qualified. The authors should revise their statements according and discuss this issue. They should also make the use-case of MINT on BCI decoding clearer and more convincing. 

We discussed, above, multiple changes and additions to the revision that were made to address these concerns. Here we briefly expand on the comment that MINT achieves “worse performance on MC_RTT, which is not based on repeated trials and is less stereotyped”. All decoders performed poorly on this task. MINT still outperformed the two traditional methods, but this was the only dataset where MINT did not also perform better (overall) than the expressive GRU and feedforward network. There are probably multiple reasons why. We agree with R3 that one likely reason is that this dataset is straining generalization, and MINT may have felt this strain more than the two machine-learning-based methods. Another potential reason is the structure of the training data, which made it more challenging to obtain library trajectories in the first place. Importantly, these observations do not support the view in Figure 1a. MINT still outperformed the Kalman and Wiener filters (whose assumptions align with Fig. 1a). To make these points we have added the following:

“Decoding was acceptable, but noticeably worse, for the MC_RTT dataset… As will be discussed below, every decode method achieved its worst estimates of velocity for the MC_RTT dataset. In addition to the impact of slower reaches, MINT was likely impacted by training data that made it challenging to accurate estimate library trajectories. Due to the lack of repeated trials, MINT used AutoLFADS to estimate the neural state during training. In principle this should work well. In practice AutoLFADS may have been limited by having only 10 minutes of training data. Because the random-target task involved more variable reaches, it may also have stressed the ability of all methods to generalize, perhaps for the reasons illustrated in Figure 1b.

The only dataset where MINT did not perform the best overall was the MC_RTT dataset, where it was outperformed by the feedforward network and GRU. As noted above, this may relate to the need for MINT to learn neural trajectories from training data that lacked repeated trials of the same movement (a design choice one might wish to avoid). Alternatively, the less-structured MC_RTT dataset may strain the capacity to generalize; all methods experienced a drop in velocity-decoding R2 for this dataset compared to the others. MINT generalizes somewhat differently than other methods, and may have been at a modest disadvantage for this dataset. A strong version of this possibility is that perhaps the perspective in Figure 1a is correct, in which case MINT might struggle because it cannot use forms of generalization that are available to other methods (e.g. generalization based on neuron-velocity correlations). This strong version seems unlikely; MINT continued to significantly outperform the Wiener and Kalman filters, which make assumptions aligned with Figure 1a.”

Comment 3. Related to 2, it may also be that MINT achieves competitive performance in offline and trial-based stereotyped decoding by overfitting to the trial structure in a given task, and thus may not generalize well to online performance due to overfitting. For example, a recent work showed that offline decoding performance may be overfitted to the task structure and may not represent online performance (Deo et al. 2023). Please discuss. 

We agree that a limitation of our study is that we do not test online performance. There are sensible reasons for this decision:

“By necessity and desire, all comparisons were made offline, enabling benchmarked performance across a variety of tasks and decoded variables, where each decoder had access to the exact same data and recording conditions.”

We recently reported excellent online performance in the cycling task with a different algorithm

(Schroeder et al. 2022). In the course of that study, we consistently found that improvements in our offline decoding translated to improvements in our online decoding. We thus believe that MINT (which improves on the offline performance of our older algorithm) is a good candidate to work very well online. Yet we agree this still remains to be seen. We have added the following to the Discussion:

“With that goal in mind, there exist three important practical considerations. First, some decode algorithms experience a performance drop when used online. One presumed reason is that, when decoding is imperfect, the participant alters their strategy which in turn alters the neural responses upon which decoding is based. Because MINT produces particularly accurate decoding, this effect may be minimized, but this cannot be known in advance. If a performance drop does indeed occur, one could adapt the known solution of retraining using data collected during online decoding [13]. Another presumed reason (for a gap between offline and online decoding) is that offline decoders can overfit the temporal structure in training data [107]. This concern is somewhat mitigated by MINT’s use of a short spike-count history, but MINT may nevertheless benefit from data augmentation strategies such as including timedilated versions of learned trajectories in the libraries”

Comment 4. Related to 2, since MINT requires firing rates to generate the library and simple averaging does not work for this purpose in the MC_RTT dataset (that does not have repeated trials), the authors needed to use AutoLFADS to infer the underlying firing rates. The fact that MINT requires the usage of another model to be constructed first and that this model can be computationally complex, will also be a limiting factor and should be clarified. 

This concern relates to the computational complexity of computing firing-rate trajectories during training. Usually, rates are estimated via trial-averaging, which makes MINT very fast to train. This was quite noticeable during the Neural Latents Benchmark competition. As one example, for the “MC_Scaling 5 ms Phase”, MINT took 28 seconds to train while GPFA took 30 minutes, the transformer baseline (NDT) took 3.5 hours, and the switching nonlinear dynamical system took 4.5 hours.

However, the reviewer is quite correct that MINT’s efficiency depends on the method used to construct the library of trajectories. As we note, “MINT is a method for leveraging a trajectory library, not a method for constructing it”. One can use trial-averaging, which is very fast. One can also use fancier, slower methods to compute the trajectories. We don’t view this as a negative – it simply provides options. Usually one would choose trial-averaging, but one does not have to. In the case of MC_RTT, one has a choice between LFADS and grouping into pseudo-conditions and averaging (which is fast). LFADS produces higher performance at the cost of being slower. The operator can choose which they prefer. This is discussed in the following section:

“For MINT, ‘training’ simply means computation of standard quantities (e.g. firing rates) rather than parameter optimization. MINT is thus typically very fast to train (Table 1), on the order of seconds using generic hardware (no GPUs). This speed reflects the simple operations involved in constructing the library of neural-state trajectories: filtering of spikes and averaging across trials. At the same time we stress that MINT is a method for leveraging a trajectory library, not a method for constructing it. One may sometimes wish to use alternatives to trial-averaging, either of necessity or because they improve trajectory estimates. For example, for the MC_RTT task we used AutoLFADS to infer the library. Training was consequently much slower (hours rather than seconds) because of the time taken to estimate rates. Training time could be reduced back to seconds using a different approach – grouping into pseudo-conditions and averaging – but performance was reduced. Thus, training will typically be very fast, but one may choose time-consuming methods when appropriate.”

Comment 5. I also find the statement in the abstract and paper that "computations are simple, scalable" to be inaccurate. The authors state that MINT's computational cost is O(NC) only, but it seems this is achieved at a high memory cost as well as computational cost in training. The process is described in section "Lookup table of log-likelihoods" on line [978-990]. The idea is to precompute the log-likelihoods for any combination of all neurons with discretization x all delay/history segments x all conditions and to build a large lookup table for decoding. Basically, the computational cost of precomputing this table is O(V^{Nτ} x TC) and the table requires a memory of O(V^{Nτ}), where V is the number of discretization points for the neural firing rates, N is the number of neurons, τ is the history length, T is the trial length, and C is the number of conditions. This is a very large burden, especially the V^{Nτ} term. This cost is currently not mentioned in the manuscript and should be clarified in the main text. Accordingly, computation claims should be modified including in the abstract. 

As discussed above, the manuscript has been revised to clarify that our statement was accurate.

Comment 6. In addition to the above technical concerns, I also believe the authors should clarify the logic behind developing MINT better. From a scientific standpoint, we seek to gain insights into neural computations by making various assumptions and building models that parsimoniously describe the vast amount of neural data rather than simply tabulating the data. For instance, low-dimensional assumptions have led to the development of numerous dimensionality reduction algorithms and these models have led to important interpretations about the underlying dynamics (e.g., fixed points/limit cycles). While it is of course valid and even insightful to propose different assumptions from existing models as the authors do here, they do not actually translate these assumptions into a new model. Without a model and by just tabulating the data, I don't believe we can provide interpretation or advance the understanding of the fundamentals behind neural computations. As such, I am not clear as to how this library building approach can advance neuroscience or how these assumptions are useful. I think the authors should clarify and discuss this point. 

As requested, a major goal of the revision has been to clarify the scientific motivations underlying MINT’s design. In addition to many textual changes, we have added figures (Figures 1a,b and 6a,b) to outline the two competing scientific perspectives that presently exist. This topic is also addressed by extensions of existing analyses and by new analyses (e.g. Figure 6c-g). 

In our view these additions have dramatically improved the manuscript. This is especially true because we think R3’s concerns, expressed above, are reasonable. If the perspective in Figure 1a is correct, then R3 is right and MINT is essentially a hack that fails to model the data. MINT would still be effective in many circumstances (as we show), but it would be unprincipled. This would create limitations, just as the reviewer argues. On the other hand, if the perspective in Figure 1b is correct, then MINT is quite principled relative to traditional approaches. Traditional approaches make assumptions (a fixed subspace, consistent neuron-kinematic correlations) that are not correct under Figure 1b.

We don’t expect R3 to agree with our scientific perspective at this time (though we hope to eventually convince them). To us, the key is that we agree with R3 that the manuscript needs to lay out the different perspectives and their implications, so that readers have a good sense of the possibilities they should be considering. The revised manuscript is greatly improved in this regard.

Comment 7. Related to 6, there seems to be a logical inconsistency between the operations of MINT and one of its three assumptions, namely, sparsity. The authors state that neural states are sparsely distributed in some neural dimensions (Figure 1a, bottom). If this is the case, then why does MINT extend its decoding scope by interpolating known neural states (and behavior) in the training library? This interpolation suggests that the neural states are dense on the manifold rather than sparse, thus being contradictory to the assumption made. If interpolation-based dense meshes/manifolds underlie the data, then why not model the neural states through the subspace or manifold representations? I think the authors should address this logical inconsistency in MINT, especially since this sparsity assumption also questions the low-dimensional subspace/manifold assumption that is commonly made. 

We agree this is an important issue, and have added an analysis on this topic (Figure 4d). The key question is simple and empirical: during decoding, does interpolation cause MINT to violate the assumption of sparsity? R3 is quite right that in principle it could. If spiking observations argue for it, MINT’s interpolation could create a dense manifold during decoding rather than a sparse one. The short answer is that empirically this does not happen, in agreement with expectations under Figure 1b. Rather than interpolating between distant states and filling in large ‘voids’, interpolation is consistently local. This is a feature of the data, not of the decoder (MINT doesn’t insist upon sparsity, even though it is designed to work best in situations where the manifold is sparse).

In addition to adding Figure 4d, we added the following (in an earlier section):

“The term mesh is apt because, if MINT’s assumptions are correct, interpolation will almost always be local. If so, the set of decodable states will resemble a mesh, created by line segments connecting nearby training-set trajectories. However, this mesh-like structure is not enforced by MINT’s operations. Interpolation could, in principle, create state-distributions that depart from the assumption of a sparse manifold. For example, interpolation could fill in the center of the green tube in Figure 1b, resulting in a solid manifold rather than a mesh around its outer surface. However, this would occur only if spiking observations argued for it. As will be documented below, we find that essentially all interpolation is local.”

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors): 

I appreciate the detailed methods section, however, more specifics should be integrated into the main text. For example on Line 238, it should additionally be stated how many minutes were used for training and metrics like the MAE which is used later should be reported here.

Thank you for this suggestion. We now report the duration of training data in the main text:

“Decoding R^2 was .968 over ~7.1 minutes of test trials based on ~4.4 minutes of training data.”

We have also added similar specifics throughout the manuscript, e.g. in the Fig. 5 legend:

“Results are based on the following numbers of training / test trials: MC\_Cycle (174 train, 99 test), MC\_Maze (1721 train, 574 test), Area2\_Bump (272 train, 92 test), MC\_RTT (810 train, 268 test).”

Similar additions were made to the legends for Fig. 6 and 8. Regarding the request to add MAE for the multitask network, we did not do so for the simple reason that the decoded variable (muscle activity) has arbitrary units. The raw MAE is thus not meaningful. We could of course have normalized, but at this point the MAE is largely redundant with the correlation. In contrast, the MAE is useful when comparing across the MC_Maze, Area2_Bump, and MC_RTT datasets, because they all involve the same scale (cm/s).

Regarding the MC_RTT task, AutoLFADS was used to obtain robust spike rates, as reported in the methods. However, the rationale for splitting the neural trajectories after AutoLFADS is unclear. If the trajectories were split based on random recording gaps, this might lead to suboptimal performance? It might be advantageous to split them based on a common behavioural state? 

When learning neural trajectories via AutoLFADS, spiking data is broken into short (but overlapping) segments, rates are estimated for each segment via AutoLFADs, and these rates are then stitched together across segments into long neural trajectories. If there had been no recording gaps, these rates could have been stitched into a single neural trajectory for this dataset. However, the presence of recording gaps left us no choice but to stitch together these rates into more than one trajectory. Fortunately, recording gaps were rare: for the decoding analysis of MC_RTT there were only two recording gaps and therefore three neural trajectories, each ~2.7 minutes in duration. 

We agree that in general it is desirable to learn neural trajectories that begin and end at behaviorallyrelevant moments (e.g. in between movements). However, having these trajectories potentially end midmovement is not an issue in and of itself. During decoding, MINT is never stuck on a trajectory. Thus, if MINT were decoding states near the end of a trajectory that was cut short due to a training gap, it would simply begin decoding states from other trajectories or elsewhere along the same trajectory in subsequent moments. We could have further trimmed the three neural trajectories to begin and end at behaviorallyrelevant moments, but chose not to as this would have only removed a handful of potentially useful states from the library.

We now describe this in the Methods:

“Although one might prefer trajectory boundaries to begin and end at behaviorally relevant moments (e.g. a stationary state), rather than at recording gaps, the exact boundary points are unlikely to be consequential for trajectories of this length that span multiple movements. If MINT estimates a state near the end of a long trajectory, its estimate will simply jump to another likely state on a different trajectory (or earlier along the same trajectory) in subsequent moments. Clipping the end of each trajectory to an earlier behaviorally-relevant moment would only remove potentially useful states from the libraries.”

Are the training and execution times in Table 1 based on pure Matlab functions or Mex files? If it's Mex files as suggested by the code, it would be good to mention this in the Table caption.

They are based on a combination of MATLAB and MEX files. This is now clarified in the table caption:

“Timing measurements taken on a Macbook Pro (on CPU) with 32GB RAM and a 2.3 GHz 8-Core Intel Core i9 processor. Training and execution code used for measurements was written in MATLAB (with the core recursion implemented as a MEX file).”

As the method most closely resembles a Bayesian decoder it would be good to compare performance against a Naive Bayes decoder. 

We agree and have now done so. The following has been added to the text:

“A natural question is thus whether a simpler Bayesian decoder would have yielded similar results. We explored this possibility by testing a Naïve Bayes regression decoder [85] using the MC_Maze dataset. This decoder performed poorly, especially when decoding velocity (R2 = .688 and .093 for hand position and velocity, respectively), indicating that the specific modeling assumptions that differentiate MINT from a naive Bayesian decoder are important drivers of MINT’s performance.”

Line 199 Typo: The assumption of stereotypy trajectory also enables neural states (and decoded behaviors) to be updated in between time bins. 

Fixed

Table 3: It's unclear why the Gaussian binning varies significantly across different datasets. Could the authors explain why this is the case and what its implications might be? 

We have added the following description in the “Filtering, extracting, and warping data on each trial” subsection of the Methods to discuss how 𝜎 may vary due to the number of trials available for training and how noisy the neural data for those trials is:

“First, spiking activity for each neuron on each trial was temporally filtered with a Gaussian to yield single-trial rates. Table 3 reports the Gaussian standard deviations σ (in milliseconds) used for each dataset. Larger values of σ utilize broader windows of spiking activity when estimating rates and therefore reduce variability in those rate estimates. However, large σ values also yield neural trajectories with less fine-grained temporal structure. Thus, the optimal σ for a dataset depends on how variable the rate estimates otherwise are.”

An implementation of the method in an open-source programming language could further enhance the widespread use of the tool. 

We agree this would be useful, but have yet not implemented the method in any other programming languages. Implementation in Python is still a future goal.

Reviewer #2 (Recommendations For The Authors): 

- Figures 4 and 5 should show the error bars on the horizontal axis rather than portraying them vertically. 

[Note that these are now Figures 5 and 6]

The figure legend of Figure 5 now clarifies that the vertical ticks are simply to aid visibility when symbols have very similar means and thus overlap visually. We don’t include error bars (for this analysis) because they are very small and would mostly be smaller than the symbol sizes. Instead, to indicate certainty regarding MINT’s performance measurements, the revised text now gives error ranges for the correlations and MAE values in the context of Figure 4c. These error ranges were computed as the standard deviation of the sampling distribution (computed via resampling of trials) and are thus equivalent to SEMs. The error ranges are all very small; e.g. for the MC_Maze dataset the MAE for x-velocity is 4.5 +/- 0.1 cm/s. (error bars on the correlations are smaller still).

Thus, for a given dataset, we can be quite certain of how well MINT performs (within ~2% in the above case). This is reassuring, but we also don’t want to overemphasize this accuracy. The main sources of variability one should be concerned about are: 1) different methods can perform differentially well for different brain areas and tasks, 2) methods can decode some behavioral variables better than others, and 3) performance depends on factors like neuron-count and the number of training trials, in ways that can differ across decode methods. For this reason, the study examines multiple datasets, across tasks and brain areas, and measures performance for a range of decoded variables. We also examine the impact of training-set-size (Figure 8a) and population size (solid traces in Fig. 8b, see R2’s next comment below). 

There is one other source of variance one might be concerned about, but it is specific to the neuralnetwork approaches: different weight initializations might result in different performance. For this reason, each neural-network approach was trained ten times, with the average performance computed. The variability around this average was very small, and this is now stated in the Methods.

“For the neural networks, the training/testing procedure was repeated 10 times with different random seeds. For most behavioral variables, there was very little variability in performance across repetitions. However, there were a few outliers for which variability was larger. Reported performance for each behavioral group is the average performance across the 10 repetitions to ensure results were not sensitive to any specific random initialization of each network.”

- For Figure 6, it is unclear whether the neuron-dropping process was repeated multiple times. If not, it should be since the results will be sensitive to which particular subsets of neurons were "dropped". In this case, the results presented in Figure 6 should include error bars to describe the variability in the model performance for each decoder considered. 

A good point. The results in Figure 8 (previously Figure 6) were computed by averaging over the removal of different random subsets of neurons (50 subsets per neuron count), just as the reviewer requests. The figure has been modified to include the standard deviation of performance across these 50 subsets. The legend clarifies how this was done.

Reviewer #3 (Recommendations For The Authors): 

Other comments: 

(1) [Line 185-188] The authors argue that in a 100-dimensional space with 10 possible discretized values, 10^100 potential neural states need to be computed. But I am not clear on this. This argument seems to hold only in the absence of a model (as in MINT). For a model, e.g., Kalman filter or AutoLFADS, information is encoded in the latent state. For example, a simple Kalman filter for a linear model can be used for efficient inference. This 10^100 computation isn't a general problem but seems MINT-specific, please clarify. 

We agree this section was potentially confusing. It has been rewritten. We were simply attempting to illustrate why maximum likelihood computations are challenging without constraints. MINT simplifies this problem by adding constraints, which is why it can readily provide data likelihoods (and can do so using a Poisson model). The rewritten section is below:

“Even with 1000 samples for each of the neural trajectories in Figure 3, there are only 4000 possible neural states for which log-likelihoods must be computed (in practice it is fewer still, see Methods). This is far fewer than if one were to naively consider all possible neural states in a typical rate- or factor-based subspace. It thus becomes tractable to compute log-likelihoods using a Poisson observation model. A Poisson observation model is usually considered desirable, yet can pose tractability challenges for methods that utilize a continuous model of neural states. For example, when using a Kalman filter, one is often restricted to assuming a Gaussian observation model to maintain computational tractability “

(2) [Figure 6b] Why do the authors set the dropped neurons to zero in the "zeroed" results of the robustness analysis? Why not disregard the dropped neurons during the decoding process? 

We agree the terminology we had used in this section was confusing. We have altered the figure and rewritten the text. The following, now at the beginning of that section, addresses the reviewer’s query: 

“It is desirable for a decoder to be robust to the unexpected loss of the ability to detect spikes from some neurons. Such loss might occur while decoding, without being immediately detected. Additionally, one desires robustness to a known loss of neurons / recording channels. For example, there may have been channels that were active one morning but are no longer active that afternoon. At least in principle, MINT makes it very easy to handle this second situation: there is no need to retrain the decoder, one simply ignores the lost neurons when computing likelihoods. This is in contrast to nearly all other methods, which require retraining because the loss of one neuron alters the optimal parameters associated with every other neuron.”

The figure has been relabeled accordingly; instead of the label ‘zeroed’, we use the label ‘undetected neuron loss’.

(3) Authors should provide statistical significance on their results, which they already did for Fig. S3a,b,c but missing on some other figures/places. 

We have added error bars in some key places, including in the text when quantifying MINT’s performance in the context of Figure 4. Importantly, error bars are only as meaningful as the source of error they assess, and there are reasons to be careful given this. The standard method for putting error bars on performance is to resample trials, which is indeed what we now report. These error bars are very small. For example, when decoding horizontal velocity for the MC_Maze dataset, the correlation between MINT’s decode and the true velocity had a mean and SD of the sampling distribution of 0.963 +/- 0.001. This means that, for a given dataset and target variable, we have enough trials/data that we can be quite certain of how well MINT performs. However, we want to be careful not to overstate this certainty. What one really wants to know is how well MINT performs across a variety of datasets, brain areas, target variables, neuron counts, etc. It is for this reason that we make multiple such comparisons, which provides a more valuable view of performance variability.

For Figure 7, error bars are unavailable. Because this was a benchmark, there was exactly one test-set that was never seen before. This is thus not something that could be resampled many times (that would have revealed the test data and thus invalidated the benchmark, not to mention that some of these methods take days to train). We could, in principle, have added resampling to Figure 5. In our view it would not be helpful and could be misleading for the reasons noted above. If we computed standard errors using different train/test partitions, they would be very tight (mostly smaller than the symbol sizes), which would give the impression that one can be quite certain of a given R^2 value. Yet variability in the train/test partition is not the variability one is concerned about in practice. In practice, one is concerned about whether one would get a similar R^2 for a different dataset, or brain area, or task, or choice of decoded variable. Our analysis thus concentrated on showing results across a broad range of situations. In our view this is a far more relevant way of illustrating the degree of meaningful variability (which is quite large) than resampling, which produces reassuringly small but (mostly) irrelevant standard errors.

Error bars are supplied in Figure 8b. These error bars give a sense of variability across re-samplings of the neural population. While this is not typically the source of variability one is most concerned about, for this analysis it becomes appropriate to show resampling-based standard errors because a natural concern is that results may depend on which neurons were dropped. So here it is both straightforward, and desirable, to compute standard errors. (The fact that MINT and the Wiener filter can be retrained many times swiftly was also key – this isn’t true of the more expressive methods). Figure S1 also uses resampling-based confidence intervals for similar reasons.

(4) [Line 431-437] Authors state that MINT outperforms other methods with the PSTH R^2 metric (trial-averaged smoothed spikes for each condition). However, I think this measure may not provide a fair comparison and is confounded because MINT's library is built using PSTH (i.e., averaged firing rate) but other methods do not use the PSTH. The author should clarify this. 

The PSTH R^2 metric was not created by us; it was part of the Neural Latents Benchmark. They chose it because it ensures that a method cannot ‘cheat’ (on the Bits/Spike measure) by reproducing fine features of spiking while estimating rates badly. We agree with the reviewer’s point: MINT’s design does give it a potential advantage in this particular performance metric. This isn’t a confound though, just a feature. Importantly, MINT will score well on this metric only if MINT’s neural state estimate is accurate (including accuracy in time). Without accurate estimation of the neural state at each time, it wouldn’t matter that the library trajectory is based on PSTHs. This is now explicitly stated:

“This is in some ways unsurprising: MINT estimates neural states that tend to resemble (at least locally) trajectories ‘built’ from training-set-derived rates, which presumably resemble test-set rates. Yet strong performance is not a trivial consequence of MINT’s design. MINT does not ‘select’ whole library trajectories; PSTH R2 will be high only if condition (c), index (k), and the interpolation parameter (α) are accurately estimated for most moments.”

Author Response

The following is the authors’ response to the current reviews.

Our answer to the final point(s) raised is as follows:

"We thank the reviewer for the comment. We checked our datasets accordingly. Typically, the n of cells showed deviations of maximally 20% from experiment to experiment (e.g. 16-24 cells per experiment). Additionally, experiments were performed using different passages of the cells. Moreover, data were validated at different time-points during the study using newly thawed cell lines."

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1 (Public Review):

Bischoff et al present a carefully prepared study on a very interesting and relevant topic: the role of ion channels (here a Ca2+-activated K+ channel BK) in regulating mitochondrial metabolism in breast cancer cells. The potential impact of these and similar observations made in other tumor entities has only begun to be appreciated. That being said, the authors pursue in my view an innovative approach to understanding breast cancer cell metabolism. Considering the following points would further strengthen the manuscript:

We thank reviewer #1 for the overall positive feedback on our study.

Methods:

(1) The authors use an extracellular Ca2+ concentration (2 mM) in their Ringer's solutions that is almost twice as high as the physiologically free Ca2+ concentration (ln 473). Moreover, the free Ca2+ concentration of their pipette solution is not indicated (ln 487).

Indeed, we utilized 2 mM of Ca2+ in the physiologic live-cell imaging buffer. This concentration could actually be a little lower than the total Ca2+ concentration (ranging usually from 2.2 to 2.6 mM) in the body, while the free Ca2+ concentration is typically half as high. Nevertheless, we find multiple studies different from ours, which utilized 2 mM for their live-cell-based experiments. Please check the following studies, which represent only a small selection:

https://doi.org/10.1038/s41598-019-49070-8

https://doi.org/10.1016/j.bpj.2020.08.045

https://doi.org/10.1016/j.redox.2022.102319

However, to ensure that the applied conditions are physiologically relevant, we reperformed experiments using MMTV-PyMT WT and MMTV-PyMT BK-KO cells and compared cytosolic Ca2+ concentrations over time in response to cell stimulation with ATP, either in the presence of 1.0 mM (Author response image 1A) or 2.0 mM extracellular Ca2+ (Author response image1B). The respective graphs are attached in the following for reviewer’s inspection. As expected, we find that the intracellular Ca2+ concentration in MMTV-PyMT WT and BK-KO cells was dependent on the extracellular Ca2+ concentration. Importantly, however, irrespective of the exact Ca2+ concentration applied, we observed a similar difference in basal cytosolic Ca2+ between MMTV-PyMT WT and BK-KO cells (Author response image1C).

Author response image 1.

Cytosolic Ca2+ concentrations over-time in the presence of 1.mM or 2.0 mM extracellular Ca2+.

Concerning the Ca2+ concentration in the patch-pipette – we are very glad that you uncovered an error in our description and apologize for the mistake. Actually, the information the reviewer is referring to was already given in the previous version of the manuscript, but unclear because a comma was shifted (see line 487 in the originally submitted manuscript). The Ca2+ concentration of the patch-pipette was 0.1 mM in the presence of 0.6 mM EGTA, which should (according to Ca-EGTA calculator, https://somapp.ucdmc.ucdavis.edu/pharmacology/bers/maxchelator/CaEGTA-NIST.htm) be equivalent to ~30 nM of free Ca2+ in the patch pipette. We corrected the mistake in the manuscript and thank the reviewer again for spotting this inaccuracy.

(2) Ca2+I measurements: The authors use ATP to elicit intracellular Ca2+ signals. Is this then a physiological stimulus for Ca2+ signaling in breast cancer? What is the rationale for using ATP? Moreover, it would be nice to see calibrated baseline values of Ca2+i.

We thank the reviewer for the comment and suggestion. Importantly, it was demonstrated recently, that all of the utilized cell lines respond to treatment with extracellular ATP with a prominent increase in Ca2+I, most probably indicating the expression of purinergic receptors, which was a prerequisite to observe ATP induced changes in [Ca2+]i.

https://doi.org/10.1038/s41419-022-05329-z,

https://doi.org/10.1093/carcin/bgt493

https://doi.org/10.1038/s41598-018-26459-5

Furthermore, ATP plays a crucial role in the tumor microenvironment, where high rates of cell death occur. Hence, ATP is of pathophysiologic relevance for the utilized cancer cell lines.

https://doi.org/10.1038/s41568-018-0037-0

https://doi.org/10.3390/cells9112496

https://doi.org/10.1002/jcp.30580

Following the suggestions by Reviewer #1 (and #2), we included calibrations of Ca2+cyto and Ca2+mito in the manuscript, by depleting the intracellular Ca2+ stores using Ionomycin in the absence of extracellular Ca2+ (EGTA) to validate the basal difference in Ca2+cyto and Ca2+mito. Additionally, Ca2+cyto was calibrated under basal and inhibitor treated conditions, and values in nM are given in the text (p. 5, lines 185-190, 193-195 and 199-200, in the tracked changes version of the MS). The new data can be found in new Figure S2F – Figure S2J and new Figure S2R – Figure S2V. Moreover, we calculated basal [Ca2+]cyto in the different BKCa pro- and deficient cell lines and under inhibitor treated conditions. We additionally added information about the pathophysiologic relevance of ATP in the tumor microenvironment in lines 175-178 in the tracked changes version of the manuscript.

(3) Membrane potential measurements: It would be nice to see a calibration of the potential measurements; this would allow us to correlate the IV relationship with membrane potential. Without calibration, it is hard to compare unless the identical uptake of the dye is shown. Does paxilline or IbTx also induce depolarization?

We thank the reviewer for the suggestion. Indeed, membrane potential calibrations/ measurements using the membrane potential sensitive dye Dibac4(3) would be interesting, however, technically hardly feasible. The reason is that the principle of the dye is based on different uptake in response to differences in membrane potential, and not ratiometric as for most other dyes/ sensors used. Considering this limitation, we decided to perform membrane potential measurements by patch-clamp analysis. Additionally, we performed these experiments upon inhibition of PM-located BKCa by IBTX. Current-clamp experiments confirmed the difference in basal membrane potential between MMTV-PyMT WT and BK-KO cells (consult new Figure S1C and lines 127-130 in the tracked changes version of the manuscript). Interestingly, IBTX treatment depolarized the PM potential to the BK-KO cell level, which validates that BK activity and PM potential are connected. In addition to this approach, we utilized our recently developed genetically encoded K+ sensors revealing basal differences in [K+]cyto between MMTV-PyMT WT and BK-KO cells. Also this difference between both genotypes was equalized by IBTX as the respective treatment increased [K+]cyto only in WT cells, which most likely explains the cause of PM depolarization (consult lines 130-135 in the tracked changes version of the manuscript and new Figure S1D and Figure S1E).

(4) Mito-potential measurements: Why did the authors use such a long time course and preincubate cells with channel blockers overnight? Why did they not perform paired experiments and record the immediate effect of the BK channel blockers in the mito potential?

We thank the reviewer for the suggestion. We performed TMRM-based experiments with MMTV-PyMT WT cells in response to short-term exposure to paxilline, which did not significantly affect the mitochondrial membrane potential, at least within 15 minutes of treatment (Author response image 2). This indicates, that further downstream processes subsequent to (mito) BKCa inhibition affect the mitochondrial membrane potential(MMP), most probably including remodeling processes of the respiratory chain, mitochondrial ion homeostasis or glycolytic activity, ultimately also delivering reduction equivalents to mitochondria. Our final goal was to validate potential differences between a BKCa pro-and deficient cell model, whereby the latter cells lacked the BKCa channel since its origination. Hence, “long-term” (~12h) BKCa inhibition as performed in our experiments rather reflects the BK-KO cell situation. Taken together with the new experiment (Author response image 2), we can now state that the effect of BK inhibition on the MMP is at least not the consequence of an acute (within minutes) channel blockade.

Author response image 2.

Mitochondrial membrane potential, as measured using TMRM, in response to acute short-term administration of 5µM paxilline, followed by mitochondrial depolarization using FCCP.

(5) MTT assays are also based on mitochondrial function - since modulation of mito function is at the core of this manuscript, an alternative method should be used.

We thank the reviewer for the important comment. We performed additional, immunofluorescence-based experiments using Ki-67 staining to assess cell proliferation rates. The newly added data can be found in the text, lines 409-412 in the tracked changes version of the manuscript and new Figure S6D-F. The results obtained confirm the MTTbased results (Fig.6H-I).

Results:

(1) Fig. 5G: The number of BK "positive" mitoplasts is surprisingly low - how does this affect the interpretation? Did the authors attempt to record mitoBK current in the "whole-mitoplast" mode? How does the mitoBK current density compare with that of the plasma membrane? Is it possible to theoretically predict the number of mitoBK channels per mitochondrion to elicit the observed effects? Can these results be correlated with the immuno-localization of mitoBK channels?

Indeed, the number of BKCa-positive mitoplasts appears low on a first view. However, as these experiments were performed in a mitoplast-attached mode, it is important to keep in mind that only a very small area of the actual mitoplast is investigated with each patch. If no channel was detected in such region, the patch was depicted as “empty”, as presented in Fig.5G, which does, however, not mean that the entire mitochondria was actually BKCa negative. Hence, the density of BKCa in the IMM might be higher than expected from our experiments. Nevertheless, already earlier results using glioblastoma cell lines – considered to be one of the cell lines mostly enriched in mitoBKCa – demonstrated a quite low density of BKCa β4 regulatory subunit in mitochondria – please see figure 2B in the following paper: 10.1371/journal.pone.0068125 – which (based on 1:1 stoichiometry of α and β subunits) also suggests that the density of the alpha subunit of BKCa might be low in this compartment.

Author response image 3.

Author response image 3: Schematic representation of mitoplast attached patch-clamp experiments

Theoretically, density predictions of mitoBK compared to PM localized BKCa would be possible if whole-mitoplast experiments were performed, however, we are unsure what added value this information would actually burst, allowing the pharmacologic modulation of structures originally located within the mitochondrial matrix. Please also consult Author response image 3. According to the most recent models, even if there are other views on this, mitoBKCa is oriented in a way, that the C-terminus with its Ca2+ binding bowl is located within the mitochondrial matrix. Hence, to allow Ca2+ sensitivity experiments of the channel, broken up (by swelling) mitoplasts are required to make the Ca2+ binding bowl accessible for Ca2+ manipulations in the bath solution. This approach does not allow us to compare the channel density to that of the PM.

Finally, to the best of our knowledge, a combination of immunofluorescence with mitoplast patch-clamp experiments is not feasible yet, and would probably be impossible due to the low density of the mitoBKCa as well as the lack of highly sensitive and specific antibodies.

(2) There are also reports about other mitoK channels (e.g. Kv1.3, KCa3.1, KATP) playing an important role in mitochondrial function. Did the authors observe them, too? Can the authors speculate on the relative importance of the different channels? Is it known whether they are expressed organ-/tumor-specifically?

Author response image 4.

Representative single channels different to mitoBKCa detected in MDAMB-453 mitoplasts.

The reviewer is right, other K+ channels have been found in mitochondria and these also play a role in tumor cells. This is also consistent with our data (Fig.5G), where we observed other channels in the mitoplasts of BCCs as well. These all four cell lines tested. According to their conductance and our expectations from literature, these channels may e.g. include mitoIKCa, mitoSKCa, mitoKATP orothers (10.1146/annurev-biophys-092622-094853). As we focused, however, on patches containing a mitoBKCa, we did not further pharmacologically characterize these channels. Two examples of channels we found in these mitoplasts besides BKCa are presented for reviewers’ inspection (Author response image 4). As our manuscript focusses on mitoBKCa, we did not further classify these channels in smaller subgroups according to their conductance, as we feel that a differentiation between BKCa (~210 pS), and channels showing a conductance ≤150pS, or a conductance ≤100 pS is sufficient. Furthermore, this additional information would dilute our story too much making it difficult for the (non-specialist) reader to follow the red thread of the study. We added respective information in the manuscript, however. Please consult lines 365-366 in the tracked changes version of the manuscript.

Reviewer #1 is right, the observed the different K+ channels might of course be organ- or tumor-specific. For example, it has been reported that the expression of K+ channels is different in various cancer cell (lines) (https://doi.org/10.2174/13816128113199990032, 10.1016/j.pharmthera.2021.107874, 10.1038/nrc3635), a fact, which also according to our study might be exploited for pharmacological manipulation, aiming to affect proliferation/apoptosis of cancer cells. Further, a recently published single-cell and spatially resolved atlas of human breast cancer implies that the expression of different K+ channels (such as mitoIKCa, mitoSKCa, mitoKATP) might even differ between cancer- and non-cancer cells within a single tumour (https://doi.org/10.1038/s41588-021-00911-1).

Reviewer #2 (Public Review):

Summary:

The large-conductance Ca2+ activated K+ channel (BK) has been reported to promote breast cancer progression, but it is not clear how. The present study carried out in breast cancer cell lines, concludes that BK located in mitochondria reprograms cells towards the Warburg phenotype, one of the metabolic hallmarks of cancer.

Strengths:

The use of a wide array of modern complementary techniques, including metabolic imaging, respirometry, metabolomics, and electrophysiology. On the whole, experiments are astute and well-designed and appear carefully done. The use of BK knock-out cells to control for the specificity of the pharmacological tools is a major strength. The manuscript is clearly written.

There are many interesting original observations that may give birth to new studies.

Weaknesses:

The main conclusion regarding the role of a BK channel located in mitochondria appears is not sufficiently supported. Other perfectible aspects are the interpretation of co-localization experiments and the calibration of Ca2+ dyes. These points are discussed in more detail in the following paragraphs:

We thank reviewer #2 for the thorough assessment of our study.

(1) May the metabolic effects be ascribed to a BK located in mitochondria? Unfortunately not, at least with the available evidence. While it is clear these cells have a BK in mitochondria (characteristic K+ currents detected in mitoplasts) and it is also well substantiated that the metabolic effects in intact cells are explained by an intracellular BK (paxilline effects absent in the BK KO), it does not follow that both observations are linked. Given that ectopic BKDEC appeared at the surface, a confounding factor is the likely expression of BK in other intracellular locations such as ER, Golgi, endosomes, etc. To their credit, authors acknowledge this limitation several times throughout the text ("...presumably mitoBK...") but not in other important places, particularly in the title and abstract.

We thank the reviewer for this important comment and amended the title and abstract, respectively. The title of the manuscript was changed to “mitoBKCa is functionally expressed in murine and human breast cancer cells and potentially contributes to metabolic reprogramming.” Additionally, we changed appropriate passages in the text, to emphasize that mitoBKCa potentially mediates the metabolic reprogramming, but other intracellular channels could also contribute to these processes.

(2) MitoBK subcellular location. Pearson correlations of 0.6 and about zero were obtained between the locations of mitoGREEN on one side, and mRFP or RFP-GPI on the other (Figs. 1G and S1E). These are nice positive and negative controls. For BK-DECRFP however, the Pearson correlation was about 0.2. What is the Z resolution of apotome imaging? Assuming an optimum optical section of 600 nm, as obtained by a 1.4 NA objective with a confocal, that mitochondria are typically 100 nm in diameter and that BK-DECRFP appears to stain more structures than mitoGREEN, the positive correlation of 0.2 may not reflect colocalization. For instance, it could be that BK-DECRFP is not just in mitochondria but in a close underlying organelle e.g. the ER. Along the same line, why did BK-RFP also give a positive Pearson? Isn´t that unexpected? Considering that BK-DEC was found by patch clamping at the plasma membrane, the subcellular targeting of the channel is suspect. Could it be that the endogenous BK-DEC does actually reside exclusively in mitochondria (a true mitoBK), but overflows to other membranes upon overexpression? Regarding immunodetection of BK in the mitochondrial Percoll preparation (Fig. S5), the absence of NKA demonstrates the absence of plasma membrane contamination but does not inform about contamination by other intracellular membranes.

Indeed, it seems that BKCa-DEC is not an exclusive mitoBKCa, at least not upon (over-/)expression in MCF-7 cells. It is known from literature, that mitochondrial K+ channels are encoded by the nuclear genome, as no obvious gene for a K+ channel is found in the mitochondrial genome. Channel proteins are synthetized by cytosolic ribosomes and likely translocated into mitochondria via the TOM/TIM system. Although some K+ channels possess a mitochondrial targeting sequence at the N-terminus, their import is mostly far from a general mechanism, and this seems also to be true for BK channels. In the case of the K+ channel Kv1.3, an even more complex scenario is hypothesized, as the channel located in the PM could be transferred to mitochondria via mitochondria-associated membranes (MAM) structures of the ER (https://doi.org/10.3390/ijms20030734). Yet, the detailed mechanism for BK shuttling to mitochondria is not fully understood. Possibly, overflow is exactly what is happening, due to very high levels of BK-DEC expression upon transfection. However, that the channel translocates to the IMM upon transfection is not surprising and was also demonstrated for other cell models including HEK293 – see e.g. 10.1038/s41598-021-904653. Unfortunately, transfection efficiency of MCF-7 is quite low compared to HEK293 – hence, quantitative statements from mito-patches upon transfection are difficult.

In order to ensure that the mitochondrial colocalization is not a matter of poor microscope resolution, we reperformed these experiments using confocal imaging on a Zeiss LSM980 with an Airyscan 2 detector, yielding z resolutions of ~ 450 nm. These experiments confirmed the increased colocalization of BKCa-DEC with mitochondria compared to BKCa lacking the DEC exon. Furthermore, this imaging at higher resolution demonstrated, that, unfortunately, colocalization might not be the best analysis, as especially fragmented mitochondria showed a clear MitoGREEN stained matrix, surrounded by red fluorescence derived from BKCaDECRFP present in the IMM (revised Fig. 1G).

To validate the results derived from immunoblotting, we additionally stained the membranes for TMX1, a marker for the ER membrane. This analysis confirmed the high purity of the mitochondrial isolation without ER-membrane contamination after percoll purification, and hence validated the presence of BKCa in the mitochondrial membrane (revised Fig. S5D). The additional information can be found in lines 156-159 in the tracked changes version of the manuscript.

(3) Calibration of fluorescent probes. The conclusion that BK blockers or BK expression affects resting Ca2+ levels should be better supported. Fluorescent sensors and dyes provide signals or ratios that need to be calibrated if comparisons between different cell types or experimental conditions are to be made. This is implicitly acknowledged here when monitoring ER Ca2+, with an elaborate protocol to deplete the organelle in order to achieve a reading at zero Ca2+.

We thank the reviewer for the important comment. Please note that at no point in the manuscript we aim to compare different cell lines concerning their intracellular Ca2+ concentration, but we only compare the same cell lines after the different treatments, as we are aware of this limitation of fluorescent probes. However, to validate the differences in intracellular Ca2+ concentrations, we calibrated the signals derived from Fura-2 and 4mtD3cpV using ionomycin in combination with cellular Ca2+ depletion/ saturation. The newly added data can be found in the text, lines 185-190, 192-195, 199-200, and 228-230 in the tracked changes version of the manuscript, as well as new Figure S2F – Figure S2J and new Figure S2R – Figure S2V

Line 203. "...solely by the expression of BKCa-DECRFP in MCF-7 cells". Granted, the effect of BKCa-DECRFP on the basal FRET ratio appears stronger than that of BK-RFP, but it appears that the latter had some effect. Please provide the statistics of the latter against the control group (after calibration, see above).

Author response image 5.

Dot blot for data shown in Figure 2I.

The reviewer is right, it seems that BKCaRFP may also affect [Ca2+]mito. However, the effect is not significant and shows a p-value of p>0.999 using Kruskal-Wallistest followed by Dunn’s multiple comparison test, due to the non-normally distributed nature of the data. p=0.0002 for ctrl vs. BKCa-DECRFP and 0.0022 for BKCaRFP vs. BKCa-DECRFP, however. We added a scatter dot-blot of the respective data as Author response image 5 for reviewer’s inspection. Additionally, first, even using a more stringent statistical test by only comparing ctrl vs BKCaRFP using Mann-Whitney test, the results are not significant, as the p-value was determined at 0.4467, and second, we performed the requested Ca2+calibration using ionomycin under these conditions, which confirmed the difference between ctrl cells and BKCa-DECRFP expressing cells, but not BKCaRFP expressing ones. Please see Figure S2V.

Reviewer #3 (Public Review):

The original research article, titled "mitoBKCa is functionally expressed in murine and human breast cancer cells and promotes metabolic reprogramming" by Bischof et al, has demonstrated the underlying molecular mechanisms of alterations in the function of Ca2+ activated K+ channel of large conductance (BKCa) in the development and progression of breast cancer. The authors also proposed that targeting mitoBKCa in combination with established anti-cancer approaches, could be considered as a novel treatment strategy in breast cancer treatment.

The paper is clearly written, and the reported results are interesting.

Strengths:

Rigorous biophysical experimental proof in support of the hypothesis.

Weaknesses:

A combinatorial synergistic study is missing.

We thank reviewer #3 for the positive summary of our study. Indeed, we propose that targeting of mitoBKCa in combination with established anti-cancer drugs may represent a novel anti-cancer treatment strategy. Unfortunately, we feel that the manuscript is very condensed already, and that adding respective required experiments and data to support this hypothesis will make the flow of the manuscript more complex or even incomprehensible. As no attempts linking mitoBKCa activity with anti-cancer therapies have been made so far, we removed the respective information from the abstract and only discuss this aspect.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

(1) Statistics: Legends have to contain information about the number of biological replicates (N) and cells analysed (n). Statistics must be calculated with the averages of the replicates.

Author response image 6.

Representative single cell responses of Fura-2 loaded MMTV-PyMT WT cells.

We thank the reviewer for the comment and added the missing details to all figure legends.

We feel that using each cell represents exactly the power of high-resolution live-cell imaging, as there is no better biological replicate than a single separated cell, which is observed by fluorescence microscopy. This analysis is also able to visualize cell-to-cell differences in the microscopy area, similarly to patch-clamp experiments, where each single cell or mitoplast patched is used as a single replicate. Please find a representative dataset derived from fluorescence microscopy of different responses of neighboring single cells in Author response image 6.

(2) Fig. 1G: This is a poor resolution figure, mostly because of its far too small size; in its current form it bears very little information.

We agree with reviewer #1 and reperformed the imaging experiments using high resolution confocal imaging and exchanged the respective images. We feel that this increased the quality of the images significantly. Unfortunately, we were not able to increase the size of the images in the main figure, hence, we added magnifications of the respective images as new Figure S1I.

(3) Fig. 1H: What do the dotted grey lines and the labels stand for?

We believe Reviewer #1 is probably referring to Figure 1G. As indicated in the figure panel and in the text, the grey dotted lines and labels indicate the colocalization scores of mtRFP and RFP-GPI with MitoGREEN, respectively. These data are also shown in Figure S1H, including error bars and statistics. We added additional information in the text to make the meaning of the lines clearer to the reader. Please consult lines 149 – 150 in the tracked changes version of the manuscript.

Reviewer #2 (Recommendations For The Authors):

(1) May the metabolic effects be ascribed to a BK located in mitochondria? Short of a way to tackle BK function and metabolism specifically in mitochondria, the conclusion may best be toned down to "intracellular BK". For the time being the term "mitoBK" appears too ambitious.

We fell that you are right and that our previous overstatement requires adaptation as a clear (100%) attribution of the observed metabolic effects solely to mitoBKCa is not definitely possible. We have therefore amended all relevant passages in the entire MS accordingly.

(2) MitoBK subcellular location. Please address the points raised in the Public Review.

As stated above we addressed the point raised in the public review accordingly (please consult new Figure S1I and revised Figure 1G).

(3) Calibration of fluorescent probes. Please provide calibrations for cytosolic and mitochondrial Ca2+, for example, the standard high Ca2+/ionophore/metabolic inhibition treatment to reach saturation followed by Ca2+ chelation to obtain zero Ca2+.

We thank the reviewer for the comment. As you can see from our response to the public review, we performed the respective experiments, and datasets were added in the manuscript.

(4) Line 203. "...solely by the expression of BKCa-DECRFP in MCF-7 cells". Granted, the effect of BKCa-DECRFP on the basal FRET ratio appears stronger than that of BK-RFP, but it appears that the latter had some effect. Please provide the statistics of the latter against the control group (after calibration, see above).

Please consult our response to the (same) comment in the public review.

(5) Line 228. The statement "Similar results were obtained in MDA-MB-453 cells" is confusing. As shown in Fig.3, pax reduced ECAR and OCR in MMTV-PyMT WT cells. As ibtx was without effect, it is suggested that intracellular BK support metabolism. However, the effect of pax on MDA cells was the opposite. Doesn´t this divergence speak against a universal role of intracellular BKs in promoting metabolism in BCCs? A similar point may be made regarding metabolomics, which showed no effects of pax on lactate and pyruvate in MMTV-PyMT WT cells but stimulation in MDA cells. Perhaps the word "promotes" in the title of the figure should be replaced by something more neutral like "affects" or "alters", as used elsewhere,

We thank the reviewer for pointing out the overstatement regarding intracellular BK functions and changed the title of the figure as suggested.

With regard to the experiments mentioned, we would like to point out the following aspects:

First, the cell lines used strongly differ in their metabolic settings under basal conditions. While both, MMTV-PyMT and MDA-MB-453 cells seem to show similar basal ECAR levels (if BKCa was present), their OCR seems to differ strongly. MMTV-PyMT cells seem to show a basal OCR which is almost at the maximum already, while MDA-MB-453 cells possess a tremendous capacity in their OCR, as observed upon mitochondrial uncoupling using FCCP. Of note, both, ECAR and OCR are indirect metabolic measures. On the one hand, ECAR measures extracellular acidification, which is accomplished by H+ along with lactate secretion. However, lactate secretion is not the only process leading to extracellular acidification, and ECAR may hence measure a variety of H+ releasing processes, including processes of vesicle secretion. On the other hand, OCR is not directly linked to ATP production, as mitochondrial complex IV is consuming O2, ATP, however, is produced by mitochondrial complex V. This becomes even more evident when having a look on OCRs after FCCP treatment – under these conditions, the H+ gradient is destroyed and ATP synthase activity is reduced, OCR, however, increases to the maximum due to increased supply of mitochondrial complex IV with H+.

Second, please note that the LC-MS-based metabolomics derive from a static single time point and not from an over-time “live” read-outs. Moreover, underlying dynamics of the parameters measured can not be assessed. Hence, as an example, increasing levels of pyruvate can e.g. indicate faster generation, or slower subsequent degradation/ metabolization. A clear in-depth statement about what is happening under basal and BKCa inhibitor treated conditions is hence not possible. The only conclusion possible to draw from these experiments is that paxilline treatment differentially affects metabolic pathways in these cells.

Based on these limitations of both methods, we decided to perform our in-depth fluorescence microscopy-based analysis, which provided strong evidence for intracellular BKCa channels on mitochondrial ATP production. Despite opposing effects of BKCa inhibition on OCR in MMTV-PyMT WT and MDA-MB-453 cells, mitochondrial ATP production was reduced, if BKCa-DECRFP was expressed/ intracellular BKCa was functional.

In line with these findings, mitoBKCa was recently described as an uncoupling protein, which could furthermore explain the differential effects of intracellular BKCa inhibition on OCR. https://doi.org/10.1038/s41598-021-90465-3

Minor

(6) Fig. 1C. Average fluorescence intensity in 6 experiments was about 20% higher in BK-KO cells relative to WT. Such a small difference is significant but should not be evident to the eye. The pictures selected for illustration appear to show a much larger difference and therefore may not be representative. If this is the case, please omit them. The same goes for the other representative pictures.

Author response image 7.

: Representative images at different brightnesses.

Please note, that the analysis of the images was done in an unbiased way using a Fiji macro. After analysis, we chose representative images, which were closest to the average.

Furthermore, we must kindly disagree with the reviewer as changes of 20% in fluorescence intensity are indeed evident to the eye (consult Author response image 7). This panels show the same image at different brightness levels with intensity differences of 20%. Hence, we feel, that all the images the reviewer was referring are representative for the values given.

(7) Line 130. The definition of "recent" is of course relative, but 10 years?

We are very glad that you have discovered this “inconsistency", and reworded the respective phrase accordingly.

(8) Line 327. "conductivity" is the property of a medium, "conductance" is the property of a component, such as a channel.

We thank the reviewer for the important comment. We revised the text accordingly.

(9) Various figures. FRET sensor data are expressed as Ratio(FRET/CFP). This is unusual, typically it should be FRET ratio (YFP/CFP), FRET ratio(mTFP/Venus), etc. Please note that the FRET partners differ between sensors.

We acknowledge the comment of the reviewer. It is correct that fluorescent proteins vary widely between the sensors (used). Please note, however, the following: The emission measured from these sensors actually represents FRET, as CFP but not YFP is directly excited. Hence, emission is FRET, not the “intrinsic” fluorescence of the YFP. This is getting more and more important to differentiate, as there are probes existing, which can also be “alternately” excited, i.e. CFP and YFP separately, which will then yield the YFP/CFP ratio (https://doi.org/10.1021/acssensors.8b01599). In case of only CFP excitation, we feel, that the term FRET/CFP is preferable over other labelings such as YFP/CFP.

(10) BK-DEC makes BCCs cells less oxidative. However, BK-DEC was first described in cardiomyocytes, which are among the most oxidative cell types. It would be useful if authors could address this apparent contradiction in the Discussion Section.

That is an exciting point that we addressed as follows in the revised MS:

First, it is important to mention that cardiac myocytes do not show a metabolic Warburg setting and are – under physiologic conditions – maintained in a high O2 environment.

Second, a recent study from our group addressed the question about the role of mitoBKCa in primary cardiac myocytes. Indeed, mitoBKCa was functionally expressed in these cells. Interestingly, under physiologic conditions, the channel did not alter (multiple) cell behaviours nor overall cardiac physiology in a mouse model. However, upon induction of ischemia/ reperfusion injury, a lack of BK increased cardiac susceptibility to cell death resulting in increased infarction size (https://doi.org/10.1161/CIRCULATIONAHA.117.028723). Hence, also in this cell model, BKCa only played a role under oxygen limited conditions/ conditions where mitochondria were not properly functioning. Thus, the results derived from cardiac myocytes support our recent findings in BCCs, as BKCa mediates BCC resistance to hypoxic stress/ makes BCCs more independent from oxidative metabolism.

Parts of this discussion were included in the revised MS. Please consult lines 490-500 in the tracked changes version of the manuscript.

Reviewer #3 (Recommendations For The Authors):

(1) The study is very well designed and most of the computational analyses were done rigorously.

We highly appreciate the positive feedback by reviewer #3.

(2) The authors should discuss the expression of BKCa in different subsets of breast cancer. Authors may also debate on the level of steroid receptors and BKCa expressions.

We thank reviewer #3 for the important suggestion and added the requested information in the discussion, lines 445-447 and 450-454 in the tracked changes version of the manuscript.

(3) In the discussion section, the authors mentioned that the MCF7 cell is the best model to study this hypothesis. Does it imply that triple-negative breast cancer cell lines express lower levels of BKCa? The authors should discuss this.

We thank the reviewer for the interesting comment; we would like to point out that the ERα-positive MCF-7 cell line was used to study experimental overexpression of BKCa at an otherwise low baseline level. This does not imply that BKCa is expressed at lower levels in TNBC cell lines; in fact a recent study showed the opposite, i.e. overexpression of BKCa in TNBC patients (10.1186/s12885-020-07071-1). Consistent with our work, the authors conclude that the channel could even be a new strategy for development of a targeted therapy in TNBC. We also added this information in the discussion, lines 450-454 in the tracked changes version of the manuscript.

(4) The authors propose that combinatorial targeting of mitoBKCa along with known breast cancer chemotherapeutics can open a new horizon in breast cancer treatment. However, the authors did not perform any experiment to show the synergistic effect as mentioned.

As already stated in the public reviews, we feel that the manuscript is very condensed already, and that adding the respective experiments and data will make the flow of the study even more complex. For the moment, we removed all information and statements linking mitoBKCa with anti-cancer treatment strategies from the abstract and only discuss this aspect. We hope that the reviewer agrees with us that an extensive analysis of the functional mitoBKCa status in the context of established breast cancer therapies must be addressed by (our) future studies.

Minor Comments:

There are several typos and grammatical errors that need further attention and rephrasing.

We thank the reviewer for the comment and revised the text accordingly.

Author response:

The following is the authors’ response to the original reviews

We thank the reviewers for their careful and positive assessment of our manuscript. Maybe our findings are best summarized in the model below, showing that KDM5 inhibition/loss mediates a viral mimicry and DNA damage response through the generation of R-loops in genomic repeats. This is a different mechanism from the more well studied double-stranded RNA-induced “viral mimicry” response. Our studies also suggest that KDM5 inhibition may have a larger therapeutic window than STING agonists, since KDM5 inhibition seemingly does not induce “viral mimicry” in normal breast epithelial cells. 

Author response image 1.

Model of viral mimicry activation. De-repression of repetitive elements may trigger dsRNA formation, which activates the RIG-1/MDA5 pathway, as well as PKR. Alternatively, derepression of these elements may induce transcription replication conflicts (TRCs), resulting in R-loop formation. R-loops can lead to DNA damage, and/or activate the cGAS/STING pathway. Both the MAVS pathway and the cGAS/STING pathway converge to activate type I interferon (IFN) responses, resulting in decreased cell fitness and/or increased immunogenicity.

We do agree with the assessment that the study would be strengthened by in vivo studies. However, there are 4 different isoforms of KDM5 (3 in females), and existing KDM5specific inhibitors do not have adequate PK/PD properties for in vivo studies. We would also like to note that most mouse studies have not been proven to accurately predict immunotherapy responses in patients. Future studies in ex vivo tumor models would strengthen the clinical relevance of these studies. In the interim, we have added some normal macrophage studies in Figure S5 and an example of studies in normal T-cells below. Such studies will also be important to ensure that future KDM5 inhibitors do not have adverse effects on the immune system. Here, we observe that KDM5 inhibition appears to have neutral or slightly reduced T cell viability with KDM5 inhibition (Author response image 2a). However, KDM5 inhibition also results in increased CD107a expression in T-cells, indicative of a more cytotoxic phenotype (Author response image 2b). These studies suggest that KDM5 inhibitors do not have significant adverse effects on T cells or macrophages (figure S5) in the normal immune environment.

Author response image 2.

KDM5 inhibition does not have significant adverse effects on T-cells. a) Fold change proliferation of T-cells from 2 different human donors (left and right panels on graph) activated with 0.25ug/ml CD3 and treated with the indicated concentrations of C48 or a positive control (CBLB) compared to vehicle controls. b. FACS plots and histograms of CD107a surface expression (x-axis) versus forward scatter (FSC, y-axis) of T-cells from 2 different humans donors activated with 0.25ug/ml or 0.5mug/ml CD3 and treated with the indicated concentrations of C48.

Specific comments and answers to Reviewer #1:

We have added some additional analysis of data from other breast cancer cell lines to strengthen our points (Figure S2f, Figure S3e, Figure S4g-h, k.) We have also uploaded all the data to Geo with the following accession numbers :

GSE296387: H3K4me3 CUT-and-Tag data

GSE296584: S9.6 CUT-and-Tag data

GSE296974: RNA-sequencing data

Responses to Reviewer #1 (Recommendations for the authors):

(1) We have not conducted genomic studies comparing KDM5 expression to retroelement activation status in the tumor data sets but recognize that this is important for future studies. Again, there are several KDM5 isoforms and looking at repeat expression in these larger data sets is complex. We have added some data correlating KDM5 expression with ISG signatures in Figure S3j-l as well as in the graph below (Author response image 3). The correlation with ISG and AP signatures is modest, but strongest for KDM5B and C in breast cancer data sets, consistent with our disruption data for these 2 isoforms. As mentioned above, we do agree that future studies of KDM5s along with a broader analysis of other epigenetic modifying enzymes over repeats in various cancer types will shed light on the role of histone modifying enzymes in suppressing “viral mimicry” in tumors.

Author response image 3.

Correlation between gene expression and IFN gene set GSVA scores in breast cancer cell lines. a) Pearson correlation score between gene expression and IFN signature (ISG) gene set variation analysis (GSVA) scores in breast cancer cell lines as reported in DepMap. Higher ranks indicate an inverse correlation between expression of the individual gene and the expression of the ISG gene set. Correlation ranks for KDM5A, B and C are highlighted. b) as in a), but comparing gene expression to antigen presentation (AP) GSVA scores.

(2) We apologize for the mislabeling in figure 2B – has been corrected in the revised version.

(3) We agree that blocking the cGAS/STING pathway, only partially rescues the ISREGFP and HLA-A, B, C phenotype in HCC1428 cells. We have added data (Figure S2f) showing that this rescue is stronger in MCF7 cells. It is possible that the MDA5/MAVS pathway may also contribute to activation of the Type I interferon response. However, we have data that MAVS plays a minor (if any) role in this context, as MAVS KO minimally decreases C48-induced ISRE-GFP activity and HLA-A, B, C surface expression in HCC1428 cells (added Figure S2g).

Furthermore, there is no significant increase in dsRNA observed (using J2 antibody as a readout in immunofluorescence experiments) with C48 treatment as compared to 5’-azacytidine treatment or ADAR K/O (data not included). However, we have not performed MAVS/PKR K/O experiments to completely rule out the involvement of the dsRNA sensing pathways.

(4) These experiments were performed in the operetta imaging system, rather than confocal imaging, and therefore we do not have such images. Quantification of RNaseH1-GFP in the whole cell is reported in the figure, as RNaseH1-GFP signal is increased in both the nucleus and the cytoplasm with C48 treatment. This is not unexpected, as our data suggest that R-loop formation occurs in repetitive regions of the genome that are de-repressed by KDM5 inhibition in the nucleus, and the RNA/DNA hybrids, generated from R-loops, may activate cGAS/STING pathway in the cytoplasm.

(5) Disruption of siXPF and siXPG is relatively toxic in itself. Complete knockouts in breast cancer cells were not viable and we partially knocked down XPF using siRNA instead. We do agree that these kinds of rescue studies need to be expanded upon in future studies, but they served as further proof of the conclusions presented here.

(6) We have provided all the data in Geo and alternative representations can be made.

(7) Unfortunately, CUT-and-Tag experiments were not performed in cells expressing siXPF and therefore we cannot provide this data. However, XPF has been previously shown to be responsible for excising R-loops from the genome, rendering them detectable by cGAS/STING in the cytoplasm (Crossley et al, 2022, referenced in the current MS). Therefore, while we demonstrate that XPF knockdown attenuates type I IFN pathway activation upon KDM5 inhibition, it may not necessarily reduce R-loop formation in retroelements; it may just prevent their excision and downstream cGAS/STING activation. We do agree that CUT-and-Tag experiments in cells treated with siXPF versus siControl will have to be performed in the future to test this hypothesis.

Responses to Reviewer #2 (Recommendations for the authors):

(1) We have modified the text as well as the figure legend to state that this is a simplistic representation of the pathway in normal cells. As stated in the introduction, these pathways can be modified in tumors. The data presented suggest that the dsRNA pathway can be activated in all breast cancer cell lines tested, whereas more variation is observed in the activation of the STING pathway.  

(2) The ADAR guides target ADAR 110 and p150 but not ADAR2. This has been clarified in the text.  

(3) The guides have been renamed in the figure as the reviewer suggests.  

(4) It has been shown by others that KDM5 can occupy the STING promoter (https://pubmed.ncbi.nlm.nih.gov/30080846/); which supports the reviewer’s suggestion that STING upregulation in HMECs may be due to increased H3K4me3 at the STING gene. However, we argue that STING upregulation is not sufficient to activate “viral mimicry” due to the absence of “tumor-specific R-loops” (due to an increase in TRC in tumor cells) in normal cells. It is interesting to note that the S9.6 signal in subtelomeric regions is increased in HMECS similar to what is observed in tumor cells. However, the S9.6 signal over other repeats is not (Author response image 4), suggesting that C48-induced increases over non-telomeric repeats are tumor specific. This suggests that the tumor-specific increases in R-loop formation, which lead to “viral mimicry” activation, are not driven by those formed in subtelomeric regions. Future studies will have to expand on these findings.

Author response image 4.

Percent of S9.6 reads that align to repetitive genome in HMEC cells. (a) % of total aligned S9.6 reads that map to subtelomeric region in HMEC cells treated with DMSO or 2.5 μM C48. (b) % of total aligned S9.6 reads that map to repetitive elements in general in HMEC cells treated as in a).

(5) Clarity on R-loop quantification has been added to the figure legend as well as in the Materials and Methods section. Mean fluorescence intensity in the whole cell (this includes both nuclear and cytoplasmic signals) was quantified together and normalized to the number of DAPI-stained nuclei per well. As mentioned above all quantified in the Operetta imaging system.

(6) We have added some data that shows that increases in H3K4me3 is observed in and around ISGs upon KDM5 inhibition (Figure S4f). However, without time course experiments it is difficult to assess whether these are direct effects of the KDM5 inhibitor or indirect effects from activation of Type I IFN (similarly to what has previously been reported with 5’-azacytidine induction of “viral mimicry”, https://pubmed.ncbi.nlm.nih.gov/26317465/).

(7) We have previously included data showing that S9.6 reads in repeats that do not display C48-mediated increases in H3K4me3 also do not increase with C48 treatment (this is now Figure S4o). In addition, we have added some data showing that repeats with increased H3K4me3 and repeats with increased transcription upon C48 treatment also have increased S9.6 reads. Repeats that display both increases in H3K4me3 and mRNA expression have even greater increases in S9.6 signal compared to repeats that have increases in either one (Figure S4m-n). Taken together, this data suggest that KDM5 inhibition increases H3K4me3 in repeats, thereby allowing for their transcription, which can increase the probability of Transcription replication conflicts (TRC) and R-loop formation at such loci.

(8) As mentioned earlier in this response, while we observe increased S9.6 reads in subtelomeric regions of HCC1428 cells upon KDM5 inhibition, we also observe this in normal HMEC cells. Since KDM5 inhibition does not induce viral mimicry in HMEC cells, this suggests that R-loops formed in subtelomeric regions do not dictate the response observed with C48 treatment in breast cancer cells.

We hope that these answers to the reviewers comments as well as the additional data provided strengthens our findings.

Author response:

The following is the authors’ response to the original reviews

Summary of Revisions

We sincerely thank the editors and reviewers for their thorough assessment and constructive feedback, which has greatly improved our manuscript. We have carefully addressed all concerns as summarized below:

In response to the requests made by Reviewer #1:

• Clarified task design and acknowledged its limitations regarding endpoint accuracy control.

• Included analysis comparing the effects of cerebellar block on within-trial versus inter-trial movements.

• Clearly defined target groupings, replacing the term “single-joint” with “movements with low coupling torques” and “multi-joint” with “movements with high coupling torques”: definitions which are now supported by a supplementary material describing the net torque data as a function of the targets.

• Added detailed descriptions of trial success criteria, based on timing, and positional constraints.

• Expanded figures illustrating the effect of the cerebellar block on movement decomposition and variability in joint space and across different target directions.

In response to the requests made by Reviewer #2:

• Included an explicit discussion highlighting why the acute reduction in muscle torque during cerebellar block is likely due to agonist weakness rather than cocontraction, emphasizing the rationale behind our torque-centric analysis.

• Clearly defined trial success criteria and included the timing and accuracy constraints used in our study.

• Clarified our rationale for grouping targets based on shoulder flexion/extension, clearly justified by interaction torque analysis.

• Revised the caption and legend of Figure 3d for clarity and included partial correlation results to account for the variability across monkeys for the analysis of reduction in hand velocity vs. coupling torque in control. 

In response to the requests made by Reviewer #3:

• Included electrophysiological validation of the accuracy of targeting the superior cerebellar peduncle from one of the monkeys used in the experiment.

• Provided new analyses comparing movement decomposition and variability between slower and faster movements within the cerebellar block condition.

• Revised manuscript text to clarify terminology and clearly explained the rationale behind target groupings and torque analyses.

• Expanded discussion sections to better explain the relationships between timing deficits, movement decomposition, trajectory variability, and faulty motor commands.

• Clarified methodological choices regarding our analysis timeframe and acknowledged limitations related to the distinction between feedforward and feedback control.

Reviewer #1 (Public review): 

Summary:

In a previous work, Prut and colleagues had shown that during reaching, high-frequency stimulation of the cerebellar outputs resulted in reduced reach velocity. Moreover, they showed that the stimulation produced reaches that deviated from a straight line, with the shoulder and elbow movements becoming less coordinated. In this report, they extend their previous work by the addition of modeling results that investigate the relationship between the kinematic changes and torques produced at the joints. The results show that the slowing is not due to reductions in interaction torques alone, as the reductions in velocity occur even for movements that are single joints. More interestingly, the experiment revealed evidence for the decomposition of the reaching movement, as well as an increase in the variance of the trajectory.

Strengths:

This is a rare experiment in a non-human primate that assessed the importance of cerebellar input to the motor cortex during reaching.

We thank the reviewer for their positive feedback on our study. We particularly appreciate their recognition of the novelty and importance of our experimental approach in non-human primates, as well as their insightful summary of our key findings.

Weaknesses:

My major concerns are described below.

If I understand the task design correctly, the monkeys did not need to stop their hand at the target. I think this design may be suboptimal for investigating the role of the cerebellum in control of reaching because a number of earlier works have found that the cerebellum's contributions are particularly significant as the movement ends, i.e., stopping at the target. For example, in mice, interposed nucleus neurons tend to be most active near the end of the reach that requires extension, and their activation produces flexion forces during the reach (Becker and Person 2019). Indeed, the inactivation of interposed neurons that project to the thalamus results in overshooting of reaching movements (Low et al. 2018). Recent work has also found that many Purkinje cells show a burst-pause pattern as the reach nears its endpoint, and stimulation of the mossy fibers tends to disrupt endpoint control (Calame et al. 2023). Thus, the fact that the current paper has no data regarding endpoint control of the reach is puzzling to me.

We appreciate the reviewer’s point that cerebellar contributions can be particularly critical near the endpoint of a reach. In our task design, monkeys were indeed required to hold at the target briefly—100 ms for Monkeys S and P, and 150 ms for Monkeys C and M—before receiving the reward. However,  given the size of the targets and the velocity of movements, it often happened that the monkeys didn’t have to stop their movements fully to obtain the reward. Importantly, we relaxed the task’s requirements (by increasing the target size and reducing the temporal constraints) to enable the monkeys to perform a sufficient number of successful trials under both the control and the cerebellar block conditions. This was necessary as we found that strict criteria regarding these parameters yielded a very low success rate in the cerebellar block condition. Nevertheless, as we appreciate now, this task design is suboptimal for studying endpoint accuracy which is an important aspect of cerebellar control. In the methods section of our revised manuscript, we have clarified this aspect of the task design and acknowledged that it is sub-optimal for examining the role of the cerebellum in end-point control (lines 475-485). The task design of our future studies will explicitly address this point more carefully.

Because stimulation continued after the cursor had crossed the target, it is interesting to ask whether this disruption had any effects on the movements that were task-irrelevant. The reason for asking this is because we have found that whereas during task-relevant eye or tongue movements the Purkinje cells are strongly modulated, the modulations are much more muted when similar movements are performed but are task-irrelevant (Pi et al., PNAS 2024; Hage et al. Biorxiv 2024). Thus, it is interesting to ask whether the effects of stimulation were global and affected all movements, or were the effects primarily concerned with the task-relevant movements.

This is an insightful suggestion. The behavioral task in the present study was designed with a focus on task-relevant, reward-associated reaching movements. Nevertheless, we also have data on the inter-trial movements (e.g., return-to-center reaches) under continued cerebellar stimulation, which were not directly associated with reward. In response to the reviewer’s comment, we compared the effects of cerebellar block on endpoint velocities between these two types of movements. We found that reductions in peak hand velocity during inter-trial movements were significantly smaller than those observed during the target directed reaches. We have updated the Results section of our manuscript (lines 125-137) and expanded our supplementary document (Supplementary Figure S1) to include this analysis. 

If the schematic in Figure 1 is accurate, it is difficult for me to see how any of the reaching movements can be termed single joint. In the paper, T1 is labeled as a single joint, and T2T4 are labeled as dual-joint. The authors should provide data to justify this.

The reviewer is correct. Movements to all targets involved both shoulder and elbow joints, but the degree to which each joint participated varied in a targetspecific manner. In our original manuscript, we used the term “single-joint” to refer to movements in which one joint was nearly stationary, resulting in minimal coupling torque at the adjacent joint. Specifically, for Targets 1 and 5, the net torque—and thus acceleration— at the elbow was negligible, causing the shoulder to experience low coupling torques (as illustrated in Figure 3c of our revised manuscript). Following this comment and  to avoid confusion, we have now explained this explicitly in the revised manuscript (lines 178-187). This is supported by Supplementary Figure S2 demonstrating the net torques at the shoulder and elbow for movements to each target. We have also replaced the term ‘singlejoint movements’  and ‘multi-joint movements’  with  ‘movements with low coupling torques’ and ‘movements with high coupling torques’ respectively in our revised manuscript (lines 178-180, 204-207, 225-227, 230-232, 305-307, and 362-365).  

Because at least part of this work was previously analyzed and published, information should be provided regarding which data are new.

While some of the same animals and stimulation protocol were presented in prior work, the inverse-dynamics modeling, the analyses exploring progressive velocity changes across trials under a cerebellar block, and the relationship of motor noise to movement velocity are newly reported in this manuscript. We have included a clear statement in the Methods section specifying which components of the dataset and analyses are entirely new (lines 582-589).

Reviewer #1 (Recommendations for the authors):

(1) Before the results are presented, it is useful to present the experimental paradigm in more detail. For example, after the center-out movement was completed, was the monkey required to hold at the target location? How did the next trial begin (re-centering movement)? Next, specify the stimulation protocol, noting that each session was divided into 3-4 blocks of stimulation and not stimulation, with each block 50-80 trials.

We have updated the results section of our revised manuscript (lines 91-104) to present the experimental paradigm in more detail according to the reviewer’s advice.

(2) Figure 1. Hand velocity does not show how the reach was completed. Did the subjects stop at the target or simply shoot through it and turn around without stopping? Why are the traces cut off?

Monkeys were indeed required to hold at the target briefly (100-150 ms) before receiving the reward. However,  given the size of the targets and the velocity of movements, it often happened that the monkeys didn’t have to stop their movements fully to obtain the reward. The hand velocity profile shown in Figure 1b and the torque profiles shown in Figures 2a and 2b correspond to the period from movement onset to the entry of the control cursor into the peripheral target which marked the end of the movement for the trial. Since the monkeys didn’t have to stop their movements fully for the trial to end, the traces appear cut off at the beginning of the deceleration/stopping phase of the movement. We have updated the captions of Figures 1b, 2a, and 2b to include this information (lines 869-872 and 882-884).  

(3) Maybe state that the data regarding reaction times are not presented because of the task design in which the go signal was predictable.

In monkeys M and C, the timing of the go signal was fixed and therefore predictable. Furthermore, they were also allowed a grace period of 200 ms before the go signal to facilitate predictive timing which often resulted in negative reaction times. However, in Monkeys S and P, the go signal was variable in timing and the monkeys were not allowed to initiate the movements before the go signal. In our previous studies (Nashef et al., 2019; Israely et al. 2025), we reported increased reaction times under cerebellar block. However, since the present study focuses specifically on execution-related motor deficits, we did not analyze reaction time data. 

(4) Please provide the data and analysis regarding the entire reach, including the period after the cursor crosses the target and returns to the center position.

We compared the peak hand velocity of the target-directed movements to the inter-trial return-to-center movements. Cerebellar block produced significantly smaller reductions in peak hand velocity during inter-trial movements compared to within-trial reaches. The results section of our revised manuscript (lines 125137) and the supplementary material (Supplementary Figure S1) have been updated accordingly. While the behavioral task in the present study was designed with a focus on task-relevant, reward-associated reaching movements, it will be interesting to examine in detail the effect of cerebellar block on spontaneous movements in a future study.

(5) Figure 5. To illustrate the decomposition of multijoint movements into a sequence of single joint movements, I suggest plotting movements in joint space (in addition to Cartesian space as you have done now). The results in Figure 5 are most interesting and thus should be expanded. Please provide this data using the format in Figure 1C, that is, as a function of direction.

Following the reviewer’s suggestion, we have plotted sample trajectories in joint-velocity (Supplementary Figures 3a and b) and position space (Supplementary Figures 4a and b) to highlight the decomposition of multi-joint movements and increased inter-trial trajectory variability respectively during the cerebellar block. Additionally, we also analyzed movement decomposition and trajectory variability as a function of target direction (Supplementary Figures 3c and 4c respectively). The corresponding text in the Results section has been updated accordingly (lines 256-261, 267-271, 277-278 and 280-288).

Reviewer #2 (Public review):

This manuscript asks an interesting and important question: what part of 'cerebellar' motor dysfunction is an acute control problem vs a compensatory strategy to the acute control issue? The authors use a cerebellar 'blockade' protocol, consisting of high-frequency stimuli applied to the cerebellar peduncle which is thought to interfere with outflow signals. This protocol was applied in monkeys performing center outreaching movements and has been published from this laboratory in several preceding studies. I found the takehome-message broadly convincing and clarifying - that cerebellar block reduces muscle activation acutely particularly in movements that involve multiple joints and therefore invoke interaction torques, and that movements progressively slow down to in effect 'compensate' for these acute tone deficits. The manuscript was generally well written, and the data was clear, convincing, and novel. My comments below highlight suggestions to improve clarity and sharpen some arguments.

We thank the reviewer for their thoughtful and constructive feedback. We are grateful for their recognition of the significance of our findings regarding acute and compensatory motor responses following a cerebellar block.

Primary comments:

(1) Torque vs. tone: Is it known whether this type of cerebellar blockade is reducing muscle tone or inducing any type of acute co-contraction that could influence limb velocity through mechanisms different than 'atonia'? If so, the authors should discuss this information in the discussion section starting around line 336, and clarify that this motivates (if it does) the focus on 'torques' rather than muscle activation. Relatedly, besides the fact that there are joints involved, is there a reason there is so much emphasis on torque per se? If the muscle is deprived of sufficient drive, it would seem that it would be more straightforward to conceptualize the deficit as one of insufficient timed drive to a set of muscles than joint force. Some text better contextualizing the choices made here would be sufficient to address this concern. I found statements like those in the introduction "hand velocity was low initially, reflecting a primary muscle torque deficit" to be lacking in substance. Either that statement is self-evident or the alternative was not made clear. Finally, emphasize that it is a loss of self-generated torque at the shoulder that accounts for the velocity deficits. At times the phrasing makes it seem that there is a loss of some kind of passive torque.

We appreciate the reviewer's emphasis on distinguishing between reduced muscle tone and altered co-contraction patterns as potential explanations for decreased limb velocity. Our focus on torques per se arises from previous studies suggesting that a core deficit in cerebellar ataxia is impaired prediction of passive coupling torques (Bastian et al., 1996). In our study, we demonstrate that motor deficits in cerebellar ataxia result in fact from both the inability to compensate for passive coupling torques and an acute insufficiency in the ability to generate active muscle torques.

The muscle torque, representing the sum of all muscle forces acting at a joint, can indeed be reduced by any of the two mechanisms: (i) co-contraction of agonist and antagonist muscles, and/or (ii) insufficient agonist muscle activity (i.e., agonist weakness). In cerebellar ataxia, co-contraction has been proposed as a simplifying strategy to stabilize stationary joints during decomposed multi-joint movements (Bastian et al., 1996). In our experiments, this strategy would likely emerge gradually following cerebellar block similar to the adaptive slowing of movements aimed at reducing inter-joint interactions. However, we found that irrespective of the magnitude of coupling torques involved, reduction in the velocity of movements also occurred immediately following cerebellar block—a pattern less consistent with gradually emerging compensatory strategies. We therefore argue that this acute onset of movement slowing was mainly driven by agonist weakness. Our argument is further supported by previous studies which attributed reduced agonist muscle activity as a cause for the slowing of voluntary movements in individuals with cerebellar lesions (Hallet et al. 1991; Wild et al., 1996). Additionally, early studies have also reported muscle weakness (asthenia) and hypotonia acutely following cerebellar injury in humans (Haines et al., 2007) and experimental lesions in animals (Luciani, 1893; Bremer et al., 1935; Fulton & Dow, 1937; Granit et al., 1955).

We have modified the discussion section of our revised manuscript (lines 366-376) to explain/clarify this. Additionally, we have also underscored that the observed velocity deficits primarily reflect a reduction of self-generated torque at the shoulder (whether acute or adaptive), rather than any reduction in passive torque (lines 350-352).

(2) Please clarify some of the experimental metrics: Ln 94 RESULTS. The success rate is used as a primary behavioral readout, but what constitutes success is not clearly defined in the methods. In addition to providing a clear definition in the methods section, it would also be helpful for the authors to provide a brief list of criteria used to determine a 'successful' movement in the results section before the behavioral consequences of stimulation are described. In particular, the time and positional error requirements should be clear.

Successful trials were defined as trials in which monkeys didn’t leave the center position before the “Go” signal and entered the peripheral target within a permitted movement time. We have updated the results (lines 91-104) and methods (lines 475-485) section of our revised manuscript to include (i) the timing criteria of each phase of the trials and (ii) the size of the peripheral targets indicating the tolerance for endpoint accuracy.  

(3) Based on the polar plot in Figure 1c, it seemed odd to consider Targets 1-4 outward and 5-8 inward movements, when 1 and 5 are side-to-side. Is there a rationale for this grouping or might results be cleaner by cleanly segregating outward (targets 2-4) and inward (targets 6-8) movements? Indeed, by Figure 3 where interaction torques are measured, this grouping would seem to align with the hypothesis much more cleanly since it is with T2,T3,and T4 where clear coupling torques deficits are seen with cerebellar block.

We acknowledge the reviewer's observation regarding the classification of targets 1 and 5 as side-to-side movements rather than strictly "outward" or "inward." In the initial section of our results, we grouped the targets based on shoulder joint movements: "outward" targets involved shoulder flexion, while "inward" targets involved shoulder extension. This classification highlighted the more pronounced effect of cerebellar block on movements requiring shoulder flexion compared to those requiring shoulder extension. For subsequent analyses, we focused on the effects of cerebellar block on movements to "outward" targets, which included directions involving low (target 1) or high (targets 2–4) coupling torques. To clarify this aspect, we have revised our manuscript to explain our definition of "outward" (targets 1–4) and "inward" (targets 5–8) target groupings based on shoulder  flexion and extension movements respectively (lines 117-120).

(4) I did not follow Figure 3d. Both the figure axis labels and the description in the main text were difficult to follow. Furthermore, the color code per animal made me question whether the linear regression across the entire dataset was valid, or would be better performed within animal, and the regressions summarized across animals. The authors should look again at this section and figure.

We have revised the legend of Figure 3d to include a detailed explanation of how the value along each axis is computed  (lines 908-920 of the revised manuscript). Please note that  the color coding of the data points is as per the target number (T1-T4) and not the monkey number (as denoted in the figure legend). Also, pooling of data across monkeys was done after confirming that data from each animal expressed a similar trend. Specifically, the correlation coefficients were all positive but statistically significant in 3 out of the 4 monkeys. Following the reviewers’ feedback, we now performed  a partial correlation analysis (which controls for the variability across monkeys) and found a significant correlation (r = 0.32, p < 0.001) between reduction in peak hand velocities during cerebellar block and the net coupling torque impulse. We have updated the manuscript to include the result of the partial correlation analysis (lines 173-176).  

(5) Line 206+ The rationale for examining movement decomposition with a cerebellar block is presented as testing the role of the cerebellum in timing. Yet it is not spelled out what movement decomposition and trajectory variability have to do with motor timing per se.

The reviewer is right and the relations between timing, decomposition and variability need to be explicitly explained. In the results  section of our revised manuscript, we have explained how decomposed movements and trajectory variability may reflect impaired temporal coordination across multiple joints—a critical cerebellar function (lines 235-244).

Reviewer #2 (Recommendations for the authors):

(1) Rephrase the findings, starting Line 232. Here the authors state, "Next, we asked whether movement decomposition was mainly due to lower hand velocities. We therefore selected a subset of control trials that matched the cerebellar block trials in their peak velocity. However, even though movement decomposition in these control trials was higher compared to all control trials, it was still significantly lower than velocity matched cerebellar block trials." I suggest inverting the final sentence to: "Movement decomposition in control trials was significantly lower than velocity-matched cerebellar block trials, even though these control trials themselves had somewhat higher decomposition indices than all control trials together." A similar issue pops up with trajectory variability below that simply requires some editing to be less clunky.

Following the reviewer’s suggestion, we have revised the sentences related to movement decomposition and trajectory variability. These sentences now reads as follows: 

(lines 267-271 in the revised manuscript): “Movement decomposition in control trials was significantly lower than velocity-matched cerebellar block trials (p < 0.001; Figure 5c), even though these control trials themselves had 11.0% (CI [5.2, 17.0], p = 0.03) higher decomposition than the mean value calculated across all control trials.” 

(lines 280-288 in the revised manuscript): “ When we compared the subset of velocitymatched control and cerebellar block trials, we found that cerebellar block trials exhibited 34.6% (CI [26.2, 43.2], p < 0.001) higher trajectory variability (Figure 5e). Normally, slower movements are also less variable due to the speed-accuracy tradeoff (Plamondon and Alimi 1997). Indeed, the trajectory variability in this subset of slower control trials was 5.5% (CI [0.9, 9.9], p = 0.02) lower than that of all control trials. In other words, despite slower movements, cerebellar block led to increased trajectory variability.”

(2) Typo: Ln 73 sequences, not sequence.

Typo error was corrected (line 75 of revised manuscript). 

Reviewer #3 (Public review):

Summary:

In their manuscript, "Disentangling acute motor deficits and adaptive responses evoked by the loss of cerebellar output," Sinha and colleagues aim to identify distinct causes of motor impairments seen when perturbing cerebellar circuits. This goal is an important one, given the diversity of movement-related phenotypes in patients with cerebellar lesions or injuries, which are especially difficult to dissect given the chronic nature of the circuit damage. To address this goal, the authors use high-frequency stimulation (HFS) of the superior cerebellar peduncle in monkeys performing reaching movements. HFS provides an attractive approach for transiently disrupting cerebellar function previously published by this group. First, they found a reduction in hand velocities during reaching, which was more pronounced for outward versus inward movements. By modeling inverse dynamics, they find evidence that shoulder muscle torques are especially affected. Next, the authors examine the temporal evolution of movement phenotypes over successive blocks of HFS trials. Using this analysis, they find that in addition to the acute, specific effects on muscle torques in early HFS trials, there was an additional progressive reduction in velocity during later trials, which they interpret as an adaptive response to the inability to effectively compensate for interaction torques during cerebellar block. Finally, the authors examine movement decomposition and trajectory, finding that even when low-velocity reaches are matched to controls, HFS produces abnormally decomposed movements and higher than expected variability in trajectory.

Strengths:

Overall, this work provides important insight into how perturbation of cerebellar circuits can elicit diverse effects on movement across multiple timescales.

The HFS approach provides temporal resolution and enables analysis that would be hard to perform in the context of chronic lesions or slow pharmacological interventions. Thus, this study describes an important advance over prior methods of circuit disruption, and their approach can be used as a framework for future studies that delve deeper into how additional aspects of sensorimotor control are disrupted (e.g., response to limb perturbations).

In addition, the authors use well-designed behavioral approaches and analysis methods to distinguish immediate from longer-term adaptive effects of HFS on behavior. Moreover, inverse dynamics modeling provides important insight into how movements with different kinematics and muscle dynamics might be differentially disrupted by cerebellar perturbation.

We thank the reviewer for their detailed assessment and thoughtful comments and greatly appreciate their positive feedback.  

Weaknesses:

The argument that there are acute and adaptive effects to perturbing cerebellar circuits is compelling, but there seems to be a lost opportunity to leverage the fast and reversible nature of the perturbations to further test this idea and strengthen the interpretation. Specifically, the authors could have bolstered this argument by looking at the effects of terminating HFS - one might hypothesize that the acute impacts on muscle torques would quickly return to baseline in the absence of HFS, whereas the longer-term adaptive component would persist in the form of aftereffects during the 'washout' period. As is, the reversible nature of the perturbation seems underutilized in testing the authors' ideas.

We agree that our approach could more explicitly exploit the rapid reversibility of high-frequency stimulation (HFS) by examining post-stimulation ‘washout’ periods. However, for the present dataset, we ended the session after the set of cerebellar block trials without using an explicit washout period. We plan to study the effect of the cerebellar block on immediate post-block washout trials in the future.    

The analysis showing that there is a gradual reduction in velocity during what the authors call an adaptive phase is convincing. That said, the argument is made that this is due to difficulty in compensating for interaction torques. Even if the inward targets (i.e., targets 68) do not show a deficit during the acute phase, these targets still have significant interaction torques (Figure 3c). Given the interpretation of the data as presented, it is not clear why disruption of movement during the adaptive phase would not be seen for these targets as well since they also have large interaction torques. Moreover, it is difficult to delve into this issue in more detail, as the analyses in Figures 4 and 5 omit the inward targets.

The reviewer is right and  movements to Targets 6–8 (inward) were seemingly unaffected despite also involving significant interaction torques. Specifically, we noted that while outward targets (2–4) tend to involve higher coupling torque impulses on average, this alone does not fully explain the differential impact of cerebellar block, as illustrated by discrepancies at the individual target level (e.g., target 7 vs. target 1). We propose two possible explanations: (1) a bias toward shoulder flexion in the effect of cerebellar block—consistent with earlier studies showing ipsilateral flexor activation or tone changes following stimulation or lesioning of the deep cerebellar nuclei; and (2) posture-related facilitation of inward (shoulder extension) movements from the central starting position. This point is addressed in the Discussion section (lines 404-433  in the revised manuscript).

The text in the Introduction and in the prior work developing the HFS approach overstates the selectivity of the perturbations. First, there is an emphasis on signals transmitted to the neocortex. As the authors state several times in the Discussion, there are many subcortical targets of the cerebellar nuclei as well, and thus it is difficult to disentangle target-specific behavioral effects using this approach. Second, the superior cerebellar peduncle contains both cerebellar outputs and inputs (e.g., spinocerebellar). Therefore, the selectivity in perturbing cerebellar output feels overstated. Readers would benefit from a more agnostic claim that HFS affects cerebellar communication with the rest of the nervous system, which would not affect the major findings of the study.

The reviewer is right that the superior cerebellar peduncle carries both descending and ascending fibers, and that cerebellar nuclei project to subcortical as well as cortical targets. Therefore, we cannot rule out the fact that the effect of HFS  may be mediated in part through pathways other than the cerebello-thalamo-cortical pathway (as mentioned in the Discussion section). However, it is also important to note that in primates the cerebellar-thalamo-cortical (CTC) pathway greatly expanded (at the expense of the cerbello-rubro-spinal tract) in mediating cerebellar control of voluntary movements (Horne and Butler, 1995). The cerebello-subcortical pathways diminished in importance over the course of evolution (Nathan and Smith, 1982, Padel et al., 1981, ten Donkelaar, 1988). Previously we found that the ascending spinocerebellar axons which enter the cerebellum through the superior cerebellar peduncle (SCP) are weakly task-related and the descending system is quite small (Cohen et al, 2017). We have clarified these points and acknowledged that HFS disrupts cerebellar communication broadly, rather than solely the cerebellothalamo-cortical pathway in the methods section of our revised manuscript (lines 531544).  

The text implies that increased movement decomposition and variability must be due to noise. However, this assumption is not tested. It is possible that the impairments observed are caused by disrupted commands, independent of whether these command signals are noisy. In other words, commands could be low noise but still faulty.

We recognize the reviewer’s concern about linking movement decomposition and trial-to-trial trajectory variability with motor noise. We interpret these motor abnormalities as a form of motor noise in the sense that they are generated by faulty motor commands. We draw our interpretation from the findings of previous research work which show that the cerebellum aids in the state estimation of the limb and subsequent generation of accurate feedforward commands. Therefore, disruption of the cerebellar output may lead to faulty motor commands resulting in the observed asynchronous joint activations (i.e., movement decomposition) and unpredictable trajectories (i.e., increased trial-to-trial variability). Both observed deficits resemble increased motor noise. This point is presented in our Discussion section (lines 436-458 of the revised manuscript),

Throughout the text, the use of the term 'feedforward control' seems unnecessary. To dig into the feedforward component of the deficit, the authors could quantify the trajectory errors only at the earliest time points (e.g., in Figure 5d), but even with this analysis, it is difficult to disentangle feedforward- and feedback-mediated effects when deficits are seen throughout the reach. While outside the scope of this study, it would be interesting to explore how feedback responses to limb perturbation are affected in control versus HFS conditions. However, as is, these questions are not explored, and the claim of impaired feedforward control feels overstated.

We agree that to strictly focus on feedforward control, we could have examined the measured variables in the first 50-100 ms of the movement which has been shown to be unaffected by feedback responses (Pruszynski et al. 2008, Todorov and Jordan 2002,  Pruszynski  and Scott 2012, Crevecoeur  et al. 2013). However, in our task, the amplitude of movements made by the monkeys was small, and therefore the response measures in the first 50-100 ms were too small for a robust estimation. Also, fixing a time window led to an unfair comparison between control and cerebellar block trials, in which velocity was significantly reduced and therefore movement time was longer.  Therefore, we used the peak velocity, torque impulse at the peak velocity, and maximum deviation of the hand trajectory as response measures. We have acknowledged this point in the methods section of our revised manuscript (lines 590-600). We have also refrained from using the term feedforward control throughout the text of our revised manuscript as suggested by the reviewer.

The terminology 'single-joint' movement is a bit confusing. At a minimum, it would be nice to show kinematics during different target reaches to demonstrate that certain targets are indeed single joint movements. More of an issue, however, is that it seems like these are not actually 'single-joint' movements. For example, Figure 2c shows that target 1 exhibits high elbow and shoulder torques, but in the text, T1 is described as a 'single-joint' reach (e.g. lines 155-156). The point that I think the authors are making is that these targets have low interaction torques. If that is the case, the terminology should be changed or clarified to avoid confusion.

Indeed, as reviewer #1 also noted, movements to targets 1 and 5 are not purely single-joint but rather have relatively low coupling torques. Movements to all targets involved both shoulder and elbow joints, but the degree to which each joint participated varied in a target-specific manner. In our original manuscript, we used the term “single-joint” to refer to movements in which one joint was largely stationary, resulting in minimal coupling torque at the adjacent joint. Specifically, for Targets 1 and 5, the net torque—and thus acceleration—at the elbow was negligible, causing the shoulder to experience low coupling torques (as illustrated in Figure 3c of our revised manuscript). Following this comment and  to avoid confusion, we have now explained this explicitly in the revised manuscript (lines 178-187). This is supported by Supplementary Figure S2 demonstrating the net torques at the shoulder and elbow for movements to each target. We have also replaced the term ‘single-joint movements’  and ‘multi-joint movements’  with  ‘movements with low coupling torques’ and ‘movements with high coupling torques’ respectively in our revised manuscript (lines 178-180, 204-207, 225-227, 230-232, 305-307, and 362-365).

The labels in Figure 3d are confusing and could use more explanation in the figure legend. In Figure 3d, it is stated that data from all monkeys is pooled. However, if there is a systematic bias between animals, this could generate spurious correlations. Were correlations also calculated for each animal separately to confirm the same trend between velocity and coupling torques holds for each animal?

We have revised the legend of Figure 3d to include a detailed explanation of how the values along each axis are computed  (lines 908-920 of the revised manuscript). Please note that the pooling of data across monkeys was done after confirming that data from each animal expressed a similar trend. Specifically, the correlation coefficients were all positive but statistically significant in 3 out of the 4 monkeys. Moreover, following the reviewers’ feedback, we also did a partial correlation analysis (which controls for the variability across monkeys) and found a significant correlation (r = 0.32, p < 0.001) between reduction in peak hand velocities during cerebellar block and the net coupling torque impulse. We have updated the manuscript to include the result of the partial correlation analysis (lines 173-176).  

In Table S1, it would be nice to see target-specific success rates. The data would suggest that targets with the highest interaction torques will have the largest reduction in success rates, especially during later HFS trials. Is this the case?

The breakdown of the percentage increase in failure rate due to cerebellar block as a function of target direction is shown in Author response image 1 inserted to this response. 

Author response image 1.

Effect of cerebellar block on failure rate. The change in failure rate for the cerebellar block trials was computed relative to the control trials per session per target. The depicted values are the mean ± 95% confidence intervals across all sessions pooled from all four monkeys. The individual means of each monkey are overlaid. Statistical significance is denoted as follows: p ≥ 0.05NS, p < 0.05*, p < 0.01**, p < 0.001*** [T1-8: Targets 1-8]

The increase in failure rate due to cerebellar block was not affected by the target direction (linear mixed model analysis,  target x trial-type interaction effect: p  = 0.44).  However, it should be noted that success/failure depends on several factors beyond just the execution related impaired limb dynamics. In a previous study (Nashef et al. 2019) we identified several causes of failure such as (i) not entering the central target in time, (ii) premature exit from the central target before the ‘go’ signal,  (iii) reaction time longer than the time permitted to reach the peripheral target after the ‘go’ signal, or (iv) not holding at the peripheral target for the required time at the end of the movement.   

Reviewer #3 (Recommendations for the authors):

(1) It would be helpful to provide some supplemental information on electrophysiological validation of the targeting in each monkey. Was any variability in targeting observed (e.g., some targeting was more effective at eliciting cortical responses)? If so, does targeting variability relate to any of the variability in behavioral effects of HFS across monkeys?

Although we currently do not have an exact measure of the proportion of fibers blocked by HFS, our targeting approach consistently elicited robust cortical responses across monkeys. Specifically, we implanted the stimulating electrode at the location that produced the maximum peak-to-peak evoked responses in the primary motor cortex. Author response image 2 in this response demonstrates that even a slight deviation (~0.5 mm) from this optimal site reduced these responses substantially.:

Author response image 2.

Evoked responses in the primary motor cortex as a function of the location of the stimulation site. [LEFT] Coronal T2-weighted MRI showing the planned trajectory to target the superior cerebellar peduncle (location marked by the tip of the arrowhead) through a round chamber suitably positioned over the skull. [RIGHT] Evoked multi-unit (300-7500 Hz) responses from one of the recording electrodes in the primary motor cortex are used to guide the stimulating electrode to the correct implant site. As the stimulating electrode was lowered deeper, maximum peak-to-peak evoked responses were obtained at a depth of 32.5 mm relative to the cortical surface. This was chosen as the implant site. Elevating or lowering the electrode by ~0.5 mm from this depth reduced the peak-to-peak response amplitude. 

(2) The emphasis in the Introduction that HFS provides direct insight into deficits seen in patients with cerebellar disease or injury is a bit overstated. Patients have very diverse etiologies, only a modest number of which might be faithfully mimicked by SCP HFS. I would suggest some text acknowledging that this is only a limited model for cerebellar disease or injury.

We agree with the reviewer that the high-frequency stimulation of the superior cerebellar peduncle provides a limited model that does not fully replicate the diverse pathologies seen in cerebellar disease or injury. In fact, in the introduction section (lines 53-59 of our revised manuscript) we have mentioned that the discrepancy in the conclusions of various clinical studies may reflect the heterogeneity of the individuals with cerebellar lesions who often have differences in lesion etiology and associated damage beyond the cerebellum itself. While this may preclude the generalization of our findings to the wider clinical population per se, our approach offers a precise and controlled method to investigate the immediate and adaptive changes in motor behavior following the disruption of cerebellar signals.

(3) Do animals with HFS show less decomposition and trajectory variability in their slower movements when compared to their faster movements? Comparisons are only made with velocity-matched control blocks, but the comparison of slower vs. faster reaches during HFS blocks would also be informative.

To answer this point we classified movements during cerebellar block as either slow or fast based on the median peak hand velocity of the cerebellar block trials per target per session. We then computed the decomposition index and trajectory variability for the fast and slow movements during cerebellar block relative to control in the same way as in Figure 5 of our manuscript (i.e., the percentage change relative to control). Our analysis revealed significantly lower movement decomposition (p < 0.001) and reduced trajectory variability (p < 0.001) for slower movements compared to faster ones within the cerebellar block condition (Author response image 3).

Author response image 3.

Effect of slow and fast movements during cerebellar block on movement decomposition and trajectory variability. [LEFT] Change in decomposition index (i.e., the proportion of the movement time during which the movement was decomposed) for slow and fast cerebellar block trials relative to all control trials. The change in median decomposition was computed per session per target and then averaged across all eight targets to arrive at one value per session. The depicted values are the mean ± 95% confidence intervals across all sessions pooled from all four monkeys. The individual means of each monkey are overlaid. [RIGHT] Change in inter-trial trajectory variability for slow and fast cerebellar block trials relative to all control trials. The trajectory variability was measured as the standard deviation of the maximum perpendicular distance of the trajectories from the Y-axis after transforming them as in Figure 5d of the main text. The change in trajectory variability for the fast and slow cerebellar block trials was then computed per session per target and averaged across all eight targets to arrive at one value per session. The depicted values are the mean ± 95% confidence intervals across all sessions pooled from all four monkeys. The individual means of each monkey are overlaid. Statistical significance is denoted as follows: p ≥ 0.05NS, p < 0.05*, p < 0.01**, p < 0.001***. [Cbl: Cerebellar block].

(4) Line 220- 'velocity' should be 'speed' or 'absolute velocity'?

The term velocity was changed to speed in  the revised manuscript (line 255).

Author response:

The following is the authors’ response to the original reviews.

eLife assessment

This study presents a valuable finding on the mechanism to promote distant metastasis in breast cancer. The evidence supporting the claims of the authors is convincing. The work will be of interest to medical biologists working on breast cancer.

Public Reviews:

Reviewer #1 (Public Review):

Strengths

The paper has shown the expression of RGS10 is related to the molecular subtype, distant metastasis, and survival status of breast cancer. The study utilizes bioinformatic analyses, human tissue samples, and in vitro and in vivo experiments which strengthen the data. RGS10 was validated to inhibit EMT through a novel mechanism dependent on LCN2 and miR-539-5p, thereby reducing cancer cell proliferation, colony formation, invasion, and migration. The study elaborated the function of RGS10 in influencing the prognosis and biological behavior which could be considered as a potential drug target in breast cancer.

Weakness

The mechanism by which the miR-539-5p/RGS10/LCN2 axis may be related to the prognosis of cancer patients still needs to be elucidated. In addition, the sample size used is relatively limited. Especially, if further exploration of the related pathways and mechanisms of LCN2 can be carried out by using organoid models, as well as the potential of RGS10 as a biomarker for further clinical translation to verify its therapeutic target effect, which will make the data more convincing.

Answer: Thank you for your comments and suggestions. In future research, we will utilize large clinical cohorts and organoid models to further explore relevant research mechanisms.

Reviewer #2 (Public Review):

Liu et al., by focusing on the regulation of G protein-signaling 10 (RGS10), reported that RGS10 expression was significantly lower in patients with breast cancer, compared with normal adjacent tissue. Genetic inhibition of RGS10 caused epithelial-mesenchymal transition, and enhanced cell proliferation, migration, and invasion, respectively. These results suggest an inhibitory role of RGS10 in tumor metastasis. Furthermore, bioinformatic analyses determined signaling cascades for RGS10-mediated breast cancer distant metastasis. More importantly, both in vitro and in vivo studies evidenced that alteration of RGS10 expression by modulating its upstream regulator miR-539-5p affects breast cancer metastasis. Altogether, these findings provide insight into the pathogenesis of breast tumors and hence identify potential therapeutic targets in breast cancer.

The conclusions of this study are mostly well supported by data. However, there is a weakness in the study that needs to be clarified.

In Figure 2A, although some references supported that SKBR3 and MCF-7 possess poorly aggressive and less invasive abilities, examining only RGS10 expression in those cells, it could not be concluded that 'RGS10 acts as a tumor suppressor in breast cancer'. It would be better to introduce a horizontal comparison of the invasive ability of these 3 types of cells using an invasion assay.

Answer: Thank you for your comments and suggestions. MDA-MB-231, SKBR3, and MCF-7 originate from triple-negative breast cancer (high invasiveness), Her-2 receptor overexpression (relatively weak invasiveness), and luminal type breast cancer (relatively weak invasiveness) separately. Previous studies have demonstrated the invasive ability of these 3 types of cells. (PMID: 34390568)

Reviewer #3 (Public Review):

Distant metastasis is the major cause of death in patients with breast cancer. In this manuscript, Liu et al. show that RGS10 deficiency elicits distant metastasis via epithelial-mesenchymal transition in breast cancer. As a prognostic indicator of breast cancer, RGS10 regulates the progress of breast cancer and affects tumor phenotypes such as epithelial-mesenchymal transformation, invasion, and migration. The conclusions of this paper are mostly well supported by data, but some analyses need to be clarified.

(1) Because diverse biomarkers have been identified for EMT, it is recommended to declare the advantages of using RGS10 as an EMT marker.

Answer: Thank you for your comments. The dysregulation of RGS protein expression has been observed to be associated with various types of cancer. (PMID: 26293348). Previous studies have shown that RGS10 knocking down can lead to chemotherapy resistance of ovarian cancer cells to paclitaxel, cisplatin, and vincristine. In colorectal tumors, the transcription of RGS10 is regulated by DNA methylation and histone deacetylation. As a key regulatory factor in the G protein signaling pathway, RGS 10 is involved in tumor development including survival, polarization, adhesion, chemotaxis, and differentiation, these hints suggest RGS10 might be a marker for EMT in breast cancer.

(2) The authors utilized databases to study the upstream regulatory mechanisms of RSG10. It is recommended to clarify why the authors focused on miRNAs rather than other epigenetic modifications.

Answer: Thank you for your comments. miRNAs are short-chain non-coding RNA molecules that bind to the target mRNA's 3 'untranslated region (3'UTR) to cause mRNA degradation or translation inhibition, thus regulating gene expression in cells. These small molecules play a crucial role in regulating the expression of cancer-related genes and can act as tumor promoters or tumor suppressors. To further improve the molecular mechanism of malignant biological behavior of breast cancer cells with RGS10, we verified that miR-539-5p might be the upstream regulation target of RGS10 through bioinformatics prediction and in-vitro experiments.

(3) The role of miR-539-5p in breast cancer has been described in previous studies. Hence, it is recommended to provide detailed elaboration on how miR-539-5p regulates the expression of RSG10.

Answer: Thank you for your comments. To verify the effect of miRNA-539-5p regulating the expression of RSG10, we transfected miR-539-5p mimic, miR-539-5p mimic NC, miR-539-5p inhibitor, miR-539-5p inhibitor NC in SKBR3 cells and MDA-MB-231 cells respectively, and verified the expression of RGS10 through RT-qPCR and Western blot experiments. The results showed that compared with the transfected miR-539-5p mimic NC or wild-type SKBR3 cells, RGS10 m RNA and protein levels were significantly reduced. On the contrary, after MDA-MB-231 cells were transfected with miR-539-5p inhibitor to inhibit the expression of miR-539-5p, RGS10 mRNA and protein levels in MDA-MB-231 cells were significantly increased (Fig. 3.4A-C, Fig. 3.5A-C). This indicates that miR-539-5p can target and regulate RGS10.

(4) To enhance the clarity and interpretability of the Western blot results, it would be advisable to mark the specific kilodalton (kDa) values of the proteins.

Answer: Thank you for your comments and suggestions. We have corrected to mark the specific kilodalton (kDa) values of the proteins in WB.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

The function of RGS10 in breast cancer was identified in the paper. However, some major issues in this paper need to be specified:

(1) From reading the introduction section and its references, RGS proteins participate in multiple essential cellular processes and may be tumor initiators or suppressors (Li et al., 2023). This article focuses on the significance of RGS10 in breast cancer, it is recommended to show how the function of RGS10 exhibits therapeutic significance in other types of cancer.

Answer: Thanks for your comments and suggestions on our findings. The dysregulation of RGS protein expression has been observed to be associated with various types of cancer. Especially in ovarian cancer cells. (PMID: 26293348). It has been found that the RGS10 expression is lower than that of normal ovarian cells. (PMID: 21044322). In addition, it has been found that knocking down RGS10 can enhance the vitality of ovarian cancer cells and promote chemoresistance by activating the Rheb GTP/mTOR signaling pathway. (PMID: 26319900). A study suggests that RGS10 mediates inflammation signaling regulation in SKOV-3 ovarian cancer cells with high expression of TNF and COX-2 after RGS10 knockdown. In colorectal tumors, RGS10 transcription is regulated by DNA methylation and histone deacetylation. (PMID: 35810565). RGS10 expression also are associated with poor prognosis in laryngeal cancer, hepatocellular carcinoma, and pediatric acute myeloid leukemia. (PMID: 32776811, PMID: 26516143, PMID: 30538250)

(2) The authors characterize RGS10 protein expression in the breast cancer cell lines MDA-MB-231, MCF7, and SKBR3 in vitro Figure 2A. However, more information would strengthen the data - e.g. information on the expression of RGS10 protein and the survival in public databases, as well as the correlation between RGS10 and Her-2 expression.

Answer: Thanks for your comments. we have checked the correlation of RGS10 expression and survival rate of Her-2 positive breast cancer patients in a public database. Although there is no significant difference in the “p” value, however, RGS10 high-expression patients have a favorable prognosis tendency than RGS10 low-expression patients after the 100th month.

Author response image 1.

(3) Regarding the current situation of clinical trials in the RGS family, the potential to develop RGS 10 for clinic translation is a driving factor for EMT.

Answer: Thank you for your comments. The RGS (G protein signal transduction regulator) gene family provides an important "braking" function for the cell receptor family of G-protein coupled receptors (GPCR). GPCR controls hundreds of important functions in systemic cells and is the largest class of drug targets, with over one-third of FDA approved drugs treating diseases by binding to GPCR and altering its activity. When GPCRs are activated by hormones or neurotransmitters, they initiate signaling cascades within host cells through signal-carrying proteins called G proteins. The function of the RGS protein is to inactivate the G protein, thereby shutting down this signaling cascade reaction, which limits G protein signal transduction and allows cells to reset and receive new incoming signals. If it were not for it, the signals triggered by GPCR would inappropriately remain on, and the signal transduction would experience dysfunction (PMID: 33007266). The potential to develop RGS10 as a driving factor of EMT is meaningful for clinic translation.

(4) In Figure 3A, the paper showed that differential gene expression revealed 70 genes were significantly upregulated in RGS10-depleted SKBR3 cells, The authors didn't show any data on the expression of other EMT-related proteins in pathway analysis.

Answer: Thank you for your comments. The enrichment analysis of RNA sequencing in RGS10-depleted SKBR3 cells suggests that high correlation factors that are associated with EMT, such as TAGLN, TNFSF10, NDUFA4L2, CCN5, PHGDH, ST3GAL5, ANG, and LCN2.

(5) In Figure 3B, the paper focuses on LCN2 in pathway analysis, however, the author did not elaborate on the significance of LCN2-related pathways in EMT.

Answer: Thank you for your comments. Some studies have the significance of LCN2-related pathways in EMT. It was confirmed that LCN2 upregulation triggered by PTEN insufficiency induces EMT to promote migration and invasion in MCF7 cells (PMID: 27466505). The activation of STAT3 contributes to an increase in LCN2 expression, which activates ERK pathway-dependent EMT, thus promoting lung metastasis in MDA-MB-231 cells in breast cancer (PMID: 33473115). The silencing of LCN2 reduced the ability of migration and invasion of SUM149 cells and the proportion of tumor stem cells, suggesting that LCN2 may mediate the invasion and metastasis of cancer cells by regulating the stemness of breast cancer cells. The biological effects of LCN2 small molecule inhibitors ZINC00640089 and ZINC00784494 targeting IBC cells have been confirmed. The siRNA-mediated silencing of LCN2 in IBC cells significantly reduces cell proliferation, viability, migration, and invasion. (PMID: 34445288).

(6) Minor: the author did not conduct a semi-quantitative analysis of the immunohistochemical results of RGS10.

Answer: Thank you for your suggestion. We would like to demonstrate the qualitative analysis of RGS10 immunohistochemistry. The semi-quantitative analysis is not required in the paper.

Reviewer #2 (Recommendations For The Authors):

The role of RGS10 was well-characterized in this study, However, some minor points need to be modified.

(1) Page 15 line 296, description of cell proliferation was missing, please modify.

Answer: Thank you for your comments. We have corrected the description of cell proliferation on Page 15 highlighted in red.

(2) In Figure 2C, the title of the Y-axis was missing.

Answer: Thank you for your comments. We have corrected the description of the Y-axis title in Figure 2C.

(3) Describe the transfection reagent that was used in this study, and incorporated into the methods section.

Answer: Thank you for your comments. We have added the description of the transfection reagent to the methods section.

(4) The manuscript needs proofreading.

Answer: Thank you for your comments. We have proofread the manuscript.

Author Response

The following is the authors’ response to the original reviews.

eLife assessment

This fundamental study provides compelling evidence to explain how chemical variations within a set of kinase inhibitors drive the selection of specific Erk2 conformations. Conformational selection plays a critical role in targeting medically relevant kinases such as Erk2 and the findings reported here open new avenues for designing small molecule inhibitors that block the active site while also steering the population of the enzyme into active or inactive conformations. Since protein dynamics and conformational ensembles are essential for enzyme function, this work will be of broad interest to those working in drug development, signal transduction, and enzymology.

Public Reviews:

Reviewer #1 (Public Review):

Summary: The authors set out to determine how chemical variation on kinase inhibitors determines the selection of Erk2 conformations and how inhibitor binding affects ERk2 structure and dynamics.

Strengths: The study is beautifully presented both verbally and visually. The NMR experiments and the HDX experiments complement each other for the study of Erk2 solution dynamics. X-ray crystallography of Erk2 complexes with inhibitors shows small but distinct structural changes that support the proposed model for the impact of inhibitor binding.

Weaknesses: A discussion of compound residence time for the different compounds and kinase constructs and how it could affect the very slow HDX rates might be helpful. For example, could any of the observed effects in Figure 4 be due to slow compound dissociation rather than slowed down kinase dynamics? What would be the implications?

Response: Rate constants for kon and koff were estimated for three inhibitors using surface plasmon resonance:

Author response table 1.

SPR estimates of Kd for selected inhibitors ranged between 0.03-3 nM. All HDX time courses involved prebinding of 20 µM inhibitor and 17 µM ERK2 for 30 min (predicted occupancy 99.9%), followed by deuteration time courses with 20 µM inhibitor and 1.7 µM ERK2. Estimated rates of dissociation were ~0.0003-0.007 s-1 and rates of binding were 20-100 s-1 for the inhibitors tested. Because the binding rates are faster than the intrinsic H-D exchange rate at pD 7 (~1 s-1), we expect ligands to rebind and form the enzyme:ligand complex faster than the free enzyme undergoes exchange. Therefore, HDX rates should mostly reflect deuteration of the inhibitor-bound enzyme for all inhibitors.

Reviewer #2 (Public Review):

Erk2 is an essential element of the MAP kinase signaling cascade and directly controls cell proliferation, migration, and survival. Therefore, it is one of the most important drug targets for cancer therapy. The catalytic subunit of Erk2 has a bilobal architecture, with the small lobe harboring the nucleotide-binding pocket and the large lobe harboring the substrate-binding cleft. Several studies by the Ahn group revealed that the catalytic domain hops between (at least) two conformational states: active (R) and inactive (L), which exchange in the millisecond time scale based on the chemical shift mapping. The R state is a signature of the double phosphorylated Erk2 (2P-Erk2), while the L state has been associated with the unphosphorylated kinase (0P-Erk2). Interestingly, the X-ray structures reveal only minimal differences between these two states, a feature that led to the conclusion that active and inactive states are structurally similar but dynamically very different. The Ahn group also found that ATP-competitive inhibitors can steer the populations of Erk2 either toward the R or the L state, depending on their chemical nature. The latter opens up the possibility of modulating the activity of this kinase by changing the chemistry of the ATP-competitive inhibitor. To prove this point, the authors present a set of nineteen compounds with diverse chemical substituents. From their combined NMR and HDX-Mass Spec analyses, fourteen inhibitors drive the kinase toward the R state, while four compounds keep the kinase hopping between the R and L states. Based on these data, the authors rationalize the effects of these inhibitors and the importance of the nature of the substituents on the central scaffold to steer the kinase activity. While all these inhibitors target the ATP binding pocket, they display diverse structural and dynamic effects on the kinase, selecting a specific structural state. Although the inhibited kinase is no longer able to phosphorylate substrates, it can initiate signaling events functioning as scaffolds for other proteins. Therefore, by changing the chemistry of the inhibitors it may be possible to affect the MAP cascade in a predictable manner. This concept, recently introduced as proof of principle, finds here its significance and practical implications. The design of the next-generation inhibitors must be taken into account for these design principles. The research is well executed, and the data support the author's conclusions.

Reviewer #3 (Public Review):

Summary: Anderson et al utilize an array of orthogonal techniques to highlight the importance of protein dynamics for the function and inhibition of the kinase ERK2. ERK2 is important for a large variety of biological functions.

Strengths: This is a thorough and detailed study that uses a variety of techniques to identify critical molecular/chemical parameters that drive ERK2 in specific states.

Weaknesses: No details rules were identified so that novel inhibitors could be designed. Nevertheless, the mode of action of these existing inhibitors is much better defined.

Response: As recommended we added a sentence to the Discussion suggesting that inhibitors that perturb the β1-β2-β3 sheet in such a way that moves helix αC and αL16 away from the binding site might confer R-state selection. We view this as a preliminary model for predicting conformation selection in ERK2.

Reviewer #1 (Recommendations For The Authors):

Maybe the authors can comment on how the HDX timescale and the NMR timescale relate to each other and how such different timescales can report on the same event. In particular, the HDX timescale appears to be on the scale on minutes to tens hours (e.g. 2P state). How would inhibitor dissociation and rebinding affect the observed HDX signal? Is it worth considering compound residence time for the different compounds/kinase states?

Response: The HDX-MS and NMR experiments report different processes therefore their timescales do not necessarily match. For native state proteins at neutral pH, HDX-MS reports fluctuations that allow solvent exposure of backbone amide N-H, reflecting conformational mobility of the main chain. This is often modeled as a two-state interconversion between “closed” (HDX protected) and “open” (HDX accessible) states. Because the µs-ms timescale of main chain fluctuations is faster than the intrinsic rate of HDX (kexch, ~1 s-1), the observed HDX rate (kobs) can be approximated by the ratio of kopen/kclosed x kexch = Kop x kexch. Therefore, kobs can be considered a thermodynamic measurement that reflects Kop.

The [methyl 13C,1H] NMR CPMG experiment that we used to identify global exchange behavior in Xiao et al (PNAS, 2014) modeled the 2P-ERK2 apoenzyme by a two-state equilibrium (L⇌R) between methyl-ILV conformers, yielding rate constants kL→R 240 s-1 and kR→L 60 s-1. Some methyls had large enough chemical shifts between L and R that they appeared as separate peaks in HMQC spectra that matched the L and R populations estimated by CPMG. In this study, the HMQC peaks shown in Figures 1, 6, and 9 are those that report shifts in L vs R populations and conformation selection for the R-state by VTX11e, BVD523 and triazolopyridine inhibitors.

Where HDX and NMR agree is in their ability to report changes in populations of L and R in 2P-ERK2. This was first shown when both HDX and NMR measurements reported perturbations at the activation loop induced by inhibitors with differential selection for the R- vs L-states (Pegram et al. PNAS, 2019). CPMG measurements then confirmed that methyl probes in the activation loop are included in the global exchange process (Iverson et al., Biochemistry, 2020). Therefore, the HDX and NMR experiments reflect shifts in the equilibrium between L and R conformers, rather than motions with specific timescales.

Reviewer #2 (Recommendations For The Authors):

I believe the paper is suitable for the special issue of Elife dedicated to protein kinases after the authors address minor concerns/comments.

a) Introduction, page 3: "[..] But within the ATP binding site, the conserved residues ...are largely overlapping." Do the authors mean that the residues are overlapping in the X-ray structures? If so, what is the rmsd among the X-ray structures?

Response: The overlap between conserved residues K52, E69, D147, N152 and D165 in 2P- and 0P-ERK2 is presented in Fig. S1C, which shows an overlay between their apoenzyme crystal structures (PDBID: 2ERK, 5UMO). The RMSD of atoms in each residue are: K52 0.63 Å (9 atoms); E69 0.15 Å (9 atoms); D147 0.055 Å (8 atoms); D165 0.88 Å (8 atoms). As recommended, this information was added to the legend to Suppl. Fig. S1.

b) Introduction, page 5: "[...] For example binding of VTX11 partially inhibits...[..]" Please provide a citation.

Response: As recommended we added a citation at end of this sentence (Pegram et al. PNAS, 2019).

c) Introduction, page 5: "[...] N-lobe deformities..." What do the authors mean by deformities? Are there frustrated conformations?

Response: We used the term “deformities” to mean conformational differences, which may be but are not necessarily due to frustration. To avoid confusion, we removed the term “deformities” and replaced it with “conformational changes”.

d) Supplementary Information. The authors report the chemical shift perturbations for several inhibitors. Does the extent of the chemical shift perturbation reflect the strength of the binding for each inhibitor? In other words, do the largest chemical shift perturbations correspond to the highest binding affinity?

Response: The concentrations used in the NMR ligand binding experiments (150 µM ERK2, 180 µM inhibitor) allow 99.9+% complex formation over the 0.03 - 3 nM range of Ki for all inhibitors. Therefore, the chemical shifts report changes in electronic environment between bound and free enzyme. These can be ascribed to first or second sphere contacts with ligand or distal allosteric effects. But they are not likely to reflect differences in binding affinity.

New Suppl. Fig. S3 now adds HMQC titrations of VTX11e and GDC0994 into 2P-ERK2, which confirm binding saturation based on the disappearance of free enzyme peaks.

e) Do the authors have any evidence for the dynamic effects of the different inhibitors? Of course, a systematic analysis of the protein dynamics by NMR will require a significant amount of time and effort beyond this work. However, did the authors measure the effects of the inhibitors on the linewidths of the methyl groups distal from the binding site?<br /> Response: As recommended, we examined linewidths of selected peaks in the presence and absence of inhibitors. The results show no significant systematic differences between bound and free ERK2. Therefore dynamic effects of different inhibitors are not indicated by the available data.

f) The authors identified the b3-aC loop as a critical element for the internal network of interactions. Can this structural element be targeted by small molecules as well?

Response: Yes, in fact the X-ray structures of 0P-ERK2 bound to the inhibitor, SCH772984, and 2P-ERK2 bound to the related compound, SCHCPD336, both show inhibitor occupying a pocket between between strand β3 and helix αC, leading to disruption of β3-αC contacts (Chaikaud et al., NSMB 2014; Pegram et al., PNAS 2019). To the extent that β3-αC contacts are important for conformation selection to the R-state, this may explain why SCH772984 favors the L-state. We revised the Discussion to add this point.

g) The authors should mention a recent paper suggesting that it is possible to control substrate-binding affinity by changing the nature of the ATP-binding inhibitors ((DOI: 10.1126/sciadv.abo0696).

Response. As recommended we added this point and citation to the Discussion.

Reviewer #3 (Recommendations For The Authors):

3.1. The manuscript is well written, but very long and sometimes repetitive. Some parts of the introduction are repeated in the result section and parts of the result section are repeated in the discussion. It will be easy to shorten the work to make it easier to read.

Response: As recommended we streamlined the Discussion to remove some of the repetitive elements, while trying to retain the main conclusions and rationale for readers who are not well versed in kinase structure.

3.2. Only specific residues are shown for the NMR spectra figures - while this is helpful to understand the concept, full spectra need to be shown to allow for direct comparison of the data quality (i.e. in supplemental material). If statements are made that measurements are done under full saturation - it should be shown that saturation is achieved in the measurements. All relaxation data should be made available - similar to CSPs.

Response: As recommended, new Suppl. Figs. S2 and S9 were added to show the full spectra of each inhibitor complex analyzed by NMR. New Suppl. Fig. S3 now adds titrations of 2P-ERK2 with VTX11e and GDC0994.The results confirm binding saturation based on the disappearance of free enzyme peaks.

3.3. No validation report was provided, nor a PDB number - so it is unclear if the crystal structures have been submitted - they need to be submitted in order to also access an mtz file, which is critical to understanding the quality of the structure (especially the ligand). This makes it difficult to assess the quality of the structures.

Response: Table S1 has been revised to show data collection and refinement parameters for PDBID: 8U8K (2PERK2:Inh#8, Fig. 8C) and 8U8J (2P-ERK2:Inh#16, Fig. 8D). RCSB validation reports are attached and PDB depositions have been approved and will be released upon VOR assignment.

Author response:

The following is the authors’ response to the previous reviews.

Recommendations for the authors:

We sincerely value the insightful and constructive feedback (italicized) provided by the reviewers, which has been instrumental in identifying areas of our manuscript that required further clarification or amendment. In response to these valuable comments, we have significantly revised the manuscript to enhance clarity and accuracy. Specifically, we have corrected an oversight related to the robot’s velocity and secondary antibody ratios, and addressed previously missing values in Figs. 3E and 4E. Importantly, these corrections did not alter the outcomes of our results. Additionally, we have enriched our manuscript with new data analyses, as reflected in Figures 1B, 1F, 2H-J, 4D, 4F-H, S1A, S1C-E, S3H, S5, and Table 1, ensuring a more comprehensive presentation of our findings. Below are our responses detailing each comment and explaining the modifications integrated into the revised manuscript.

Reviewer 1:

(1) To address the question of whether PAG photostimulation biases the cells that respond to the robot, a counterbalanced experiment, in which the BLA activity is initially recorded during the foraging vs. robot test and the PAG stimulation happens at the end of the session, should have been performed.

In our study, we investigated fear behavior and BLA cell responses to intrinsic dPAG photostimulation (320 pulses) in naïve animals, followed by their reactions to an extrinsic predatory robot. We recognize the reviewer's concern regarding the potential  influence of initial dPAG photostimulation on BLA neuron responses to the robot. We address this issue in our discussion (pg. 13) as follows: “However, it is crucial to consider the recent discovery that optogenetic stimulation of CA3 neurons (3000 pulses) leads to gain-of-function changes in CA3-CA3 recurrent (monosynaptic) excitatory synapses (Oishi et al., 2019). Although there is no direct connection between dPAG neurons and the BLA (Vianna and Brandao 2003, McNally, Johansen, and Blair 2011, Cameron et al. 1995), and no studies have yet demonstrated gain-of-function changes in polysynaptic pathways to our knowledge, the potential for our dPAG photostimulation (320 pulses) to induce similar changes in amygdalar neurons, thereby enhancing their sensitivity to predatory threats, cannot be dismissed.”

(2) In Figure 3, it is unclear which criteria (e.g. response latency, minimum Z score, spike fidelity) was used to identify the BLA neurons that were indirectly activated by PAG stimulation. A graphic containing at least the distribution of the response latencies for each BLA neuron after PAG laser activation is needed.

We have specified the criteria for determining the responsiveness of BLA neurons to dPAG stimulation on page 22. This involves analyzing the first 500-ms post-stimulation across five 0.1-s bins. Units were classified as ‘stim cells’ if they showed z-scores greater than 3 (z > 3) in any of the bins during the initial 500-ms period post-stimulation. Neurons activated by both pellet procurement and dPAG stimulation were not included in the 'stim cell' category. Additionally, we have included a graphic in the revised manuscript (Fig. S3C) that presents the distribution of response latencies of BLA neurons to dPAG stimulation.

(3) To strengthen the claim that it is a BLA-PVT-PAG circuit that carries information about predatory threat, a new experiment using CTB and cFos could be used to demonstrate that PAG neurons that project to PVT are recruited during the robot exposure.

Our study primarily aimed to explore the transmission of threat signals between the dPAG and BLA. We acknowledge that our evidence for the PVT’s intermediary role, derived from CTB injections in the BLA and subsequent CTB+cFos co-labeling analysis in the PVT (Fig. 4G and 4H), is limited. Accordingly, we have moderated the emphasis on the PVT’s involvement in both the abstract and introduction. We now present the PVT’s role as a promising direction for future research in the discussion section of our revised manuscript.

(4) In Fig 2, the authors' interpretation is that photostimulation of PAG neurons elicits fleeing responses in the rats. However, there is a vast literature demonstrating that the PAG is also involved in nociception. Although this is recognized by the authors in the first part of the introduction and briefly described in the discussion, the authors should more explicitly explain that PAG stimulation produces analgesia and thus is unlikely to underlie the escaping responses observed. This may not be intuitive for a broader audience.

We appreciate the reviewer's insightful suggestion to elaborate on the PAG involvement in nociception and analgesia, as supported by the literature. While our initial manuscript acknowledged these functions, we have now expanded our discussion to address the PAG’s multifaceted roles (pg. 12): “As mentioned in the introduction, the dPAG is recognized as part of the ascending nociceptive pathway to the BLA (De Oca et al. 1998, Gross and Canteras 2012, Herry and Johansen 2014, Kim, Rison, and Fanselow 1993, Ressler and Maren 2019, Walker and Davis 1997). The dPAG is also implicated in non-opioid analgesia (e.g., Bagley and Ingram 2020, Cannon et al. 1982, Fields 2000). However, it is essential to emphasize that, despite its roles in pain modulation, the primary behavior observed in dPAG-stimulated, naive rats foraging for food in an open arena was goal-directed escape to the safe nest, underscoring the dPAG’s critical function in survival behaviors.” Note that this aligns with human studies on PAG stimulation (e.g., Carrive and Morgan 2012, Magierek et al. 2003), particularly those by Amano et al. (Amano et al. 1982), which reported patients feeling an urge to escape, similar to being chased, upon PAG stimulation.

(5) To truly demonstrate the functional links between the PAG and BLA, more experiments are needed. For example, one could record from BLA neurons during the robot surge while performing optogenetic inhibition of the PAG neurons. There is also no evidence that activity in the indirect pathway that connects the PAG to the BLA is indispensable for the expression of defensive responses towards the robot (e.g., causality tests using chemogenetic or optogenetic inactivation).

We agree that incorporating optogenetic inhibition of PAG neurons while simultaneously recording from BLA neurons during a robot surge would strengthen the evidence for the functional connectivity between the PAG and BLA. Such an experiment would necessitate the transfection and photoinhibition of a wide array of dPAG neurons responsive to predatory threats. This procedure is technically more viable in transgenic mouse models, given their suitability for genetic manipulation. In light of this, and in response to the suggestions in the Joint Public Review, we have revised the abstract, introduction, and discussion to offer a more cautious interpretation of our findings. This revision reflects a careful consideration of both the evidence and the limitations inherent in our study (pg. 13): “While our findings demonstrate that opto-stimulation of the dPAG is sufficient to trigger both fleeing behavior and increased BLA activity, we have not established that the dPAG is necessary for the BLA’s response to predatory threats. To establish causality, it is essential to conduct experiments such as optogenetic inhibition to determine whether the dPAG is indispensable for activating BLA neurons and initiating escape behavior in the face of threats. The complexity of targeting the dPAG, which includes its dorsomedial, dorsolateral, lateral, and ventrolateral subdivisions (e.g., Bandler, Carrive, and Zhang 1991, Bandler and Keay 1996, Carrive 1993), suggests the need for future studies using transgenic mouse models. Should inactivation of the dPAG negate the BLA's response to predatory threats, it would underscore the dPAG's central role in this defensive mechanism. Conversely, if BLA responses remain unaffected by dPAG inactivation, this could indicate the existence of multiple pathways for antipredatory defense mechanisms.”

(6) The manuscript lacks information about the number of rats and trials that were used across the experiments (e.g. Fig 2G-J). In some occasions, the authors start the experiments with a specific number of animals and then reduce the N by half without providing a rationale (e.g. Fig. 3). Equally confusing is the experimental timeline. For example: a) Were the pre-robot, robot, and post-robot sessions always performed within the same day? b) It was described that microdrivable arrays were used, but did the same rats experienced the robot test more than one time? c) How many bins were used for normalization during the Z-score calculation and when were the data binned at 100 ms versus 1 s? d) How many trials were used for each analysis? For example, to identify robot cells, did the authors establish a minimum number of trials per animal to calculate the peristimulus time histograms? Having a significant number of trials is critical to make sure that the observed neuronal responses are replicable across the trials. e) How was the neuronal activity related to "pellet retrieval" aligned during robot sessions? Was the activity aligned with the moment in which the rat touches the pellet or when the animal returns to the nest with the pellet? f) How did the authors control for trials in which the rat consumed the pellets in the same local vs. those in which they returned to the nest to eat it? All these points are extremely important for future replicability.

We apologize for any confusion caused by the initial lack of detail in our experimental procedures. The revised manuscript has been updated with comprehensive methodological details:  

(i) The study involved thirteen rats (ChR2, n = 9; EYFP, n = 4), subjected to dPAG stimulation using fixed light parameters (473 nm, 20 Hz, 10-ms pulse width, 2 s duration) during Long and Short pellet distance trials (refer to Fig. 2E-G). The stimulation intensity was adjusted to each animal's response (fleeing behavior), ranging from 1-3 mW. Additional testing occurred over multiple days, with incremental adjustments to stimulation parameters (intensity, frequency, duration) after confirming normal baseline foraging behavior (Fig. 2H-J, at x = 0). These details are now clearly depicted in the manuscript.

(ii) The primary objective was to investigate BLA neuron responses to dPAG opto-stimulation. Six rats were initially tested, with three later assessed for their reactions to dPAG stimulation in the presence of an actual predator, to gauge behavioral effects.

(iii) Regarding the experimental timeline:

a) Pre-robot, robot, and post-robot sessions were conducted successively on the same day.

b) Sessions with the robot predator were repeated until habituation occurred or when unit recordings were deemed invalid due to microdrive limitations or the absence of unit detection. Throughout these sessions, the success rate for pellet retrieval remained consistently low. Specifically, the mean success rate for the dPAG recordings was 2.803% + 1.311. For the BLA recordings, animals did not succeed in retrieving pellets during any of the robot trials. To provide a more detailed account of the methodology, the manuscript has been updated to include the number of recording days and the units recorded in the "Behavioral Procedures" section.

c) As described in Materials and Methods, unit recording data were binned at 0.1-s intervals and normalized against a 5-s pre-event baseline (50 bins). For statistical analyses in Figure 1F’s rightmost column, 1-s bins were used to simplify post-hoc analysis corrections.

d) Each recording session consisted of 5-15 trials. Trials were excluded if rats attempted to procure the pellet within 10 s post-dPAG stimulation or robot activation, ensuring accurate characterization of unit responsiveness. Consequently, the number of trials varied among subjects.

e) Pellet retrieval was indicated by the animal entering a designated zone 19 cm from the pellet, driven by hunger.

f) Animals were trained to retrieve pellets and return to their nest for consumption prior to robot testing sessions, as elaborated in the “Baseline foraging” section.

(7) In the abstract, the authors mention that predictive cues are ambiguous during naturalistic predatory threats, but it is not clear what do they mean by ambiguous. In addition, in the introduction section, the authors describe that the present study will investigate how the dPAG and BLA communicate threat signals. However, the author should clarify right in the beginning that these two regions are not monosynaptically connected with each other and cite the proper references.

The abstract’s original sentence, “…where predictive cues are ambiguous and do not afford reiterative trial-and-error learning…” has been refined to “…characterized by less explicit cues and the absence of reiterative trial-and-error learning events …” This adjustment more accurately reflects that cues in natural settings often lack the clear and consistent quality of those in controlled experimental settings, which is necessary for the straightforward process of trial-and-error learning.

Regarding the dPAG and BLA connectivity, the revised introduction (pg. 5) now states: “Considering the lack of direct monosynaptic projections between dPAG and BLA neurons (Vianna and Brandao 2003, McNally, Johansen, and Blair 2011, Cameron et al. 1995), we utilized anterograde and retrograde tracers in the dPAG and BLA, respectively. This was complemented by c-Fos expression analysis following exposure to predatory threats. Our anatomical findings suggest that the paraventricular nucleus of the thalamus (PVT) may be part of a network that conveys predatory threat information from the dPAG to the BLA.”

(8) In the introduction section, the authors should clarify that the US information is conveyed from the PAG to BLA via the lateral thalamus (posterior intralaminar nucleus, medial geniculate nucleus) or dorsal midline thalamus (paraventricular nucleus of the thalamus). The statement regarding how "the PAG functions as part of the ascending pain transmission pathway, providing footshock US information to the BLA" is misleading because the PAG does not send monosynaptic projections directly to the BLA.

The revised text (pg. 3) now reads: “…suggest that the dPAG is part of the ascending US pain transmission pathway to the BLA, the presumed site for CS-US association formation (De Oca et al. 1998, Gross and Canteras 2012, Herry and Johansen 2014, Kim, Rison, and Fanselow 1993, Ressler and Maren 2019, Walker and Davis 1997). This pathway is thought to be mediated through the lateral and dorsal-midline thalamus regions, including the posterior intralaminar nucleus and paraventricular nucleus of the thalamus (Krout and Loewy, 2000; McNally, Johansen, and Blair, 2011; Yeh, Ozawa, and Johansen, 2021; but see Brunzell and Kim, 2001).”

(9) The author's assumption that threat information flows from the PAG to the BLA, rather than BLA to PAG, based on electrical stimulation and lesion experiments performed in previous studies is problematic for at least three reasons: a) Electrical stimulation can activate fibers of passage as well as presynaptic neurons antidromically. b) The lesion approach may not have targeted 100% of the neurons in PAG, which extends anatomically along the antero-posterior axis of the midbrain for several millimeters in rats. This observation also disagrees with more recent studies using optogenetics and imaging tools demonstrating that the PAG is the downstream target of the BLA-CeA pathway. c) The authors cited prior reports describing the role of the amygdala-PAG pathway in dampening the US response and providing a negative signal to the PAG. However, a series of previous studies demonstrating that the PAG serves as the downstream target of the central nucleus of the amygdala for the expression of defensive response are completely ignored by the authors. Here are just some examples: Massi et al, 2023, PMID: 36652513; Tovote et al 2016, PMID: 27279213; Penzo et al, 2014 PMID: 24523533).

We recognize the complexities in interpreting findings from electrical stimulation and lesion studies. Our prior work (Kim et al. 2013) supports the conclusion that predatory threat information directionally flows from the dPAG to the BLA, as evidenced by distinct behavioral outcomes from experimental manipulations of dPAG and BLA. Specifically, dPAG stimulation-induced fleeing behavior was blocked by BLA lesions (as well as muscimol inactivation), whereas BLA stimulation-induced fleeing was unaffected by dPAG or combined dPAG+vPAG lesions (refer to Fig. 5A), suggesting a flow from dPAG to BLA. Our manuscript further clarifies that dPAG optostimulation results confirmed that escape behavior in foraging rats, induce by dPAG electrical stimulation (Kim et al. 2013), was activated by intrinsic dPAG neurons rather than by fibers of passage or current spread to other brain regions.  

Furthermore, the PAG’s anatomical and functional diversity, with distinct segments along its longitudinal axis associated with different defensive behaviors, reinforces our conclusions. The dPAG is implicated in flight responses, while the vPAG is associated with freezing behavior (e.g., Bandler and Shipley 1994, Kim, Rison, and Fanselow 1993, Lefler, Campagner, and Branco 2020, Morgan, Whitney, and Gold 1998). The critiques' referenced studies primarily focus on the BLA-CeA-vPAG circuit's role in freezing during Pavlovian fear conditioning, contrasting with our emphasis on the dPAG-PVT-BLA circuit and its mediation in escape behavior in response to naturalistic predatory threats.

We also note that different invasive procedures can yield varying behavioral outcomes. For example, both acute (e.g., optogenetic and muscimol inactivation) and chronic (e.g., surgical ablation) manipulations within the same brain circuit have shown diverse effects across species (Otchy et al. 2015). Moreover, optogenetics comes with its own set of conceptual and technical challenges (Adamantidis et al. 2015), including the difficulty of targeting, quantifying and photo-inhibiting 100% of PAG neurons. Despite the limitations of each technique, our collective evidence from lesions, inactivation, electrical stimulation (Kim et al. 2013), optostimulation, and single-unit recordings (the present study) supports the premise that the dPAG acts upstream of the BLA in processing predatory threat information.

(10) In the discussion, the authors suggest that the PVT may be the interface between the PAG and the BLA for the expression of antipredatory defensive behavior during their foraging vs. robot test, but previous studies looking at the role of PVT in antipredator defensive behavior and/or approach-avoidance conflict tasks are not cited and discussed in the manuscript (Engelke et al, 2021, PMID: 33947849; Choi et al 2019, PMID: 30979815; Choi and McNally 2017, PMID: 28193686).

We thank the reviewer for pointing out these pivotal studies, which we have carefully reviewed and integrated into the revised manuscript (pg. 14): “These results, in conjunction with previous research on the roles of the dPAG, PVT, and BLA in producing flight behaviors in naïve rats (Choi and Kim 2010, Daviu et al. 2020, Deng, Xiao, and Wang 2016, Kim et al. 2013, Kim et al. 2018, Kong et al. 2021, Ma et al. 2021, Reis et al. 2021), the anterior PVT’s involvement in cat odor-induced avoidance behavior (Engelke et al. 2021), and the PVT’s regulation of behaviors motivated by both appetitive and aversive stimuli (Choi and McNally 2017, Choi et al. 2019), suggest the involvement of the dPAGàPVTàBLA pathways in antipredatory defensive mechanisms, particularly as rats leave the safety of the nest to forage in an open arena (Figure 4I) (Reis et al. 2023).”  

(11) The authors use the expression "looming robot predator" in many cases throughout the manuscript. However, it is unclear whether the defensive responses observed in the rats are elicited by the looming stimulus produced by the movement of the robot towards the rats. The authors describe that rats do not respond to a stationary robot, but would the sound produced by the movement of the robot elicit defensive responses? Would non-approaching lateral or dorsoventral movements (not associated with looming) be sufficient to induce defensive behavior in the rats? There is a vast literature in the field about defensive behaviors induced by looming stimuli. The authors should empirically demonstrate that the escaping responses induced by the robot are mediated by looming or refrain to use the looming terminology to avoid confusion.

Our use of "looming robot predator" is based on empirical evidence from a prior parametric study, which identified the forward, or 'looming,' motion of the Robogator as the key stimulus eliciting a flight response in rats (Kim, Choi, and Lee 2016). This reaction significantly decreased when the robot moved backward from the same starting position, producing a similar sound, and was absent when the robot remained stationary. This suggests that neither sound alone nor the mere presence of a novel object provokes goal-directed escape behavior (Kong et al. 2021). This aligns with studies indicating that simulated looming stimuli, like an expanding disk, induce flight or freezing responses in mice (De Franceschi et al. 2016, Yilmaz and Meister 2013).

It should be noted that the 2013 study by Yilmaz & Meister (Yilmaz and Meister 2013) on the looming disk paradigm showed that not all mice responded to the stimuli (e.g., Figs. 2A and 3A), with those that did exhibiting rapid habituation by the second exposure. This contrasts with our predatory robot paradigm (Choi and Kim 2010), where all rats consistently fled from the looming robotic predator across multiple trials, underscoring the critical role of looming motion in simulating predator attacks that trigger flight behavior in rats.

Thus, the term "looming" accurately captures the nature of the robot's movement and its effect on eliciting defensive responses in rats. Nonetheless, should the editors agree with the reviewer's suggestion to minimize potential confusion, we are willing to substitute "looming" with "approaching," although we consider the terms to be synonymous in the context of our study.

(12) If the authors are citing the Rescorla-Wagner model, they should include at least one additional sentence to explain it, as many people in the field are not familiar with this model.

In response to the request for clarification on the Rescorla-Wagner model, we have added an explanatory sentence (pg. 4): “Fundamentally, the negative feedback circuit between the amygdala and the dPAG serves as a biological implementation of the Rescorla–Wagner (1972) model, a foundational theory of associative learning that emphasizes the importance of prediction errors in reinforcement (i.e., US), as applied to FC (Fanselow 1998).”

(13) The authors need to include the normality test used to determine whether a parametric or non-parametric statistical analysis was the most appropriate test for each experiment.

We have included the outcomes of the normality tests, detailed in Table S1.

(14) In Fig. 1F, the authors show a representative PAG neuron with peristimulus-time histogram and rasters reaching frequencies higher than 100 Hz and sustained firing rates of >50 Hz following robot activation. The authors should include a firing rate analysis (e.g., average firing rate and maximum firing rate before and after robot activation) of the 22 robot-responsive PAG neurons recorded during the session to clarify whether this high firing rate, which is atypical in other brain regions, is commonly observed in the PAG. Showing the isolated waveforms of some representative neurons would help to clarify whether the activity is being recorded from a single-isolated unit instead of multiple units within the same channel.

In response to the critique, we have expanded our analysis to include both average and maximum firing rates before and after robot activation for the 22 robot-responsive PAG neurons. This detailed firing rate analysis, illustrating their distribution, has been incorporated into the revised manuscript (refer to Figure S1C and S1D). Furthermore, to alleviate concerns regarding the identification of single-unit activity versus potential multi-unit recordings, we have included peri-event raster plots and waveforms for two additional representative neurons in Figure 1F.

(15) In Figure 2, the authors should indicate when the recordings are performed on anesthetized vs. freely-moving awake animals.

In the original manuscript, we specified that the optrode recordings depicted in Figure 2B were conducted on anesthetized rats. To enhance clarity and directly address the critique, we have now clearly indicated this condition in Figure 2A as well.

(16) The optogenetic stimulation parameters used in Fig 2H indicate that 0.5 mW was sufficient to induce behavioral changes. This is surprising because most optogenetic experiments in the field use much higher intensities (> 5mW). If much lower intensities are sufficient to drive PAG-mediated behaviors, this may be a very important observation that should be conveyed to the field. I recommend the reviewers clarify if they in fact used 0.5 mW and then discuss that the laser intensity used in the experiments was 10X lower than that required for other brain regions

In our study, we indeed observed that 0.5 mW of dPAG stimulation increased the latency to procure the pellet without completely preventing the action. Notably, at 1 mW, more than half of the animals (n = 5/9 rats; Fig. 2H) and at 3 mW, all rats (9/9) failed to procure the pellet and fled from the foraging area to the nest (Fig. 2G). These results indicate that even lower intensities were sufficient to elicit behavioral changes through dPAG stimulation in a large foraging arena, highlighting the dPAG's sensitivity to optogenetic manipulation. This finding is consistent with our earlier research on dPAG electrical stimulation, which required significantly lower intensities to provoke defensive behaviors compared to the BLA. Specifically, the stimulation intensity needed for aversive behavior in the dPAG was substantially lower (dPAG: 65.0 ± 6.85 µA) than for the BLA (BLA: 275.0 ± 24.44 µA) (Kim et al. 2013). Furthermore, Deng et al. (Deng, Xiao, and Wang 2016) showed that 1 mW of blue light could elicit a 60% freezing response, with 2 mW triggering flight behavior within a latency of 0.6 seconds.

(17) In Fig 2 G-J, how many animals are being used per group and how was the sequence of the experiments performed? This is very important for replicability.

A total of three rats were utilized for the robot testing experiments depicted in Fig. 2 G-J. The experimental sequence for these animals consisted of successive pre-stimulation, stimulation, post-stimulation, and robot sessions. We have updated the manuscript to include this information.

(18) For the photostimulation of PAG neurons in Figs. 2 and 3, the authors need to clarify if the same parameters of laser stimulation used during the anesthetized recordings were also used during the behavioral tests. Also, the wavelength corresponding to the blue laser should be 473 nm instead of 437 nm.

We thank the reviewer for identifying the error. We confirm that the opto-stimulation parameters (473 nm, 10-ms pulse width, 2 s duration) were consistently applied across both anesthetized recordings and behavioral tests. This consistency has been explicitly stated in the revised manuscript to ensure clarity regarding our experimental approach.

(19) In Fig. 3I, how was the representative trials selected? Instead of picking up the most representative trials, the authors should demonstrate the response of the cell during the entire session.

In response to the critique, we clarify that the color-coded PETH shown in Fig. 3I represents averaged BLA activity across a comprehensive set of trials. This includes 8 pre-stimulation, 10 stimulation, and 8 post-stimulation trials for the robot-activated sessions, with a similar distribution for non-stimulated sessions. This approach was chosen to provide a representative overview of the cell's response throughout the entire session. To address the request for more detailed data, we have added traditional PETHs to the revised manuscript (see Fig. S3H), which depict the cell's response across all trials.

(20) Fig 4 D should demonstrate a colabeling between the anterograde PAG fibers in the PVT and the retrogradely labeled neurons from BLA instead of PAG fibers only.

We wish to clarify that Fig. 4D is intended to show the distribution of dPAG terminals within the midline thalamic nuclei, as noted in prior research (Krout and Loewy 2000). Although dPAG terminals are distributed throughout the midline thalamus, our observations have specifically highlighted a notable increase in c-Fos expression within the paraventricular nucleus of the thalamus (PVT) in rats subjected to the robotic predator stimulus, in contrast to those in the foraging-only control condition (Fig. 4E). Addressing the reviewer's point, we direct attention to Fig. 4G, which includes images labeled "Robot-experienced" and "Merge." This figure demonstrates a subset of PVT neurons that were retrogradely labeled with CTB injected into the BLA, anterogradely labeled with AAV injected into the dPAG, and activated (as indicated by c-Fos expression) in response to the robotic predator. This provides specific colabeling evidence between anterograde PAG fibers in the PVT and retrogradely labeled neurons from the BLA, directly addressing the critique.

(21) The resolution of the cFos images is very low and makes it hard to appreciate.

We have updated Figs. 4F and 4G with high-resolution versions to ensure the details are more clearly visible. Furthermore, should there be a need for even greater clarity, we are prepared to supply the images as TIFF files, which are known for preserving high image quality.

Reviewer 2:

(1) The text is clearly written, and I appreciated the inclusion of interesting citations, such as the one about paintings by cavemen. The authors also do a good job of discussing the underlying theoretical framework and the figures are easy to understand. Although the topic is very interesting, the amount of novel work is somewhat low. Figure 1 shows that dPAG cells are activated by the predator, and this has been shown by many prior reports. Similarly, Figure 2 shows that dPAG activation creates defensive responses, and this too has been shown by many prior reports.

We appreciate the reviewer’s positive remarks. We acknowledge the rich body of research documenting dPAG neuronal activation by various predator cues such as odors (e.g., fox urine) (Lu et al. 2023), and scenarios involving anesthetized or spontaneously moving rat/cat predators, either physically partitioned or harness-restrained (Bindi et al. 2022, Deng, Xiao, and Wang 2016, Esteban Masferrer et al. 2020). Nevertheless, our study distinguishes itself by examining dPAG neuronal responses to a robotic predator, uniquely designed to replicate consistent looming motions across multiple trials and subjects within an environment that simulates natural foraging conditions, inclusive of a safe nest (cf. Choi and Kim, 2010). This approach allowed us to not only reveal the immediate activation of dPAG neurons in response to a rapidly approaching predator but also to explore the consequent fleeing behavior towards safety, thereby providing new insights into the dPAG's role in mediating goal-directed defensive responses in a more ecologically-relevant setting. Furthermore, our investigation extends beyond these findings to assess the impact of dPAG activation on BLA neuronal responses and their functional connectivity during predator-prey interactions, offering a fresh perspective on the neural circuits that support survival behaviors in animals when confronted with naturalistic threats.

(2) The results in Figure 3 are novel and interesting, but the characterization of BLA activity is incomplete. For example, what are the percentages of BLA cells that are inhibited or activated by all major behaviors observed? These behaviors include approach to pellet, escape from robot, freezing, stretch-attend postures, etc. These same analyses should also be added to dPAG activity in Figure 1. How does BLA single cell encoding of these behaviors relate to their responsivity to dPAG stimulation? And, finally, it is unclear what is the significance of BLA correlated synchronous firing. Is the animal more or less likely to be performing certain behaviors when correlated BLA firing occurs?

Our analysis, as presented in Figs. 3I, 3K, and S3D-F, selectively focused on BLA cell responses during distinct behaviors such as approaching a pellet and escaping from the robot. These behaviors were selected because their precise temporal markers allow for accurate correlation with BLA cell activity, building on the findings of our previous research (Kim et al. 2018, Kong et al. 2021).

The robot's motion, programmed to advance a fixed distance before retreating to its starting position, is designed to repeatedly elicit foraging, thus facilitating analysis of neural changes during conflict situations involving food approach and predator avoidance. However, this also leads to the rapid diminution of freezing and stretch-attend postures inside the nest as animals quickly adapt to the robot's movement pattern, rendering a time-stamped analysis of these behaviors unfeasible under our experimental conditions. While the inclusion of these behaviors in our analysis would be insightful, especially in extended interaction scenarios where the robot advances to the nest opening and remains before returning in a less predictable manner, such conditions would likely reduce foraging behavior due to increased fear, deviating from our study's primary objective of elucidating the interactions between the dorsal periaqueductal gray (dPAG) and the basolateral amygdala (BLA) functions.

Regarding the significance of BLA correlated synchronous firing, our findings, particularly in Figures 3M-O and S4, demonstrate significant synchronous activity among BLA neuronal pairs during encounters with the robot, as opposed to pre-stim, stim, and post-stim sessions. This synchrony is notably prominent among neurons responsive to dPAG stimulation, indicating that BLA neurons involved in processing dPAG signals may play a crucial role in enhancing BLA network coherence to effectively manage predatory threat information (pg. 13).

(3) In Figure 4, the authors identify the PVT as a potential region that can mediate dPAG to BLA communication via anatomical tracing. However, functional assays are missing. For example, if the PVT is inhibited chemogenetically, does this result in a smaller number of BLA cells that are activated by dPAG stimulation? Does activation of the dPAG-PVT or the PVT-BLA projections cause defensive behaviors? Functionally showing that the dPAG-PVT-BLA circuit controls defensive actions would be a major advance in the field and would greatly enhance the significance of this paper. It would also provide an anatomical substrate to support the view that the BLA is downstream of the dPAG, which was first demonstrated by the authors in their elegant 2013 PNAS paper.

We appreciate the reviewer’s constructive critique and valuable suggestions on the necessity for functional validation of the dPAG-PVT-BLA circuit's involvement in mediating defensive behaviors. In light of these comments, we have carefully considered and included a discussion on the importance of these proposed experiments as a direction for future research in our manuscript revision (also see response to Reviewer 1’s critique #5).

Our initial work in 2013 (Kim et al. 2013) laid the groundwork for identifying BLA neurons responsive to dPAG stimulation and suggested the PVT as a potential relay in this neural circuit. Recognizing the limitations of our current study, which does not include direct functional assays, we have adjusted our manuscript to convey the speculative aspect of the dPAG-PVT-BLA circuit’s role more accurately. Moreover, we have enriched our discussion by citing relevant studies that lend support to our proposed circuit mechanism. These references serve to place our findings within the broader context of existing research and highlight the imperative for subsequent studies to empirically confirm the functional significance of the dPAG-PVT-BLA pathway in driving defensive behaviors.

Reviewer 3:

(1) The Introduction refers to a negative feedback amygdala-dPAG from a study of the Johansen group, but in this case, the authors were referring to the ventrolateral and not the dorsal PAG.

We thank the reviewer for pointing out the need to distinguish between the dPAG and vPAG regions in our introduction. While Johansen et al. (2010) investigated the roles of PAG (including both dPAG and vPAG regions; see their Supplementary Figs. 4, 5, and 10), the differentiation between their specific contributions to the amygdala's negative feedback mechanism was not explicitly detailed in their initial publication. This distinction was further elaborated upon in later work by the same group (Yeh, Ozawa, and Johansen 2021), which specifically illuminated the dPAG's role in conditioned fear memory formation and its neural pathways to the PVT that influence fear learning. To reflect this nuanced understanding, we have revised our introduction (pg. 3): “In parallel, Johansen et al. (2010) found that pharmacological inhibition of the PAG, encompassing both dPAG and vPAG regions, diminishes the behavioral and neural responses in the amygdala elicited by periorbital shock US, thereby impairing the acquisition of auditory FC.”

(2) In the experiments recording dPAG in response to the predator threat, the authors mentioned cells activated by the predator threat, referred to as "robot cells." Were these cells inhibited in response to threat?

In the Result and Materials and Methods sections, we report that 23.4% (22 out of 94) of dPAG neurons, termed “robot cells,” showed a significant increase in firing rates (z > 3) within a latency of less than 500 ms during exposure to the looming robot threat, but not during the pre- and post-robot sessions. These cells are highlighted in Figures 1E-G. In contrast, we identified only a single unit exhibiting a decrease in activity (z-score < -3) in response to the robot threat. Given the overwhelming prevalence of cells with excitatory responses to the threat, our discussions and analyses have primarily centered on these excited cells. Nevertheless, to ensure a full depiction of our observations, we have included data on the inhibited unit in the revised manuscript, specifically in Figure S1E.

(3) The authors claim that tetrodes were implanted in the dorsal PAG; however, the electrodes' tips shown in the figures are positioned more ventrally in the lateral PAG (see Figures 1B, S5A).

The PAG is anatomically organized into dorsomedial (dmPAG), dorsolateral (dlPAG), lateral (lPAG), and ventrolateral (vlPAG) columns along the rostro-caudal axis of the aqueduct. The designation "dorsal PAG" (dPAG) traditionally encompasses the dmPAG, dlPAG, and lPAG regions, a classification supported by extensive track-tracing, neurochemical, and immunohistochemical evidence (e.g., (Bandler, Carrive, and Zhang 1991, Bandler and Keay 1996, Carrive 1993)). As Bandler and Shipley (Bandler and Shipley 1994) summarized, “These findings suggest that what has been traditionally called the 'dorsal PAG' (a collective term for regions dorsal and lateral to the aqueduct), consists of three anatomically distinct longitudinal columns: dorsomedial and lateral columns…and a dorsolateral column…" Similarly, Schenberg et al. (Schenberg et al. 2005) clarified in their review that, “According to this parcellation...the defensive behaviors (freezing, flight or fight) and aversion-related responses (switch-off behavior) were ascribed to the DMPAG, DLPAG, and LPAG (usually named the ‘dorsal’ PAG).” In our study, electrode placements were strictly within these specified dPAG regions. The electrode tip locations depicted in Figures 1B and S5A correspond with the -6.04 mm template (left panel below) from Paxinos & Watson’s atlas (Paxinos and Watson 1998), situated anteriorly to the emergence of the  vlPAG (right panel below). To enhance clarification in our manuscript, we provide a detailed definition of the dPAG that includes the dmPAG, dlPAG,  and lPAG, and support our electrode placement rationale with references to established literature (pg. 5).

Author response image 1.

(4) It would be nice to include a series of observations applying inhibitory tools (i.e., optogenetic photo inhibition) in the dPAG and BLA and see how they affect the behavioral responses in the 'approach food-avoid predator' paradigm. Moreover, it would be interesting to explore how inhibiting the dPAG to PVT pathway influences the flee response during the robot surge.

We appreciate the suggestion to explore the effects of optogenetic inhibition in the dPAG and BLA on behavioral responses within the 'approach food-avoid predator' paradigm, as well as the potential impact of inhibiting the dPAG to PVT pathway on flee responses during robot surge incidents. As mentioned in our response to Reviewer 1’s critique #5, the application of optogenetic inhibition necessitates transfecting, quantifying, and photoinhibiting a comprehensive set of dPAG neurons activated by predatory threats. This approach is more viable in future studies that can leverage transgenic mouse models for their genetic tractability. Following the Joint Public Review’s recommendations, we have revised our manuscript to ensure a more measured interpretation of our data, carefully balancing the evidence from tracer studies against the limitations of our current methodology.

Furthermore, referencing Reviewer 1’s critique #9, it is important to consider that various invasive techniques can yield different behavioral outcomes. For instance, research by Olveczky and colleagues (Otchy et al. 2015) demonstrated that acute manipulations (i.e., optogenetic and muscimol inactivation) and chronic surgical ablation of the same brain circuit can produce distinct effects in rats and finches. Despite these methodological constraints, our collective results from lesion, inactivation, electrical stimulation (Kim et al. 2013), optostimulation, and single-unit recording (present) studies cohesively suggest that the dPAG functions upstream of the BLA in processing predatory threat signals.

(5) The authors should also examine whether 'synaptic' appositions exist between the anterogradely labeled terminals from the dPAG and the double labeled CTB and cFOS neurons in the PVT.

We appreciate the suggestion to investigate the presence of synaptic appositions, which could potentially offer valuable insights into the synaptic connections and functional interactions within this neural circuit. However, due to the specialized nature of electron microscopy required for these examinations and the extensive resources it entails, this line of inquiry falls beyond the scope of our current study. We hope to address this aspect in future studies, where we can dedicate the necessary resources and expertise to conducting these intricate analyses.

(6) It is odd to see the projection fields shown in Fig. 4D, where the projection to the PVT looks much sparser compared to other targets in the thalamus and hypothalamus. If the projection to the PVT has such an important function, why does it seem so weak? This should be discussed. Also, because the projection to the PVT seems sparse, the authors should consider alternative paths like the one involving the cuneiform nucleus. The cuneiform nucleus is an important region responding to looming shadows with strong bidirectional links to the dorsolateral periaqueductal gray, providing strong projections to the rostral PVT.

The perceived scarcity of the dPAG-PVT pathway might not reflect its functional significance accurately. The PVT's small size could make its projections appear less dense in broad anatomical studies. To address this, we have updated Figure 4D with a high-resolution image that offers a detailed view of the PVT region. This enhancement (refer to the updated Fig. 4, bottom) more accurately depicts the projection density within the PVT. It is also critical to consider that the functional impact of neural pathways is not solely dependent on the quantity of projecting neurons. For instance, work by Deisseroth and colleagues (Rajasethupathy et al. 2015) has shown that even relatively sparse monosynaptic projections from the anterior cingulate cortex to the hippocampus can exert significant effects on neural circuit dynamics. Additionally, we have expanded our discussion to consider the potential roles of other circuits, such as the cuneiform nucleus, in driving the behavioral responses observed in our study (pg. 15): “Given the recent significance attributed to the superior colliculus in detecting innate visual threats (Lischinsky and Lin 2019, Wei et al. 2015, Zhou et al. 2019) and the cuneiform nucleus in the directed flight behavior of mice (Bindi et al. 2023, Tsang et al. 2023), further exploration into the communication between these structures and the dPAG-BLA circuitry is warranted.”

(7) Finally, in the Discussion, it would be nice to comment on how the BLA mediates flee responses. Which pathways are likely involved?

This excellent suggestion has been incorporated in the discussion (pg. 15): “Future studies will also need to delineate the downstream pathways emanating from the BLA that orchestrate goal-directed flight responses to external predatory threats as well as internal stimulations from the dPAG/BLA circuit. Potential key structures include the dorsal/posterior striatum, which has been associated with avoidance behaviors in response to airpuff in head-fixed mice (Menegas et al. 2018) and flight reactions triggered by auditory looming cues (Li et al. 2021). Additionally, the ventromedial hypothalamus (VMH) has been implicated in flight behaviors in mice, evidenced by responses to the presence of a rat predator (Silva et al. 2013) and upon optogenetic activation of VMH Steroidogenic factor 1 (Kunwar et al. 2015) or the VMH-anterior hypothalamic nucleus pathway (Wang, Chen, and Lin 2015). Investigating the indispensable role of these structures in flight behavior could involve lesion or inactivation studies. Such interventions are anticipated to inhibit flight behaviors elicited by amygdala stimulation and predatory threats, confirming their critical involvement. Conversely, activating these structures in subjects with an inactivated or lesioned amygdala, which would typically inhibit fear responses to external threats (Choi and Kim 2010), is expected to induce fleeing behavior, further elucidating their functional significance.”

Adamantidis, A., S. Arber, J. S. Bains, E. Bamberg, A. Bonci, G. Buzsaki, J. A. Cardin, R. M. Costa, Y. Dan, Y. Goda, A. M. Graybiel, M. Hausser, P. Hegemann, J. R. Huguenard, T. R. Insel, P. H. Janak, D. Johnston, S. A. Josselyn, C. Koch, A. C. Kreitzer, C. Luscher, R. C. Malenka, G. Miesenbock, G. Nagel, B. Roska, M. J. Schnitzer, K. V. Shenoy, I. Soltesz, S. M. Sternson, R. W. Tsien, R. Y. Tsien, G. G. Turrigiano, K. M. Tye, and R. I. Wilson. 2015. "Optogenetics: 10 years after ChR2 in neurons--views from the community."  Nat Neurosci 18 (9):1202-12. doi: 10.1038/nn.4106.

Amano, K., T. Tanikawa, H. Kawamura, H. Iseki, M. Notani, H. Kawabatake, T. Shiwaku, T. Suda, H. Demura, and K. Kitamura. 1982. "Endorphins and pain relief. Further observations on electrical stimulation of the lateral part of the periaqueductal gray matter during rostral mesencephalic reticulotomy for pain relief."  Appl Neurophysiol 45 (1-2):123-35.

Bagley, E. E., and S. L. Ingram. 2020. "Endogenous opioid peptides in the descending pain modulatory circuit."  Neuropharmacology 173:108131. doi: 10.1016/j.neuropharm.2020.108131.

Bandler, R., P. Carrive, and S. P. Zhang. 1991. "Integration of somatic and autonomic reactions within the midbrain periaqueductal grey: viscerotopic, somatotopic and functional organization."  Prog Brain Res 87:269-305. doi: 10.1016/s0079-6123(08)63056-3.

Bandler, R., and K. A. Keay. 1996. "Columnar organization in the midbrain periaqueductal gray and the integration of emotional expression."  Prog Brain Res 107:285-300. doi: 10.1016/s0079-6123(08)61871-3.

Bandler, R., and M. T. Shipley. 1994. "Columnar organization in the midbrain periaqueductal gray: modules for emotional expression?"  Trends Neurosci 17 (9):379-89. doi: 10.1016/0166-2236(94)90047-7.

Bindi, R. P., C. C. Guimaraes, A. R. de Oliveira, F. F. Melleu, M. A. X. de Lima, M. V. C. Baldo, S. C. Motta, and N. S. Canteras. 2023. "Anatomical and functional study of the cuneiform nucleus: A critical site to organize innate defensive behaviors."  Ann N Y Acad Sci 1521 (1):79-95. doi: 10.1111/nyas.14954.

Bindi, R. P., R. G. O. Maia, F. Pibiri, M. V. C. Baldo, S. L. Poulter, C. Lever, and N. S. Canteras. 2022. "Neural correlates of distinct levels of predatory threat in dorsal periaqueductal grey neurons."  Eur J Neurosci 55 (6):1504-1518. doi: 10.1111/ejn.15633.

Cameron, A. A., I. A. Khan, K. N. Westlund, and W. D. Willis. 1995. "The efferent projections of the periaqueductal gray in the rat: a Phaseolus vulgaris-leucoagglutinin study. II. Descending projections."  J Comp Neurol 351 (4):585-601. doi: 10.1002/cne.903510408.

Cannon, J. T., G. J. Prieto, A. Lee, and J. C. Liebeskind. 1982. "Evidence for opioid and non-opioid forms of stimulation-produced analgesia in the rat."  Brain Res 243 (2):315-21. doi: 10.1016/0006-8993(82)90255-4.

Carrive, P, and M. M. Morgan. 2012. "Periaqueductal Gray." In The Human Nervous System, edited by J. K.; Paxinos Mai, G., 367-400. London: Academic Press.

Carrive, P. 1993. "The periaqueductal gray and defensive behavior: functional representation and neuronal organization."  Behav Brain Res 58 (1-2):27-47. doi: 10.1016/0166-4328(93)90088-8.

Choi, E. A., P. Jean-Richard-Dit-Bressel, C. W. G. Clifford, and G. P. McNally. 2019. "Paraventricular Thalamus Controls Behavior during Motivational Conflict."  J Neurosci 39 (25):4945-4958. doi: 10.1523/JNEUROSCI.2480-18.2019.

Choi, E. A., and G. P. McNally. 2017. "Paraventricular Thalamus Balances Danger and Reward."  J Neurosci 37 (11):3018-3029. doi: 10.1523/JNEUROSCI.3320-16.2017.

Choi, J. S., and J. J. Kim. 2010. "Amygdala regulates risk of predation in rats foraging in a dynamic fear environment."  Proc Natl Acad Sci U S A 107 (50):21773-7. doi: 10.1073/pnas.1010079108.

De Franceschi, G., T. Vivattanasarn, A. B. Saleem, and S. G. Solomon. 2016. "Vision Guides Selection of Freeze or Flight Defense Strategies in Mice."  Curr Biol 26 (16):2150-4. doi: 10.1016/j.cub.2016.06.006.

De Oca, B. M., J. P. DeCola, S. Maren, and M. S. Fanselow. 1998. "Distinct regions of the periaqueductal gray are involved in the acquisition and expression of defensive responses."  J Neurosci 18 (9):3426-32. doi: 10.1523/JNEUROSCI.18-09-03426.1998.

Deng, H., X. Xiao, and Z. Wang. 2016. "Periaqueductal Gray Neuronal Activities Underlie Different Aspects of Defensive Behaviors."  J Neurosci 36 (29):7580-8. doi: 10.1523/JNEUROSCI.4425-15.2016.

Engelke, D. S., X. O. Zhang, J. J. O'Malley, J. A. Fernandez-Leon, S. Li, G. J. Kirouac, M. Beierlein, and F. H. Do-Monte. 2021. "A hypothalamic-thalamostriatal circuit that controls approach-avoidance conflict in rats."  Nat Commun 12 (1):2517. doi: 10.1038/s41467-021-22730-y.

Esteban Masferrer, M., B. A. Silva, K. Nomoto, S. Q. Lima, and C. T. Gross. 2020. "Differential Encoding of Predator Fear in the Ventromedial Hypothalamus and Periaqueductal Grey."  J Neurosci 40 (48):9283-9292. doi: 10.1523/JNEUROSCI.0761-18.2020.

Fanselow, M. S. 1998. "Pavlovian conditioning, negative feedback, and blocking: mechanisms that regulate association formation."  Neuron 20 (4):625-7. doi: 10.1016/s0896-6273(00)81002-8.

Fields, H. L. 2000. "Pain modulation: expectation, opioid analgesia and virtual pain."  Prog Brain Res 122:245-53. doi: 10.1016/s0079-6123(08)62143-3.

Gross, C. T., and N. S. Canteras. 2012. "The many paths to fear."  Nat Rev Neurosci 13 (9):651-8. doi: 10.1038/nrn3301.

Herry, C., and J. P. Johansen. 2014. "Encoding of fear learning and memory in distributed neuronal circuits."  Nat Neurosci 17 (12):1644-54. doi: 10.1038/nn.3869.

Kim, E. J., O. Horovitz, B. A. Pellman, L. M. Tan, Q. Li, G. Richter-Levin, and J. J. Kim. 2013. "Dorsal periaqueductal gray-amygdala pathway conveys both innate and learned fear responses in rats."  Proc Natl Acad Sci U S A 110 (36):14795-800. doi: 10.1073/pnas.1310845110.

Kim, E. J., M. S. Kong, S. G. Park, S. J. Y. Mizumori, J. Cho, and J. J. Kim. 2018. "Dynamic coding of predatory information between the prelimbic cortex and lateral amygdala in foraging rats."  Sci Adv 4 (4):eaar7328. doi: 10.1126/sciadv.aar7328.

Kim, J. J., J. S. Choi, and H. J. Lee. 2016. "Foraging in the face of fear: Novel strategies for evaluating amygdala functions in rats." In Living without an amygdala, edited by D. G. Amaral and R. Adolphs, 129-148. The Guilford Press.

Kim, J. J., R. A. Rison, and M. S. Fanselow. 1993. "Effects of amygdala, hippocampus, and periaqueductal gray lesions on short- and long-term contextual fear."  Behav Neurosci 107 (6):1093-8. doi: 10.1037//0735-7044.107.6.1093.

Kong, M. S., E. J. Kim, S. Park, L. S. Zweifel, Y. Huh, J. Cho, and J. J. Kim. 2021. "'Fearful-place' coding in the amygdala-hippocampal network."  Elife 10. doi: 10.7554/eLife.72040.

Krout, K. E., and A. D. Loewy. 2000. "Periaqueductal gray matter projections to midline and intralaminar thalamic nuclei of the rat."  J Comp Neurol 424 (1):111-41. doi: 10.1002/1096-9861(20000814)424:1<111::aid-cne9>3.0.co;2-3.

Kunwar, P. S., M. Zelikowsky, R. Remedios, H. Cai, M. Yilmaz, M. Meister, and D. J. Anderson. 2015. "Ventromedial hypothalamic neurons control a defensive emotion state."  Elife 4. doi: 10.7554/eLife.06633.

Lefler, Y., D. Campagner, and T. Branco. 2020. "The role of the periaqueductal gray in escape behavior."  Curr Opin Neurobiol 60:115-121. doi: 10.1016/j.conb.2019.11.014.

Li, Z., J. X. Wei, G. W. Zhang, J. J. Huang, B. Zingg, X. Wang, H. W. Tao, and L. I. Zhang. 2021. "Corticostriatal control of defense behavior in mice induced by auditory looming cues."  Nat Commun 12 (1):1040. doi: 10.1038/s41467-021-21248-7.

Lischinsky, J. E., and D. Lin. 2019. "Looming Danger: Unraveling the Circuitry for Predator Threats."  Trends Neurosci 42 (12):841-842. doi: 10.1016/j.tins.2019.10.004.

Lu, B., P. Fan, M. Li, Y. Wang, W. Liang, G. Yang, F. Mo, Z. Xu, J. Shan, Y. Song, J. Liu, Y. Wu, and X. Cai. 2023. "Detection of neuronal defensive discharge information transmission and characteristics in periaqueductal gray double-subregions using PtNP/PEDOT:PSS modified microelectrode arrays."  Microsyst Nanoeng 9:70. doi: 10.1038/s41378-023-00546-8.

Magierek, V., P. L. Ramos, N. G. da Silveira-Filho, R. L. Nogueira, and J. Landeira-Fernandez. 2003. "Context fear conditioning inhibits panic-like behavior elicited by electrical stimulation of dorsal periaqueductal gray."  Neuroreport 14 (12):1641-4. doi: 10.1097/00001756-200308260-00020.

McNally, G. P., J. P. Johansen, and H. T. Blair. 2011. "Placing prediction into the fear circuit."  Trends Neurosci 34 (6):283-92. doi: 10.1016/j.tins.2011.03.005.

Menegas, W., K. Akiti, R. Amo, N. Uchida, and M. Watabe-Uchida. 2018. "Dopamine neurons projecting to the posterior striatum reinforce avoidance of threatening stimuli."  Nat Neurosci 21 (10):1421-1430. doi: 10.1038/s41593-018-0222-1.

Morgan, M. M., P. K. Whitney, and M. S. Gold. 1998. "Immobility and flight associated with antinociception produced by activation of the ventral and lateral/dorsal regions of the rat periaqueductal gray."  Brain Res 804 (1):159-66. doi: 10.1016/s0006-8993(98)00669-6.

Otchy, T. M., S. B. Wolff, J. Y. Rhee, C. Pehlevan, R. Kawai, A. Kempf, S. M. Gobes, and B. P. Olveczky. 2015. "Acute off-target effects of neural circuit manipulations."  Nature 528 (7582):358-63. doi: 10.1038/nature16442.

Paxinos, G., and C. Watson. 1998. The Rat Brain in Stereotaxic Coordinates. San Diego: Academic Press.

Rajasethupathy, P., S. Sankaran, J. H. Marshel, C. K. Kim, E. Ferenczi, S. Y. Lee, A. Berndt, C. Ramakrishnan, A. Jaffe, M. Lo, C. Liston, and K. Deisseroth. 2015. "Projections from neocortex mediate top-down control of memory retrieval."  Nature 526 (7575):653-9. doi: 10.1038/nature15389.

Ressler, R. L., and S. Maren. 2019. "Synaptic encoding of fear memories in the amygdala."  Curr Opin Neurobiol 54:54-59. doi: 10.1016/j.conb.2018.08.012.

Schenberg, L. C., R. M. Povoa, A. L. Costa, A. V. Caldellas, S. Tufik, and A. S. Bittencourt. 2005. "Functional specializations within the tectum defense systems of the rat."  Neurosci Biobehav Rev 29 (8):1279-98. doi: 10.1016/j.neubiorev.2005.05.006.

Silva, B. A., C. Mattucci, P. Krzywkowski, E. Murana, A. Illarionova, V. Grinevich, N. S. Canteras, D. Ragozzino, and C. T. Gross. 2013. "Independent hypothalamic circuits for social and predator fear."  Nat Neurosci 16 (12):1731-3. doi: 10.1038/nn.3573.

Tsang, E., C. Orlandini, R. Sureka, A. H. Crevenna, E. Perlas, I. Prankerd, M. E. Masferrer, and C. T. Gross. 2023. "Induction of flight via midbrain projections to the cuneiform nucleus."  PLoS One 18 (2):e0281464. doi: 10.1371/journal.pone.0281464.

Vianna, D. M., and M. L. Brandao. 2003. "Anatomical connections of the periaqueductal gray: specific neural substrates for different kinds of fear."  Braz J Med Biol Res 36 (5):557-66. doi: 10.1590/s0100-879x2003000500002.

Walker, D. L., and M. Davis. 1997. "Involvement of the dorsal periaqueductal gray in the loss of fear-potentiated startle accompanying high footshock training."  Behav Neurosci 111 (4):692-702. doi: 10.1037//0735-7044.111.4.692.

Wang, L., I. Z. Chen, and D. Lin. 2015. "Collateral pathways from the ventromedial hypothalamus mediate defensive behaviors."  Neuron 85 (6):1344-58. doi: 10.1016/j.neuron.2014.12.025.

Wei, P., N. Liu, Z. Zhang, X. Liu, Y. Tang, X. He, B. Wu, Z. Zhou, Y. Liu, J. Li, Y. Zhang, X. Zhou, L. Xu, L. Chen, G. Bi, X. Hu, F. Xu, and L. Wang. 2015. "Processing of visually evoked innate fear by a non-canonical thalamic pathway."  Nat Commun 6:6756. doi: 10.1038/ncomms7756.

Yeh, L. F., T. Ozawa, and J. P. Johansen. 2021. "Functional organization of the midbrain periaqueductal gray for regulating aversive memory formation."  Mol Brain 14 (1):136. doi: 10.1186/s13041-021-00844-0.

Yilmaz, M., and M. Meister. 2013. "Rapid innate defensive responses of mice to looming visual stimuli."  Curr Biol 23 (20):2011-5. doi: 10.1016/j.cub.2013.08.015.

Zhou, Z., X. Liu, S. Chen, Z. Zhang, Y. Liu, Q. Montardy, Y. Tang, P. Wei, N. Liu, L. Li, R. Song, J. Lai, X. He, C. Chen, G. Bi, G. Feng, F. Xu, and L. Wang. 2019. "A VTA GABAergic Neural Circuit Mediates Visually Evoked Innate Defensive Responses."  Neuron 103 (3):473-488 e6. doi: 10.1016/j.neuron.2019.05.027.

Author response:

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

SARS-CoV-2 infection induces syncytia formation, which promotes viral transmission. In this paper, the authors aimed to understand how host-derived inflammatory cytokines IL-1α/β combat SARS-CoV-2 infection.

Strengths:

First, they used a cell-cell fusion assay developed previously to identify IL-1α/β as the cytokines that inhibit syncytia formation. They co-cultured cells expressing the spike protein and cells expressing ACE2 and found that IL-1β treatment decreased syncytia formation and S2' cleavage.

Second, they investigated the IL-1 signaling pathway in detail, using knockouts or pharmacological perturbation to understand the signaling proteins responsible for blocking cell fusion. They found that IL-1 prevents cell-cell fusion through MyD88/IRAK/TRAF6 but not TAK1/IKK/NF-κB, as only knocking out MyD88/IRAK/TRAF6 eliminates the inhibitory effect on cell-cell fusion in response to IL-1β. This revealed that the inhibition of cell fusion did not require a transcriptional response and was mediated by IL-1R proximal signaling effectors.

Third, the authors identified RhoA/ROCK activation by IL-1 as the basis for this inhibition of cell fusion. By visualizing a RhoA biosensor and actin, they found a redistribution of RhoA to the cell periphery and cell-cell junctions after IL-1 stimulation. This triggered the formation of actin bundles at cell-cell junctions, preventing fusion and syncytia formation. The authors confirmed this molecular mechanism by using constitutively active RhoA and an inhibitor of ROCK.

Diverse Cell types and in vivo models were used, and consistent results were shown across diverse models. These results were convincing and well-presented.

Weaknesses:

As the authors point out in the discussion, whether IL-1-mediated RhoA activation is specific to viral infection or regulates other RhoA-regulated processes is unclear. We would also require high-magnification images of the subcellular organization of the cytoskeleton to appreciate the effect of IL-1 stimulation.

Thanks for the suggestions. We tested the role of IL-1β in other RhoA-regulated processes, and found that IL-1β-mediated RhoA activation also reduced cell migration in a cell scratch assay (see Author response image 1). We also provided high-magnification images in the revised Figures 4 and 5, as well as their respective figure supplements.

Author response image 1.

(A) Cell scratch assay images of HEK293T cells treated with PBS or IL-1β. (B) Quantification of cell migration in (A).

Reviewer #2 (Public Review):

Summary:

In this study, Zheng et al investigated the role of inflammatory cytokines in protecting cells against SARS-CoV-2 infection. They demonstrate that soluble factors in the supernatants of TLR-stimulated THP1 cells reduce fusion events between HEK293 cells expressing SARS-CoV-2 S protein and the ACE2 receptor. Using qRT-PCR and ELISA, they demonstrate that IL-1 cytokines are (not surprisingly) upregulated by TLR treatment in THP1 cells. Further, they convincingly demonstrate that recombinant IL-1 cytokines are sufficient to reduce cell-to-cell fusion mediated by the S protein. Using chemical inhibitors and CRISPR knock-out of key IL-1 receptor signaling components in HEK293 cells, they demonstrate that components of the myddosome (MYD88, IRAK1/4, and TRAF6) are required for fusion inhibition, but that downstream canonical signaling (i.e., TAK1 and NFKB activation) is not required. Instead, they provide evidence that IL-1-dependent non-canonical activation of RhoA/Rock is important for this phenotype. Importantly, the authors demonstrate that expression of a constitutively active RhoA alone is sufficient to inhibit fusion and that chemical inhibition of Rock could reverse this inhibition. The authors followed up these in vitro experiments by examining the effects of IL-1 on SARS-COV-2 infection in vivo and they demonstrate that recombinant IL-1 can reduce viral burden and lung pathogenesis in a mouse model of infection. However, the contribution of the RhoA/Rock pathway and inhibition of fusion to IL-1-mediated control of SARS-CoV-2 infection in vivo remains unclear.

Strengths:

(1) The bioluminescence cell-cell fusion assay provides a robust quantitative method to examine cytokine effects on viral glycoprotein-mediated fusion.

(2) The study identifies a new mechanism by which IL-1 cytokines can limit virus infection.

(3) The authors tested IL-1 mediated inhibition of fusion induced by many different coronavirus S proteins and several SARS-CoV-2 strains.

Weaknesses:

(1) The qualitative assay demonstrating S2 cleavage and IL-1 mediated inhibition of this phenotype is extremely variable across the data figures. Sometimes it appears like S2 cleavage (S2') is reduced, while in other figures immunoblots show that total S2 protein is decreased. Based on the proposed model the expectation would be that S2 abundance would be rescued when cleavage is inhibited.

In our present manuscript, IL-1-mediated changes of the full-length spike showed some variation between authentic SARS-CoV-2 infection model and HEK293T-S + HEK293T-ACE2 coculture model, while IL-1 inhibited S2’ cleavage accompanied by a reduction of S2 subunit in both models.

In the authentic SARS-CoV-2 infection model, we observed that IL-1 inhibited S2' cleavage accompanied with a reduction in both S2 subunit and full-length spike protein. This is likely because the S2 subunit and full-length spike protein in this model are not only from infected cells, but also from intracellular viral particles. IL-1 inhibited SARS-CoV-2 induced cell-cell fusion and reduced the viral load in host cells, therefore the abundance of S2 subunit and full-length spike proteins were both reduced.

In the HEK293T-based co-culture model, IL-1 inhibited S2' cleavage accompanied with a reduction in S2 subunit, while the full-length spike protein was more or less rescued. Based on our previous study, R685A and ΔRRAR spike mutants cannot generate the S2 subunit, but still generated S2′ fragment to induce cell-cell fusion, and the S2' fragment produced from R685A and ΔRRAR spike mutants were only slightly reduced compared to wild-type spike protein, suggesting that the S2' fragment is mainly derived from the full-length spike directly, and to a minimal extent from the S2 subunit (Fig. 4B and 4G, PMID: 34930824). Thus, inhibition of S2’ cleavage by IL-1 mainly rescued the full-length spike protein.

(2) The text referencing Figure 1H suggests that TLR-stimulated THP-1 cell supernatants "significantly" reduce syncytia, but image quantification and statistics are not provided to support this statement.

Thanks for pointing out this issue. We have provided fluorescence image quantification and statistics in the revised version of our manuscript (Figure 1D, Figure 1-figure supplement 1A, Figure 1H-1I, Figure 2H-2I, Figure 1-figure supplement 1D-1E, Figure 1-figure supplement 1H-1I, Figure 2-figure supplement 1C-1D, Figure 2-figure supplement 2B-2E, Figure 2-figure supplement 2G-2H, Figure 2-figure supplement 6A-6B, Figure 2-figure supplement 7F-7G).

(3) The authors conclude that because IL-1 accumulates in TLR2-stimulated THP1 monocyte supernatants, this cytokine accounts for the ability of these supernatants to inhibit cell-cell fusion. However, they do not directly test whether IL-1 is required for the phenotype. Inhibition of the IL-1 receptor in supernatant-treated cells would help support their conclusion.

Thanks for the suggestion. Accordingly, we performed experiment and found that IL-1RA treatment reduced the inhibitory effect of PGN-stimulated THP-1 cell culture supernatant on cell-cell fusion, suggesting that IL-1 is required for the inhibition. This result has been added in our revised manuscript (Figure 2J and Figure2-figure supplement 4C).

(4) Immunoblot analysis of IL-1 treated HEK293 cells suggests that this cytokine does not reduce the abundance of ACE2 or total S protein in cells. However, it is possible that IL-1 signaling reduces the abundance of these proteins on the cell surface, which would result in a similar inhibition of cell-cell fusion. The authors should confirm that IL-1 treatment of their cells does not change Ace2 or S protein on the cell surface.

Thanks for the suggestion. Accordingly, we applied Wheat Germ Agglutinin (WGA) to stain cell surface in HKE293T cells and observed that IL-1β treatment did not change ACE2 or Spike protein on the cell surface. This result has been added in our revised manuscript (Figure 5-figure supplement 3A-D).

(5) In Figure 5A, expression of constitutively active RhoA appears to have profound effects on how ACE2 runs by SDS-PAGE, suggesting that RhoA may have additional effects on ACE2 biology that might account for the decreased cell-cell fusion. This phenotype should be addressed in the text and explored in more detail.

Thanks for pointing out this. We also noticed that the occurrence of cell-cell fusion reduced the amount of ACE2, whereas inhibition of cell-cell fusion restored the ACE2 abundance. Take the original Figure 5A (revised Figure 4-figure supplement 2B) as example, the increased ACE2 protein should be attributed to the decreased cell-cell fusion upon RhoA-CA transfection, as Spike binding with ACE2 leads to clathrin- and AP2-dependent endocytosis, resulting in ACE2 degradation in the lysosome (PMID: 36287912).

In addition, we have examined the potential effect of RhoA-CA on ACE2, and found that RhoA-CA did not affect ACE2 expression, nor Spike binding to ACE2 (revised Figure 5-figure supplement 2E); it did not affect ACE2 distribution on cell surface either (revised Figure 5-figure supplement 2F and G).

(6) The experiments linking IL-1 mediated restriction of SARS-COV-2 fusion to the control of virus infection in vivo are incomplete. The reported data demonstrate that recombinant IL-1 can restrict virus replication in vivo, but they fall short of confirming that the in vitro mechanism described (reduced fusion) contributes to the control of SARS-CoV2 replication in vivo. A critical piece of data that is missing is the demonstration that the ROCK inhibitor phenocopies IL-1RA treatment of SARS-COV-2 infected mice (viral infection and pathology).

Thanks for this suggestion. Accordingly, we applied the ROCK inhibitor in vivo to confirm its role in SARS-CoV-2-infected mice, and found similar phenotype as the IL-1RA treatment experiment. That is to say, Y-26732 treatment prevented the formation of IL-1β-induced actin bundles at cell-cell junctions, thus promoted syncytia formation and further viral transmission in vivo (revised Figure 7).

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

I suggest providing single-channel images in a supplementary figure for the live-cell images in Figures 4 and 5. Higher magnification images would also help distinguish the subcellular details of the cytoskeleton organization.

Thanks for the suggestion. We have provided the single channel images and higher magnification images in the revised Figures 4 and 5, as well as their respective figure supplements.

In Figure 4, the authors showed that IL-1 activates RhoA and induces the accumulation of activated RhoA at the cell-cell junctions. They also showed that IL-1 promotes the formation of actin bundles at cell-cell junctions. However, the authors have not shown any connection between RhoA and actin yet, but in lines 263-264, they claim that actin bundle formation is induced by RhoA. Evidence for this part was shown in later results, but at this moment, it is lacking. The same applies to lines 282-284; I think this conclusion that IL-1-induced actin bundle formation is through the RhoA-ROCK pathway should come after showing how RhoA affects actin bundle formation at cell-cell junctions. To this end, I suggest moving Supplementary Figures S12B and S12D to the main figure, as they provide strong evidence of the IL-1-RhoA-ROCK-actin pathway.

We appreciate these valuable comments. As suggested, we have moved the respective supplementary figures to the main figures to support our findings in the revised manuscript (Figure 4E and Figure 4-figure supplement 2B; Figure 5C and Figure 5-figure supplement 2A), the text has also been adjusted accordingly.

Author response:

The following is the authors’ response to the previous reviews.

eLife assessment

This important study advances our understanding of how past and future information is jointly considered in visual working memory by studying gaze biases in a memory task that dissociates the locations during encoding and memory tests. The evidence supporting the conclusions is convincing, with state-of-the-art gaze analyses that build on a recent series of experiments introduced by the authors. This work, with further improvements incorporating the existing literature, will be of broad interest to vision scientists interested in the interplay of vision, eye movements, and memory.

We thank the Editors and the Reviewers for their enthusiasm and appreciation of our task, our findings, and our article. We also wish to thank the Reviewers for their constructive comments that we have embraced to improve our article. Please find below our point-by-point responses to this valuable feedback, where we also state relevant revisions that we have made to our article.

In addition, please note that we have now also made our data and code publicly available.

Reviewer 1, Comments:

In this study, the authors offer a fresh perspective on how visual working memory operates. They delve into the link between anticipating future events and retaining previous visual information in memory. To achieve this, the authors build upon their recent series of experiments that investigated the interplay between gaze biases and visual working memory. In this study, they introduce an innovative twist to their fundamental task. Specifically, they disentangle the location where information is initially stored from the location where it will be tested in the future. Participants are tasked with learning a novel rule that dictates how the initial storage location relates to the eventual test location. The authors leverage participants' gaze patterns as an indicator of memory selection. Intriguingly, they observe that microsaccades are directed toward both the past encoding location and the anticipated future test location. This observation is noteworthy for several reasons. Firstly, participants' gaze is biased towards the past encoding location, even though that location lacks relevance to the memory test. Secondly, there's a simultaneous occurrence of an increased gaze bias towards both the past and future locations. To explore this temporal aspect further, the authors conduct a compelling analysis that reveals the joint consideration of past and future locations during memory maintenance. Notably, microsaccades biased towards the future test location also exhibit a bias towards the past encoding location. In summary, the authors present an innovative perspective on the adaptable nature of visual working memory. They illustrate how information relevant to the future is integrated with past information to guide behavior.

Thank you for your enthusiasm for our article and findings as well as for your constructive suggestions for additional analyses that we respond to in detail below.

This short manuscript presents one experiment with straightforward analyses, clear visualizations, and a convincing interpretation. For their analysis, the authors focus on a single time window in the experimental trial (i.e., 0-1000 ms after retro cue onset). While this time window is most straightforward for the purpose of their study, other time windows are similarly interesting for characterizing the joint consideration of past and future information in memory. First, assessing the gaze biases in the delay period following the cue offset would allow the authors to determine whether the gaze bias towards the future location is sustained throughout the entire interval before the memory test onset. Presumably, the gaze bias towards the past location may not resurface during this delay period, but it is unclear how the bias towards the future location develops in that time window. Also, the disappearance of the retro cue constitutes a visual transient that may leave traces on the gaze biases which speaks again for assessing gaze biases also in the delay period following the cue offset.

Thank you for raising this important point. We initially focused on the time window during the cue given that our central focus was on gaze-biases associated with mnemonic item selection. By zooming in on this window, we could best visualize our main effects of interest: the joint selection (in time) of past and future memory attributes.

At the same time, we fully agree that examining the gaze biases over a more extended time window yields a more comprehensive view of our data. To this end, we have now also extended our analysis to include a wider time range that includes the period between cue offset (1000 ms after cue onset) and test onset (1500 ms after cue onset). We present these data below. Because we believe our future readers are likely to be interested in this as well, we have now added this complementary visualization as Supplementary Figure 4 (while preserving the focus in our main figure on the critical mnemonic selection period of interest).

Author response image 1.

Supplementary Figure 4. Gaze biases in extended time window as a complement to Figure 1 and Supplementary Figure 2. This extended analysis reveals that while the gaze bias towards the past location disappears around 600 ms after cue onset, the gaze bias towards the future location persists (panel a) and that while the early (joint) future bias occurs predominantly in the microsaccade range below 1 degree visual angle, the later bias to the future location incorporates larger eye movement that likely involve preparing for optimally perceiving the anticipated test stimulus (panel b).

This extended analysis reveals that while the gaze bias towards the past location disappears around 600 ms after cue onset (consistent with our prior reports of this bias), the gaze bias towards the future location persists. Moreover, as revealed by the data in panel b above, while the early (joint) future bias occurs predominantly in the microsaccade range below 1 degree visual angle, the later bias to the future location incorporates larger eye movement that likely involve preparing for optimally perceiving the anticipated test stimulus.

We now also call out these additional findings and figure in our article:

Page 2 (Results): “Gaze biases in both axes were driven predominantly by microsaccades (Supplementary Fig. 2) and occurred similarly in horizontal-to-vertical and vertical-tohorizontal trials (Supplementary Fig. 3). Moreover, while the past bias was relatively transient, the future bias continued to increase in anticipation of the of the test stimulus and increasingly incorporated eye-movements beyond the microsaccade range (see Supplementary Fig. 4 for a more extended time range)”.

Moreover, assessing the gaze bias before retro-cue onset allows the authors to further characterize the observed gaze biases in their study. More specifically, the authors could determine whether the future location is considered already during memory encoding and the subsequent delay period (i.e., before the onset of the retro cue). In a trial, participants encode two oriented gratings presented at opposite locations. The future rule indicates the test locations relative to the encoding locations. In their example (Figure 1a), the test locations are shifted clockwise relative to the encoding location. Thus, there are two pairs of relevant locations (each pair consists of one stimulus location and one potential test location) facing each other at opposite locations and therefore forming an axis (in the illustration the axis would go from bottom left to top right). As the future rule is already known to the participants before trial onset it is possible that participants use that information already during encoding. This could be tested by assessing whether more microsaccades are directed along the relevant axis as compared to the orthogonal axis. The authors should assess whether such a gaze bias exists already before retro cue onset and discuss the theoretical consequences for their main conclusions (e.g., is the future location only jointly used if the test location is implicitly revealed by the retro cue).

Thank you – this is another interesting point. We fully agree that additional analysis looking at the period prior to retrocue onset may also prove informative. In accordance with the suggested analysis, we have therefore now also analysed the distribution of saccade directions (including in the period from encoding to retrocue) as a function of the future rule (presented below, and now also included as Supplementary Fig. 5). Complementary recent work from our lab has shown how microsaccade directions can align to the axis of memory contents during retention (see de Vries & van Ede, eNeuro, 2024). Based on this finding, one may predict that if participants retain the items in a remapped fashion, their microsaccades may align with the axis of the future rule, and this could potentially already happen prior to cue onset.

These complementary analyses show that saccade directions are predominantly influenced by the encoding locations rather than the test locations, as seen most clearly by the saccade distribution plots in the middle row of the figure below. To obtain time-courses, we categorized saccades as occurring along the axis of the future rule or along the orthogonal axis (bottom row of the figure below). Like the distribution plots, these time course plots also did not reveal any sign of a bias along the axis of the future rule itself.

Importantly, note how this does not argue against our main findings of joint selection of past and future memory attributes, as for that central analysis we focused on saccade biases that were specific to the selected memory item, whereas the analyses we present below focus on biases in the axes in which both memory items are defined; not only the cued/selected memory item.

Author response image 2.

Supplementary Figure 5. Distribution of saccade directions relative to the future rule from encoding onset. (Top panel) The spatial layouts in the four future rules. (Middle panel) Polar distributions of saccades during 0 to 1500 ms after encoding onset (i.e., the period between encoding onset and cue onset). The purple quadrants represent the axis of the future rule and the grey quadrants the orthogonal axis. (Bottom panel) Time courses of saccades along the above two axes. We did not observe any sign of a bias along the axis of the future rule itself.

We agree that these additional results are important to bring forward when we interpret our findings. Accordingly, we now mention these findings at the relevant section in our Discussion:

Page 5 (Discussion): “First, memory contents could have directly been remapped (cf. 4,24–26) to their future-relevant location. However, in this case, one may have expected to exclusively find a future-directed gaze bias, unlike what we observed. Moreover, using a complementary analysis of saccade directions along the axis of the future rule (cf. 24), we found no direct evidence for remapping in the period between encoding and cue (Supplementary Fig. 5)”.

Reviewer 2, Comments:

The manuscript by Liu et al. reports a task that is designed to examine the extent to which "past" and "future" information is encoded in working memory that combines a retro cue with rules that indicate the location of an upcoming test probe. An analysis of microsaccades on a fine temporal scale shows the extent to which shifts of attention track the location of the location of the encoded item (past) and the location of the future item (test probe). The location of the encoded grating of the test probe was always on orthogonal axes (horizontal, vertical) so that biases in microsaccades could be used to track shifts of attention to one or the other axis (or mixtures of the two). The overall goal here was then to (1) create a methodology that could tease apart memory for the past and future, respectively, (2) to look at the time-course attention to past/future, and (3) to test the extent to which microsaccades might jointly encode past and future memoranda. Finally, some remarks are made about the plausibility of various accounts of working memory encoding/maintenance based on the examination of these time courses.

Strengths:

This research has several notable strengths. It has a clear statement of its aims, is lucidly presented, and uses a clever experimental design that neatly orthogonalizes "past" and "future" as operationalized by the authors. Figure 1b-d shows fairly clearly that saccade directions have an early peak (around 300ms) for the past and a "ramping" up of saccades moving in the forward direction. This seems to be a nice demonstration the method can measure shifts of attention at a fine temporal resolution and differentiate past from future-oriented saccades due to the orthogonal cue approach. The second analysis shown in Figure 2, reveals a dependency in saccade direction such that saccades toward the probe future were more likely also to be toward the encoded location than away from the encoded direction. This suggests saccades are jointly biased by both locations "in memory".

Thank you for your overall appreciation of our work and for highlighting the above strengths. We also thank you for your constructive comments and call for clarifications that we respond to below.

Weaknesses:

(1) The "central contribution" (as the authors characterize it) is that "the brain simultaneously retains the copy of both past and future-relevant locations in working memory, and (re)activates each during mnemonic selection", and that: "... while it is not surprising that the future location is considered, it is far less trivial that both past and future attributes would be retained and (re)activated together. This is our central contribution." However, to succeed at the task, participants must retain the content (grating orientation, past) and probe location (future) in working memory during the delay period. It is true that the location of the grating is functionally irrelevant once the cue is shown, but if we assume that features of a visual object are bound in memory, it is not surprising that location information of the encoded object would bias processing as indicated by microsaccades. Here the authors claim that joint representation of past and future is "far less trivial", this needs to be evaluaed from the standpoint of prior empirical data on memory decay in such circumstances, or some reference to the time-course of the "unbinding" of features in an encoded object.

Thank you. We agree that our participants have to use the future rule – as otherwise they do not know to which test stimulus they should respond. This was a deliberate decision when designing the task. Critically, however, this does not require (nor imply) that participants have to incorporate and apply the rule to both memory items already prior to the selection cue. It is at least as conceivable that participants would initially retain the two items at their encoded (past) locations, then wait for the cue to select the target memory item, and only then consider the future location associated with the target memory item. After all, in every trial, there is only 1 relevant future location: the one associated with the cued memory item. The time-resolved nature of our gaze markers argues against such a scenario, by virtue of our observation of the joint (simultaneous) consideration of past and future memory attributes (as opposed to selection of past-before-future). These temporal dynamics are central to the insights provided by our study.

In our view, it is thus not obvious that the rule would be applied at encoding. In this sense, we do not assume that the future location is part of both memory objects from encoding, but rather ask whether this is the case – and, if so, whether the future location takes over the role of the past location, or whether past and future locations are retained jointly.

Our statements regarding what is “trivial” and what is “less trivial” regard exactly this point: it is trivial that the future is considered (after all, our task demanded it). However, it is less trivial that (1) the future location was already available at the time of initial item selection (as reflected in the simultaneous engagement of past and future locations), and (2) that in presence of the future location, the past location was still also present in the observed gaze biases.

Having said that, we agree that an interesting possibility is that participants remap both memory items to their future-relevant locations ahead of the cue, but that the past location is not yet fully “unbound” by the time of the cue. This may trigger a gaze bias not only to the new future location but also to the “sticky” (unbound) past location. We now acknowledge this possibility in our discussion (also in response to comment 3 below) where we also suggest how future work may be able to tap into this:

Page 6 (Discussion): “In our study, the past location of the memory items was technically irrelevant for the task and could thus, in principle, be dropped after encoding. One possibility is that participants remapped the two memory items to their future locations soon after encoding, and had started – but not finished – dropping the past location by the time the cue arrived. In such a scenario, the past signal is merely a residual trace of the memory items that serves no purpose but still pulls gaze. Alternatively, however, the past locations may be utilised by the brain to help individuate/separate the two memory items. Moreover, by storing items with regard to multiple spatial frames (cf. 37) – here with regard to both past and future visual locations – it is conceivable that memories may become more robust to decay and/or interference. Also, while in our task past locations were never probed, in everyday life it may be useful to remember where you last saw something before it disappeared behind an occluder. In future work, it will prove interesting to systematically vary to the delay between encoding and cue to assess whether the reliance on the past location gradually dissipates with time (consistent with dropping an irrelevant feature), or whether the past trace remains preserved despite longer delays (consistent with preserving utility for working memory).”

(2) The authors refer to "future" and "past" information in working memory and this makes sense at a surface level. However, once the retrocue is revealed, the "rule" is retrieved from long-term memory, and the feature (e.g. right/left, top/bottom) is maintained in memory like any other item representation. Consider the classic test of digit span. The digits are presented and then recalled. Are the digits of the past or future? The authors might say that one cannot know, because past and future are perfectly confounded. An alternative view is that some information in working memory is relevant and some is irrelevant. In the digit span task, all the digits are relevant. Relevant information is relevant precisely because it is thought be necessary in the future. Irrelevant information is irrelevant precisely because it is not thought to be needed in the immediate future. In the current study, the orientation of the grating is relevant, but its location is irrelevant; and the location of the test probe is also relevant.

Thank you for this stimulating reflection. We agree that in our set-up, past location is technically “task-irrelevant” while future location is certainly “task-relevant”. At the same time, the engagement of the past location suggests to us that the brain uses past location for the selection – presumably because the brain uses spatial location to help individuate/separate the items, even if encoded locations are never asked about. Therefore, whether something is relevant or irrelevant ultimately depends on how one defines relevance (past location may be relevant/useful for the brain even if technically irrelevant from the perspective of the task). In comparison, the use of “past” and “future” may be less ambiguous.

It is also worth noting how we interpret our findings in relation to demands on visual working memory, inspired by dynamic situations whereby visual stimuli may be last seen at one location but expected to re-appear at another, such as a bird disappearing behind a building (the example in our introduction). Thus, past for us does not refer to the memory item perse (like in the digit span analogue) but, rather, quite specifically to the past location of a dynamic visual stimulus in memory (which, in our experiment, was operationalised by the future rule, for convenience).

(3) It is not clear how the authors interpret the "joint representation" of past and future. Put aside "future" and "past" for a moment. If there are two elements in memory, both of which are associated with spatial bindings, the attentional focus might be a spatial average of the associated spatial indices. One might also view this as an interference effect, such that the location of the encoded location attracts spatial attention since it has not been fully deleted/removed from working memory. Again, for the impact of the encoded location to be exactly zero after the retrieval cue, requires zero interference or instantaneous decay of the bound location information. It would be helpful for the authors to expand their discussion to further explain how the results fit within a broader theoretical framework and how it fits with empirical data on how quickly an irrelevant feature of an object can be deleted from working memory.

Thank you also for this point (that is related to the two points above). As we stated in our reply to comment 1 above, we agree that one possibility is that the past location is merely “sticky” and pulls the task-relevant future bias toward the past location. If so, our time courses suggest that such “pulling” occurs only until approximately 600 ms after cue onset, as the past bias is only transient. An alternative interpretation is that the past location may not be merely a residual irrelevant trace, but actually be useful and used by the brain.

For example, the encoded (past) item locations provide a coordinate system in which to individuate/separate the two memory items. While the future locations also provide such a coordinate system, the brain may benefit from holding onto both coordinate systems at the same time, rendering our observation of joint selection in both frames. Indeed, in a recent VR experiment in which we had participants (rather than the items) rotate, we also found evidence for the joint use of two spatial frames, even if neither was technically required for the upcoming task (see Draschkow, Nobre, van Ede, Nature Human Behaviour, 2022). Though highly speculative at this stage, such reliance on multiple spatial frames may make our memories more robust to decay and/or interference. Moreover, while past location was never explicitly probed in our task, in daily life the past location may sometimes (unexpectedly) become relevant, hence it may be useful to hold onto it, just in case. Thus, considering the past location merely as an “irrelevant feature” (that takes time to delete) may not do sufficient justice to the potential roles of retaining past locations of dynamic visual objects held in working memory.

As also stated in response to comment 1 above, we now added these relevant considerations to our Discussion:

Page 5 (Discussion): “In our study, the past location of the memory items was technically irrelevant for the task and could thus, in principle, be dropped after encoding. One possibility is that participants remapped the two memory items to their future locations soon after encoding, and had started – but not finished – dropping the past location by the time the cue arrived. In such a scenario, the past signal is merely a residual trace of the memory items that serves no purpose but still pulls gaze. Alternatively, however, the past locations may be utilised by the brain to help individuate/separate the two memory items. Moreover, by storing items with regard to multiple spatial frames (cf. 37) – here with regard to both past and future visual locations – it is conceivable that memories may become more robust to decay and/or interference. Also, while in our task past locations were never probed, in everyday life it may be useful to remember where you last saw something before it disappeared behind an occluder. In future work, it will prove interesting to systematically vary to the delay between encoding and cue to assess whether the reliance on the past location gradually dissipates with time (consistent with dropping an irrelevant feature), or whether the past trace remains preserved despite longer delays (consistent with preserving utility for working memory).”

Reviewer 3, Comments:

This study utilizes saccade metrics to explore, what the authors term the "past and future" of working memory. The study features an original design: in each trial, two pairs of stimuli are presented, first a vertical pair and then a horizontal one. Between these two pairs comes the cue that points the participant to one target of the first pair and another of the second pair. The task is to compare the two cued targets. The design is novel and original but it can be split into two known tasks - the first is a classic working memory task (a post-cue informs participants which of two memorized items is the target), which the authors have used before; and the second is a classic spatial attention task (a pre-cue signal that attention should be oriented left or right), which was used by numerous other studies in the past. The combination of these two tasks in one design is novel and important, as it enables the examination of the dynamics and overlapping processes of these tasks, and this has a lot of merit. However, each task separately is not new. There are quite a few studies on working memory and microsaccades and many on spatial attention and microsaccades. I am concerned that the interpretation of "past vs. future" could mislead readers to think that this is a new field of research, when in fact it is the (nice) extension of an existing one. Since there are so many studies that examined pre-cues and post-cues relative to microsaccades, I expected the interpretation here to rely more heavily on the existing knowledge base in this field. I believe this would have provided a better context of these findings, which are not only on "past" vs. "future" but also on "working memory" vs. "spatial attention".

Thank you for considering our findings novel and important, while at the same time reminding us of the parallels to prior tasks studying spatial attention in perception and working memory. We fully agree that our task likely engages both attention to the (past) memory item as well as spatial attention to the upcoming (future) test stimulus. At the same time, there is a critical difference in spatial attention for the future in our task compared with ample prior tasks engaging spatial cueing of attention for perception. In our task, the cue never directly cues the future location. Rather, it exclusively cues the relevant memory item. It is the memory item that is associated with the relevant future location, according to the future rule. This integration of the rule-based future location into the memory representation is distinct from classical spatial-attention tasks in which attention is cued directly to a specific location via, for example, a spatial cue such as an arrow.

Thus, if we wish to think about our task as engaging cueing of spatial attention for perception, we have to at least also invoke the process of cueing the relevant location via the appropriate memory item. We feel it is more parsimonious to think of this as attending to both the past and future location of a dynamic visual object in working memory.

If we return to our opening example, when we see a bird disappear behind a building, we can keep in working memory where we last saw it, while anticipating where it will re-appear to guide our external spatial attention. Here too, spatial attention is fully dependent on working-memory content (the bird itself) – mirroring the dynamic semng in our study. Thus, we believe our findings contribute a fresh perspective, while of course also extending established fields. We now contextualize our finding within the literature and clarify our unique contribution in our revised manuscript:

Page 5 (Discussion): “Building on the above, at face value, our task may appear like a study that simply combines two established tasks: tasks using retro-cues to study attention in working memory (e.g.,2,31-33) and tasks using pre-cues to study orienting of spatial attention to an upcoming external stimulus (e.g., 31,32,34–36). A critical difference with common pre-cue studies, however, is that the cue in our task never directly informed the relevant future location. Rather, as also stressed above, the future location was a feature of the cued memory item (according to the future rule), and not of the cue itself. Note how this type of scenario may not be uncommon in everyday life, such as in our opening example of a bird flying behind a building. Here too, the future relevant location is determined by the bird – i.e. the memory content – itself.”

Reviewer 2, Recommendations:

It would be helpful to set up predictions based on existing working memory models. Otherwise, the claim that the joint coding of past/future is "not trivial" is simply asserted, rather than contradicting an existing model or prior empirical results. If the non-trivial aspect is simply the ability to demonstrate the joint coding empirical through a good experimental design, make it clear that this is the contribution. For example, it may be that prevailing models predict exactly this finding, but nobody has been able to demonstrate it cleanly, as the authors do here. So the non-triviality is not that the result contradicts working memory models, but rather relates to the methodological difficulty of revealing such an effect.

Thank you for your recommendation. First, please see our point-by-point responses to the individual comments above, where we also state relevant changes that we have made to our article, and where we clarify what we meant with “non trivial”. As we currently also state in our introduction, our work took as a starting point the framework that working memory is inherently about the past while being for the future (cf. van Ede & Nobre, Annual Review of Psychology, 2023). By virtue of our unique task design, we were able to empirically demonstrate that visual contents in working memory are selected via both their past and their future-relevant locations – with past and future memory attributes being engaged together in time. With “not trivial” we merely intend to make clear that there are viable alternatives than the findings we observed. For example, past could have been replaced by the future, or it could have been that item selection (through its past location) was required before its future-relevant location could be considered (i.e. past-before-future, rather than joint selection as we reported). We outline these alternatives in the second paragraph of our Discussion:

Page 5 (Discussion): “Our finding of joint utilisation of past and future memory attributes emerged from at least two alternative scenarios of how the brain may deal with dynamic everyday working memory demands in which memory content is encoded at one location but needed at another.

First, [….]”

Our work was not motivated from a particular theoretical debate and did not aim to challenge ongoing debates in the working-memory literature, such as: slot vs. resource, active vs. silent coding, decay vs. interference, and so on. To our knowledge, none of these debates makes specific claims about the retention and selection of past and future visual memory attributes – despite this being an important question for understanding working memory in dynamics everyday semngs, as we hoped to make clear by our opening example.

Reviewer 3, Recommendations:

I recommend that the present findings be more clearly interpreted in the context of previous findings on working memory and attention. The task design includes two components - the first (post-cue) is a classic working memory task and the second (the pre-cue) is a classic spatial attention design. Both components were thoroughly studied in the past and this previous knowledge should be better integrated into the present conclusions. I specifically feel uncomfortable with the interpretation of past vs. future. I find this framework to be misleading because it reads like this paper is on a topic that is completely new and never studied before, when in fact this is a study on the interaction between working memory and spatial attention. I recommend the authors minimize this past-future framing or be more explicit in explaining how this new framework relates to the more common terminology in the field and make sure that the findings are not presented in a vacuum, as another contribution to the vibrant field that they are part of.

Thank you for these recommendations. Please also see our point-by-point responses to the individual comments above. Here, we explained our logic behind using the terminology of past vs. future (in addition, see also our response to point 2 or reviewer 2). Here, we also stated relevant changes that we have made to our manuscript to explain how our findings complement – but are also distinct from – prior tasks that used pre-cues to direct spatial attention to an upcoming stimulus. As we explained above, in our task, the cue itself never contained information about the upcoming test location. Rather, the upcoming test location was a property of the memory item (given the future rule). Hence, we referred to this as a “future attribute” of the cued memory item, rather than as the “cued location” for external spatial attention. Still, we agree the future bias likely (also) reflects spatial allocation to the upcoming test array, and we explicitly acknowledge this in our discussion. For example:

Page 5 (Discussion): “This signal may reflect either of two situations: the selection of a future-copy of the cued memory content or anticipatory attention to its the anticipated location of its associated test-stimulus. Either way, by the nature of our experimental design, this future signal should be considered a content-specific memory attribute for two reasons. First, the two memory contents were always associated with opposite testing locations, hence the observed bias to the relevant future location must be attributed specifically to the cued memory content. Second, we cued which memory item would become tested based on its colour, but the to-be-tested location was dependent on the item’s encoding location, regardless of its colour. Hence, consideration of the item’s future-relevant location must have been mediated by selecting the memory item itself, as it could not have proceeded via cue colour directly.”

Page 6 (Discussion): “Building on the above, at face value, our task may appear like a study that simply combines two established tasks: tasks using retro-cues to study attention in working memory (e.g.,2,31-33) and tasks using pre-cues to study orienting of spatial attention to an upcoming external stimulus (e.g., 31,32,34–36). A critical difference with common pre-cue studies, however, is that the cue in our task never directly informed the relevant future location. Rather, as also stressed above, the future location was a feature of the cued memory item (according to the future rule), and not of the cue itself. Note how this type of scenario may not be uncommon in everyday life, such as in our opening example of a bird flying behind a building. Here too, the future relevant location is determined by the bird – i.e. the memory content – itself.”

Author response:

The following is the authors’ response to the original reviews.

Reviewer 1 (Public reviews):

Summary

Howard et al. performed deep mutational scanning on the MC4R gene, using a reporter assay to investigate two distinct downstream pathways across multiple experimental conditions. They validated their findings with ClinVar data and previous studies. Additionally, they provided insights into the application of DMS results for personalized drug therapy and differential ligand responses across variant types.

Strengths

They captured over 99% of variants with robust signals and investigated subtle functionalities, such as pathway-specific activities and interactions with different ligands, by refining both the experimental design and analytical methods.

Weaknesses

While the study generated informative results, it lacks a detailed explanation regarding the input library, replicate correlation, and sequencing depth for a given number of cells. Additionally, there are several questions that it would be helpful for authors to clarify.

(1) It would be helpful to clarify the information regarding the quality of the input library and experimental replicates. Are variants evenly represented in the library? Additionally, have the authors considered using long-read sequencing to confirm the presence of a single intended variant per construct? Finally, could the authors provide details on the correlation between experimental replicates under each condition?

Are variants evenly represented in the library?

We strive to achieve as evenly balanced library as possible at every stage of the DMS process (e.g., initial cloning in E. coli through integration into human cells). Below is a representative plot showing the number of barcodes per amino acid variant at each position in a given ~60 amino acid subregion of MC4R, which highlights how evenly variants are represented at the E. coli cloning stage.

Author response image 1.

We also make similar measurements after the library is integrated into HEK293T cell lines, and see similarly even coverage across all variants, as shown in the plot below:

Author response image 2.

Additionally, have the authors considered using long-read sequencing to confirm the presence of a single intended variant per construct?

We agree long-read sequencing would be an excellent way to confirm that our constructs contain a single intended variant. However, we elected for an alternate method (outlined in more detail in Jones et al. 2020) that leverages multiple layers of validation. First, the oligo chip-synthesized portions of the protein containing the variants are cloned into a sequence-verified plasmid backbone, which greatly decreases the chances of spuriously generating a mutation in a different portion of the protein. We then sequence both the oligo portion and random barcode using overlapping paired end reads during barcode mapping to avoid sequencing errors and to help detect DNA synthesis errors. At this stage, we computationally reject any constructs that have more than one variant. Given this, the vast majority of remaining unintended variants would come from somatic mutations introduced by the E. coli cloning or replication process, which should be low frequency. We have used our in-house full plasmid sequencing method, OCTOPUS, to sample and spot check this for several other DMS libraries we have generated using the same cloning methods. We have found variants in the plasmid backbone in only ~1% of plasmids in these libraries. Our statistical model also helps correct for this by accounting for barcode-specific variation. Finally we believe this provides further motivation for having multiple barcodes per variant, which dilutes the effect of any unintended additional variants.

Finally, could the authors provide details on the correlation between experimental replicates under each condition?

Certainly! In general, the Gs reporter had higher correlation between replicates than the Gq system (r ~ 0.5 vs r ~ 0.4). The plots below, which have been added as a panel to Supplementary Figure 1, show two representative correlations at the RNA-seq stage of read counts for barcodes between the low a-MSH conditions.

We added the following text to reference this panel:

(see Methods > Sequence processing for barcode expression): “The correlation (r) of barcode readcounts between replicates was ~0.5 and ~0.4 for the Gs and Gq assays, respectively (Supplementary Fig. 1E).”

One important advantage of our statistical model is that it’s able to leverage information from barcodes regardless of the number of replicates they appear in.

(2) Since the functional readout of variants is conducted through RNA sequencing, it seems crucial to sequence a sufficient number of cells with adequate sequencing saturation. Could the authors clarify the coverage depth used for each RNA-seq experiment and how this depth was determined? Additionally, how many cells were sequenced in each experiment?

The text has been added in the manuscript as follows:

(in Methods > Running DMS Assays): “Given the seeding density (~17x10⁶ cells per 150 mm replicate dish), time from seeding to collection, and doubling time of HEK293T cells, approximately 25.5x10⁶ cells were collected per replicate. This translates to approximately 30-60x cellular coverage per amino acid variant in each replicate.”

(in Methods > Sequence processing for barcode expression): “Total mapped reads per replicate at the RNA-seq stage were as follows:

- Gs/CRE: 9.1-18.2 million mapped reads, median=12.3

- Gq/UAS: 8.6-24.1 million mapped reads, median=14.5

- Gs/CRE+Chaperone: 6.4-9.5 million mapped reads, median=7.5”

The median read counts per sample per barcode were 8, 10, and 6 reads for Gs/CRE, Gq/UAS, and Gs/CRE+Chaperone assays, respectively. The median number of barcodes per variant across all samples (the “median of medians”) were 56 for Gs/CRE, 28 for Gq/UAS, and 44 for Gs/CRE+Chaperone.”

(3) It appears that the frequencies of individual RNA-seq barcode variants were used as a proxy for MC4R activity. Would it be important to also normalize for heterogeneity in RNA-seq coverage across different cells in the experiment? Variability in cell representation (i.e., the distribution of variants across cells) could lead to misinterpretation of variant effects. For example, suppose barcode_a1 represents variant A and barcode_b1 represents variant B. If the RNA-seq results show 6 reads for barcode_a1 and 7 reads for barcode_b1, it might initially appear that both variants have similar effect sizes. However, if these reads correspond to 6 separate cells each containing 1 copy of barcode_a1, and only 1 cell containing 7 copies of barcode_b1, the interpretation changes significantly. Additionally, if certain variants occupy a larger proportion of the cell population, they are more likely to be overrepresented in RNA sequencing.

We account for this heterogeneity in several ways. First, as shown above (see Response to Reviewer 1, Question 1), we aim to have even representation of variants within our libraries. Second, we utilize compositional control conditions like forskolin or unstimulated conditions to obtain treatment-independent measurements of barcode abundance and, consequently, of mutant-vs-WT effects that are due to compositional rather than biological variability. We expect that variability observed under these controls is due to subtle effects of molecular cloning, gene expression, and stochasticity. Using these controls, we observe that mutant-vs-WT effects are generally close to zero in these normalization conditions (e.g., in untreated Gq, see Supplementary Figure 3) as compared to treated conditions. For example, pre-mature stops behave similar to WT in normalization conditions. This indicates that mutant abundance is relatively homogenous. Where there are barcode-dependent effects on abundance, we can use information from these conditions to normalize that effect. Finally, our mixed-effect model accounts for barcode-specific deviations from the expected mutant effect (e.g., a “high count” barcode consistently being high relative to the mean).

(4) Although the assay system appears to effectively represent MC4R functionality at the molecular level, we are curious about the potential disparity between the DMS score system and physiological relevance. How do variants reported in gnomAD distribute within the DMS scoring system?

Figure 2D shows DMS scores (variant effect on Gs signaling) relative to human population frequency for all MC4R variants reported in gnomAD as of January 8, 2024.

(5) To measure Gq signaling, the authors used the GAL4-VPR relay system. Is there additional experimental data to support that this relay system accurately represents Gq signaling?

The full Gq reporter uses an NFAT response element from the IL-2 promoter to regulate the expression of the GAL4-VPR relay. In this system, the activation of Gq signaling results in the activation of the NFAT response element, and this signal is then amplified by the GAL4-VPR relay. The NFAT response element has been previously well-validated to respond to the activation of Gq signaling (e.g., Boss, Talpade, and Murphy 1996). We will have added this reference to the text (see Results> Assays for disease-relevant mechanisms) to further support the use of the Gq assay.

(6) Identifying the variants responsive to the corrector was impressive. However, we are curious about how the authors confirmed that the restoration of MC4R activity was due to the correction of the MC4R protein itself. Is there a possibility that the observed effect could be influenced by other factors affected by the corrector? When the corrector was applied to the cells, were any expected or unexpected differential gene expression changes observed?

While we do not directly measure whether Ipsen-17 has effects on other signaling processes, previous work has shown that Ipsen-17 treatment does not indirectly alter signaling kinetics such as receptor internalization (Wang et al., 2014). Furthermore, our analysis methods inherently account for this by normalizing variant effects to WT signaling levels. Any observed rescue of a given variant inherently means that the variant is specifically more responsive to Ipsen-17 than WT, and the fact that different variants exhibit different levels of rescue is reassuring that the mechanism is on target to MC4R. Lastly, Ipsen-17 is known to be an antagonist of alpha-MSH activity and is thought to bind directly to the same site on MC4R (Wang et al., 2014).

We have revised text in the Methods section as follows (see Running DMS Assays) to better articulate this : “For chaperone experiments, cells were washed 3x with 10 mL DMEM to remove Ipsen 17 prior to agonist stimulation as it has been shown to be an antagonist of α-MSH activity and is thought to bind directly to the same site on MC4R (Wang et al. 2014).”

(7) As mentioned in the introduction, gain-of-function (GoF) variants are known to be protective against obesity. It would be interesting to see further studies on the observed GoF variants. Do the authors have any plans for additional research on these variants?

We agree this would be an excellent line of inquiry, but due to changes in company priorities we unfortunately do not have any plans for additional research on these variants.

Reviewer 2 (Public reviews):

Overview

In this manuscript, the authors use deep mutational scanning to assess the effect of ~6,600 protein-coding variants in MC4R, a G protein-coupled receptor associated with obesity. Reasoning that current deep mutational scanning approaches are insufficiently precise for some drug development applications, they focus on articulating new, more precise approaches. These approaches, which include a new statistical model and innovative reporter assay, enable them to probe molecular phenotypes directly relevant to the development of drugs that target this receptor with high precision and statistical rigor.

They use the resulting data for a variety of purposes, including probing the relationship between MC4R's sequence and structure, analyzing the effect of clinically important variants, identifying variants that disrupt downstream MC4R signaling via one but not both pathways, identifying loss of function variants are amenable to a corrector drug and exploring how deep mutational scanning data could guide small molecule drug optimization.

Strengths

The analysis and statistical framework developed by the authors represent a significant advance. In particular, the study makes use of barcode-level internally replicated measurements to more accurately estimate measurement noise.

The framework allows variant effects to be compared across experimental conditions, a task that is currently hard to do with rigor. Thus, this framework will be applicable to a large number of existing and future deep mutational scanning experiments.

The authors refine their existing barcode transcription-based assay for GPCR signaling, and develop a clever "relay" new reporter system to boost signaling in a particular pathway. They show that these reporters can be used to measure both gain of function and loss of function effects, which many deep mutational scanning approaches cannot do.

The use of systematic approaches to integrate and then interrogate high-dimensional deep mutational scanning data is a big strength. For example, the authors applied PCA to the variant effect results from reporters for two different MC4R signaling pathways and were able to discover variants that biased signaling through one or the other pathway. This approach paves the way for analyses of higher dimensional deep mutational scans.

The authors use the deep mutational scanning data they collect to map how different variants impact small molecule agonists activate MC4R signaling. This is an exciting idea, because developing small-molecule protein-targeting therapeutics is difficult, and this manuscript suggests a new way to map small-molecule-protein interactions.

Weaknesses

The authors derive insights into the relationship between MC4R signaling through different pathways and its structure. While these make sense based on what is already known, the manuscript would be stronger if some of these insights were validated using methods other than deep mutational scanning.

Likewise, the authors use their data to identify positions where variants disrupt MC4R activation by one small molecule agonist but not another. They hypothesize these effects point to positions that are more or less important for the binding of different small molecule agonists. The manuscript would be stronger if some of these insights were explored further.

Impact

In this manuscript, the authors present new methods, including a statistical framework for analyzing deep mutational scanning data that will have a broad impact. They also generate MC4R variant effect data that is of interest to the GPCR community.

Recommendations for the authors:

(1) Page 7 - the Gq reporter relay system is clever. Could the authors include the original data showing that the simpler design didn't work at all, or at least revise the text to say more precisely what "not suitable due to weak SNR" means?

We added a panel (D) to Supplementary Figure 2 showing that the native NFAT reporter was ~10x weaker than the CRE reporter, and the relay system amplified the NFAT signal to be comparable to the CRE reporter:

(2) Page 7 - Even though the relay system gives some signal, it's clearly less sensitive/higher background than Gs. How does that play out in the quantitative analysis?

—AND—

(4) Page 10 - The Gq library had fewer barcodes per variant, and, as noted above, the Gq reporter doesn't work quite as well as the Gs one. It would be nice if the authors could comment on how these aspects of the Gq experiments affected data quality/power to detect effects.

Due to the reviewer's excellent suggestion, we updated Supplementary Figure 2B to better contextualize the quantitative effects of the difference in signal to noise ratio of the Gq versus the Gs reporter system (see changes below). These distributions show the Z-statistic for testing either each stop mutation (red) or all possible coding variants against WT. Thus, a |Z| > 1.96 corresponds to a p = 0.05 in a two-sided Wald Test. We can see that in the Gs reporter, 95% of the stops are nominally significantly different from WT (visualized above with the majority of the red distribution being < -1.96). Alternatively, only 64% of stops are nominally significantly different from WT in Gq. This implies that it will be more difficult to detect effects in the Gq system, especially those less severe than stops.

In addition to the overall signal to noise ratio being less in the Gq system, there were also less barcodes per variant (28 vs 56 barcodes per variant on average for Gq vs Gs). As demonstrated in Supplementary Figure 2C, the error bars on our estimates are related to the number of barcodes per variant (Standard Error ~ 1 / sqrt(Number of Barcodes), as shown in the plot below). This suggests that our estimates of mutant effects will be less certain in the Gq library than the Gs library. For example, the average standard error in the Gq library was 0.260 which was ~1.58 times larger than the Gs library's 0.165. Finally, we believe this further reiterates the power of our statistical framework, as it naturally enables formalized hypothesis testing that takes these errors into account when making comparisons both within reporters and across reporters.

(3) Page 9 - it would be nice to see the analysis framework applied to a few existing datasets from other types of assays, to really judge its performance. That's not the main point of this paper, and it's fine, but it would be lovely!

We agree with the reviewer and hope others apply our framework to their problems to further refine its utility and applicability! To that end, we’ve open-sourced it under a permissive license to help encourage the community to use it. Part of the challenge in applying it to other existing datasets is that few DMS experiments leverage variant-level replication through barcodes. While we re-analyzed an older DMS data from Jones et al. 2020 to produce the distributions in Supplementary Figure 2b, a more thorough comparison is outside the scope of this paper. That said, we have two additional manuscripts in preparation that leverage this framework to analyze DMS data in different proteins and assay types.

(5) Page 10 - In discussing the relationship of the data to ClinVar and AM, the authors use qualitative comparisons like "majority" and "typically." Just giving numbers would better help the reader appreciate how the data compare.

We added specific proportions for these statements to the text for the ClinVar and AlphaMissense comparisons as follows:

(See Results > Comprehensive Deep Mutational Scanning of MC4R): “For example, the majority (63.3%, 31/49) of human MC4R variants classified as pathogenic or likely pathogenic in ClinVar (Landrum et al., 2014) lead to a significant reduction of Gs signaling under low α-MSH stimulation conditions (significance threshold: false discovery rate (FDR) < 1%; Fig. 2C). Variants that are significantly loss-of-function in this condition are rarer in the human population, and more common human variants have no significant effect on MC4R function (significance threshold: FDR < 1%; Fig. 2D). Loss-of-function variants by our DMS assay are also typically (e.g., AlphaMissense: 93.4%, 1894/2028) predicted to be deleterious by commonly used variant effect predictors like AlphaMissense (Cheng et al., 2023) and popEVE (Orenbuch et al., 2023) (Supplementary Fig. 5).”

(6) Pages 10-12, Figures 2C, E. The data look really nice, but the correlation with clinvar and the Huang data is not perfect (e.g. many pathogenic variants are classified as WT and partial LoF variants too). Can the authors comment on this discrepancy? For ClinVar, they should say when ClinVar was accessed and also how they filtered variants. I would recommend using variants with at least 1 star. Provided they did use high-quality clinical classifications, do they think the classifications are wrong, or their data? The same goes for Huang.

—AND—

(7) Page 13 - similar to previous comments, I'm curious about the 5 path/likely path ClinVar variants that are not LoF in the assay. Are they high noise/fewer barcodes? Or does the assay just miss some aspect of human biology?

ClinVar data was accessed on January 5, 2024 (see Methods: Comparison to human genetics data and variant effect predictors). No annotation quality filtering was performed, and we have revised the text as follows to clarify this:

(see Methods > Comparison to Human Genetics Data and Variant Effect Predictors): “Pathogenicity classifications of MC4R missense and nonsense variants were obtained from ClinVar (Landrum et al., 2014) on January 5, 2024, and all available annotations were included in the analysis regardless of ClinVar review status metric.”

A substantial proportion of the discrepancy between our data and ClinVar is, as the reviewer suggests, likely due to low quality ClinVar annotations. Of the five variants that the reviewer notes were reported as pathogenic/likely pathogenic but did not result in loss of protein function in any of our DMS assays, two (V50M and V166I) have been reclassified in ClinVar to uncertain or conflicting interpretation since we accessed annotations in early 2024. An additional two of the five discrepant variants (Q43K and S58C) currently have 0 star ratings to support their pathogenic/likely pathogenic annotation. The remaining discrepant variant (S94N) has a 1 star rating supporting an annotation of “likely pathogenic.

The Huang et al. paper did an admirably thorough job of aggregating variant annotations from more than a dozen primary literature sources that each reported functional validation data for small panels of variants. However, one inherent limitation of this approach is that the resulting annotation classes are based on experiments that were carried out using inconsistent methods and/or scoring criteria. For example, classifications in the Huang et al. paper are based on an inconsistent mix of functional assay types (e.g., Gs signaling, Gq signaling, protein cell surface expression, etc.), and different variants were tested in different cell types (e.g., HEK293T, CHO, Cos-7, etc.). In principle, DMS assays should provide a more accurate assessment of the relative quantitative differences between alleles since each variant was tested using identical experimental conditions and analysis parameters.

That being said, while very good, our assays are likely missing or only indirectly reporting on at least some aspects of MC4R biology. For example, in addition to Gs and Gq signaling, MC4R interfaces with β-arrestin. Variants that are protective against obesity-related phenotypes have been shown to increase recruitment of β-arrestin to MC4R, and we did not directly assess this function.

(8) Page 15, Fig 3C - The three variants they highlight all have paradoxical changes in bias as a-MSH dose is increased (e.g. the bias inverts). I'm not a GPCR expert, but this seems interesting and a little weird. Perhaps the authors could comment on it?

We agree this is an interesting observation that deserves further study, but unfortunately is outside the scope of our priorities at the moment. As noted, all three highlighted variants in this region have a biased basal activity, and this bias inverts upon stimulation. While we don’t have a good explanation for why this would be the case, this phenomenon has been previously observed for 158R (Paisdzior et al., 2020). Our DMS data emphasizes how diverse biased effects can be and further highlights the importance of characterizing these effects. It would be interesting if further studies could elucidate the mechanistic basis for this behavior and how it may be related to G protein coupling in this region.

(9) Page 16 - I'm not familiar with the A21x1 formalism. For the general reader, maybe the authors could introduce this formalism.

Given the shared structural topology of GPCRs, others have developed a variety of numbering schemes to refer to where various variants are to allow more direct comparisons between different GPCRs. We use the GPCRDB.org numbering scheme (e.g., F202^5x4) as it takes experimentally determined structures into account. Roughly speaking, the number preceding the “x” corresponds to which transmembrane domain (one through seven) or region the residue is located in. The numbers following the “x” correspond to where that residue is located in that region relative to a structurally conserved residue that is always assigned 50. For example F202^5x48 means that F202 is located in the 5th transmembrane helix and is 2 residues before the most conserved M204^5x50. We updated the text to clarify this accordingly:

(see Results > Structural Insights into Biased Signaling): “Upon ligand binding, W258 (W258^6x48 in https://gpcrdb.org/ nomenclature, where 6 corresponds to the 6th transmembrane helix and 48 denotes 258 is 2 residues before the most conserved residue in that helix (Isberg et al., 2015)) of the conserved CWxP motif undergoes a conformational rearrangement that is translated to L133^3x36 and I137^3x40, of the conserved PIF motif (MIF in melanocortin receptors).”

(10) Page 17, Figure 3A - Since 137, 254, and 140 are not picked out on the structure, I have no idea where they are. If the authors want to show readers these residues, perhaps they could be annotated or a panel added. Since ~1 entire page of the manuscript is dedicated to this cascade, it might make sense to add a panel. Just amplifying the comment above as regards position 79, others were discussed in that paragraph but not highlighted.

We updated Supplementary Fig. 6C,D to label all of the listed residues on the protein structure for easy reference.

Author response:

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

HMGCS1, 3-hydroxy-3-methylglutaryl-CoA synthase1 is predicted to be involved in Acetyl-CoA metabolic process and mevalonate-cholesterol pathway. To induce diet-induced diabetes, they fed wild-type littermates either a standard chow (Control) or a high fat-high sucrose (HFHG) diet, where the diet composition consisted of 60% fat, 20% protein, and 20% carbohydrate (H10060, Hfkbio, China). The dietary regimen was maintained for 14 weeks. Throughout this period, body weight and fasting blood glucose (FBG) levels were measured on a weekly basis. Although the authors induced diabetes with a diet also rich in fat, the cholesterol concentration or metabolism was not investigated. After the treatment, were the animals with endothelial dysfunction? How was the blood pressure of the animals?

Thank you for your comments and kind suggestions. We have conducted a study on the impact of HFHG diet on the serum levels of total cholesterol(T-CHO) in mice over a 14-week period. Our findings indicated that the HFHG diet significantly elevated T-CHO levels in the serum of mice (Supplementary Figure 5E). Additionally, HFHG diet was associated with an increased in blood pressure (Figure 5F) and it exacerbated the progression of endothelial dysfunction in mice (Figure 5H-L).

Strengths:

To explore the potential role of circHMGCS1 in regulating endothelial cell function, the authors cloned exons 2-7 of HMGCS1 into lentiviral vectors for ectopic overexpression of circHMGCS1 (Figure S2). The authors could use this experiment as a concept proof and investigate the glucose concentration in the cell culture medium. Is the pLV-circ HMGCS1 transduction in HUVEC increasing the glucose release? (Line 163)

In the manuscript, we utilized a DMEM culture medium containing 4500 mg/L glucose. Given that the HUVEC cell culture is glucose-dependent for its metabolic processes, it was challenging to precisely evaluate the relationship between pLV-circHMGCS1 transduction and the glucose concentration in the medium.

Weaknesses:

(1) Pg 20. The cells were transfected with miR-4521 mimics, miR-inhibitor, or miR-NC and incubated for 24 hours. Subsequently, the cells were treated with PAHG for another 24 hours. Were the cells transfected with lipofectanine? The protocol or the lipofectamine kit used should be described. The lipofectamine protocol suggests using an incubation time of 72 hours. Why did the authors incubate for only 24 hours? If the authors did the mimic and inhibitor curves, these should be added to the supplementary figures. Please, describe the miRNA mimic and antagomir concentration used in cell culture.

For detailed transfection methods of miRNA mimic and its inhibitor, please refer to “Transfection of miRNA mimic or inhibitor” (Line 587) in the revised Experimental Section. We employed the Hieff Trans®siRNA/miRNA in vitro transfection reagent (yeason, China, 40806ES03), with a transfection duration of 48h. The miR-4521 content in HUVEC post-transfection was quantified using qRT-PCR. The transfection of the miR-4521 mimic for 48h notably enhanced its expression in HUVEC (Supplementary Figure 3B), whereas the transfection of the miR-4521 inhibitor for the same duration significantly suppressed its expression (Supplementary Figure 3C). The concentration used for both miRNA mimic and inhibitor transfection was 50 nM. In the revised manuscript, we have corrected the transfection time and clarified that we did not utilize miRNA antagomirs in our experiments.

(2) Pg 20, line 507. What was the miR-4521 agomiR used to treatment of the animals?

miRNA agomir serves as a valuable experimental tool for elucidating miRNA function, used to simulate the overexpression of a specific miRNA. miRNA agomir is a chemically modified RNA molecule identical in sequence to the target miRNA, engineered for enhanced stability and transfection efficacy. Utilizing miRNA agomir enables the overexpression of the target miRNA, facilitating the investigation of miRNA functions and mechanism in vivo. In our study, we have employed miRNA mimic for cellular studies and miRNA agomir in vivo applications to achieve high expression of miRNA (Fu et al, 2019).

(3) Figure 1B. The results are showing the RT-qPCR for only 5 circRNA, however, the results show 48 circRNAs were upregulated, and 18 were downregulated (Figure S1D). Why were the other cicRNAs not confirmed? The circRNAs upregulated with high expression are not necessarily with the best differential expression comparing control vs. PAHG groups. Furthermore, Figure 1A and S1D show circRNAs downregulated also with high expression. Why were these circRNAs not confirmed?

Our study aims to the identification of potential biomarkers for endothelial dysfunction in type 2 diabetes, To the end, we focused on circRNAs that exhibited significant upregulation following PAHG treatment. In our sequencing data, the p-values for these top upregulated circRNAs were notably below the threshold of 0.001, prompting their selection for further validation. We employed qRT-PCR to ascertain the consistency of their expression levels with the RNA-sequencing findings. Among these, circHMGCS1 was identified as a promising candidate with regulatory potential in endothelial dysfunction. Additionally, circRNAs that were significantly downregulated will be the subject of our ongoing research endeavors.

(4) Figure 1B shows the relative circRNAs expression. Were host genes expressed in the same direction?

circRNAs are generated from specific exons or introns of their host genes, either individually or in combination, and the main function of circRNA depends on its non-coding RNA characteristics. The expression levels of circRNAs is not necessarily correlated with those of their host genes, and similarly, the function of circRNAs do not inherently relate to the functions of the host genes (Kristensen et al, 2019; Liu & Chen, 2022). Consequently, the data presented in Figure 1B were primarily aimed at validating the accuracy of circRNA-seq. Although we did not conduct host gene expression analysis for the identified circRNAs, our subsequent results indicated that the overexpression of circHMGCS1 did not influence the expression levels of HMGCS1 (Figure 2A).

(5) Line 128. The circRNA RT-qPCR methodology was not described. The methodology should be described in detail in the Methods Session.

The only difference between the circRNA RT-qPCR method and other gene detection is that random primers need to be used for reverse transcription during the reverse transcription process. Unlike linear RNAs that possess a 3' polyA tail, which allows for the use of oligo(dT) primers, circRNAs require random primers to initiate the reverse transcription process. Beyond this distinction, the other processes are no different from the common qRT-PCR process. We have revised the Isolation of RNA and miRNA for quantitative Real Time-PCR (qRT-PCR) analysis method in the revised version (Line 695).

(6) Line 699. The relative gene expression was calculated using the 2-ΔΔCt method. This is not correct, the expression for miRNA and gene expression are represented in percentage of control.

We initially employed the 2^-ΔΔCt method to ascertain the relative gene expression levels. Subsequently, we scaled all values by a factor of 100 to amplify the visual representation of the observed variations, thereby enhancing the visualization of the data.

(7) Line 630. Detection of ROS for tissue and cells. The methodology for tissue was described, but not for cells.

We have added the detailed description of the cellular ROS detection methods in the revised manuscript as follows:

For ROS detection in cells, the treated cells were washed once by PBS, then 20 μM DHE was added, and incubated at 37°C for 30 min away from light, then washed three times by PBS and then colorless DMEM medium was added, followed by fluorescence microscopy for observation (Line 640-643).

(8) Line 796. RNA Fluorescent In Situ Hybridization (RNA-FISH). Figure 1F shows that the RNA-Fluorescence in situ hybridization (RNA-FISH) confirmed the robust expression of cytoplasmic circHMGCS1 in HUVECs (Figure 1F). However, in the methods, lines 804 and 805 described the probes targeting circMAP3K5 and miR-4521 were applied to the sections. Hybridization was performed in a humid chamber at 37C overnight. Is it correct?

We have made a correction in the revised manuscript. The accreted description is "the probes targeting circHMGCS1 and miR-4521 were applied to the sections"(Line816).

(9) Line 14. Fig 1-H. The authors discuss qRT-PCR demonstrated that circHMGCS1 displayed a stable half-life exceeding 24 h, whereas the linear transcript HMGCS1 mRNA had a half-life less than 8 h (Figure 1H). Several of the antibodies may contain trace amounts of RNases that could degrade target RNA and could result in loss of RNA hybridization signal or gene expression. Thus, all of the solutions should contain RNase inhibitors. The HMGCS1 mRNA expression could be degraded over the incubation time (0-24hs) leading to incorrect results. Moreover, in the methods is not mentioned if the RNAse inhibitor was used. Please, could the authors discuss and provide information?

This experiment was performed in cell culture as described in our Experimental Methods (Line 753), where we added actinomycin D directly into the cell culture well plates, and the cells remained in a healthy state during this treatment. We did not directly extract mRNA from cells for this experiment. Additionally, all solutions utilized throughout the whole experiment were prepared using Rnase-free water, ensuring that the integrity of the mRNA.

(10) Further experiments demonstrated that the overexpression of circHMGCS1 stimulated the expression of adhesion molecules (VCAM1, ICAM1, and ET-1) (Figures 2B and 2C), suggesting that circHMGCS1 is involved in VED. How were these genes expressed in the RNA-seq?

In the manuscript, we only focused exclusively on circRNA and miRNA sequencing, and not perform mRNA sequencing, Consequently, we employed qRT-PCR and Western blot to assess the expression alterations of ET-1, ICAM1, and VCAM1 at gene and protein level. The findings revealed that the overexpression of circHMGCS1 significantly upregulated the expression of adhesion molecules (VCAM1, ICAM1, and ET-1).

(11) Line 256. By contrast, the combined treatment of circHMGCS1 and miR-4521 agomir did not significantly affect the body weight and blood glucose levels. OGTT and ITT experiments demonstrated that miR-4521 agomir considerably enhanced glucose tolerance and insulin resistance in diabetic mice (Figures 5C, 5D, and Figures S5B and S5C). Why did the miR-4521 agomir treatment considerably enhance glucose tolerance and insulin resistance in diabetic mice, but not the blood glucose levels?

Our results showed that miR-4521 agomir could effectively suppress the increase of body weight and blood glucose in mice (Figure 5A-B).

(12) In the experiments related to pull-down, the authors performed Biotin-coupled miR-4521 or its mutant probe, which was employed for circHMGCS1 pull-down. This result only confirms the Luciferase experiments shown in Figure 4A. The experiment that the authors need to perform is pull-down using a biotin-labeled antisense oligo (ASO) targeting the circHMGCS1 backsplice junction sequence followed by pulldown with streptavidin-conjugated magnetic beads to capture the associated miRNAs and RNA binding proteins (RBPs). Also, the ASO pulldown assay can be coupled to miRNA RT-qPCR and western blotting analysis to confirm the association of miRNAs and RBPs predicted to interact with the target circRNA.

This point is correct. As suggested, we utilized a biotin-labeled circHMGCS1 probe for pull down experiments. Because circRNA-miRNA interactions are mainly mediated by the RNA-induced silencing complex, which includes Argonaute 2 (AGO2), we examined the levels of miR-4521 and AGO2 in the capture meterial. Our results demonstrated that circHMGCS1 significantly captured miR-4521 in the cells, with a concomitant acquisition of AGO2. These findings have been integrated into the revised manuscript (Supplementary Figures 4D and 4E).

(13) In Figure 5, the authors showed that the results suggest that miR-4521 can inhibit the occurrence of diabetes, whereas circHMGCS1 specifically dampens the function of miR-4521, weakening its protective effect against diabetes. In this context, what are the endogenous target genes for the miR-4521 that could be regulating diabetes?

In this study, we focused on the role of miR-4521 in endothelial function. Our animal experiments involving ARG1 knockdown revealed that the reduction of ARG1 expression resulted in the inability of miR-4521 to modulate the progression of type 2 diabetes. Consequently, ARG1 is likely an endogenous target gene of miR-4521, potentially implicated in the regulation of diabetes.

(14) In the western blot of Figure 5, the β-actin band appears to be different from the genes analyzed. Was the same membrane used for the four proteins? The Ponceau S membrane should be provided.

As described in our experimental methodology (Western blot analysis), we have utilized PVDF membranes for our Western blot experiments. β-actin, recognized for its high expression and specificity as a housekeeping gene, yields distinct bands with minimal background noise. This property can lead to the migration β-actin from the spot wells to both sides during electrophoresis. So much so that it is not aligned with the lane shown by the target gene. And the other 3 genes can see the phenomenon of obvious lane because their expression is not as high as β-actin. We replaced β-actin with a similar background in the revised manuscript (Figure 5L).

(15) Why did the authors use AAV9, since the AAV9 has a tropism for the liver, heart, skeletal muscle, and not to endothelial vessels?

AAV9 has garnered significant interest as a gene delivery vector due to its extensive tissue penetration, minimal immunogenicity, and stable gene expression profile. Its application in cardiovascular disease research and therapy has been widely reported (Barbon et al, 2023; Yao et al, 2018; Zincarelli et al, 2008). Meanwhile, we employed AAV9 for gene delivery via the tail vein injection in mice, and as shown in Figure 5J and Figure 7Q, we observed GFP signals carried by AAV9 in the thoracic aorta of mice. These findings suggest that AAV9 possesses the capability to infect endothelial cells effectively.

Reviewer #2 (Public Review):

Summary:

The authors observed an aggravated vascular endothelial dysfunction upon overexpressing circHMGCS1 and inhibiting miR-4521. This study discovered that circHMGCS1 promotes arginase 1 expression by sponging miR-4521, which accelerated the impairment of vascular endothelial function.

Strengths:

The study is systematic and establishes the regulatory role of the circHMGCS1-miR-4521 axis in diabetes-induced cardiovascular diseases.

Weaknesses:

(1) The authors selected the miR-4521 as the target based on their reduced expression upon circHMGCS1 overexpression. Since the miRNA level is downregulated, the downstream target gene is expected to be upregulated even in the absence of circRNA. The changes in miRNA expression opposite to the levels of target circRNA could be through Target RNA-Directed MicroRNA Degradation. In addition, miRNA can also be stabilized by circRNAs. Hence, selecting miRNA targets based on opposite expression patterns and concluding miRNA sponging by circRNA needs further evidence of direct interactions.

Thank you for your positive comments and kind suggestions.

As suggested by Public Reviewer #1 (12), we employed a biotin-tagged circHMGCS1 to capture miR-4521 and AGO2 in HUVECs (Supplementary Figures 4D and 4E), and Dual luciferase assays have confirmed that miR-4521 can bind to circHMGCS1 directly. Furthermore, RNA pull down and RIP assays have demonstrated the direct binding capability of circHMGCS1 for miR-4521. Collectively, these findings underscore the direct interaction between circHMGCS1 and miR-4521.

(2) The majority of the experiments were performed with an overexpression vector which can generate a lot of linear RNAs along with circRNAs. The linear RNAs produced by the overexpression vectors can have a similar effect to the circRNA due to sequence identity.

In our manuscript, the employed vectors incorporate reverse repeat sequences that facilitate efficient circularization of circRNAs. This design ensures robust circular shearing upon the insertion of circRNA sequences into the polyclonal sites, thereby enhancing the overexpression of circRNAs (Supplementary Figure 2). Moreover, we used lentiviral virus as a vector for circRNA overexpression, not direct plasmid transfection. As demonstrated in Figure 2A, upon overexpression of circHMGCS1, we observed a significant upregulation in circHMGCS1 levels compared to the pLV-circNC and Control groups. Notably, the expression levels of the linear HMGCS1 mRNA did not exhibit significant alterations.

(3) There is a lack of data of circHMGCS1 silencing and its effect on target miRNA & mRNAs.

According to your suggestion, we employed shRNA to knockdown circHMGCS1 in HUVEC, and qRT-PCR was used to assess the expression levels of miR-4521 and ARG1. The knockdown of circHMGCS1 significantly inhibit the expression of circHMGCS1 in HUVEC without obviously affecting the levels of HMGCS1 mRNA. We then selected circHMGCS1 shRNA1 for further investigation. We observed that the knockdown of circHMGCS1 resulted in an upregulation of miR-4521 and a downregulation of ARG1 expression.

Author response image 1.

The impact of circHMGCS1 knockdown on ARG1 and miR-4521 expression levels in HUVEC. The cells were transfected with either circHMGCS1 shRNA1 or circHMGCS1 shRNA2, and the expressions levels of circHMGCS1 and HMGCS1 (A), miR-4521 (B) and ARG1 (C and D) in HUVECs were detected by qRT-PCR and Western blot. n=3 in each group. *p < 0.05, **p < 0.01. All significant difference was determined by one-way ANOVA followed by Bonferroni multiple comparison post hoc test, error bar indicates SD.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

I suggest improving the discussion based on the literature.

(1) Line 131. .... (hsa_circ_0008621, 899 nt in length, identified as circHMGCS1 in subsequent studies because of its host gene being HMGCS1). Please, provide the reference.

We appreciate the valuable comments. We have made changes for improvement, which is add in Line 133(Liang et al, 2021).

(2) The authors conclude that both in vitro and in vivo data suggest that the miR-4521 or circHMGCS1 fails to regulate the effect of diabetes-induced VED in the absence of ARG1. Therefore, ARG1 may serve as a promising VED biomarker, and circHMGCS1 and miR-4521 play a key role in regulating diabetes-induced VED by ARG1. In this context, they should re-evaluate whether this is the best title. "Circular RNA HMGCS1 sponges miR-4521 to aggravate type 2 diabetes-induced vascular endothelial dysfunction"

This manuscript initiates its exploration with circRNA as the focal point of study (Figure 1 and Figure 2), It then delves into the miRNAs associated with circRNA and elucidates their interactions (Figure 3, Figure 4 and Figure 5). Subsequently, the manuscript identifies the target genes of miRNA and validates the regulatory effects of circRNA and miR-4521 on ARG1 (Figure 6). The study culminates with the application of the ceRNA theory to confirm the significance of ARG1 in the functional interplay between circHMGCS1 and miR-4521 (Figure 7). These findings throughout the manuscript are dedicated to uncovering the pivotal roles of circHMGCS1 and miR-4521 in modulating vascular endothelial function. Notably, the interaction between circHMGCS1 and miR-4521 represents a novel discovery of our research. Therefore, we aim to emphasize the critical function of circHMGCS1 and miR-4521 in the regulation of vascular endothelial dysfunction in type 2 diabetes within the manuscript.

Reviewer #2 (Recommendations For The Authors):

I have a few suggestions for improving the study further.

(1) Although the experiments suggest the role of circHMGCS1, miR-4521 in vascular endothelial function, the direct regulation or interaction of circHMGCS1-miR-4521-ARG1 is unclear. A rescue experiment that checks the effect of circHMGCS1 silencing with/without inhibition of miR-4521 on ARG1 expression must be performed to prove the circHMGCS1- miR-4521 regulatory axis.

Thank you very much for your constructive comments.

According to your suggestion, we utilized shRNA to effectively knockdown circHMGCS1 in HUVEC, Subsequent expression analysis via qRT-PCR was conducted to assess the levels of miR-4521 and ARG1. The knockdown of circHMGCS1 significantly reduced the expression of circHMGCS1 in HUVEC without influencing the expression of the host gene HMGCS1. Concurrently, the knockdown of circHMGCS1 resulted in an upregulation of miR-4521 (Supplementary Figure 4B) and a downregulation of ARG1 (Figure 6P and 6Q). In our manuscript, the upregulation in ARG1 expression caused by circHMGCS1 overexpression was reduced by miR-4521, and the downregulation in ARG1 expression caused by miR-4521 overexpression was also reversed by circHMGCS1. When miR-4521 was knocked down, the expression of ARG1 increased, and circHMGCS1 abrogated its regulatory effect on the expression of ARG1. Collectively, these findings indicate that the interplay between circHMGCS1 and miR-4521 significantly influences ARG1 expression.

Author response image 2.

The impact of circHMGCS1 knockdown on ARG1 and miR-4521 expression levels in HUVEC. The cells were transfected with either circHMGCS1 shRNA1 or circHMGCS1 shRNA2, and the expressions levels of circHMGCS1 and HMGCS1 (A), miR-4521 (B) and ARG1 (C and D) in HUVECs were detected by qRT-PCR and Western blot. n=3 in each group. *p < 0.05, **p < 0.01. All significant difference was determined by one-way ANOVA followed by Bonferroni multiple comparison post hoc test, error bar indicates SD.

(2) It is unclear how the authors arrived at the circHMGCS1-miR-4521 pair. The pull down of circHMGCS1 followed by qPCR enrichment analysis of all target miRNAs must be performed to select the target miRNA.

In this manuscript, we identified the expression of miRNA under PAHG treatment through miRNA sequencing, and then further screened out 4 miRNAs with potential binding sites to circHMGCS1 utilizing the miRanda database. Subsequently, we employed qRT-PCR and Western blot analysis to confirm the regulatory influence of miR-4521 on endothelial function (Figure 3). Following this, RIP, RNA pull down, dual luciferase and RNA-FISH experiments were conducted to map the interaction between circHMGCS1 and miR-4521 (Figure 4), the direct interaction between circHMGCS1 and miR-4521 was further substantiated through overexpression and knockdown studies (Figures 5-7). while the reviewer's method may offer a more direct validation, our methodology initially involved a database-driven screening of candidate miRNAs with the potential to target and bind circHMGCS1, followed by experimental validation of these interactions. Both methodologies are capable of establishing the interaction sites between circHMGCS1 and miR-4521.

(3) Since the back splicing is not that efficient, the linear RNA from the overexpression construct may produce many linear RNAs with miRNA binding sites. The effect seen in the case of overexpression experiments needs to consider the level of linear and circular HMGCS1 produced by the vector.

In this manuscript, the vector's multiple cloning site is flanked by inverted repeat sequences that facilitate efficient circRNA looping. This design enables the inserted sequence to form a stable loop and undergo circularization upon transcription, leading to the overexpression of circRNA (Supplementary Figure 2). For the validation of circular RNA, we employed divergent primers that straddle the circRNA splicing junction. These primers are specific for circRNA amplification and do not amplify the corresponding linear RNA, as demonstrated in Figure 2A. Upon overexpression of circHMGCS1, we observed a significant increase in circHMGCS1 levels compared to the empty vector and Control groups, while there was no significant change in the expression level of HMGCS1 mRNA.

(4) As miR-4521 has multiple miRNA binding sites on circHMGCS1, it is not very clear which sites were mutated in circHMGCS1-MUT.

We have made corrections to Supplementary Figure 4C. Utilizing the miRanda algorithm, we identified 10 potential binding sites for miR-4521 on circHMGCS1. Subsequently, we selected the site with the highest binding affinity for mutational analysis (miR-4521 binding positions 3-15, circHMGCS1 binding positions 260-281, binding rate 91.67%, binding ability -17.299999 kCal/Mol). We employed a dual-luciferase assay to confirm the direct interaction between circHMGCS1 and miR-4521.

(5) Since the ceRNA network works efficiently in an equimolar concentration of the regulatory molecules, providing the copy number of circHMGCS1, miR-4521, and target mRNAs would be helpful.

We employed qRT-PCR to ascertain the absolute quantification of mRNA copy numbers, following established methodologies (Nolan et al, 2006; Wagatsuma et al, 2005; Zhang et al, 2009). Our qRT-PCR data reveal that the circHMGCS1 mRNA copy number is 2343±529. In comparison, the ARG1 mRNA copy number stands at 88±27, while the miR-4521 copy number is significantly higher, recorded at 36277±9407.

Author response image 3.

The distribution of copy numbers for circHMGCS1, miR-4521 and ARG1 in HUVECs.

(6) The yellow highlighted "cyclization-mediated sequence-F & R" does not seem to be complementary sequences. The method section may include the details of the vectors and cloning strategies for the overexpression constructs.

The figure below illustrates the schematic representation of the complementary structure between the upstream and downstream sequences that facilitate circRNA circularization. This strategic pairing is designed to enhance the circularization efficiency of circRNA while concurrently suppressing mRNA synthesis (Liang & Wilusz, 2014). Details of this design have been integrated into the experimental method (Line539). The specific additions are as follows:

The circHMGCS1 sequence [NM_001098272: 43292575-43297268], the splice site AG/GT and ALU elements were inserted into the pCDH-circRNA-GFP vector (upstream ALU: AAAGTGCTGAGATTACAGGCGTGAGCCACCACCCCCGGCCCACTTTTTGTAAAGGTACGTACTAATGACTTTTTTTTTATACTTCAG, downstream ALU: GTAAGAAGCAAGGAAAAGAATTAGGCTCGGCACGGTAGCTCACACCTGTAATCCCAGCA). The restriction enzyme sites selected were EcoRI and NotI.

Author response image 4.

(7) Since circHMGCS1 is a multi-exonic circRNA that can undergo alternative splicing and divergent primers only validate the backsplice junction, the full-length sequence of mature circHMGCS1 needs to be checked by circRNA-RCA PCR followed by Sanger sequencing.

In compliance with your guidance, we have enriched the revised manuscript with additional data. Specifically, we have included the full-length nucleic acid electrophoresis diagram of circHMGCS1 in Supplementary Figure 1F, the Sanger sequencing results in Supplementary Figure 1G, and a comparative analysis of the circHMGCS1 sequences obtained from Sanger sequencing with those referenced in the circBase database, presented in Supplementary Figure 1H.

Reference:

Barbon, E., C. Kawecki, S. Marmier, A. Sakkal, F. Collaud, S. Charles, G. Ronzitti, C. Casari, O.D. Christophe, C.V. Denis, P.J. Lenting, and F. Mingozzi. 2023. Development of a dual hybrid AAV vector for endothelial-targeted expression of von Willebrand factor. Gene Ther. 30: 245-254.

Fu, Y., J. Chen, and Z. Huang. 2019. Recent progress in microRNA-based delivery systems for the treatment of human disease. ExRNA. 1: 24.

Kristensen, L.S., M.S. Andersen, L.V.W. Stagsted, K.K. Ebbesen, T.B. Hansen, and J. Kjems. 2019. The biogenesis, biology and characterization of circular RNAs. Nat Rev Genet. 20: 675-691.

Liang, D., and J.E. Wilusz. 2014. Short intronic repeat sequences facilitate circular RNA production. Genes Dev. 28: 2233-2247.

Liang, J., X. Li, J. Xu, G.M. Cai, J.X. Cao, and B. Zhang. 2021. hsa_circ_0072389, hsa_circ_0072386, hsa_circ_0008621, hsa_circ_0072387, and hsa_circ_0072391 aggravate glioma via miR-338-5p/IKBIP. Aging (Albany NY). 13: 25213-25240.

Liu, C.X., and L.L. Chen. 2022. Circular RNAs: Characterization, cellular roles, and applications. Cell. 185: 2016-2034.

Nolan, T., R.E. Hands, and S.A. Bustin. 2006. Quantification of mRNA using real-time RT-PCR. Nat Protoc. 1: 1559-1582.

Wagatsuma, A., H. Sadamoto, T. Kitahashi, K. Lukowiak, A. Urano, and E. Ito. 2005. Determination of the exact copy numbers of particular mRNAs in a single cell by quantitative real-time RT-PCR. J Exp Biol. 208: 2389-2398.

Yao, C., T. Veleva, L. Scott, Jr., S. Cao, L. Li, G. Chen, P. Jeyabal, X. Pan, K.M. Alsina, I.D. Abu-Taha, S. Ghezelbash, C.L. Reynolds, Y.H. Shen, S.A. Lemaire, W. Schmitz, F.U. Müller, A. El-Armouche, N. Tony Eissa, C. Beeton, S. Nattel, X.H.T. Wehrens, D. Dobrev, and N. Li. 2018. Enhanced Cardiomyocyte NLRP3 Inflammasome Signaling Promotes Atrial Fibrillation. Circulation. 138: 2227-2242.

Zhang, X.X., T. Zhang, M. Zhang, H.H. Fang, and S.P. Cheng. 2009. Characterization and quantification of class 1 integrons and associated gene cassettes in sewage treatment plants. Appl Microbiol Biotechnol. 82: 1169-1177.

Zincarelli, C., S. Soltys, G. Rengo, and J.E. Rabinowitz. 2008. Analysis of AAV serotypes 1-9 mediated gene expression and tropism in mice after systemic injection. Mol Ther. 16: 1073-1080.

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public Review): 

Summary: 

This study explores the neural control of muscle by decomposing the firing activity of constituent motor units from the grid of surface electromyography (EMG) in the Tibialis (TA) Anterior and Vastus Lateralis (VL) during isometric contractions. The study involves extensive samples of motor units across the broadest range of voluntary contraction intensities up to 80% of MVC. The authors examine the rate coding of the population of motor units, which describes the instantaneous firing rate of each motor unit as a function of muscle force. This relationship is characterized by a natural logarithm function that delineates two distinct phases: an initial phase with a steep acceleration in firing rate, particularly pronounced in low-threshold motor units, and a subsequent modest linear increase in firing rate, more significant in high-threshold motor units. 

Strengths: 

The study makes a significant contribution to the field of neuromuscular physiology by providing a detailed analysis of motor unit behavior during muscle contractions in a few ways.

(1) The significance lies in its comprehensive framework of motor unit activity during isometric contractions in a broad range of intensities, providing insights into the non-linear relationship between the firing rate and the muscle force. The extensive sample of motor units across the pool confirms the observation in animal studies in which the spinal motoneuron exhibits a discharge consisting of distinct phases in response to synaptic currents, under the influence of persistent inward currents. As such, it is now reasonable to state the human motor units across the pool are also under the control of gain modulation via some neuromodulatory effects in addition to synaptic inputs arising from ionotropic effects.

(2) The firing scheme across the entire motoneuron pool revealed in this study reconciles the discrepancy in firing organization under debate; i.e., whether it is 'onion skin' like or not (Heckman and Enoka 2012). The onion skin like model states that the low threshold motor units discharge higher than high threshold motor units and have been held for a long time because the firing behaviors were examined in a partial range of contraction force range due to technical limitations. This reconciliation is crucial because it is fundamental to modelling the organization of motor unit recruitment and rate coding to achieve a desired force generation to advance our understanding of motor control.

(3) The extensive data collection with a novel blind source separation algorithm on the expanded number of channels of surface EMG signal provides a robust dataset that enhances the reliability and validity of findings, setting a new standard for empirical studies in the field. 

Collectively, this study fills several knowledge gaps in the field and advances our understanding of the mechanism underlying the isometric force generation.

We thank the reviewer for their positive appreciation of our work.

Weaknesses: 

Although the findings and claims based on them are mostly well aligned, some accounts of the methods and claims need to be clarified.

(1) The authors examine the input-output function of a motor unit by constructing models, using force as an input and discharge rate as an output. It sounds circular, or the other way around to use the muscle force as an input variable, because the muscle force is the result of motor unit discharges, not the cause that elicits the discharges. More specifically, as a result of non-linear interactions of synchronous and/or asynchronous discharges of a population of a given motoneuron pool that give rise to transient increase/maintenance in twitch force, the gross muscle force is attained. I acknowledge that it is extremely challenging experimentally to measure synaptic currents impinging upon the spinal motoneurons in human subjects and the author has an assumption that the force could be used as a proxy of synaptic currents. However, it is necessary to explicitly provide the caveats and rationale behind that. Force could be used as the input variable for modelling.

Force is indeed used in this study as a proxy of the common excitatory synaptic currents as their direct measurement is not possible in vivo in humans. It is worth noting that this approach has been extensively used in the past by many groups to study rate coding (e.g., Monsters & Chan, De Luca’s, Heckman’s, and Fuglevand’s groups). Heckman’s, Gorassini’s, Fuglevand’s groups and others have considered the non-linearities in the relation between motor unit firing rates and muscle force in humans as an indicator of the impact of neuromodulation on motor unit behaviour and changes of the intrinsic properties of motoneurons.

One could also use the cumulative spike train as a more direct estimate of common excitatory inputs, assuming that it is possible to identify a group of motor units not influenced by PICs, as done when selecting a reference low-threshold motor neuron in the delta F method (Gorassini et al., 1998), or the cumulative spike train of low-threshold motor neurons (Afsharipour et al., 2020). However, this approach was not possible in our study as we did not have the same units across contractions to estimate cumulative spike trains. It was therefore not possible to pool the data across contractions as we did to generate force/firing rate relations on the widest range of force.

We added a sentence in the discussion to highlight this limitation (P19, L470):

‘This result must be confirmed with a more direct proxy of the net synaptic drive, such as the firing rate of a reference low-threshold motor neuron used in the delta F method (Gorassini et al., 1998), or the cumulative spike train of low-threshold motor neurons (Afsharipour et al., 2020)’.

(2) The authors examine the firing organizations in TA and VL in this study without explicit purposes and rationale for choosing these muscles. The lack of accounts makes it hard for the readers to interpret the data presented, particularly in terms of comparing the results from the different muscles.

We wanted to compare the rate coding of pools of motor units from proximal (VL) and distal (TA) muscles within the lower limb. Indeed, distal and proximal muscles exhibit differences in rate coding and spatial recruitments (De Luca et al., 1982, J Physiol), potentially due to different levels of recurrent inhibition (Cullheim & Kellerth, 1978, J Physiol; Rossi & Mazzocchio, 1991, Exp Brain Res; Edgley et al., 2021, J Neurosci) or different levels of neuromodulation depending on their involvement (or not) in postural control (Hoonsgaard et al., 1988, J Physiol; Kim et al., 2020, J Neurophysiol).

We added a paragraph at the beginning of the result section to support our muscle choice (P6; L137): ‘16 participants performed either isometric dorsiflexion (n = 8) or knee extension tasks (n = 8) while we recorded the EMG activity of the tibialis anterior (TA - dorsiflexion) or the vastus lateralis (VL – knee extension) with four arrays of 64 surface electrodes (256 electrodes per muscle). The motoneuron pools of these two muscles of the lower limb receive a large part of common input (Laine et al., 2015; Negro et al., 2016a), constraining the recruitment of their motor units in a fixed order across tasks. They are therefore good candidates for an accurate description of rate coding. Moreover, we wanted to determine whether differences in rate coding observed between proximal and distal muscles in the upper limb (De Luca et al., 1982) were also present in the lower limb.’.

Another factor that guided our muscle choice was the low risk of crosstalk. For this, we verified with ultrasound that our arrays of 256 electrodes only covered the muscle of interest, staying away from the neighbouring muscles. This was possible as superficial muscles from the leg are bulkier than those from the upper limb. Given the small diameter of each electrode (2 mm), it is unlikely that the motor units from the neighbouring muscles were in the recorded muscle volume (Farina et al., 2003, IEEE Trans Biomed Eng)

(3) In the methods, the author described the manual curation process after applying the blind source separation algorithm. For the readers to understand the whole process of decomposition and to secure rigor and robustness of the analyses, it would be necessary to provide details on what exact curation is performed with what criteria. 

The manual curation of EMG decomposition with blind source separation is different from what is classically done with intramuscular EMG and template-matching algorithms. 

In short, our decomposition algorithm uses fast independent component analysis (fastICA) to retrieve motor unit spike trains from the EMG signals. For this, it iteratively optimises a set of weights, i.e., a separation vector, for each motor unit. The projection of the EMG signals on this separation vector generates a sparse motor unit pulse train, with most of its samples close to zero and only a few samples close to one (Figure 1B). The discharge times are estimated from this motor unit pulse train using a peak detection function and a k-mean classification with two classes to separate the high peaks (spikes) from the low peaks (noise and other motor units).

The manual curation consists of inspecting the automatic detection of the peaks of the motor unit pulse train and manually add missed peaks (missed discharge times) or remove wrongly detected peaks. Then, the separation vector is updated using the correct discharge times and the motor unit pulse train recalculated. This procedure generally improves the distance between the discharge times and the noise, which confirm the accuracy of the manual curation. If that’s not the case, the motor unit is discarded from the analyses.

We added a section on manual editing in the methods (P23, L615):

‘At the end of these automatic steps, all the motor unit pulse trains and identified discharge times were visually inspected, and manual editing was performed to correct the false identification of artifacts or the missed discharge times (Del Vecchio et al., 2020; Hug et al., 2021; Avrillon et al., 2023). The manual editing consisted of i) removing the spikes causing erroneous discharge rates (outliers), ii) adding the discharge times clearly separated from the noise, iii) recalculating the separation vector, iv) reapplying the separation vector on the entire EMG signals, and v) repeating this procedure until the selection of all the discharge times is achieved. The manual editing of potential missed discharge times and falsely identified discharge times was never immediately accepted. Instead, the procedure was consistently followed by the application of the updated motor unit separation vector on the entire EMG signals to generate a new motor unit pulse train. Then, the manual editing was only accepted when the silhouette value increased or stayed well above the threshold of 0.9 quantified with the silhouette value (Negro et al., 2016b). Only these motor units were retained for further analysis.’

(4) In Figure 3, the early recruited units tend to become untraceable in the higher range of contraction. This is more pronounced in the muscle VL. This limitation would ambiguate the whole firing curve along the force axis and therefore limitation and the applicability in the different muscles needs to be discussed. 

The loss of low threshold motor units in the higher range of contractions was caused either by the decrease in signal-to-noise ratio for small motor units when many larger ones are recruited, or by the cancellation of the surface action potentials of the small units in the interference electromyographic signal, or by the recruitment of a motor unit with a very similar spatio-temporal filter (an example is shown in the figure below). In the latter case, the motor unit pulse train contains peaks that represent the discharge times of both motor units (green and red dots in the simulated example below), making them undistinguishable by the operator during manual editing.

Author response image 1.

This was discussed in the results (P7; L190):

‘On average, we tracked 67.1 ± 10.0% (25th–75th percentile: 53.9 – 80.1%) of the motor units between consecutive contraction levels (10% increments, e.g., between 10% and 20% MVC) for TA and 57.2 ± 5.1% (25th–75th percentile: 46.6 – 68.3%) of the motor units for VL (Figure S2). There are two explanations for the inability to track all motor units across consecutive contraction levels. First, some motor units are recruited at higher targets only. Second, it is challenging to track small motor units beyond a few contraction levels due to a lower signal-to-noise ratio for the small motor units when larger motor units are recruited, or signal cancellation (Keenan et al., 2005; Farina et al., 2014a).’

However, we believe that it had a limited impact on the output of the paper, as the non-linear portion of the rate coding/force relation due to the persistent inward currents occurs during the first seconds after recruitment, before plateauing (for a review see Binder et al., 2020, Physiology).

(5) It is unclear how commonly the notion "the long-held belief that rate coding is similar across motor units from the same pool" is held among the community without a reference. Different firing organizations have been modelled and discussed in the seminal paper by Fuglevand et al. (1993) and as far as I understand, the debate has not converged to a specific consensus. As such, any reference would be required to support the claim the notion is widely recognized.

In the paper of Fuglevand et al., (1993, J Neurophysiol), all the motor units had the same rate coding pattern relative to the excitatory input, though they changed the slope of the relations and the saturation threshold of motor units between simulations. This is similar to the paper of De Luca & Contessa (2012, J Neurophysiol), where the equation used to simulate the rate coding was non-linear, but consistent across motor units.  

We added these citations to the text:

‘Overall, we found that motor units within a pool exhibit distinct rate coding with changes in force level (Figure 2 and 3), which contrasts with the long-held belief that rate coding is similar across motor units from the same pool (Fuglevand et al., 1993; De Luca and Contessa, 2012).’

(6) The authors claim that the firing behavior as a function of force is well characterized by a natural logarithmic function, which consists of initial steep acceleration followed by a modest increase in firing rate. Arguably the gain modulation in firing rate could be attributed to a neuromodulatory effect on the spinal motoneuron, which has been suggested by a number of animal studies. However, the complexity of the interactions between ionotropic and neuromodulatory inputs to motoneurons may require further elucidation to fully understand the mechanisms of neural control; it is possible to consider the differential acceleration among different threshold motor units as a differential combinatory effect of ionotropic and neuromodulatory inputs, but it is not trivially determined how differentially or systematically the inputs are organized. Likewise, the authors make an account for the difference in firing rate between TA and VL in terms of different amounts or balances of excitatory and inhibitory inputs to the motoneuron pool, but again this could be explained by other factors, such as a different extent of neuromodulatory effects. To determine the complexity of the interactions, further studies will be warranted.

We appreciate the reviewer’s view on this point, as we indeed only indirectly inferred the combination of neuromodulatory and ionotropic inputs to motoneurons in this study. A more direct manipulation of the sources of neuromodulatory and ionotropic inputs will be required in the future to directly highlight the mechanisms responsible for these variations in rate coding within pools. However, it is also worth noting that the acceleration in firing rate, the increase in firing rate during the ramp up, and the hysteresis between ramps up and downs have been used to infer the distribution of ionotropic and neuromodulatory inputs from the firing rate/force relations (Johnson et al., 2017; Beauchamp et al., 2023; Chardon et al., 2023). This approach has been validated with hundreds of thousands of simulations using a biophysical model of motor neurons (Chardon et al., 2023). There is also a series of studies in humans showing how the absence of neuromodulation modulated via inhibitory inputs (Revill & Fuglevand, 2017) or medication blocking serotonin receptors (Goodlich et al., 2023) impact the non-linearity of the firing rate/force relation. Therefore, we are confident that the differences observed within and between pools are linked to different distribution of excitatory/inhibitory inputs and neuromodulation.

We added a sentence in the discussion to highlight this point (P18; L435):

‘Taken together, these results show how ionotropic and neuromodulatory inputs to motoneurons uniquely combine to generate distinct rate coding across the pool, even if a more direct manipulation of the sources of neuromodulatory and ionotropic inputs will be required to directly estimate their interactions.’

(7) It is unclear with the account " ... the bandwidth of muscle force is < 10Hz during isometric contraction" in the manuscript alone, and therefore, it is difficult to understand the following claim. It appears very interesting and crucial for motor unit discharge and force generation and maintenance because it would pose a question of why the discharge rate of most motor units is higher than 10Hz, despite the bandwidth being so limited, but needs to be elaborated.

We described the slow fluctuations in smoothed firing rates associated with the variations in force observed during isometric contractions. The bandwidth of muscle force is lower than 10Hz due to the contractile properties of muscle tissues (Baldissera et al., 1998, J Physiol). Having an average firing rate higher than this bandwidth enables the pool of motor neurons to effectively transmit the common inputs (the main discriminant of muscle force) over this bandwidth without distortion (Farina et al., 2014, J Physiol). Increasing the firing rate beyond the muscle bandwidth also increases the power of the spike train at the direct current frequency (frequency equal to 0) since this power is related to the number of spikes per second. Thus, increasing the firing rate well beyond the muscle bandwidth still has a clear effect in force. To illustrate this point, note that electrical stimuli delivered at 100 Hz can lead to an increase in muscle force.

Reviewer #2 (Public Review):  

Summary: 

The motivation for this study is to provide a comprehensive assessment of motor unit firing rate responses of entire pools during isometric contractions. The authors have used new quantitative methods to extract more unique motor units across contractions than prior studies. This was achieved by recording muscle fibre action potentials from four high-density surface electromyogram (HDsEMG) arrays (Caillet et al., 2023), quantifying residual EMG comparing the recorded and data-based simulation (Figure 1A-B), and developing a metric to compare the spatial identification for each motor unit (Figure 1D-E). From identified motor units, the authors have provided a detailed characterization of recruitment and firing rate responses during slow voluntary isometric contractions in the vastus lateralis and tibialis anterior muscles up to 80% of maximum intensity. In the lower limb, it is interesting how lower threshold motor units have firing rate responses that saturate, whereas higher threshold units that presumably produce higher muscle contractile forces continue to increase their firing rate. In many ways, these results agree with the rate coding of motor units in the extensor digitorum communis muscle (Monster and Chan, 1977). The paper is detailed, and the analyses are well explained. However, there are several points that I think should be addressed to strengthen the paper.

We thank the reviewer for their positive appreciation of our work.

General comments: 

(1) The authors claim they have measured the complete rate coding profiles of motor units in the vastus lateralis and tibialis anterior muscles. However, this study quantified rate coding during slow and prolonged voluntary isometric contractions whereas the function of rate coding during movements (Grimby and Hannerz, 1977) or more complex isometric contractions (Cutsem and Duchateau, 2005; Marshall et al., 2022) remains unexplored. For example, supraspinal inputs may not scale the same way across low and higher threshold motor units, or between muscles (Devanne et al., 1997), making the response of firing rates to increasing isometric contraction force less clear. 

We agree with the reviewer that rate coding strategies may vary with the velocity and the type of contractions (Duchateau & Enoka, 2008, J Physiol). It is thus likely that the firing rate would increase during the first milliseconds of fast contractions, with the occurrence of doublets (Cutsem and Duchateau, 2005, J Physiol; Del Vecchio et al., 2019, J Physiol), or that motor unit firing rate may be lower during lengthening than shortening contractions (Duchateau & Enoka, J Physiol). 

However, the decomposition of EMG signals in non-stationary conditions remains challenging, and is still limited to slow varying patterns of force (Chen et al., 2000, Oliveira & Negro, 2021, Mendez Guerra et al., 2024, Yeung et al., 2024). Future methodological developments will be required to expand our findings to other patterns of force.

Conceptually, the authors focus on the literature on intrinsic motoneurone properties, but in vivo, other possibilities are that descending supraspinal drive, spinal network dynamics, and afferent inputs have different effects across motor unit sizes, muscles, and types of contractions. Also, the influence from local muscles that act as synergists (e.g., vastii muscles for the vastus lateralis, and peroneal muscles that evert the foot for the tibialis anterior) or antagonists (coactivation during higher contraction intensities would stiffen the joint) may provide differential forms of proprioceptive feedback across motor pools. 

The reviewer is right that differences in spinal network dynamics and afferent inputs may explain the differences in rate coding observed between the two muscles. Indeed, computational models have shown how the pattern of inhibitory inputs may affect the increase in firing rate during linear increase in force (Powers & Heckman, 2017, J Neurophysiol; Chardon et al., 2023, Elife). Specifically, the difference observed between proportional inhibitory inputs vs. a push pull pattern mirror the differences observed here between the TA (push-pull like pattern) and the VL (proportional pattern). This difference may reflect the impact of various pathways of inhibition, such as reciprocal inhibition or recurrent inhibition from homonymous motor units or motor units from synergistic muscles. 

These points have been further discussed in the manuscript (P19; L475):

‘The increase in firing rate was also significantly greater for TA motor units than for those in VL. This difference may reflect a varying balance between excitatory/inhibitory synaptic inputs and neuromodulation due to multiple spinal circuits (Heckman and Binder, 1993; Heckman et al., 2008; Johnson et al., 2017; Powers and Heckman, 2017; Chardon et al., 2023; Škarabot et al., 2023). Specifically, the strength of recurrent and reciprocal inhibitory inputs to motoneurons innervating VL and TA, and their proportional or inverse covariation with excitatory inputs, respectively, may explain the differences in rate limiting and maximal firing rates (Heckman and Binder, 1993; Heckman et al., 2008; Johnson et al., 2017; Powers and Heckman, 2017; Chardon et al., 2023; Škarabot et al., 2023). Thus, the motor units from the VL may receive more recurrent inhibition than those of distal muscles, though direct evidence of these differences remains to be found in humans (Windhorst, 1996). Interestingly, similar differences in rate coding were previously observed between proximal and distal muscles of the upper limb (De Luca et al., 1982). However, other muscles that serve different functions within the human body, such as muscles from the face, have different rate coding characteristics with much higher firing rates (Kirk et al., 2021). Future work should investigate those muscles and other to reveal the myriads of rate coding strategies in human muscles.’

(2) The evidence that the entire motor unit pool was recorded per muscle is not clear. There appears to be substantial residual EMG (Figure 1B), signal cancellation of smaller motor units (lines 172-176), some participants had fewer than 20 identified motor units, and contractions never went above 80% of MVC. Also, to my understanding, there remains no gold-standard in awake humans to estimate the total motor unit number in order to determine if the entire pool was decomposed. 

The reviewer is right that we did not decode the full pool of motor units. As indicated in the initial version of the manuscript (e.g. title, introduction), we considered that we identified an extensive sample of motor units representative of the dynamic of the pool. This claim was supported by the identification of motor units with recruitment thresholds ranging from 0 to 75% of the maximal force. 

This statement was in the introduction (P4; L109): ‘We were able to identify up to ~200 unique active motor units per muscle and per participant in two human muscles in vivo, yielding extensive samples of motor units that are representative of the entire motoneuron pools (Caillet et al., 2023a).’

Furthermore, using four HDsEMG arrays also raises questions about how some channels were placed over non-target muscles, and if motor units were decomposed from surrounding synergists.

A factor that guided our muscle choice was the low risk of crosstalk. For this, we verified with ultrasound that our arrays of 256 electrodes only covered the muscle of interest, staying away from the neighbouring muscles. This was possible as superficial muscles from the leg are bulkier than those from the upper limb. Given the small diameter of each electrode (2 mm), it is unlikely that the motor units from the neighbouring muscles were in the recorded muscle volume.

(3) The authors claim (Abstract L51; Discussion L376) that a commonly held view in the field is that rate coding is similar across motor units from the same pool. Perhaps this is in reference to some studies that have carefully assessed lower threshold motor units during lower force ramp contractions (e.g., Fuglevand et al., 2015; Revill and Fuglevand, 2017). However, a more complete integration of the literature exploring motor unit firing rate responses during rapid isometric contractions, comparing different muscles and contraction intensities would be helpful. From Figure 3, the range of rate coding in the tibialis anterior (~7-40 Hz) is greater than the vastus lateralis (~5-22 Hz) muscle across contraction levels. In agreement with other studies, the range of rate coding within some muscles is different than others (Kirk et al., 2021) and during maximal intensity (Bellemare et al., 1983) or rapid contractions (Desmedt and Godaux, 1978). Likewise, within a motor pool, there is a diversity of firing rate responses across motor units of different sizes as a function of isometric force (Monster and Chan, 1977; Desmedt and Godaux, 1977; Kukula and Clamann, 1981; Del Vecchio et al., 2019; Marshall et al., 2022). A strength of this paper is how firing rate responses are quantified across a wide range of motor unit recruitment thresholds and between two muscles. I suggest improving clarity for the general reader, especially in the motivation for testing two lower limb muscles, and elaborating on some of the functional implications.

We thank the reviewer for his input on this question. We have added references to these works and lines of research in the discussion:

(P18; L449): ‘In addition, rate coding patterns should also vary with the pattern of contractions, with fast contractions lowering the range of recruitment thresholds within motoneuron pools (Desmedt and Godaux, 1977b, 1979; van Bolhuis et al., 1997). The variability in rate coding observed here between motor units from the same pool could lead to small deviations from the size principle sometimes observed between pairs of units during isometric contractions with various patterns of force (Desmedt and Godaux, 1979; Marshall et al., 2022) or during the derecruitment phase (Bracklein et al., 2022).’ (P19; L487): ‘However, other muscles that serve different functions within the human body, such as muscles from the face, have different rate coding characteristics with much higher firing rates (Kirk et al., 2021). Future work should investigate those muscles and other to reveal the myriads of rate coding strategies in human muscles.’

In addition to the responses above, we have added a section at the beginning of the results to motivate the choice of the muscles (P6; L137):

‘16 participants performed either isometric dorsiflexion (n = 8) or knee extension tasks (n = 8) while we recorded the EMG activity of the tibialis anterior (TA - dorsiflexion) or the vastus lateralis (VL – knee extension) with four arrays of 64 surface electrodes (256 electrodes per muscle). The motoneuron pools of these two muscles of the lower limb receive a large part of common input (Laine et al., 2015; Negro et al., 2016a), constraining the recruitment of their motor units in a fixed order across tasks. They are therefore good candidates for an accurate description of rate coding. Moreover, we wanted to determine whether differences in rate coding observed between proximal and distal muscles in the upper limb (De Luca et al., 1982) were also present in the lower limb.’.

Reviewer #3 (Public Review): 

Summary: 

This is an interesting manuscript that uses state-of-the-art experimental and simulation approaches to quantify motor unit discharge patterns in the human TA and VL. The non-linear profiles of motor unit discharge were calculated and found to have an initial acceleration phase followed by an attenuation phase. Lower threshold motor units had a larger gain of the initial acceleration whereas the higher threshold motor unit had a higher gain in the attenuation phase. These data represent a technical feat and are important for understanding how humans generate and control voluntary force. 

Strengths: 

The authors used rigorous, state-of-the-art analyses to decompose and validate their motor unit data during a wide range of voluntary efforts.

The analyses are clearly presented, applied, and visualized. 

The supplemental data provides important transparency. 

We thank the reviewer for their positive appreciation of our work.

Weaknesses: 

The number of participants and muscles tested are quite small - particularly given the constraints on yield. It is unclear if this will translate to other motor pools. The justification for TA and VL should be provided.

One strength of our study is to provide relations between key-parameters of rate coding (acceleration in firing rate, increase in firing rate, hysteresis) and the recruitment thresholds of motor units within two different pools, and for each individual participant. These relations were consistent across all the participants (Figures 2 to 4), making us confident that increasing the sample size would not change the conclusions of the study.

It is likely that the differences observed here between the VL and TA will also appear between other muscles of the leg, due to differences in the arrays of excitatory and inhibitory inputs they receive, the pattern of inhibitory inputs during increases in force (recurrent/reciprocal inhibition), and different levels of neuromodulation (Johnson et al., 2017, J Neurophysiol; Beauchamp et al., 2023; J Neural Eng). We have added a paragraph in the results to motivate our choice of muscles (P6; L137):

‘16 participants performed either isometric dorsiflexion (n = 8) or knee extension tasks (n = 8) while we recorded the EMG activity of the tibialis anterior (TA - dorsiflexion) or the vastus lateralis (VL – knee extension) with four arrays of 64 surface electrodes (256 electrodes per muscle). The motoneuron pools of these two muscles of the lower limb receive a large part of common input (Laine et al., 2015; Negro et al., 2016a), constraining the recruitment of their motor units in a fixed order across tasks. They are therefore good candidates for an accurate description of rate coding. Moreover, we wanted to determine whether differences in rate coding observed between proximal and distal muscles in the upper limb (De Luca et al., 1982) were also present in the lower limb.’.

While an impressive effort was made to identify and track motor units across a range of contractions, it appears that a substantial portion of muscle force was not identified. Though high-intensity contractions are challenging to decompose - the authors are commended for their technical ability to record population motor unit discharge times with recruitment thresholds up to 75% of a participant's maximal voluntary contractions. However previous groups have seen substantial recruitment of motor units above 80% and even 90% maximum activation in the soleus. Given the innervation ratios of higher threshold motor units, if recruitment continued to 100%, the top quartile would likely represent a substantial portion of the traditional fast-fatigable motor units. It would be highly interesting to understand the recruitment and rate coding of the highest threshold motor units, at a minimum I would suggest using terms other than "entire range" or "full spectrum of recruitment thresholds"

Motor units were indeed identified between 0 and 80% of the maximal force in this study. This is due to the requirements of the decomposition algorithm that needs sustained and stable contraction to converge toward a set of separation vectors that generate sparse spike trains. Thus, it was not possible for our participants to sustain contractions above 80%MVC without generating fatigue.

However, it is important to note that only a few motor units are recruited above 80% of the maximal force in the TA (Van Cutsem et al., 1998, J Physiol), as well as in other muscles of the lower limb (Oya et al., 2009, J Physiol; Aeles et al., 2020, J Neurophysiol). Thus, we may have only missed a few motor units recruited above 80% of the maximal force. Nevertheless, we removed the terms ‘full spectrum of recruitment thresholds’ and ‘entire range’ from the manuscript to now read ‘most of the spectrum of recruitment thresholds observed in humans.’.

The quantification of hysteresis using torque appears to make self-evident the observation that lower threshold motor units demonstrate less hysteresis with respect to torque. If there is motor unit discharge there will be force. I believe this limitation goes beyond the floor effects discussed in the manuscript. Traditionally, individuals have used the discharge of a lower threshold unit as the measure on which to apply hysteresis analyses to infer ion channel function in human spinal motoneurons.

We agree with the reviewer that the hysteresis is classically estimated using the firing rate of a ‘reporter unit’ with the delta F method (introduced in humans by Gorassini et al..), or most recently with the advances in motor unit identification using the cumulative spike train of the identified motor unit. The researchers use this data as a proxy of the synaptic drive, and compare their values at recruitment and derecruitment thresholds of the ‘test unit’. 

As mentioned above in response to reviewer 1, this approach was not possible in our study as we did not have the same units across contractions to estimate cumulative spike trains. It was therefore not possible to pool the data across contractions as we did here to generate force/firing rate relations on the widest range of force. This limitation is now highlighted in the discussion section (P19; L470): ‘This result must be confirmed with a more direct proxy of the net synaptic drive, such as the firing rate of a reference low-threshold motor neuron used in the delta F method (Gorassini et al., 1998), or the cumulative spike train of low-threshold motor neurons (Afsharipour et al., 2020).’.

The main findings are not entirely novel. See Monster and Chan 1977 and Kanosue et al 1979. 

We agree with the reviewer that the results of the paper are remarkably aligned with previous experimental findings in humans, in animals, or with in vitro and in silico models. However, we believe that our study shows in humans the incredible variety of rate coding patterns within a pool of motor units that span most of the spectrum of recruitment thresholds observed in humans. It also highlights the variability of rate coding patterns between motor neurons that have a similar recruitment threshold. Finally, we observe differences between pools of motor neurons innervating two different muscles in the lower limb, mirroring what has been done in the past in the upper limb muscle. 

Recommendations for the authors:  

Reviewer #1 (Recommendations For The Authors): 

The wording 'decode' across the manuscript may sound somewhat unsuitable for the context, because 'decode' would involve interpreting the signals and activities to understand how they relate to specific variables or proxies of behavior. Here in this study it does not necessarily involve the interpretation, but sounds to be used for decomposing the signal into the constituent motor units. As such, it might be appropriate to use other words such as decompose, read out, or extract.

‘Decode’ was removed from the manuscript to now read motor unit ‘identification’

Reviewer #2 (Recommendations For The Authors): 

Figures 1 and 2 are informative and interesting. Figures 3 and 4 are harder to interpret. For example, in Figure 4, data plotted along the diagonal is overplotted and not as informative.

For the sake of clarity, we separated the lines of the fits and the scatter plots in in the right panels in Figure 3. In Figure 4, we remove the scatter plots and only reported the lines of the fits for each participant. 

Do you think the different durations of the isometric plateau across contraction intensities influenced motor unit derecruitment? Longer duration in lower threshold motor units would have resulted in a larger effect of PICs?

We did not find an effect of the duration of the plateau on the derecruitment threshold. Notably, a computational study found that the duration of the plateau may impact the delta F, due to the combination of PICs, spike threshold accommodation and spike frequency adaptation (Revill & Fuglevand, 2011, J Neurophysiol). However, we did not use the delta F value here to estimate the effect of PICs on the hysteresis. 

L703. For the measure of firing rate hysteresis the difference between recruitment and derecruitment was calculated, but why not use the delta-F method? This is more commonly used to assess hysteresis as a rough estimate of intrinsic dynamics.

As further discussed above, this approach was not possible in our study as we did not have the same units across contractions to estimate cumulative spike trains. It was therefore not possible to pool the data across contractions as we did here to generate force/firing rate relations on the widest range of force.

This was mentioned in the discussion (P19; L470):

‘This result must be confirmed with a more direct proxy of the net synaptic drive, such as the firing rate of a reference low-threshold motor neuron used in the delta F method (Gorassini et al., 1998), or the cumulative spike train of low-threshold motor neurons (Afsharipour et al., 2020).’

L144. The standard deviation seems high. Some participants had fewer than 20 motor units and your number of participants per muscle was eight, could you state the complete range?

A table was added in the results section to indicate the yields of the decomposition per contraction.

If other studies are able to randomly sample motor units with intramuscular electrodes does this also represent an estimate of rate coding from the 'entire' pool? One criticism of HDsEMG arrays is that they are biased towards decomposing superficial larger motor units and in the male sex. 

The decomposition of EMG signals recorded with arrays of surface electrodes is indeed biased toward the identification of motor units with the larger action potentials in the signal (large and superficial; Farina & Holobar, 2016, Proceedings of IEEE). We took advantage of the latter limitation by performing successive contractions at different levels of force with the objective to identify the last recruited motor units (larger units according to the size principle), while tracking the smaller ones. In that way, we were able to sequentially identify motor units recruited from 0% to 75% of the maximal force. A similar approach could be applied to selective intramuscular electrodes. However, because identifying motor units up to maximal force requires a highly selective pair of fine wires or needle electrodes, the procedure described above should be repeated hundreds of times to reach the same samples as those obtained in our study.

L151-161. The ratio between simulated and decomposed surface EMG reached 55% for the TA and 70% for the VL. How does this provide support that the "entire" MU pool was sampled?

As said above, we do not identify all the motor units during each contraction, but rather the larger ones with the larger action potentials within the EMG signals. However, we used here a sequential approach to identify new motor units during each trial while tracking smaller units. In that way, we were able to sequentially identify on average 130 motor units per muscle.

To avoid any confusion, we removed the references to ‘entire’ pools in the manuscript.  

L266. How is it possible that in some participants no motor units were recruited below 5% of MVC? Do the authors suspect they produced force from synergist muscles or that the decomposition failed to identify these presumably smaller and deeper motor units?

This mostly results from the limitations of the decomposition algorithm. In these participants, it is likely that the decomposition was biased toward motor units only active during the plateau of force or recruited at the end of the ramp.

Figure 2B. Do the higher threshold motor units with linear responses receive more inhibitory input (coactivation) or are devoid of large PIC effects?

Were antagonist muscles recorded? During higher contraction intensities, greater antagonist coactivation in some trials or participants may have linearized the firing rate profiles (e.g., Revill and Fuglevand, 2017).

L427. This is a neat finding that higher threshold motor units are less likely to have the functional  hallmark of a strong PIC effect and may therefore be more representative of extrinsic inputs. Could this be an advantage to increase the precision of stronger contractions or reduce the fatigability of muscle fibres during repeated strong contractions?

Synaptic contacts with Renshaw cells (Fyffe, 1991, J Neurophysiol) and Ia inhibitory interneurons (Heckman & Binder, 1991, J Neurophysiol) are widespread within pools of motor units, which induces homogeneously distributed inhibitory inputs. However, the amplitude of these inhibitory inputs can increase with muscle force. We found that the EMG amplitude of the soleus and the gastrocnemius medialis recorded with bipolar EMG during the dorsiflexion increased with the force. Therefore, the higher inhibitory at higher force may also contribute to the linearisation of the force/firing rate relations observed with high threshold motor neurons, as suggested by Revill and Fuglevand (2017, J Physiol). 

We discussed this point in the new version of the manuscript (P17; L415):

‘The level of recurrent and reciprocal inhibition has also probably increased with the increase in force during the ramp up, progressively blunting the effect of persistent inward currents for late-recruited motor units (Kuo et al., 2003; Hyngstrom et al., 2007; Revill and Fuglevand, 2017). This may also explain the larger percentage of high-threshold motor units with a linear fit for the firing rate/force relation (Figure 2), as the integration of larger inhibitory inputs should linearise the firing rate/force relation (Revill and Fuglevand, 2017).’. 

In Figure 2B, it makes sense that linear firing rate responses occur later in the ramp contraction when myotendinous slack is lower. Do the authors think contractile dynamics are matched to the firing rate profiles?

To our knowledge, there is no direct data on the link between the linearity of the force/firing rate relation and the stiffness of the tendon. A recent work from Mazzo et al. (2021, J Physiol) has shown that repeated stretches of calf muscles, which induce a decrease in their stiffness, induced an increase in motor unit firing rate at low levels of forces. This indicates that the contractile properties of the muscle may potentially also impact the profile of rate coding when considered as function of force. 

We added this point in the discussion (P20; L512):

‘On a different note, the steep increase in firing rate over the first percentages of the ramp-up may also enable the motor units to produce the required level of force despite having a more compliant muscletendon unit (Mazzo et al., 2021).’

L371. It is likely that Marshall et al., 2022, recorded over 100 unique motor units from the same animal.

The reviewer is right that Marshall may have identified hundreds of motor units across sessions in one non-human primate. However, there is no ways to verify this statement as they used fine wire electrodes inserted in different locations in each session, which made it impossible to verify the uniqueness of each identified unit. Conversely, we verified in our study that all the motor units were unique using the distribution of their surface action potentials across the 236 surface electrodes.

L378. What do the authors mean by "rate coding is similar"? I find this statement confusing. Is this regarding the absolute firing rate range, response to force increases, hysteresis, or how they scale with contraction intensity?

This statement was removed from the discussion to avoid any confusion.

Reviewer #3 (Recommendations For The Authors): 

The authors may want to consider other mechanisms of the linearization of discharge rates of medium and high threshold motor units. Monica's work may suggest that, over time, there is a subthreshold activation of the PIC, which serves to linearize the eventual suprathreshold activation underlying repetitive discharge. Additionally, Andy has shown that inhibitory drive from cutaneous inputs can linearize the initial acceleration of low threshold motor units - cutaneous inputs, or even Ib inputs, may be greater later in the contraction and serve to linearize discharge rates. 

We thank the reviewer for their input on the discussion, where we now discuss this point:

‘The level of recurrent and reciprocal inhibition has also probably increased with the increase in force during the ramp up, progressively blunting the effect of persistent inward currents for late-recruited motor units (Kuo et al., 2003; Hyngstrom et al., 2007; Revill and Fuglevand, 2017). This may also explain the larger percentage of high-threshold motor units with a linear fit for the firing rate/force relation (Figure 2), as the integration of larger inhibitory inputs should linearise the firing rate/force relation (Revill and Fuglevand, 2017).’. 

Lines 433 - intrinsic properties, in particular the afterhyperpolarization, will likely influence maximal discharge rate and provide a ceiling to the change in firing rate.

This point is now discussed in the draft (P17; L428):

‘This difference may be explained by smaller excitatory synaptic inputs onto low- than high-threshold motoneurons (Powers and Binder, 2001; Heckman and Enoka, 2012), lower synaptic driving potential of the dendritic membrane (Powers and Binder, 2000; Cushing et al., 2005; Fuglevand et al., 2015), and longer and larger afterhyperpolarisation phase in low- than high-threshold motoneurons (Bakels and Kernell, 1993; Gardiner, 1993; Deardorff et al., 2013; Caillet et al., 2022).’

The actual yield per contraction is not entirely clear. Figure S2 is quite nice in this regard, but a table with this and other information on it may be helpful. This would help with the beginning of the abstract and discussion when it is stated that, on average over 100 motor units were identified per person. 

We added a table in the results to give the number of motor units identified per contraction.

Are the thin film units represented in S2 and S3?

Only motor units identified from signals recorded with arrays of surface electrodes are presented in figures S2 and S3.

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public Review):

Summary:

This important work advances our understanding of sperm motility regulation during fertilization by uncovering the midpiece/mitochondria contraction associated with motility cessation and structural changes in the midpiece actin network as its mode of action. The evidence supporting the conclusion is solid, with rigorous live cell imaging using state-of-the-art microscopy, although more functional analysis of the midpiece/mitochondria contraction would have further strengthened the study. The work will be of broad interest to cell biologists working on the cytoskeleton, mitochondria, cell fusion, and fertilization. Strengths: The authors demonstrate that structural changes in the flagellar midpiece F-actin network are concomitant to midpiece/mitochondrial contraction and motility arrest during sperm-egg fusion by rigorous live cell imaging using state-of-art microscopy.

Response P1.1: We thank the reviewer for her/his positive assessment of our manuscript.

Weaknesses:

Many interesting observations are listed as correlated or in time series but do not necessarily demonstrate the causality and it remains to be further tested whether the sperm undergoing midpiece contraction are those that fertilize or those that are not selected. Further elaboration of the function of the midpiece contraction associated with motility cessation (a major key discovery of the manuscript) would benefit from a more mechanistic study.

Response P1.2: We thank the reviewer for this point. We have toned down some of our statements since some of the observations are indeed temporal correlations. We will explore some of these possible connections in future experiments. In addition, we have now incorporated additional experiments and possible explanations about the function of the midpiece contraction.

Reviewer #2 (Public Review): 

(1) The authors used various microscopy techniques, including super-resolution microscopy, to observe the changes that occur in the midpiece of mouse sperm flagella. Previously, it was shown that actin filaments form a double helix in the midpiece. This study reveals that the structure of these actin filaments changes after the acrosome reaction and before sperm-egg fusion, resulting in a thinner midpiece. Furthermore, by combining midpiece structure observation with calcium imaging, the authors show that changes in intracellular calcium concentrations precede structural changes in the midpiece. The cessation of sperm motility by these changes may be important for fusion with the egg. Elucidation of the structural changes in the midpiece could lead to a better understanding of fertilization and the etiology of male infertility. The conclusions of this manuscript are largely supported by the data, but there are several areas for improvement in data analysis and interpretation. Please see the major points below.

Response P2.1: We thank the reviewer for the positive comments.

(2) It is unclear whether an increased FM4-64 signal in the midpiece precedes the arrest of sperm motility. in or This needs to be clarified to argue that structural changes in the midpiece cause sperm motility arrest. The authors should analyze changes in both motility and FM4-64 signal over time for individual sperm.

Response P2.2 : We have conducted single cell experiments tracking both FM4-64 and motility as the reviewer suggested (Supplementary Fig S1). We have observed that in all cases, cells gradually diminished the beating frequency and increased FM4-64 fluorescence in the midpiece until a complete motility arrest is observed. A representative example is shown in this Figure but we will reinforce this concept in the results section.

(3) It is possible that sperm stop moving because they die. Figure 1G shows that the FM464 signal is increased in the midpiece of immotile sperm, but it is necessary to show that the FM4-64 signal is increased in sperm that are not dead and retain plasma membrane integrity by checking sperm viability with propidium iodide or other means.

Response P2.3: This is a very good point. In our experiments, we always considered sperm that were motile to hypothesize about the relevance of this observation. We have two types of experiments: 

(1) Sperm-egg Fusion: In experiments where sperm and eggs were imaged to observe their fusion, sperm were initially moving and after fusion, the midpiece contraction (increase in FM4-64 fluorescence was observed) indicating that the change in the midpiece (that was observed consistently in all fusing cells analyzed), is part of the process. 

(2) Sperm that underwent acrosomal exocytosis (AE): we have observed two behaviours as shown in Figure 1: 

a) Sperm that underwent AE and they remain motile without midpiece contraction (they are alive for sure); 

b) Sperm that underwent AE and stopped moving with an increase in FM464 fluorescence. We propose that this contraction during AE is not desired because it will impede sperm from moving forward to the fertilization site when they are in the female reproductive tract. In this case, we acknowledge that the cessation of sperm motility may be attributed to cellular death, potentially correlating with the increased FM4-64 signal observed in the midpiece of immotile sperm that have undergone AE. To address this hypothesis, we conducted image-based flow cytometry experiments, which are well-suited for assessing cellular heterogeneity within large populations.

Author response image 1 illustrates the relationship between cell death and spontaneous AE in noncapacitated mouse sperm, where intact acrosomes are marked by EGFP. Cell death was evaluated using Sytox Blue staining, a dye that is impermeable to live cells and shows affinity for DNA. AE was assessed by the absence of EGFP in the acrosome. 

Author response image 1a indicates a lack of correlation between Sytox and EGFP fluorescence. Two populations of sperm with EGFP signals were found (EGFP+ and EGFP-), each showing a broad distribution of Sytox signal, enabling the distinction between cells that retain plasma membrane integrity (live sperm: Sytox-) and those with compromised membranes (dead cells: Sytox+). The observed bimodal distribution of EGFP signal, regardless of live versus dead cell populations, indicates that the fenestration of the plasma membrane known to occur during AE is a regulated process that does not necessarily compromise the overall plasma membrane integrity. 

These observations are reinforced by the single-cell examples in Author response image 1b, where we were able to identify sperm in four categories: live sperm with intact acrosome (EGFP+/Sytox-), live sperm with acrosomal exocytosis (EGFP-/Sytox-), dead sperm with intact acrosome (EGFP+/Sytox+), and dead sperm with AE (EGFP-/Sytox+). Note the case of AE (lacking EGFP signal) which bears an intact plasma membrane (lacking Sytox Blue signal). Author response image 2 shows single-cell examples of the four categories observed with confocal microscopy to reinforce the observations from Author response image 1a.

Author response image 1.

Fi. Image based flow cytometry analysis (ImageStream Merk II), of non-capacitated mouse sperm, showing the distribution of EGFP signal (acrosome integrity) against Sytox Blue staining (cell viability).  (A) The quadrants show: Sytox Blue + / EGFP low (17.6%), Sytox Blue + / EGFP high (40.1%), Sytox Blue - / EGFP high (20.2%), and Sytox Blue - / EGFP low (21.7%). Each quadrant indicates the percentage of the total sperm population exhibiting the corresponding staining pattern. Axes are presented in a log10 scale of arbitrary units of fluorescence.  (B) Representative single-cell images corresponding to the four categorized sperm populations from the flow cytometry analysis in panel (A). The top row displays sperm with compromised plasma membrane integrity (Sytox Blue +), showing low (left) and high (right) EGFP signals. The bottom row shows sperm with intact plasma membrane (Sytox Blue -), displaying high (left) and low (right) EGFP signal. It is worth noting that when analyzing the percentages in (A), we observed that the data also encompass a population of headless flagella, which was present in all observed categories. Therefore, the percentages should be interpreted with caution.

Author response image 2.

Confocal Microscopy Examples of AE and cell viability. The top row features sperm with compromised plasma membrane integrity (Sytox Blue +) and high EGFP expression; the second row displays sperm with compromised membrane and low EGFP expression; the third row illustrates sperm with intact membrane (Sytox Blue -) and high EGFP expression; the bottom row shows sperm with intact membrane and low EGFP expression. 

Author response images 3-5 provide insight into the relationship between FM4-64 and Sytox Blue fluorescence intensities in non-capacitated sperm (CTRL, Author response image 3), capacitated sperm and acrosome exocytosis events stimulated with 100 µM progesterone (PG, Author response image 4), and capacitated sperm stimulated with 20 µM ionomycin (IONO, Author response image 5). Two populations of sperm with Sytox Blue signals were clearly distinguished (Sytox+ and Sytox-), enabling the discernment between live and dead sperm. Interestingly, the upper right panels of Author response images 3A, 4A, and 5A (Sytox Blue+ / FM4-64 high) consistently show a positive correlation between FM4-64 and Sytox Blue. This observation aligns with the concern raised by Reviewer 2, suggesting that compromised membranes due to cell death provide more binding sites for FM4-64. 

Nonetheless, the lower panels of Author response images 3A, 4A and 5A (Sytox Blue-) show no correlation with FM4-64 fluorescence, indicating that this population can exhibit either low or high FM4-64 fluorescence. As expected, in stark contrast with the CTRL case, the stimulation of AE with PG or IONO in capacitated sperm increased the population of live sperm with high FM4-64 fluorescence (Sytox Blue+ / FM4-64 high: CTRL: 7.85%, PG: 8.73%, IONO: 13.5%). 

Single-cell examples are shown in Author response images 3B, 4B, and 5B, where the four categories are represented: dead sperm with low FM4-64 fluorescence (Sytox Blue+ / FM4-64 low), dead sperm with high FM4-64 fluorescence (Sytox Blue+ / FM4-64 high), live sperm with low FM4-64 fluorescence (Sytox Blue- / FM4-64 low), and live sperm with high FM4-64 fluorescence (Sytox Blue- / FM4-64 high). 

Author response image 3.

Relationship between cell death and FM4-64 fluorescence in capacitated sperm without inductor of RA. Image-based flow cytometry analysis of non-capacitated mouse sperm loaded with FM464 and Sytox Blue dyes, with one and two minutes of incubation time, respectively. (A) The quadrants show: Sytox Blue+ / FM4-64 low (13.3%), Sytox Blue+ / FM4-64 high (49.8%), Sytox Blue- / FM4-64 low (28.1%), and Sytox Blue- / FM4-64 high (7.85%). Each quadrant indicates the percentage of the total sperm population exhibiting the corresponding staining pattern. Axes are presented on a log10 scale of arbitrary units of fluorescence. (B) Representative single-cell images corresponding to the four categorized sperm populations from the flow cytometry analysis in panel (A).

Author response image 4.

Relationship between cell death and FM4-64 fluorescence capacitated sperm stimulated with progesterone. Image-based flow cytometry analysis of non-capacitated mouse sperm loaded with FM4-64 and Sytox Blue dyes, with one and two minutes of incubation time, respectively. (A) The quadrants show: Sytox Blue+ / FM4-64 low (9.04%), Sytox Blue+ / FM4-64 high (61.6%), Sytox Blue- / FM4-64 low (19.7%), and Sytox Blue- / FM4-64 high (8.73%). Each quadrant indicates the percentage of the total sperm population exhibiting the corresponding staining pattern. Axes are presented on a log10 scale of arbitrary units of fluorescence. (B) Representative single-cell images corresponding to the four categorized sperm populations from the flow cytometry analysis in panel (A)

Author response image 5.

Relationship between cell death and FM4-64 fluorescence capacitated sperm stimulated with ionomycin. Image-based flow cytometry analysis of non-capacitated mouse sperm loaded with FM464 and Sytox Blue dyes, with one and two minutes of incubation time, respectively. (A) The quadrants show: Sytox Blue+ / FM4-64 low (4.52%), Sytox Blue+ / FM4-64 high (60.6%), Sytox Blue- / FM4-64 low (20.5%), and Sytox Blue- / FM4-64 high (13.5%). Each quadrant indicates the percentage of the total sperm population exhibiting the corresponding staining pattern. Axes are presented on a log10 scale of arbitrary units of fluorescence. (B) Representative single-cell images corresponding to the four categorized sperm populations from the flow cytometry analysis in panel (A).

Based on the data presented in Author response images 1 to 6, we derive the following conclusions summarized below:

(1) There is no direct relationship between cell death (Sytox Blue-) and AE (EGFP) (Author response images 1 and 2).

(2) There is bistability in the FM4-64 fluorescent intensity. Before reaching a certain threshold, there is no correlation between FM4-64 and Sytox Blue signals, indicating no cell death. However, after crossing this threshold, the FM4-64 signal becomes correlated with Sytox Blue+ cells, indicating cell death (Author response images 4-6).

(3) The Sytox Blue- population of capacitated sperm is sensitive to AE stimulation with progesterone, leading to the expected increase in FM4-64 fluorescence.

Therefore, while the FM4-64 signal alone is not a definitive marker for either AE or cell death, it is crucial to use additional viability assessments, such as Sytox Blue, to accurately differentiate between live and dead sperm in studies of acrosome exocytosis and sperm motility. In the present work, we did not use a cell viability marker due to the complex multicolor, multidimensional fluorescence experiments. However, cell viability was always considered, as any imaged sperm was chosen based on motility, indicated by a beating flagellum. The determination of whether selected sperm die during or after AE remains to be elucidated. The results presented in Figure 2 and Supplementary S1 show examples of motile sperm that experience an increase in FM4-64 fluorescence.

All this information is added to the manuscript (Supplementary Figure 1D).

(4) It is unclear how the structural change in the midpiece causes the entire sperm flagellum, including the principal piece, to stop moving. It will be easier for readers to understand if the authors discuss possible mechanisms.

Response P2.4: As requested, we have incorporated a possible explanation in the discussion section (see line 644-656). We propose three possible hypotheses for the cessation of sperm motility, which can be attributed to the simultaneous occurrence of various events:

(1) Rapid increase in [Ca2+]i levels: A rapid increase in [Ca2+]i levels may trigger the activation of Ca2+ pumps within the flagellum. This process consumes local ATP levels, disrupting glycolysis and thereby depleting the energy required for motility.

(2) Reorganization of the actin cytoskeleton: Alterations in the actin cytoskeleton can lead to changes in the mechanical properties of the flagellum, impacting its ability to move effectively.

(3) Midpiece contraction: Contraction in the midpiece region can potentially interfere with mitochondrial function, impeding the energy production necessary for sustained motility.

(5) The mitochondrial sheath and cell membrane are very close together when observed by transmission electron microscopy. The image in Figure 9A with the large space between the plasma membrane and mitochondria is misleading and should be corrected. The authors state that the distance between the plasma membrane and mitochondria approaches about 100 nm after the acrosome reaction (Line 330 - Line 333), but this is a very long distance and large structural changes may occur in the midpiece. Was there any change in the mitochondria themselves when they were observed with the DsRed2 signal?

Response P2.5: The authors appreciate the reviewer’s observation regarding the need to correct the image in Figure 9A, as the original depiction conveys a misleading representation of the spatial relationship between the mitochondrial sheath and the plasma membrane. This figure has been corrected to accurately reflect a more realistic proximity, while keeping in mind that it is a cartoonish representation.

Regarding the comments about the distances mentioned between former lines 330 and 333, the measurement was not intended to describe the gap between the plasma membrane and the mitochondria but rather the distance between F-actin and the plasma membrane. 

Author response image 6 shows high-resolution scanning electron microscopy (SEM) of two sperm fixed with a protocol tailored to preserve plasma membranes (ref), where the insets clearly show the flagellate architecture in the midpiece with an intact plasma membrane covering the mitochondrial network. A non-capacitated sperm with an intact acrosome is shown in panel A, and a capacitated sperm that has experienced AE is shown in panel B.

Notably, the results depicted in Author response image 6 demonstrate that, irrespective of the AE status, the distance between the plasma membrane and mitochondria consistently remains less than 20 nm, thus confirming the close proximity of these structures in both physiological states. As Reviewer 2 pointed out, if there is no significant difference in the distance between the plasma membrane and mitochondria, then the observed structural changes in the actin network within the midpiece should somehow alter the actual deposition of mitochondria within the midpiece. Figure 5D-F shows that midpiece contraction is associated with a decrease in the helical pitch of the actin network; the distance between turns of the actin helix decreases from  l = 248  nm to  l = 159  nm. This implies a net change in the number of turns the helix makes per 1 µm, from 4 to 6 µm-1.

Author response image 6.

SEM image showing the proximity between plasma membrane and mitochondria. Scale bar 100 nm.

Additionally, a structural contraction can be observed in Figure 5D-F, where the radius of the helix decreases by about 50 nm. To clarify this point, we sought to measure the deposition of individual DsRed2 mitochondria using computational superresolution microscopy—FF-SRM (SRRF and MSSR), Structured Illumination Microscopy (SIM), or a combination of both (SIM + MSSR), in 2D. Author response image 7 shows that these three approaches allow the observation of individual DsRed mitochondria; however, the complexity of their 3D arrangement, combined with the limited space between mitochondria (as seen in Author response image 6), precludes a reliable estimation of mitochondrial organization within the midpiece. To overcome these challenges, we decided to study the midpiece architecture via SEM experiments on non-capacitated versus capacitated sperm stimulated with ionomycin to undergo the AE.

Author response image 7.

Organization of mitochondria observed via FF-SRM and SIM. Scale bar 2 µm. F.N: Fluorescence normalized. F: Frequency

Author response image 8 presents a single-cell comparison of the midpiece architecture in noncapacitated (NC) and acrosome-intact (AI) versus acrosome-reacted (AR) sperm, along with measurements of the midpiece diameter throughout its length. Notably, the diameter of the midpiece increases from the base of the head to more distal regions, ranging from 0.45 nm to 1.10 µm (as shown in Author response images 7 and 8). A significant correlation between the diameter of the flagellum and its curvature was observed (Author response image 9), suggesting a reorganization of the midpiece due to shearing forces. This is further exemplified in Author response images 8 and 9, which provide individual examples of this phenomenon.

Author response image 8.

Comparison of the midpiece architecture in acrosome-intact and acrosome-reacted sperm using scanning electron microscopy (SEM).

As expected, the overall diameter of the midpiece in AI sperm was larger than in AR sperm, with measurements of 0.731 ± 0.008 µm for AI and 0.694 ± 0.007 µm for AR (p = 0.013, Kruskal-Wallis test n > 100, N = 2), as shown in Author response image 10. Additionally, this Author response image 7 indicates that the reorganization of the midpiece architecture involves a change in the periodicity of the mitochondrial network, with frequencies shifting from fNC to fEA mitochondria per micron.  

Author response image 9.

Comparison of the midpiece architecture in acrosome-intact (A) and acrosome-reacted (B) sperm using scanning electron microscopy (SEM).

Collectively, the structural results presented in Figure 5 and Author response images 6 to 10 demonstrate that the AE involves a comprehensive reorganization of the midpiece, affecting its diameter, pitch, and the organization of both the actin and mitochondrial networks. All this information is now incorporated in the new version of the paper (Figure. 2F)

Author response image 10.

Quantification of the midpiece diameter of the sperm flagellum in acrosome-intact and acrosome-reacted sperm analyzed by scanning electron microscopy (SEM). Data is presented as mean ± SEM. Kruskal-Wallis test was employed,  p = 0.013 (AI n=85 , AR n=72).

(6) In the TG sperm used, the green fluorescence of the acrosome disappears when sperm die. Figure 1C should be analyzed only with live sperm by checking viability with propidium iodide or other means.

Response P2.6: We concur with Reviewer 2 that ideally, any experiment conducted for this study should include an intrinsic cell viability test. However, the current research employs a wide array of multidimensional imaging techniques that are not always compatible with, or might be suboptimal for, simultaneous viability assessments. In agreement with the reviewer's concerns, it is recognized that the data presented in Figure 1C may inherently be biased due to cell death. Nonetheless, Author response image 1 demonstrates that the relationship between AE and cell death is more complex than a straightforward all-or-nothing scenario. Specifically, Author response image 1C illustrates a case where the plasma membrane is compromised (Sytox Blue+) yet maintains acrosomal integrity (EGFP+). This observation contradicts Reviewer 1's assertion that "the green fluorescence of the acrosome disappears when sperm die," as discussed more comprehensively in response P2.3.

In light of these observations, we have meticulously revisited the entire manuscript to address and clarify potential biases in our results due to cell death. Consequently, Author response image 5 and its detailed description have been incorporated into the supplementary material of the manuscript to contribute to the transparency and reliability of our findings.

Reviewer #3 (Public Review):

(1) While progressive and also hyperactivated motility are required for sperm to reach the site of fertilization and to penetrate the oocyte's outer vestments, during fusion with the oocyte's plasma membrane it has been observed that sperm motility ceases. Identifying the underlying molecular mechanisms would provide novel insights into a crucial but mostly overlooked physiological change during the sperm's life cycle. In this publication, the authors aim to provide evidence that the helical actin structure surrounding the sperm mitochondria in the midpiece plays a role in regulating sperm motility, specifically the motility arrest during sperm fusion but also during earlier cessation of motility in a subpopulation of sperm post acrosomal exocytosis. The main observation the authors make is that in a subpopulation of sperm undergoing acrosomal exocytosis and sperm that fuse with the plasma membrane of the oocyte display a decrease in midpiece parameter due to a 200 nm shift of the plasma membrane towards the actin helix. The authors show the decrease in midpiece diameter via various microscopy techniques all based on membrane dyes, bright-field images and other orthogonal approaches like electron microscopy would confirm those observations if true but are missing. The lack of additional experimental evidence and the fact that the authors simultaneously observe an increase in membrane dye fluorescence suggests that the membrane dyes instead might be internalized and are now staining intracellular membranes, creating a false-positive result. The authors also propose that the midpiece diameter decrease is driven by changes in sperm intracellular Ca2+ and structural changes of the actin helix network. Important controls and additional experiments are needed to prove that the events observed by the authors are causally dependent and not simply a result of sperm cells dying.

Response P3.1: We appreciate the reviewer's observations and critiques. In response, we have expanded our experimental approach to include alternative methodologies such as mathematical modeling and electron microscopy, alongside further fluorescence microscopy studies. This diversified approach aims to mitigate potential interpretation artifacts and substantiate the validity of our observations regarding the contraction of the sperm midpiece. Additionally, we have implemented further control experiments to fortify the credibility and robustness of our findings, ensuring a more comprehensive and reliable set of results.

First, we acknowledge the concerns raised by Reviewer 2 regarding the interpretation of the magnitude of the observed contraction of the sperm flagellum's midpiece (see response P2.5). Specifically, we believe that the assertion that "... there is a decrease in midpiece parameter due to a 200 nm shift of the plasma membrane towards the actin helix" stated by reviewer 3 needs careful examination. We recognize that the fluorescence microscopy data provided might not conclusively support such a substantial shift. Our live cell imaging and superresolution microscopy experiments indicate that there is a significant decrease in the diameter of the sperm flagellum associated with AE. This is supported by colocalization experiments where FM4-64-stained structures (fluorescing upon binding to membranes) are observed moving closer to Sir-Actinlabeled structures (binding to F-actin). Quantitatively, Figure S5 describes the spatial shift between FM4-64 and Sir-Actin signals, narrowing from a range of 140-210 nm to 50-110 nm (considering the 2nd and 3rd quartiles of the distributions). The mean separation distance between both signals changes from 180 nm in AI cells to 70 nm in AR cells, a net shift of 110 nm. This observation suggests caution regarding the claim of a "200 nm shift of the plasma membrane towards the actin cortex." 

Moreover, the concerns raised by Reviewer #3 about the potential internalization of membrane dyes, which might create a false-positive result by staining intracellular membranes, offer an alternative mechanism to explain a shift of up to 100 nm. This perspective is also supported by the critique from Reviewer #2 regarding the substantial distance (about 100 nm) between the plasma membrane and mitochondria post-acrosome reaction:  “The authors state that the distance between the plasma membrane and mitochondria approaches about 100 nm after the acrosome reaction (…), but this is a very long distance and large structural changes may occur in the midpiece”. These insights have prompted us to refine our methodology and interpretation of the data to ensure a more accurate representation of the underlying biological processes.

Author response image 11 shows a first principles approach in two spatial dimensions to explore three scenarios where a membrane dye, such as FM4-64, stains structures at and within the midpiece of a sperm flagellum, but yet does not result in a net change of diameter. Author response image 11A-C illustrates three theoretical arrangements of fluorescent dyes: Model 1 features two rigid, parallel structures that mimic the plasma membrane surrounding the midpiece of the flagellum. Model 2 builds on Model 1 by incorporating the possibility of dye internalization into structures located near the membrane, suggesting a slightly more complex interaction with nearby membranous intracellular structures. Model 3 represents an extreme scenario where the fluorescent dyes stain both the plasma membrane and internal structures, such as mitochondrial membranes, indicating extensive dye penetration and binding. Author response image 11D-F displays the convolution of the theoretical fluorescent signals from Models 1 to 3 with the theoretical point spread function (PSF) of a fluorescent microscope, represented by a Gaussian-like PSF with a sigma of 19 pixels (approximately 300 nm). This process simulates how each model's fluorescence would manifest under microscopic observation, showing subtle differences in the spatial distribution of fluorescence among the models. Author response image 11G-I reveals the superresolution images obtained through Mean Shift Super Resolution (MSSR) processing of the models depicted in Author response image 11D-F.

By analyzing the three scenarios, it becomes clear that the signals from Models 2 and 3 shift towards the center compared to Model 1, as depicted in Author response image 11J. This shift in fluorescence suggests that the internalization of the dye and its interaction with internal structures might significantly influence the perceived spatial distribution and intensity of fluorescence, thereby impacting the interpretation of structural changes within the midpiece. Consequently, the experimentally observed contraction of up to 100 nm in  could represent an actual contraction of the sperm flagellum's midpiece, a relocalization of the FM4-64 membrane dyes to internal structures, or a combination of both scenarios.

To discern between these possibilities, we implemented a scanning electron microscopy (SEM) approach. The findings presented in Figure 5 and Author response images 7 to 9 conclusively demonstrate that the AE involves a comprehensive reorganization of the midpiece. This reorganization affects its diameter, which changes by approximately 50 nm, as well as the pitch and the organization of both the actin and mitochondrial networks. This data corroborates the structural alterations observed and supports the validity of our interpretations regarding midpiece dynamics during the AE.

Author response image 11.

Modeling three scenarios of midpiece staining with membrane fluorescent dyes.

Secondly, we wish to clarify that in some of our experiments, we have utilized changes in the intensity of FM4-64 fluorescence as an indirect measure of midpiece contraction. This approach is supported by a linear inverse correlation between these variables, as illustrated in Figure S2D. It is important to note that this observation is correlative and indirect; therefore, our data does not directly substantiate the claim that "in a subpopulation of sperm undergoing AE and sperm that fuse with the plasma membrane of the oocyte, there is a decrease in midpiece parameter due to a 200 nm shift of the plasma membrane towards the actin helix". Specifically, we have not directly measured the distance between the plasma membrane and actin cortex in experiments involving gamete fusion.

All the concerns highlighted in this Response P1.1 have been addressed and incorporated into the manuscript. This addition aims to provide comprehensive insight into the experimental observations and methodologies used, ensuring that the data is transparent and accessible for thorough review and replication.

Editor Comment:

As the authors can see from the reviews, the reviewers had quite different degrees of enthusiasm, thus discussed extensively. The major points in consensus are summarized below and it is highly recommended that the authors consider their revisions.

(1) Causality of midpiece contraction with motility arrest is not conclusively supported by the current evidence. Time-resolved imaging of FM4-64 and motility is needed and the working model needs to be revised with two scenarios - whether the sperm contracting indicates a fertilizing sperm or sperm to be degenerated.

(2) The rationale for using FM4-64 as a plasma membrane marker is not clear as it is typically used as an endo-membrane marker, which is also related to the discrepancy of Fluo-4 signal diameter vs. FM4-64 (Figure 4E). The viability of sperm with increased FM4-64 needs to be demonstrated.

(3) The mechanism of midpiece contraction in motility cessation along the whole flagellum is not discussed.

(4) The use of an independent method to support the changes in midpiece diameter/structural changes such as DsRed (transgenic) or TEM.

(5) The claim of Ca2+ change needs to be toned down.

Response Editor: We thank the editor and the reviewers for their thorough and positive assessment of our work and the constructive feedback to further improve our manuscript. Please find below our responses to the reviewers’ comments. We have addressed all these points in the current version. Briefly,

(1) Time resolved images to show the correlation between FM4-64 fluorescence increase and the motility was incorporated

(2) The rationale for using FM4-64 was added.

(3) The mechanism of midpiece contraction was discussed in the paper

(4) An independent method was included to support our conclusions (SEM and other markers not based on membrane dyes)

(5) The results related to the calcium increase were toned down.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

(1) To claim midpiece actin polymerization/re-organization is required for AE, demonstrating that AE does not occur in the presence of actin depolymerizing drugs (e.g., Latrunculin A, Cytochalasin D) would be necessary since the current data only shows the association/correlation. Was the block of AE by actin depolymerization observed?

Response R1.1: We agree with the reviewer but unfortunately, since actin polymerization and or depolymerization in the head are important for exocytosis, we cannot use this experimental approach to dissect both events. Addition of these inhibitors block the occurrence of AE (PMID: 12604633).

(2) Please provide the rationale for using FM4-64 to visualize the plasma membrane since it has been reported to selectively stain membranes of vacuolar organelles. What is the principle of increase of FM4-64 dye intensity, other than the correlation with midpiece contraction? For example, in lines 400-402: the authors mentioned that 'some acrosomereacted moving sperm within the perivitelline space had low FM4-64 fluorescence in the midpiece (Figure 6C). After 20 minutes, these sperm stopped moving and exhibited increased FM4-64 fluorescence, indicating midpiece contraction (Figure 6D).' While recognizing the increase of FM4-64 dye intensity can be an indicator of midpiece contraction, without knowing how and when the intensity of FM4-64 dye changes, it is hard to understand this observation. Please discuss.

Response R1.2: FM4-64 is an amphiphilic styryl fluorescent dye that preferentially binds to the phospholipid components of cell membranes, embedding itself in the lipid bilayer where it interacts with phospholipid head groups. Due to its amphiphilic nature, FM dyes primarily anchor to the outer leaflet of the bilayer, which restricts their internalization. It has been demonstrated that FM4-64 enters cells through endocytic pathways, making these dyes valuable tools for studying endocytosis.

Upon binding, FM4-64's fluorescence intensifies in a more hydrophobic environment that restricts molecular rotation, thus reducing non-radiative energy loss and enhancing fluorescence. These photophysical properties render FM dyes useful for observing membrane fusion events. When present in the extracellular medium, FM dyes rapidly reach a chemical equilibrium and label the plasma membrane in proportion to the availability of binding sites.

In wound healing studies, for instance, the fluorescence of FM4-64 is known to increase at the wound site. This increase is attributed to the repair mechanisms that promote the fusion of intracellular membranes at the site of the wound, leading to a rise in FM4-64 fluorescence. Similarly, an increase in FM4-64 fluorescence has been reported in the heads of both human and mouse sperm, coinciding with AE. In this scenario, the fusion between the plasma membrane and the acrosomal vesicle provides additional binding sites for FM4-64, thus increasing the total fluorescence observed in the head. This dynamic response of FM4-64 makes it an excellent marker for studying these cellular processes in real-time.

This study is the first to report an increase in FM4-64 fluorescence in the midpiece of the sperm flagellum. Figures 5 and Author response images 6 to 9 demonstrate that during the contraction of the sperm flagellum, structural rearrangements occur, including the compaction of the mitochondrial sheath and other membranous structures. Such contraction likely increases the local density of membrane lipids, thereby elevating the local concentration of FM4-64 and enhancing the probability of fluorescence emission. Additionally, changes in the microenvironment such as pH or ionic strength during contraction might further influence FM4-64’s fluorescence properties, as detailed by Smith et al. in the Journal of Membrane Biology (2010). The photophysical behavior of FM4-64, including changes in quantum yield due to tighter membrane packing or alterations in curvature or tension, may also contribute to the increased fluorescence observed. Notably, Figure S2 indicates that other fluorescent dyes like Memglow 700, Bodipy-GM, and FM1-43 also show a dramatic increase in their fluorescence during the midpiece contraction. Investigating whether the compaction of the plasma membrane or other mesoscale processes occur in the midpiece of the sperm flagellum could be a valuable area for future research. The use of fluorescent dyes such as LAURDAN or Nile Red might provide further insights into these membrane dynamics, offering a more comprehensive understanding of the biochemical and structural changes during sperm motility and gamete fusion events.

(3) As the volume of the whole midpiece stays the same while the diameter decreases along the whole midpiece (midpiece contraction), the authors need to describe what changes in the midpiece length they observe during the contraction. Was the length of the midpiece during the contraction measured and compared before and after contraction?

Response R1.3: As requested, we have measured the length of the midpiece in AI and AR sperm. As shown in Author response image 12 (For review purposes only), no statistically significant differences were observed. 

Author response image 12.

Midpiece length measured by the length of mitochondrial DsRed2 fluorescence in EGFP-DsRed2 sperm. Measurements were done before (acrosome-intact) and after (acrosome-reacted) acrosome exocytosis and midpiece contraction. Data is presented as the mean ± sem of 14 cells induced by 10 µM ionomycin. Paired t-test was performed, resulting in no statistical significance. 

(4) Most of all, it is not clear what the midpiece, thus mitochondria, contraction means in terms of sperm bioenergetics and motility cessation. Would the contraction induce mitochondrial depolarization or hyperpolarization, increase or decrease of ATP production/consumption? It will be great if this point is discussed. For example, an increase in mitochondrial Ca2+ is a good indicator of mitochondrial activity (ATP production).

Response R1.4: That is an excellent point. We have discussed this idea in the discussion (line 620-624). We are currently exploring this idea using different approaches because we also think that these changes in the midpiece may have an impact in the function of the mitochondria and perhaps, in their fate once they are incorporated in the egg after fertilization. 

(5) The authors claimed that Ca2+ signal propagates from head to tail, which is the opposite of the previous study (PMID: 17554080). Please clarify if it is a speculation. Otherwise, please support this claim with direct experimental evidence (e.g., high-speed calcium imaging of single cells).

Response R1.5: In that study, it was claimed that a [Ca2+]i  increase that propagates from the tail to the head occurs when CatSper is stimulated. They did not evaluate the occurrence of AE when monitoring calcium.

Our data is in agreement with our previous results (PMID: 26819478) that consistently indicated that only the[Ca2+]i  rise originating in the sperm head is able to promote AE. 

(6) Figure 4E: Please explain how come Fluo4 signal diameter can be smaller than FM4-64 dye if it stains plasma membrane (at 4' and 7').

Response R1.6: When colocalizing a diffraction-limited image (Fluo4) with a super-resolution image (FM4-64), discrepancies in signal sizes and locations can become apparent due to differences in resolution. The Fluo4 signal, being diffraction-limited, adheres to a resolution limit of approximately 200-300 nanometers under conventional light microscopy. This limitation causes the fluorescence signal to appear broader and less defined. Conversely, super-resolution microscopy techniques, such as SRRF (Super-Resolution Radial Fluctuations), achieve resolutions down to tens of nanometers, allowing FM4-64 to reveal finer details at the plasma membrane and display potentially smaller apparent sizes of stained structures. Although both dyes might localize to the same cellular regions, the higher resolution of the FM4-64 image allows it to show a more precise and smaller diameter of the midpiece of the flagellum compared to the broader, less defined signal of Fluo4. To address this, the legend of Figure 4E has been slightly modified to clarify that the FM4-64 image possesses greater resolution. 

(7) Figure 5D-G: the midpiece diameter of AR intact cells was shown ~ 0.8 um or more in Figure 2, while now the radius in Figure 5 is only 300 nm. Since the diameter of the whole midpiece is nearly uniform when the acrosome is intact, clarify how and what brings this difference and where the diameter/radius measurement is done in each figure.

Response R1.7: The difference resides in what is being measured. In Figure 2, the total diameter of the cell is measured, through the maximum peaks of FM4-64 fluorescence which is a probe against plasma membrane. As for Figure 5, the radius shown makes reference to the radius of the actin double helix within the midpiece. To that end, cells were fixed and stained with phalloidin, a F-actin probe.

Minor points

(8) Figure S1 title needs to be changed. The "Midpiece contraction" concept is not introduced when Figure S1 is referred to.

Response R1.8: This was corrected in the new version.

(9) Reference #19: the authors are duplicated.

Response R1.9: This was corrected in the new version.

(10) Line 315-318: sperm undergoing contraction -> sperm undergoing AR/AE?

Response R1.10: This was corrected in the new version.

(11) Line 3632 -> punctuation missing.

Response R1.11: Modified as requested.

(12) Movie S7: please add an arrow to indicate the spermatozoon of interest.

Response R1.12:  The arrow was added as suggested.

(13) Line 515: One result of this study was that the sperm flagellum folds back during fusion coincident with the decrease in the midpiece diameter. The authors did not provide an explanation for this observation. Please speculate the function of this folding for the fertilization process.

Response R1.13: As requested, this is now incorporated in the discussion. We speculate that the folding of the flagellum during fusion further facilitates sperm immobilization because it makes it more difficult for the flagellum to beat. Such processes can enhance stability and increase the probability of fusion success. Mechanistically, the folding may occur as a consequence of the deformation-induced stress that develops during the decrease of midpiece diameter. 

Reviewer #2 (Recommendations For The Authors):

(1) Figure 2C, D, E. Does "-1" on the X-axis mean one minute before induction? If so, the diameter is already smaller and FM4-64 fluorescence intensity is higher before the induction in the spontaneous group. Does the acrosome reaction already occur at "-1" in this group?

Response R2.1: Yes, “-1” means that the measurements of the diameter/FM4-64 fluorescence was done one minute before the induction. And it is correct that the diameter is smaller and FM464 fluorescence higher in the spontaneous group because these sperm underwent acrosome exocytosis before the induction, that is, spontaneously.

(2) Figure 3D. Purple dots are not shown in the graph on the right side.

Response R2.2: Modified as requested.

(3) Lines 404-406. "These results suggest that midpiece contraction and motility cessation occur only after acrosome-reacted sperm penetrate the zona pellucida". Since midpiece contraction and motility cessation also occur before the passage through the zona pellucida (Figure 9B), "only" should be deleted.

Response R2.3: Modified as requested.

Reviewer #3 (Recommendations For The Authors):

(1) Do the authors have a hypothesis as to why the observed decrease in midpiece parameter results in cessation of sperm motility? It would be beneficial for the manuscript to include a paragraph about potential mechanisms in the discussion.

Response R3.1: As requested, a potential mechanism has been proposed in the discussion section (line 644-656).

(2) Since the authors propose in Gervasi et al. 2018 that the actin helix might be responsible for the integrity of the mitochondrial sheath and the localization of the mitochondria, is it possible that the proposed change in plasma membrane diameter and actin helix remodeling for example alters the localization of the mitochondria? TEM should be able to reveal any associated structural changes. In its current state, the manuscript lacks experimental evidence supporting the author's claim that the "helical actin structure plays a role in the final stages of motility regulation". The authors should either include additional evidence supporting their hypothesis or tone down their conclusions in the introduction and discussion.

Response  R3.3: We agree with the reviewer. This is an excellent point. As suggested by this reviewer as well as the other reviewers, we have performed SEM to observe the changes in the midpiece observed after its contraction for two main reasons. First, to confirm this observation using a different approach that does not involve the use of membrane dyes. As shown in Author response image 6-10, we have observed that in addition to the midpiece diameter, there is a reorganization of the mitochondria sheet that is also suggested by the SIM experiments. These observations will be explored with more experiments to confirm the structural and functional changes that mitochondria undergo during the contraction. We are currently investigating this phenomenon, These results are now included in the new Figure  2F.

(3) In line 134: The authors write: 'Some of the acrosome reacted sperm moved normally, whereas the majority remained immotile". Do the authors mean that a proportion of the sperm was motile prior to acrosomal exocytosis and became immotile after, or were the sperm immotile to begin with? Please clarify.

Response R3.4: This statement is based on the quantification of the motile sperm after induction of AE within the AR population (Fig. 1C). 

(4) The authors do not provide any experimental evidence supporting the first scenario. In video 1 a lot of sperm do not seem to be moving to begin with, only a few sperm show clear beating in and out of the focal plane. The highlighted sperm that acrosome-reacted upon exposure to progesterone don't seem to be moving prior to the addition of progesterone. In contrast, the sperm that spontaneously acrosome react move the whole time. In video 1 this reviewer was not able to identify one sperm that stopped moving upon acrosomal exocytosis. Similarly in video 3, although the resolution of the video makes it difficult to distinguish motile from non-motile sperm. In video 2 the authors only show sperm that are already acrosome reacted. Please explain and provide additional evidence and statistical analysis supporting that sperm stop moving upon acrosomal exocytosis.

Response R3.5: In videos 1 and 3, the cells are attached to the glass with concanavalin-A, this lectin makes sperm immotile (if well attached) because both the head and tail stick to the glass. The observed motility of sperm in these videos is likely due to them not being properly attached to the glass, which is completely normal. On the contrary, in videos 2 and 4, sperm are attached to the glass with laminin. This is a glycoprotein that only binds the sperm to the glass through its head, that is why they move freely.

(5) Could the authors provide additional information about the FM4-64 fluorescent dye?

What is the mechanism, and how does it visualize structural changes at the flagellum? Since the whole head lights up, does that mean that the dye is internalized and now stains additional membranes, similar to during wound healing assays (PMID 20442251, 33667528). Or is that an imaging artifact? How do the authors explain the correlation between FM4-64 fluorescence increase in the midpiece and the observed change in diameter? Does FM4-64 have solvatochromatic properties?

Response R3.6: We appreciate the insightful queries posed by Reviewer 3, which echo the concerns initially brought forward by Reviewer 1. For a detailed explanation of the mechanism of FM4-64 dye, how we interpret  it, visualizes structural changes in the flagellum, and its behavior during cellular processes, please refer to our detailed response in Response R1.2. In brief, FM464 is a lipophilic styryl dye that preferentially binds to the outer leaflets of cellular membranes due to its amphiphilic nature. Upon binding, the dye becomes fluorescent, allowing for the visualization of membrane dynamics. The increase in fluorescence in the sperm head or midpiece likely results from the dye’s accumulation in areas where membrane restructuring occurs, such as during AE or in response to changes in the flagellum structure.

Regarding the specific questions about internalization and whether FM4-64 stains additional membranes similarly to what is observed in wound healing assays, it's important to note that FM4-64 can indeed be internalized through endocytosis and subsequently label internal vesicular structures. Additionally, FM4-64 may experience changes in its fluorescence as a result of fusion events that increase the lipid content of the plasma membrane, as observed in studies cited (PMID 20442251, 33667528). This characteristic makes FM4-64 valuable not only for outlining cell membranes but also for tracking the dynamics of both internal and external membrane systems, particularly during cellular events that involve significant membrane remodeling, such as wound healing or AE.

Concerning whether the increased fluorescence and observed changes in diameter are artifacts or reflect real biological processes, the correlation observed likely indicates actual changes in the midpiece architecture through molecular mechanisms that remain to be further elucidated. The data presented in Figures 5 and Author response images 6-10 support that this increase in fluorescence is not merely an artifact but a feature of how FM4-64 interacts with its environment. 

Finally, regarding the solvatochromatic properties of FM4-64, while the dye does show changes in its fluorescence intensity in different environments, its solvatochromatic properties are generally less pronounced than those of dyes specifically designed to be solvatochromatic. FM464's fluorescence changes are more a result of membrane interaction dynamics and dye concentration than of solvatochromatic shifts. 

(6) For the experiment summarized in Figure S1, did the authors detect sperm that acrosome-reacted upon exposure to progesterone and kept moving? This reviewer is wondering how the authors reliably measure FM4-64 fluorescence if the flagellum moves in and out of the focal plane. If the authors observe sperm that keep moving, what was the percentage within a sperm population and how did FM4-64 fluorescence change?

Response R3.6: We did identify sperm that underwent acrosome reaction upon exposure to progesterone and continued to exhibit movement. However, due to the issue raised by the reviewer regarding the flagellum going out of focus, we opted to quantify the percentage of sperm that were adhered to the slide (using laminin). This approach allows for the observation of flagellar position over time, facilitating an easy assessment of fluorescence changes. The percentage of sperm that maintained movement after AE is depicted in Figure 1C.

(7) In Figure S1B it doesn't look like the same sperm is shown in all channels or time points, the hook shown in the EGFP channel is not always pointing in the same direction. If FM4-64 is staining the plasma membrane, how do the authors explain that the flagellum seems to be more narrow in the FM4-64 channel than in the brightfield and DsRed2 channel?

Response 3.7: It is the same sperm, but due to technical limitations images were sequentially acquired. For example, for time 5 minutes after progesterone, all images in DIC were taken, then all images in the EGFP channel, then DsRed2* and finally FM4-64. The reason for this was to acquire images as fast as possible, particularly in DIC images which were then processed to get the beat frequency.

Regarding the flagellum that seems to be more narrow in the FM4-64 channel compared to the BF or DsRed2 channel, the explanation is related to the fact that intensity of the DsRed2 signal is stronger than the other two. This higher signal may have increased the amount of photons captured by the detector.

(8) Overall, it would be beneficial to include statistics on how many sperm within a population did change FM4-64 fluorescence during AE and how many did not, in addition to information about motility changes and viability. Did the authors exclude that the addition of FM4-64 causes cell death which could result in immotile sperm or that only dying sperm show an increase in FM4-64 fluorescence?

Response 3.8: The relationship between cell death and the increase in FM4-64 fluorescence is widely discussed in Response P2.3. In our experiments, we always considered sperm that were motile to hypothesize about the relevance of this observation. We have two types of experiments: 

(1) Sperm-egg Fusion: In experiments where sperm and eggs were imaged to observe their fusion, sperm were initially moving and after fusion, the midpiece contraction (increase in FM4-64 fluorescence was observed) indicating that the change in the midpiece (that was observed consistently in all fusing cells analyzed), is part of the process. 

(2) Sperm that underwent AE: we have observed two behaviours as shown in Figure 1: 

a) Sperm that underwent AE and they remain motile without midpiece contraction (they are alive for sure); 

b) Sperm that underwent AE and stopped moving with an increase in FM464 fluorescence. We propose that this contraction during AE is not desired because it will impede sperm from moving forward to the fertilization site when they are in the female reproductive tract. In this case, we acknowledge that the cessation of sperm motility may be attributed to cellular death, potentially correlating with the increased FM4-64 signal observed in the midpiece of immotile sperm that have undergone AE. To address this hypothesis, we conducted image-based flow cytometry experiments, which are well-suited for assessing cellular heterogeneity within large populations.

Regarding the relationship between the increase in FM4-64 and AE, we have always observed that AE is followed by an increase in FM4-64 in the head in mice (PMID: 26819478) as well as in human (PMID: 25100708) sperm. This was originally corroborated with the EGFP sperm. However, not all the cells that undergo AE increase the FM4-64 fluorescence in the midpiece.

(9) The authors report that a fraction of sperm undergoes AE without a change in FM4-64 fluorescence (Figure 1F). How does the [Ca2+]i change in those cells? Again statistics on the distribution of a certain pattern within a population in addition to showing individual examples would be very helpful.

Response 3.9: A recent work shows that an initial increase in [Ca2+]i  is required to induce changes in flagellar beating necessary for hyperactivation (Sánchez-Cárdenas et al., 2018). However, when [Ca2+]i  increases beyond a certain threshold, flagellar motility ceases. These conclusions are based on single-cell experiments in murine sperm with different concentrations of the Ca2+ ionophore, A23187. The authors reported that complete loss of motility was observed when using ionophore concentrations higher than 1 μM. In contrast, spermatozoa incubated with 0.5 μM A23187 remained motile throughout the experiment. Once the Ca2+ ionophore is removed, the sperm would reduce the concentration of this ion to levels compatible with motility and hyperactivation (Navarrete et al., 2016). However, some of the washed cells did not recover mobility in the recorded time window (Sánchez-Cárdenas et al., 2018). These results would indicate that due to the increase in [Ca2+]i  induced by the ionophore, irreversible changes occurred in the sperm flagellum that prevented recovery of mobility, even when the ionophore was not present in the recording medium. 

Taking into account our results, one possible scenario to explain this irreversible change would be the contraction of the midpiece. Our results demonstrate that the increase in [Ca2+]i observed in the midpiece (whether by induction with progesterone, ionomycin or occurring spontaneously) causes the contraction of this section of the flagellum and its subsequent immobilization. 

(10) While the authors results show that changes in [Ca2+]i correlate with the observed reduction of the midpiece diameter, they do not provide evidence that the structural changes are triggered by Ca2+i influx. It could just be a coincidence that both events spatially overlap and that they temporarily follow each other. The authors should either provide additional evidence or tone down their conclusion.

Response 3.10: We agree with the reviewer. As suggested, we have toned down our conclusion.

(11) Are the authors able to detect the changes in the midpiece diameter independent from FM4-64 or other plasma membrane dyes? An alternative explanation could be that the dyes are internalized due to cell death and instead of staining the plasma membrane they are now staining intracellular membranes, resulting in increased fluorescence and giving the illusion that the midpiece diameter decreased. How do the authors explain that the Bodipy-GM1 Signal directly overlaps with DsRed2 and SIR-actin, shouldn't there be some gap? Since the rest of the manuscript is based on that proposed decrease in midpiece diameter the authors should perform orthogonal experiments to confirm their observation.

Response 3.11: As requested by the reviewer, we have not used new methods to visualize the change in sperm diameter in the midpiece. In neither of them, a membrane dye was used. First, we have performed immunofluorescence to detect a membrane protein (GLUT3). Second, we have used scanning electron microscopy. The results are now incorporated in the new Figure 2FG. In both experiments, a change in the midpiece diameter was observed. Please, also visit responses P2.5 and Author response images 8 to 10.  

Regarding the overlap between the signal of Bodipy GM1 (membrane) and the fluorescence of DsRed2 (mitochondria) and Sir-Actin (F-actin), it is only observed in acrosomereacted sperm, not in acrosome-intact sperm (Figure S4). In our view, these structures become closed after midpiece contraction, and the resolution of the images is insufficient to distinguish them clearly. This issue is also evident in Figure 5B. Therefore, we conducted additional experiments using more powerful super-resolution techniques such as STORM (Figures 5D-F).

(12) The proposed gap of 200 nM between the actin helix and the plasma membrane, has been observed by TEM? Considering that the diameter of the mouse sperm midpiece is about 1 um, that is a lot of empty space which leaves only about 600 nm for the rest of the flagellum. The axoneme is 300 nm and there needs to be room for the ODFs and the mitochondria. Please explain.

Response 3.12: Unfortunately, the filament of polymerized actin cannot be observed by TEM. Furthermore, we were discouraged from trying other approaches, such as utilizing phalloidin gold, because for some reason, it does not work properly.

In our view, the 200 nm gap between the actin cytoskeleton and the plasma membrane is occupied by the mitochondria (that is the size that it is frequently reported based on TEM; see https://doi.org/10.1172/jci.insight.166869).

(13) The results provided by the authors do not convince this reviewer that the actin helix moves, either closer to the plasma membrane or toward the mitochondria, the observed differences are minor and not confirmed by statistical analysis.

Response 3.13: As requested, the title of that section was changed. Moreover, our conclusion is exactly as the reviewer is suggesting: “Since the results of the analysis of SiR-actin slopes were not conclusive, we studied the actin cytoskeleton structure in more detail”. This conclusion is based on the statistical analysis shown in Figure S5D-E.

(14) The fluorescence intensity of all plasma membrane dyes increases in all cells chosen by the authors for further analysis. Could the increase in SiR-Actin fluorescence be explained by a microscopy artifact instead of actin helix remodeling? Alternatively, can the authors exclude that the observed increase in SIR-Actin might be an artifact caused by the increase in FM4-64 fluorescence? Since the brightness in the head similarly increases to the fluorescence in the flagellum the staining pattern looks suspiciously similar. Did the authors perform single-stain controls?

Response 3.14: We had similar concerns when we were doing the experiments using SiR-actin. Although we have performed single stain controls to make sure that the actin helix remodelling occurs during the midpiece contraction, we have performed experiments using higher resolution techniques such as STORM using a different probe to stain actin (Phalloidin).

(15) Should actin cytoskeleton remodeling indeed result in a decrease of actin helix diameter, what do the authors propose is the underlying mechanism? Shouldn't that result in changes in mitochondrial structure or location and be visible by TEM? This reviewer is also wondering why the authors focus so much on the actin helix, while the plasma membrane based on the author's results is moving way more dramatically.

Response 3.15: This raises an intriguing point. Currently, we lack an understanding of the underlying mechanism driving actin remodeling, and we are eager to conduct further experiments to explore this aspect. For instance, we are investigating the potential role of Cofilin in remodeling the F-actin network. Initial experiments utilizing STORM imaging have revealed the localization of Cofilin in the midpiece region, where the actin helix is situated.

Regarding mitochondria, thus far, we have not uncovered any evidence suggesting that acrosome reaction or fusion with the egg induces a rearrangement of these organelles within the structure. The rationale for investigating polymerized actin in depth stems from the fact that, alongside the axoneme and other flagellar structures such as the outer dense fibers and fibrous sheet, these are the sole cytoskeletal components present in that particular tail region.

(14) The fact that the authors observe that most sperm passing through the zona pellucida, which requires motility, display high FM4-64 fluorescence, doesn't that contradict the authors' hypothesis that midpiece contraction and motility cessation are connected? Videos confirming sperm motility and information about pattern distribution within the observed sperm population in the perivitelline space should be provided.

Response 3.14: We believe it is a matter of time, as depicted in Figure 1D, our model shows that first the cells lose the acrosome, present motility and low FM4-64 fluorescence in the midpiece (pattern II) and after that, they lose motility and increase FM4-64 fluorescence in the midpiece (pattern III). That is why, we think that when sperm pass the zona pellucida they present pattern II and after some time they evolve into pattern III. 

(15) In the experiments summarized in Figure 8, did all sperm stop moving? Considering that 74 % of the observed sperm did not display midpiece contraction upon fusion, again doesn't that contradict the authors' hypothesis that the two events are interdependent? Similarly, in earlier experiments, not all acrosome-reacted sperm display a decrease in midpiece diameter or stop moving, questioning the significance of the event. If some sperm display a decrease in midpiece diameter and some don't, or undergo that change earlier or later, what is the underlying mechanism of regulation? The observed events could similarly be explained by sperm death: Sperm are dying × plasma membrane integrity changes and plasma membrane dyes get internalized × [Ca2+]i simultaneously increases due to cell death × sperm stop moving.

Response 3.15: The percentage of sperm that did not exhibit midpiece contraction in Fig.8B is 26%, not 74%, indicating that it does not contradict our hypothesis. However, this still represents a significant portion of sperm that remain unchanged in the midpiece, leaving room for various explanations. For instance, it's possible that: i) the change in fluorescence was not detected due to the event occurring after the recording concluded, or ii) in some instances, this alteration simply does not occur. Nevertheless, we did not track subsequent events in the oocyte, such as egg activation, to definitively ascertain the success of fusion. Incorporation of the dye only manifests the initiation of the process.

(16) The authors propose changes in Ca2+ as one potential mechanism to regulate midpiece contraction, however, the Ca2+ measurements during fusion are flawed, as the authors write in the discussion, by potential Ca2+ fluorophore dilution. Considering that the authors observe high Ca2+ in all sperm prior to fusion, could that be a measuring artifact? Were acrosome-intact sperm imaged with the same settings to confirm that sperm with low and high Ca2+ can be distinguished? Should [Ca2+]i changes indeed be involved in the regulation of motility cessation during fusion, could the authors speculate on how [Ca2+]i changes can simultaneously be involved in the regulation of sperm hyperactivation?

Response 3.16: We agree with the reviewer that our experiments using calcium probes are not conclusive for many technical problems. We have toned down our conclusions in the new version of the manuscript.

(17) 74: AE takes place for most cells in the upper segment of the oviduct, not all of them.

Please correct.

Response 3.17: Corrected in the new version.

(18) 88: Achieved through, or achieved by, please correct.

Response 3.18: Corrected in the new version.

(19) 243: Acrosomal exocytosis initiation by progesterone, please specify.

Response 3.19: Modified in the new version.

(20) 277: "The actin cytoskeleton approaches the plasma membrane during the contraction of the midpiece" is misleading. The author's results show the opposite.

Response 3.20: As suggested, this statement was modified.

(21) 298: Why do the authors find it surprising that the F-actin network was unchanged in acrosome-intact sperm that do not present a change in midpiece diameter?

Response 3.21: The reviewer is right. The sentence was modified.

(22) Figures 5D,F: The provided images do not support a shift in the actin helix diameter.

Response 3.22: The shift in the actin helix diameter is provided in Figure 5E and 5G.

(23) Figure S5C: The authors should show representative histograms of spontaneously-, progesterone induced-, and ionomycin-induced AE. Based on the quantification the SiRactin peaks don't seem to move when the AR is induced by progesterone.

Response 3.23: As requested, an ionomycin induced sperm is incorporated.

(24) 392: Which experimental evidence supports that statement?

Response 3.24: A reference was incorporated. 

Reference 13 is published, please update. Response 3.25: updated as requested.

Author Response

The following is the authors’ response to the original reviews.

First, the authors would like to thank the reviewers and editors for their thoughtful comments. The comments were used to guide our revision, which is substantially improved over our initial submission. We have addressed all comments in our responses below, through a combination of clarification, new analyses and new experimental data.

Reviewer #1 (Public Review):

In this manuscript, the authors identified and characterized the five C-terminus repeats and a 14aa acidic tail of the mouse Dux protein. They found that repeat 3&5, but not other repeats, contribute to transcriptional activation when combined with the 14aa tail. Importantly, they were able to narrow done to a 6 aa region that can distinguish "active" repeats from "inactive" repeats. Using proximal labeling proteomics, the authors identified candidate proteins that are implicated in Dux-mediated gene activation. They were able to showcase that the C-terminal repeat 3 binds to some proteins, including Smarcc1, a component of SWI/SNF (BAF) complex. In addition, by overexpressing different Dux variants, the authors characterized how repeats in different combinations, with or without the 14aa tail, contribute to Dux binding, H3K9ac, chromatin accessibility, and transcription. In general, the data is of high quality and convincing. The identification of the functionally important two C-terminal repeats and the 6 aa tail is enlightening. The work shined light on the mechanism of DUX function.

A few major comments that the authors may want to address to further improve the work:

We thank the reviewer for their efforts and constructive comments, which have guided our revisions.

1) The summary table for the Dux domain construct characteristics in Fig. 6a could be more accurate. For example, C3+14 clearly showed moderate weaker Dux binding and H3K9ac enrichment in Fig 3c and 3e. However, this is not illustrated in Fig. 6a. The authors may consider applying statistical tests to more precisely determine how the different Dux constructs contribute to DNA binding (Fig. 3c), H3K9ac enrichment (Fig. 3e), Smarcc1 binding (Fig. 5e), and ATAC-seq signal (Fig. 5f).

We thank the reviewer for this comment, and agree that there were some modest differences in construct characteristics that were not captured in the Summary Table (6a). To better reflect the differences between constructs, we added additional dynamic range to our depiction/scoring, and believe that the new scoring system provides sufficient qualitative range to capture the difference without imposing a statistical approach.

2) Another concern is that exogenous overexpressed Dux was used throughout the experiments. The authors may consider validating some of the protein-protein interactions using spontaneous or induced 2CLCs (where Dux is expressed).

We agree that it would be helpful to determine endogenous DUX interaction with our BioID candidates. Here, we attempted co-IPs for endogenous DUX protein with the DUX antibody and were unsuccessful, which indicated that the DUX antibody is useful for detection but not efficient in the primary IP. This is why we utilized the mCherry tag for DUX IP experiments, which worked exceptionally well.

3) It could be technically challenging, but the authors may consider to validate Dux and Smarcc1 interaction in a biologically more relevant context such as mouse 2-cell embryos where both proteins are expressed. Whether Smarcc1 binding will be dramatically reduced at 4-cell embryos due to loss of Dux expression?

While we agree that it would be interesting to validate the in vivo interaction of DUX and SMARCC1 in the early embryo, it is not technically feasible for us to conduct the experiment, as the IP would require thousands of two-cell embryos, and we have the issue of poor co-IP quality with the DUX antibody.

Reviewer #2 (Public Review):

In this manuscript, Smith et al. delineated novel mechanistic insights into the structure-function relationships of the C-terminal repeat domains within the mouse DUX protein. Specifically, they identified and characterised the transcriptionally active repeat domains, and narrowed down to a critical 6aa region that is required for interacting with key transcription and chromatin regulators. The authors further showed how the DUX active repeats collaborate with the C-terminal acidic tail to facilitate chromatin opening and transcriptional activation at DUX genomic targets.

Although this study attempts to provide mechanistic insights into how DUX4 works, the authors will need to perform a number of additional experiments and controls to bolster their claims, as well as provide detailed analyses and clarifications.

We thank this reviewer for their constructive comments, and have conducted several new analyses, additional experiments and clarifications – which have strengthened the manuscript in several locations. Highlights include a statistical approach to the similarity of mouse repeats to themselves and to orthologs (Figure S1d) and clarified interpretations, a wider dynamic range to better reflect changes in DUX construct behaviors (Figure 6a), and additional data on construct behavior, including ‘inactive’ constructs (e.g C1+14aa in Figure 1a,d, new ATAC-seq in Figure S1g), and active constructs such as C3+C5+14aa and C3+C514aa (in Figure S1b).

Reviewer #3 (Public Review):

Dux (or DUX4 in human) is a master transcription factor regulating early embryonic gene activation and has garnered much attention also for its involvement in reprogramming pluripotent embryonic stem cells to totipotent "2C-like" cells. The presented work starts with the recognition that DUX contains five conserved c. 100-amino acid carboxy-terminal repeats (called C1-C5) in the murine protein but not in that of other mammals (e.g. human DUX4). Using state-of-the-art techniques and cell models (BioID, Cut&Tag; rescue experiments and functional reporter assays in ESCs), the authors dissect the activity of each repeat, concluding that repeats C3 and C5 possess the strongest transactivation potential in synergy with a short C-terminal 14 AA acidic motif. In agreement with these findings, the authors find that full-length and active (C3) repeat containing Dux leads to increased chromatin accessibility and active histone mark (H3K9Ac) signals at genomic Dux binding sites. A further significant conclusion of this mutational analysis is the proposal that the weakly activating repeats C2 and C4 may function as attenuators of C3+C5-driven activity.

By next pulling down and identifying proteins bound to Dux (or its repeat-deleted derivatives) using BioID-LC/MS/MS, the authors find a significant number of interactors, notably chromatin remodellers (SMARCC1), a histone chaperone (CHAF1A/p150) and transcription factors previously (ZSCAN4D) implicated in embryonic gene activation.

The experiments are of high quality, with appropriate controls, thus providing a rich compendium of Dux interactors for future study. Indeed, a number of these (SMARCC1, SMCHD1, ZSCAN4) make biological sense, both for embryonic genome activation and for FSHD (SMCHD1).

A critical question raised by this study, however, concerns the function of the Dux repeats, apparently unique to mice. While it is possible, as the authors propose, that the weak activating C1, C2 C4 repeats may exert an attenuating function on activation (and thus may have been selected for under an "adaptationist" paradigm), it is also possible that they are simply the result of Jacobian evolutionary bricolage (tinkering) that happens to work in mice. The finding that Dux itself is not essential, in fact appears to be redundant (or cooperates with) the OBOX4 factor, in addition to the absence of these repeats in the DUX protein of all other mammals (as pointed out by the authors), might indeed argue for the second, perhaps less attractive possibility.

In summary, while the present work provides a valuable resource for future study of Dux and its interactors, it fails, however, to tell a compelling story that could link the obtained data together.

We appreciated the reviewer’s views regarding the high quality of the work and our generation of an important dataset of DUX interactors. We also appreciate the comments provided to improve the work, and have performed and included in the revised version a set of clarifications, additional analyses and additional experiments that have served to reinforce our main points and provide additional mechanistic links. We also agree that more remains to be done to understand the function and evolution of repeats C1, C2 and C4.

Reviewer #1 (Recommendations For The Authors):

1) For immuno-blots, authors may indicate the expected bands to help readers better understand the results.

Agreed, and we have included the predicted molecular weight of proteins in the Figure Legends. We note that our work shows that the C-terminal domains confer anomalous migration in SDS-PAGE.

2) Fig. 5b, a blot missing for the mCherry group?

Figure 5b is a volcano blot, so we believe the reviewer is referring to Figure 5d, which is a coimmunoprecipitation experiment between SMARCC1 and mCherry-tagged DUX constructs. However, we are unsure of the comment as an anti mCherry sample is present in that panel.

3) Line 99-100, Fig. S1d, it seems that repeat2, but not repeat3, is more similar to human DUX4 C-terminal region.

This comment and one by another reviewer have prompted us to re-examine the similarities of the DUX repeats, and we have new analyses (Figure S1d) and an alternative framing in the manuscript as a result. We have expanded on this in our response to Reviewer #2, point #1 – and direct the reviewer there for our expanded treatment.

4) There are a few references are misplaced. For example, line 48, the studies that reported the role of Dux in inducing 2CLCs should be from Hendrickson et al., 2017, De Iaco et al., 2017, and Whiddon et al., 2017. The authors may want to double check all references.

Thanks for pointing these out. These issues have been corrected in the manuscript.

5) In the materials & methods section, a few potential errors are noticed. For example, concentrations of PD0325901 and CHIR99021 in mESC medium appear ~1000-fold higher than standards.

Thanks – corrected.

Reviewer #2 (Recommendations For The Authors):

Major Points

1) Line 99 - The authors claimed that the "human DUX4 C-terminal region is most similar to the 3rd repeat of mouse DUX", but based on Supp. Fig. 1d, the human DUX4 C-term should be most similar to the 2nd repeat of mouse DUX. If this is indeed the case, it will undermine the rest of this study, since the authors claim that the 3rd repeat is transcriptionally active, whereas the 2nd repeat is transcriptionally inactive, and the bulk of this study largely focused on how the active repeats, not the inactive repeats, are critical in recruiting key transcriptional and chromatin regulators to induce the embryonic gene expression program.

We thank the reviewer for their comments here. Since submission,and as mentioned above for reviewer #1 we have revisited the issue of similarity of the DUX4 C-terminal region to the mouse C-terminal repeats, with a BLAST-based approach that is more rigorous and informed by statistics – which is in Author response table 1 and now in the manuscript as Figure S1d, and has affected our interpretation. Our prior work involved a simple % identity comparison table and we now appreciate that some of the similarity analyses did not meet statistical significance, and therefore we are unable to draw certain conclusions. We make the appropriate modifications in the text. For example, we no longer state that the DUX4 C-terminus appears to be most similar to mouse repeats 3 and 5. This does not affect the main conclusions of the paper regarding interactions of the C-terminus with chromatin-related proteins, only our speculation on which repeat might have represented the original single repeat in the mouse – an issue we think of some interest, but did not rise to the level of mentioning in the original or current abstract.

Author response table 1.

Parameters: PAM250 matrix. Gap costs of existence: 15 and extension: 3. Numbers represent e-value of each pairwise comparison

*No significant similarities found (>0.05).

2) In Supp Fig 1d, it seems that the rat DUX4 C-terminal region is most similar to the 4th repeat of mouse DUX, which according to the author is supposedly transcriptionally inactive. This weakens the authors justification that the 3rd or 5th repeat is likely the "parental repeat for the other four", and further echoes my concern in point 1 where the human DUX4 C-term is most similar to the 2nd (inactive) repeat of mouse DUX.

The reviewer’s point is well taken and is addressed in point #1 above.

3) In Fig. 1d, the authors showed that DUX4-containing C3 and C5, but lacking acidic tail, can promote MERVL::GFP expression, albeit to a slightly lower extent compared to FL. However, in Fig. 2b, C3 or C5 alone (lacking acidic tail) completely failed to promote MERVL::GFP expression. However, in the presence of the acidic tail, both versions were able to promote MERVL::GFP expression, similar to that of FL. The latter would suggest that it is the acidic tail that is crucial for MERVL::GFP expression, and this does not quite agree with Fig 1b, where C12345 (lacking acidic tail) was able to promote MERVL::GFP expression. Although C12345 did not activate MERVL to a similar level as FL, it is clearly proficient, compared to C3 or C5 alone (lacking acidic tail) where there is no increase in MERVL at all. Additional constructs will be helpful to clarify these points. For example, 'C3+C5 minus acidic tail' and 'HD1+HD2+acidic tail only' constructs.

We agree that constructs such as those mentioned would add to the work. First, we have done the additional construct HD1+HD2+14aa tail, which is presented as ΔC12345+14aa in Figure 2a and in S2a. Additionally, we performed experiments on the requested C3+C5+14aa and C3+C5Δ14aa (see samples 6 and 7 in Author response image 1, which are now included in Supplemental Figure 2b). The results reinforce our hypothesis of an additive effect toward DUX target gene activation by increasing C-terminal repeats and including the 14aa tail.

Author response image 1.

4) Related to the above, the flow cytometry data for the MERVL::GFP reporter as presented in Figures 1 and 2, as well as in Supp. Fig. 2, show a considerably large difference in the %GFP|mCherry for the FL construct, ranging from ~6-26%. This makes it difficult to convince the reader which of the different DUX domain constructs cannot or can partially induce GFP|mCherry signal when compared to FL, and hence it is tough to definitively ascertain the exact contribution of each of the 5 C-terminal repeats with high confidence, as it appears that there exists a significant amount of variability in this MERVL::GFP reporter system. The authors need to address this issue since this is their primary method to elucidate the transcriptional activity of each of the mouse DUX repeat domains.

We note that with the Dux-/- cell lines we used throughout the timeline of the study, the percent of %GFP|mCherry expression progressively and slowly decreased – possibly due to slow/modest epigenetic silencing of the reporter. However, we always used the full-length DUX construct to establish the dynamic range. We emphasize that the relative differences between constructs over multiple cell line replicates remained relatively consistent. However, we elected to show absolute values in each experiment, rather than simply normalizing the full-length to 100% and showing relative.

5) Lines 140-142 - The authors claimed that the functional difference between the transcriptionally active and inactive repeats could be narrowed down to a "6aa region which is conserved between repeats C3 and C5, but not conserved in C1, C2 and C4". Assuming the 6aa sequence is DPLELF, why does C1C3a elicit almost twice the intensity of GFP|mCherry signal compared to C3C1c, despite both constructs having the exact same 6aa sequence?

Indeed, C1C3a and C3C1c both containing the ‘active’ DPL sequence but having different relative levels of %GFP|mCherry. This is consistent with these sequences having a positive role in DUX target gene regulation – but likely in combination with other other regions which potentiate its affect, possibly through interacting proteins or post-translational modifications.

Why does DPLEPL (the intermediate C3C1b construct) induce a similar extent of GFP|mCherry signal as the FL construct, even though the former includes 3aa from a transcriptionally inactive repeat? In contrast, GSLELF (the other intermediate C1C3b construct) that also includes 3aa from a transcriptionally inactive repeat is almost completely deficient in inducing any GFP|mCherry signal. Why is that so? Is DPL the most crucial sequence? It will be important to mutate these 3 (or the above 6) residues on FL DUX4 to examine if its transcriptional activity is abolished.

These are interesting points. DPL does appear to be the most important region in the mouse DUX repeats. However, DPL is not shared in the C-terminus of human DUX4. Notably, the DUX4 C-terminus is sufficient to activate the mouse MERVL::GFP reporter when cloned to mouse homeodomains (see Author response image 2, second sample) and other DUX target genes (initially published in Whiddon et al. 2017). One clear possibility is that the DPL region is helping to coordinate the additive effects of multiple DUX repeats, which only exist in the mouse protein.

Author response image 2.

6) Line 154 - The intermediate DUX domain construct C1C3b occupied a different position on the PCA plot from the C1C3c construct that does not contain any of the critical 6aa sequence, as shown in Fig. 2e. However, both these constructs appear to be similarly deficient in inducing any GFP|mCherry signal, as seen in Fig. 2c. Why is that so?

The PCA plot assesses the impact on the whole transcriptome and not just the MERVL::GFP reporter, suggesting the 3aa region has transcriptional effects on the genome beyond what is detected in the MERVL::GFP reporter.

7) To strengthen the claim that "Chromatin alterations at DUX bindings sites require a transcriptionally active DUX repeat", the authors should also perform CUT&Tag for constructs containing transcriptionally inactive DUX repeats (e.g. C1+14aa), and show that such constructs fail to occupy DUX binding sites, as well as are deficient in H3K9ac accumulation.

This is a good comment. We elected to control this with constructs containing or lacking an active repeat. Although we have not pursued this by CUT&TAG, we have examined the impact of DUX constructs with inactive repeats (including the requested C1+14aa, new Figure S1g) by ATAC-seq (see #12, ATAC-seq section, below), and observe no chromatin opening, suggesting that the lack of transcriptional activity is rooted in the inability to open chromatin.

8) It would be good if the authors could also include CUT&Tag data for some of the C1C3 chimeric constructs that were used in Fig. 2, since the authors argued that the minimal 6aa region is sufficient to activate many of the DUX target genes. This would also strengthen the authors’ case that the transcriptionally active, not inactive, repeats are critical for binding at DUX binding sites and ensuring H3K9ac occupancy.

We agree that these would be helpful, and have examined the inactive repeats in transcription and ATAC-seq formats during revision (new data in Figures 1d and S1g), but not yet the CUT&TAG format.

9) Line 213 - "SMARCA4" should have been "SMARCA5"? Based on Fig. 4d, SMARCA5 is picked up in the BirA*-DUX interactome, not SMARCA4.

Thanks – corrected.

10) Lines 250-252 - The authors compared the active BirA-C3 against the inactive BirA-C1 to elucidate the interactome of the transcriptionally active C3 repeat, as illustrated in Fig. 5c. They found 12 proteins more enriched in C1 and 154 proteins in C3. This information should be presented clearly as a separate tab in Supp Table 2. What are the proteins common to both constructs, i.e. enriched to a similar extent? Do they include chromatin remodellers too? Although the authors sought to identify differential interactors between the 2 constructs, it is also meaningful to perform 2 separate comparisons - active BirA-C3 against BirA alone control, and inactive BirA-C1 against BirA alone control - like in Fig. 4d, so as to more accurately define whether the active C3 repeat, and not the inactive C1 repeat, interacts with proteins involved in chromatin remodeling.

We thank the reviewer for this comment, and we have modified the manuscript by adding a second sheet in Supplementary Table 2 including the results for enriched proteins in BirA-C1 vs. C3. Additionally, due to limitations of annotation between BirA alone and BirA*-C3 being sequenced in different mass spectrometry experiments, it is difficult to quantitatively compare the two datasets with pairwise comparisons.

11) Fig 5d: The authors mentioned in the legend that endogenous IP was performed for SMARCC1. However, in line 266, they stated Flag-tagged SMARCC1. Is SMARCC1 overexpressed? The reciprocal IP should also be presented. More importantly, C1 constructs (e.g. C1+14aa and C1Δ14aa) should also be included.

To clarify, Figure 4e used exogenously overexpressed FLAG-SMARCC1 in HEK-293T cells to confirm the results of the full-length DUX BioID experiment. Figure 5d was performed with overexpressed DUX construct, but involved endogenous SMARCC1 in mESCs. This has now been made clearer in the revised manuscript.

12) For both the SMARCC1 CUT&Tag and ATAC-seq experiments shown in Figures 5e and 5f respectively, the authors need to include DUX derivatives that contain transcriptionally inactive repeats with and without the 14aa acidic tail, i.e. C1+14aa and C1Δ14aa, and show that these constructs prevent the binding/recruitment of SMARCC1 to DUX genomic targets, and correspondingly display a decrease in chromatin accessibility. Only then can they assert the requirement of the transcriptionally active repeat domains for proper DUX protein interaction, occupancy and target activation.

We agree that examination of an inactive repeat in certain approaches would improve the manuscript. Importantly, we have now included C1+14 in our ATAC-seq experiments, and in Author response image 3 two individual replicates, which constitute a new Figure S1g. Compared to the transcriptionally active DUX constructs, which see opening at DUX binding sites, we do not see chromatin opening at DUX binding sites with transcriptionally inactive C1+14.

Author response image 3.

13) To prove that DUX-interactors are important for embryonic gene expression, it will be important to perform loss of function studies. For instance, will the knockdown/knockout of SMARCC1 in cells expressing the active DUX repeat(s) lead to a loss of DUX target gene occupancy and activation?

We agree that it would be interesting to better understand SMARCC1 cooperation with DUX function in the embryo, but we believe this is beyond the scope of this paper.

Minor Points

1) Lines 124-126 - What is the reason/rationale for why the authors used one linker (GGGGS2) for constructs with a single internal deletion, but 2 different linkers (GGGGS2 and GAGAS2) for constructs with 2 internal deletions?

With Gibson cloning, there are homology overhang arms for each PCR amplicon that are required to be specific for each overlap. Additionally, each PCR amplicon needs to be specific enough from one another so that all inserts (up to 5 in this manuscript) are included and oriented in the right order. The linker sequences were included in the homology arm overlaps, so the nucleotide sequences for each linker needed to be specific enough to include all inserts. This is a general rule to Gibson cloning. Additionally, both GGGGS2 and GAGAS2 are common linker sequences used in molecular biology and the amino acids structures are similar to one another, suggesting there is no functional difference between linkers.

2) Line 704 - 705: In the figure legend, the authors stated that 'Constructs with a single black line have the linker GGGGS2 and constructs with two black lines have linkers with GGGGS2 and GAGAS2, respectively.'. This was not obvious in the figures.

Constructs used for flow and genomics experiments that are depicted in Figure 2, Supplementary Figure 2, Figure 3, Figure 4, and Figure 5 have depicted black lines where deletions are present. Where these deletions are present, there are linkers in order to preserve spacing and mobility for the protein.

3) Line 160 - Clusters #1 and #2 are likely written in the wrong order. It should have been "activating the majority of DUX targets in cluster #2, not cluster #1" and "failed to activate those in cluster #1, not cluster #2", based on the RNA-seq heatmap in Fig. 2f.

We thank the reviewer for this comment, and the error has been corrected in the manuscript.

4) Line 188 - Delete the word "of" in the following sentence fragment: "DUX binding sites correlating with the of transcriptional".

Thanks – corrected.

5) Line 191 - Delete the word "aids" in the following sentence fragment: "important for conferring H3K9ac aids at bound".

Thanks – corrected.

6) Line 711 - "C1-C3 a,b,d" should be "C1-C3 a,b,c".

Thanks – corrected.

7) Lines 711-712 - The colors "pink to blue" and "blue to pink" are likely written in the wrong order. Based on Fig. 2c, the blue to pink bar graphs should represent C1-C3 a,b,c in that order, and likewise the pink to blue bar graphs should represent C3-C1 a,b,c in that order.

Thanks – corrected.

8) There is an overload of data presented in Fig. 2c, such that it is difficult to follow which part of the figure represents each data segment as written in the figure legend. It is recommended that the data presented here is split into 2 sub-figures.

Figure 2c has a supporting figure in Supplementary Figure 2b. While there is both a graphical depiction of the constructions and the data both in the main panel of Figure 2C, we have depicted it as so to be as clear as possible for the reader to interpret the complexity and presentence of amino acids in each of the constructs.

9) Line 717 - "following" is misspelt.

Thanks – corrected.

10) Lines 720-721 - "(Top)" and "(Bottom)" should be replaced with "(Left)" and "(Right)", as the 2 bar graphs presented in Fig. 2d are placed side by side to each other, not on the top and bottom.

Thanks – corrected.

11) Lines 725 and 839 - "Principle" is misspelt. It should be "Principal".

Thanks – corrected.

12) In Figures 3d and 3e, the sample labeled "C3+14_1" should be re-labeled to "C3+14", in accordance with the other sub-figures. Additionally, for the sake of consistency, "aa" should be appended to the relevant constructs, e.g. "C3+14aa" and "C3Δ14aa".

Thanks – corrected.

13) Line 773 - Were the DUX domain constructs over-expressed for 12hr (as written in the figure legend) or 18hr (as labeled in Fig. 5d)?

Thanks – corrected.

14) Related to minor point 19 above, is there a reason/rationale for why some of the experiments used 12hr over-expression of DUX domain constructs (e.g. for CUT&TAG in Fig. 3), whereas in other experiments 18hr over-expression was chosen instead (e.g. flow cytometry for MERVL::GFP reporter in Figures 1 and 2, and co-IP validations of BirA*-DUX interactions in Fig. 4)?

Thanks for the opportunity to explain. In this work, experiments that reported on proteins that are translated following DUX gene activation (e.g. MERVL:GFP via flow) were done at 18hr to allow for enough time for transcription and translation of GFP (or other DUX target genes). For experiments that report on the impact of DUX on chromatin and transcription, such as RNA-seq, CUT&Tag, and ATAC-seq, we induced DUX domain constructs for 12 hours.

15) Line 804 - "ΔHDs" is missing between "C2345+14aa" and "ΔHD1".

Thanks – corrected.

16) In Fig. 5c, "Chromatin remodelers" is misspelt.

Thanks – corrected.

17) There is no reference in the manuscript to the proposed model that is presented in Fig. 6b.

Thanks – corrected.

Reviewer #3 (Recommendations For The Authors):

Given the uncertainty of the function of the Dux peptide repeats in mice, could it not also be possible that the underlying repeated nature of the (coding) DNA? That is, could these DNA repeats exert a regulatory function on Dux transcription itself (also given the dire consequences of misregulated DUX4 expression as seen in FSHD, for example).

Yes, it remains possible that the internal coding repeats within Dux are playing a role in locus regulation, and might be interesting to examine. However, we consider this question as being outside the scope of the current paper.

Finally, it would be interesting to know whether these repeats are, in fact, present in all mouse species. Already no longer present in rat, do they exist, or not, in more "distant" mice, e.g. M. caroli?

Determining whether all mouse strains contain C-terminal repeats in DUX is a question we also considered. However, Dux and its orthologs are present in long and very complex repeat arrays that are not present in the sequencing data or annotation in other mouse strains. Therefore, we are not unable to answer this question from existing sequencing data. Answering would require a considerable genome sequencing and bioinformatics effort, or alternatively a considerable effort aimed at cloning ortholog cDNAs from 2-cell embryos.

Minor points:

line 169: here it seems, in fact, that the 'inactive' C2, C4 repeats are more similar to each other (my calculation: 91 and 96% identity at the protein and DNA level, respectively) than the active C3 and C5 repeats (82 and 89% identity, resp.), the outlier being C1.

Thanks for this comment, which was mentioned by other reviewers as well and has been addressed through new statistical analyses and interpretation (see new Figure S1d).

line 191: I'm not sure this sentence parses correctly ("...14AA tail is important for conferring H3K9Ac aids at bound sites...")

We thank the reviewer for this comment, and we have corrected the sentence in the manuscript.

Author response:

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

The researchers demonstrated that when cytokine priming is combined with exposure to pathogens or pathogen-associated molecular patterns, human alveolar macrophages and monocyte-derived macrophages undergo metabolic adaptations, becoming more glycolytic while reducing oxidative phosphorylation. This metabolic plasticity is greater in monocyte-derived macrophages than in alveolar macrophages.

Strengths:

This study presents evidence of metabolic reprogramming in human macrophages, which significantly contributes to our existing understanding of this field primarily derived from murine models.

Weaknesses:

The study has limited conceptual novelty.

We acknowledge that the study has limited conceptual novelty, however, the current manuscript provides the field with evidence of the changes in the phenotype and functions of human macrophages in response to IFN-γ or IL-4 which is currently lacking in the literature. Moreover, our data shows for the first time that human airway macrophages change their function in response to IFN-γ.  

Reviewer #2 (Public Review):

Summary:

The authors aimed to functionally characterize primary human airway macrophages and monocytederived macrophages, correlating their glycolytic shift in metabolism. They conducted this macrophage characterization in response to type II interferon and IL-4 priming signals, followed by different stimuli of irradiated Mycobacterium tuberculosis and LPS.

Strengths:

(1) The study employs a thorough measurement of metabolic shift in metabolism by assessing extracellular acidification rate (ECAR) and oxygen consumption rate (OCR) of differentially polarized primary human macrophages using the Seahorse XFe24 Analyzer.

(2) The effect of differential metabolic shift on the expression of different surface markers for macrophage activation is evaluated through immunofluorescence flow cytometry and cytokine measurement via ELISA.

(3) The authors have achieved their aim of preliminarily characterizing the glycolysis-dependent cytokine profile and activation marker expression of IFN-g and IL-4 primed primary human macrophages.

(4) The results of the study support its conclusion of glycolysis-dependent phenotypical differences in cytokine secretion and activation marker expression of Ams and MDMs.

Weaknesses:

(1) The data are presented in duplicates for cross-analyses.

(2) The data presented supports a distinct functional profile of airway macrophages (Ams) compared to monocyte (blood)-derived macrophages (MDMs) in response to the same priming signals. However, the study does not attempt to explore the underlying mechanism for this difference.

(3) The study is descriptive in nature, and the results validate IFN-g-mediated glycolytic reprogramming in primary human macrophages without providing mechanistic insights.

(1) We acknowledge the data is presented in duplicate for cross-analyses. This duplication allowed us to examine both (A) the effect of IFN-γ or IL-4 on primary human airway and monocyte derived macrophages in the presence or absence of distinct stimulations and (B) to directly compare the fold change in function occurring in the AM with the changes in the MDM.

(2 & 3) We acknowledge that our study is descriptive however, by inhibiting glycolysis using 2DG we have demonstrated that increased flux through glycolysis is mechanistically required to mediate enhanced cytokine responses in both primary human AM and MDM primed with IFN-γ. However, we acknowledge that we have not determined the differential molecular mechanisms downstream of IFNγ in the AM versus the MDM. IFN-γ promotes both pro- and anti-inflammatory cytokines in AM and this was reduced by inhibiting glycolysis with 2DG. This identifies glycolysis as a key mechanistic pathway which can be therapeutically targeted in AM to modulate inflammation. Mechanistic studies on human AM are limited due to low number of AM retrieved from BAL samples. Nevertheless, the differences between AM and MDM identified in the current study indicate that future mechanistic studies are warranted to identify why IFN-γ promotes IL-10 in AM and not MDM, and, why TNF is differentially regulated by glycolysis in the two macrophage subpopulations, for example.  

Reviewer #3 (Public Review):

Summary:

In this manuscript, the authors explore the contribution of metabolism to the response of two subpopulations of macrophages to bacterial pathogens commonly encountered in the human lung, as well as the influence of priming signals typically produced at a site of inflammation. The two subpopulations are resident airway macrophages (AM) isolated via bronchoalveolar lavage and monocyte-derived macrophages (MDM) isolated from human blood and differentiated using human serum. The two cell types were primed using IFNγ and Il-4, which are produced at sites of inflammation as part of initiation and resolution of inflammation respectively, followed by stimulation with either irradiated Mycobacterium tuberculosis (Mtb) or LPS to simulate interaction with a bacterial pathogen. The authors use human cells for this work, which makes use of widely reported and thoroughly described priming signals, as well as model antigens. This makes the observations on the functional response of these two subpopulations relevant to human health and disease. To examine the relationship between metabolism and functional response, the authors measure rates of oxidative phosphorylation and glycolysis under baseline conditions, primed using IFNγ or IL-4, and primed and stimulated with Mtb or LPS.

Strengths:

• The data indicate that both populations of macrophages increase metabolic rates when primed, but MDMs decrease their rates of oxidative phosphorylation after IL-4 priming and bacterial exposure while AMs do not.

• It is demonstrated that glycolysis rates are directly linked to the expression of surface molecules involved in T-cell stimulation and while secretion of TNFα in AM is dependent on glycolysis, in MDM this is not the case. IL-1β is regulated by glycolysis only after IFN-γ priming in both MDM and AM populations. It is also demonstrated that Mtb and LPS stimulation produces responses that are not metabolically consistent across the two macrophage populations. The Mtb-induced response in MDMs differed from the LPS response, in that it relies on glycolysis, while this relationship is reversed in AMs. The difference in metabolic contributions to functional outcomes between these two macrophage populations is significant, despite acknowledgement of the reductive nature of the system by the authors.

• The observations that AM and MDM rely on glycolysis for the production of cytokines during a response to bacterial pathogens in the lung, but that only MDM shift to Warburg Metabolism, though this shift is blocked following exposure to IL-4, are supported by the data and a significant contribution the study of the innate immune response.

Weaknesses:

• It is unclear whether changes in glycolysis and oxidative phosphorylation in primed cells are due to priming or subsequent treatments. ECAR and OCR analyses were therefore difficult to interpret.

All data sets have been presented and analysed relative to both unprimed unstimulated to show both the effect of priming and subsequent stimulation. A second analysis was subsequently conducted where each data set was normalised to its own baseline in terms of percentage change. Therefore, each of unprimed, IFN-γ and IL-4 primed cells were set to 100% in order to assess the effect of stimulation independent of the baseline priming effect. For clarity we have removed the following line:

“Percentage change for ECAR and OCR was calculated from the respective baseline of each data set to visualise the differential ability of IFN-γ, IL-4 primed or unprimed AM to respond to stimulation (Figure S1C,D).”

We have amended the text in the manuscript (lines 164-173) to “Since IFN-γ priming increased cellular energetics in the AM at baseline, we calculated percent change in ECAR and OCR from the baseline rate of each group in order to assess if IFN-γ or IL-4 primed AM have altered capacity to change their metabolism in response to stimulation (Figure 1C,D). This was carried out to equalise all the primed data sets at baseline before stimulation (Figure S1C, S1D).  These data indicate that whilst the peak of glycolysis is elevated in IFN-γ primed AM (Figure 1A), all AM have a similar capacity to increase glycolysis upon stimulation when baseline differences in metabolism were adjusted for the effects of cytokine priming (Figure 1C). IFN-γ increased the percent change in OCR of AM in response to both bacterial stimuli compared to the unstimulated IFN-γ primed control (Figure 1D). These data indicate that priming AM alters the metabolic baselines of human tissue resident macrophages and not their ability to respond to bacterial stimuli.”

• The data may not support a claim that AM has greater "functional plasticity" without a direct comparison of antigen presentation. Moreover, MDM secrete more IL-1β than AM. The claim that AM "have increased ability to produce all cytokines assayed in response to Mtb stimulation" does not appear to be supported by the data.

Our data suggests that the MDM are more phenotypically plastic (in terms of their ability to alter expression of cell surface markers in response to cytokine cues), whereas AM have a greater ability to alter cytokine production, our measure of functional plasticity. We have now defined the use of the terms ‘functional plasticity’ and ‘phenotypic plasticity’ in the context of our paper in lines 6063. To consider different culture and plating requirements of MDM versus AM, cytokine production was analysed relative to the average of the unprimed Mtb or LPS control of the respective MDM or AM. This allowed us to draw more accurate comparisons between the two macrophage populations by examining their relative ability to increase their cytokine production (expressed as fold change) rather than defining this functional plasticity only in terms of concentrations of cytokine produced in culture.  

We have therefore added the following sentence into the conclusion of the manuscript. “Cumulatively, the data presented herein suggests that the MDM maybe more phenotypically plastic than the AM, while the AM have enhanced functional plasticity in their ability to modulate cytokine production after exposure Th1 and Th2 cytokines.”

We have edited the discussion (lines 421-423) to clarify the following "have increased ability to produce all cytokines assayed in response to Mtb stimulation" and changed it to “stimulated with Mtb have significantly more production of IL-1β, TNF and IL-10 compared with unprimed controls. This is in contrast with IFN-γ primed MDM which only upregulate TNF compared to their unprimed controls.”   

• The claim that AM are better for "innate training" via IFNγ may not be consistent with increased IL1β and a later claim that MDM have increased production and are "associated with optimal training."

We have removed the word “better” and now simply state that AM are a tractable target to induce innate training in the human lung.

• Statistical analyses may not appropriately support some of the conclusions.

We have consulted with a statistician. Please see response to reviewer 3 recommendations for authors point 1 below.  

• AM populations would benefit from further definition-presumably this is a heterogenous, mixed population.

AM are routinely >97% CD68+CD14+ used in the current study (Author response image 1). However, we acknowledge that tissue resident macrophages represent a spectrum of phenotypes. Given limitations in cell numbers from primary human AM derived from BALF, we have not attempted to define the function of discreet subpopulations of AM.

• The term "functional plasticity" could also be more stringently defined for the purposes of this study.

We are terming functional plasticity to be the macrophages’ ability to alter their production of cytokines in response to external cues like IFN-γ and IL-4 whereas phenotypic plasticity is measured based on ability to alter the cell surface expression of activation markers.  We have now defined this in the manuscript (lines 60-63).

Author response image 1.

Expression of macrophage markers on AM. 

Conclusion:

Overall, the authors succeed in their goals of investigating how inflammatory and anti-inflammatory cytokine priming contributes to the metabolic reprogramming of AM and MDM populations. Their conclusions regarding the relationship between cytokine secretion and inflammatory molecule expression in response to bacterial stimuli are supported by the data. The involvement of metabolism in innate immune cell function is relevant when devising treatment strategies that target the innate immune response during infection. The data presented in this paper further our understanding of that relationship and advance the field of innate immune cell biology.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

(1)  Authors are suggested to provide rationale for their choice of cytokines as IFN-gamma and IL-4. This will be useful for the readers.

We have updated the following sentence (line 44-46) in the manuscript to add more rationale for the choice of IFN-γ and IL-4.  “There is a paucity of data on the role of metabolism in response to Th1 or Th2 microenvironments induced by cytokines-such as IFN-γ or IL-4 respectively, in human macrophages, especially in tissue resident macrophages, such as AM.”

(2)  Authors have shown the final outcome of metabolic reprogramming in terms of expression of HLADR and CD-40, and cytokine release. What pathways/receptors are activated or associated with IL-4 and IFN-gamma priming as a first line of response?

The relationship between IFN-γ or IL-4 induced expression of CD40 is established in haematological cell lines and fibroblasts as well as APC, with roles for the JAK/STAT pathways and upregulation of IRFs defined (1-3). Similarly, the relationship between exogenous IFN-γ and upregulation of HLA-DR expression on human monocytes or endothelial cells is established (4, 5). Whist our work does not outline the signalling pathways downstream of Th1 or Th2 cytokine priming, we have shown for the first time that glycolysis mechanistically underpins the shift in phenotype and function observed in human macrophages upon priming with IFN-γ or IL-4.

(3)  What are the intracellular signals leading to glycolytic shift?

One of the most likely mechanisms that under pin the shift to glycolytic metabolism is the stabilisation of HIF-1α mediated by activation of mTOR (see response below and rebuttal figure 2).  

(4)  Additional evidence is required to show Warburg effect such as stabilization and activation of HIF1alpha.

We acknowledge that we have not shown the activation and stabilisation of HIF-1α, however, we have provided functional evidence of increased glycolysis with concomitant decreased oxidative phosphorylation indicative of Warburg metabolism.

In order to address this gap in evidence we have reworded the manuscript to describe this functional change to “Warburg-like metabolism” throughout the manuscript. In addition, we have undertaken Western Blotting to provide evidence of mTOR activation when cells are primed with IFN-γ (Author response image 2).

Author response image 2.

IFN-γ activates mTOR in primary human monocytes. Monocytes were isolated from healthy donor PBMC using magnetic separation. Monocytes were left untreated (-), stimulated with rapamycin as a negative control (Rap; 50 nM), IFN-γ (10 ng/ml) or IFN-γ and rapamycin simultaneously (IFN-γ + Rap) for 15 minutes. Phosphorylation of S6 was used as a readout of mTOR activation and measured by western blot using β-actin as a control with a blot (A) and (b) densitometry results are shown as the relative expression of pS6: β-actin from. Graphs show data of n=1 of unprimed (black dot) vs IFN-γ primed (red) with and without rapamycin. ImageLab (Bio-Rad) software was used to perform densitometric analysis. 

(5)  What is the importance of showing percentage change vs fold change in figure 1 (1C vs 1A)?

All data sets have been presented and analysed relative to both unprimed unstimulated to show the effect of first priming and subsequent stimulation (Figure 1A). A second analysis was subsequently conducted where each data set was normalised to its own baseline in terms of percentage change (Figure 1C). Therefore, each of unprimed, IFN-γ or IL-4 primed cells were set to 100% to assess the effect of stimulation independent of the pre-existing effect of priming on the baseline metabolism. For clarity we have removed the following line:

“Percentage change for ECAR and OCR was calculated from the respective baseline of each data set to visualise the differential ability of IFN-γ, IL-4 primed or unprimed AM to respond to stimulation (Figure S1C,D).”

We have amended the text (lines 164-173) in the manuscript to “Since IFN-γ priming increased cellular energetics in the AM at baseline, we calculated percent change in ECAR and OCR from the baseline rate of each group in order to assess if IFN-γ or IL-4 primed AM have altered capacity to change their metabolism in response to stimulation (Figure 1C,D). This was carried out to equalise all the primed data sets at baseline before stimulation (Figure S1C, S1D).  These data indicate that whilst the peak of glycolysis is elevated in IFN-γ primed AM (Figure S1A), all AM have a similar capacity to increase glycolysis upon stimulation when baseline differences in metabolism were adjusted for the effects of cytokine priming (Figure 1C). IFN-γ increased the percent change in OCR of AM in response to both bacterial stimuli compared to the unstimulated IFN-γ primed control (Figure 1D). These data indicate that priming AM alters the metabolic baselines of human tissue resident macrophages and not their ability to respond to bacterial stimuli.”

(6)  Why IL-4 primed cells have lower glycolysis than unprimed control cells even in absence of pathogen in Figure 1A?

IL-4 primed AM do not have statistically significant changes in glycolysis compared with unprimed control cells in the absence of stimulation.  

Reviewer #2 (Recommendations For The Authors):

The manuscript entitled "Human airway macrophages are metabolically reprogrammed by IFN-γ resulting in glycolysis dependent functional plasticity" by Cox et al., characterizes glycolytic-linked cytokine secretion and surface receptor expression of primary human airway macrophages (AM) and monocyte-derived macrophages (MDM). The authors primed the primary macrophages with type II interferon (IFN-γ) or interleukin-4 (IL-4) into Th1 and Th2 polarized states. This was followed by measurement of the shift in macrophage metabolism to glycolysis (ECAR measurement) and/or oxidative phosphorylation (OCR measurement) in response to lipopolysaccharide and irradiated Mycobacterium tuberculosis. The authors then utilize 2-DG (an inhibitor of glycolysis) to show the reliance of glycolytic shift in metabolism to drive the expression of different macrophage activation markers in MDMs and cytokine secretion in AMs.

Significance:

The study provides important validation of IFN-γ-mediated glycolytic shift and its correlated functionalities in primary human macrophage populations.

Highlights: The study characterizes glycolytic-linked cytokine secretion and expression of macrophage activation markers in primary human resident (lung) and monocyte (blood)-derived macrophages. The study also shows data in support of IFN-γ alone in mediating glycolytic reprogramming of human primary macrophages.

Limitations:

The study lacks novelty and does not provide any new or different information in relation to IFN-γmediated glycolytic shift in the metabolism of human macrophages.

Major comments:

(1) The authors have relied on irradiated Mycobacterium tuberculosis (Mtb) and LPS stimulation to measure different correlates of macrophage functions. Additionally, the authors have discussed their results with irradiated Mtb with that of infection with live Mtb. There are also recent reports that show Mtb infection limiting glycolytic reprogramming in murine and human macrophages (PMID: 31914380) in contrast to their observation with irradiated Mtb. The authors should also include live Mtb infection or other replicative live bacterium for the induction of surface activation markers and cytokine release in their setup.

We thank the reviewer for this suggestion; however, this is beyond the scope of the current study which was to assess AM and MDM in the context of immune stimulation in a reductive manner using TLR4 ligand LPS and a more complete whole bacteria stimulation. The selected bacterial ligands were employed in the study to allow us to model an optimal macrophage host response. This minimises the confounding variable of live bacteria which can perturb cellular metabolism and immune responses, which we have highlighted in the discussion. Since both LPS and irradiated Mtb induced similar metabolic and phenotypic profiles, it is likely that the effects of priming are maintained with diverse stimuli.  

(2) The authors should add a quantitative measure (like extracellular lactate secretion or ECAR level) for the extent of glycolytic inhibition by the use of 5 mM 2-DG in their setup.

We would like to draw the attention of the reviewer to the data represented in supplementary figure 2B, demonstrating that 2DG lowers ECAR at 5mM at both 1 and 24 h post stimulation with iH37Rv by an average of approximately 40%. In addition, we have acknowledged that inhibition with 5 mM 2DG does not fully inhibit glycolysis as outlined in the study limitations (lines 477-480).  

(3) Percent change and fold change have been used to show the same or similar result in Fig. 1 and 2. Whereas, supplementary Fig. 1 shows absolute ECAR/OCR values in addition to fold change. The authors can plot either fold change or percent change in different measurements to avoid confusion. For example, do ECAR changes upon LPS stimulation in Fig. 1A and 1C come from the same dataset? One of the data points in percent change shows a decrease in percent ECAR change under no cytokine control, whereas all the data points in fold change show an increase.

We have addressed this comment above in response to reviewer 1 point 5 (recommendations for the authors).

We thank the reviewer for highlighting this single error in the data points for percent change. We have fixed this data point which was a result of a calculation error. All data throughout the manuscript has now been rechecked.   

Minor comments:

(1) The manuscript for review should be line-marked for referencing and commenting during review.

We have now included line-marking on the manuscript.  

(2) The authors can depict marker legends differently for all figures. In all figures, circles to squares or triangles represent treatment/stimulation with iH37Rv or LPS. The authors can depict this as circles to squares/triangles in contrast to different legends.

We have changed the legend to include a more detailed description of data represented inserting additional information regarding the colours and symbols represented in the figures.  

(3) Describe bars in supplementary figure 1A - 1H in its legend?

We thank the reviewer for highlighting this oversight, we have amended the legend to state “error bars represent standard deviation”

(4) Discuss the significant increase in CD86 expression in IFN-γ and IL-4 primed unstimulated AMs in Fig. 3E.

We have updated the results section to state that IFN-γ increased the expression of CD86 when isolated in the absence of bacterial stimulations in Fig. 3E (lines 271-272). There is no significant increase in CD86 by IL-4 primed unstimulated AM. IL-4 primed human AM only upregulated CD86 when treated with 2DG or in the presence of stimulation.  

(5) Contrary to Fig. 2, the data points of unstimulated cells in Fig. 4 vary for different treatment conditions (no cytokine, IFN-γ, and IL-4) for each cytokine measurement. What is the difference between unstimulated cells in Fig. 4 (for each cytokine) from that of Fig. 2 (for each receptor MFI)?

Unstimulated cells change their surface activation markers and phenotype in response to IFN-γ and IL-4 in Fig. 2. For Fig. 4, IFN-γ and IL-4 are not sufficient to induce cytokine secretion in the absence of stimulation with bacterial ligands.  

(6) The methodology for seeding and treatment of cells is reemphasized for almost all results. Defining macrophage priming and stimulation of macrophages in the method section and once at the start of results should be fine.

Plating happens differently for Seahorse compared to the flow cytometric phenotyping and ELISA for cytokine production. For clarity we have stated and reemphasized the seeding and treatment of cells throughout the results section.  

(7) Clarify "IL-4 reduced glycolysis in response to LPS stimulation" in relation to the results depicted in Fig. 1A and 1C. Similarly, clarify "IL-4 resulting in reduced IL-1β and IL-10 production" in relation to Fig. 4E.

For clarity we have added the following lines (157-160, 164-170) to the manuscript:  

“IL-4 primed iH37Rv stimulated AM increased ECAR to similar extent as unprimed controls (Figure 1A; left). Conversely, IL-4 primed AM stimulated with LPS AM did not increase their ECAR to the same extent as controls (Figure 1A; right), suggesting that IL-4 reduces the AM ability to increase ECAR in response to LPS stimulation.”   

“Since IFN-γ priming increased cellular energetics in the AM at baseline, we calculated percent change in ECAR and OCR from the baseline rate of each group in order to assess if IFN-γ or IL-4 primed AM have altered capacity to change their metabolism in response to stimulation (Figure 1C,D). This was carried out to equalise all the primed data sets at baseline before stimulation (Figure S1C, S1D). These data indicate that whilst the peak of glycolysis is elevated in IFN-γ primed AM (Figure S1A), all AM have a similar capacity to increase glycolysis upon stimulation when baseline differences in metabolism were adjusted for the effects of cytokine priming (Figure 1C).”

For clarity we have amended the sentence the reviewer has highlighted (lines 214-215): “IL-4 primed AM had reduced fold change in glycolysis upon stimulation with LPS compared with controls”.

Since IFN-γ priming induced large effect sizes, we statistically analysed the IL-4 primed and unprimed data sets in the absence of the IFN-γ primed data sets to determine how IL-4 influenced macrophage function. The only data where this resulted in any statistical significance was in response to cytokine production. We have now clarified this in the methods and relevant figure legends by stating, “Statistically significant differences were determined using two-way ANOVA with a Tukey post-test (AD); *P≤0.05, **P≤0.01, ***P≤0.001, ****P≤0.0001 or #P≤0.05, ##P≤0.01 (where IFN-γ primed data sets were excluded for post-test analysis to analyse statistical differences between no cytokine and IL4 treated data sets).

To further clarify this, we have amended the text of the manuscript (lines 307-310) to reflect this. “All stimulated AM secreted IL-10 regardless of priming (Figure 4E). IFN-γ significantly enhanced iH37Rv induced IL-10 in AM compared to unprimed or IL-4 primed comparators (Figure 4E). IL-4 priming of human AM significantly reduced IL-10 production in response to iH37Rv compared with unprimed AM (Figure 4E). LPS strongly induced IL-10 production in unprimed MDM, which was significantly attenuated by either IFN-γ or IL-4 priming (Figure 4F).”  

(8) Clarify whether data points in unstimulated, iH37Rv stimulated, and LPS-stimulated control cells in Fig. 3A - 3F are from independent experiments from those in Fig. 2A - 2F? The distribution of data points of control (no 2-DG treatment) in Fig. 3 is highly similar to the corresponding data points in Fig. 2. Similarly, provide clarification for similarity in Fig. 5A - 5F and Fig. 4A - 4F.

The data illustrated in figure 2 and 3 are from one very large dataset, as are the data in figures 4 and 5. This large experiment was designed to test the effect of priming macrophages with IFN- or IL-4 (in the presence or absence of stimulation), and also to determine if the differential responses elicited due to priming were dependent on glycolysis (by inhibiting with 2DG). For clarity and transparency, the same stimulated dataset is repeated in both figures. Given the size and complexity of the experiment, we chose to present the data this way to aid the reader.  

(9) Clarify the statement "where data was reanalyzed in the absence of IFN-γ" in the section pertaining to Statistical analysis. The authors should clearly mention nature of biological and technical replicates for each experiment in its figure legend. The authors should also confirm multiple comparison correction in all 2-way ANOVA tests done in each figure legend."

We have amended the text (lines 133-136) to clarify this point “P-values of ≤0.05 were considered statistically significant and denoted with an asterisk. Alternatively, P-values of ≤0.05 were denoted with a hashtag where data was analysed in the absence of IFN-γ primed data sets, to analyse statistical differences between no cytokine and IL-4 treated data sets.”  

Figures represent biological replicates (which are the average of technical replicates, presented as a single data point). This is indicated by the following sentence in each figure legend: “Each linked data point represents the average of technical duplicates for one individual biological donor”.  

Each legend has been amended to include the multiple comparison post-test applied.

(10) Discuss the differences and similarities of IFN-γ driven metabolic reprogramming of primary murine macrophages with the results of this study relative to cytokine secretion and activation marker expression.

We have added additional discussion and detail comparing human and murine macrophages in lines 381-382, 403, 407 and 412-415 of the manuscript.

(11) The repetitive data plots of similar results can be significantly reduced to improve the interpretation of the results.

The benefit of the plotting the data in this way is for a clearer understanding and representation of the data. The repetitive data plots allow the benefit of being able to first delineate the effect of priming and priming plus stimulation and then, separately, to further examine the differences in AM versus MDM. The repetition of the primed data points then allows of the reader to determine the effect of inhibiting glycolysis with 2DG on unprimed and primed macrophages (with and without stimulation).   

Reviewer #3 (Recommendations For The Authors):

The methods used and data reported in this manuscript contribute to our understanding of the role of metabolism in programming of macrophages during priming. Suggestions for improving the presentation and interpretation of results include:

• Consult with a statistician regarding analyses of the multiple conditions used during these assays. The use of repeated statistical analyses with different comparison groups in the same figure/data set seems atypical and should either be amended or fully justified in the text. Also, use of two-way vs. one-way ANOVA should be evaluated and clarified.

We have now consulted a statistician. We have amended the text (lines 133-136) to clarify this point “P-values of ≤0.05 were considered statistically significant and denoted with an asterisk. Alternatively, P-values of ≤0.05 were denoted with a hashtag where data was analysed in the absence of IFN-γ primed data sets, to analyse statistical differences between no cytokine and IL-4 treated groups.”  

There are two variables in the data sets; cytokine priming as well as stimulation status therefore we opted for a two-way ANOVA rather than a One-way ANOVA. There are three stimulation groups: unstimulated, Mtb-stimulated and LPS-stimulated. Cytokine priming also has three groups: no cytokine, IFN-y, or IL-4. There are two variables (priming and stimulation), each with 3 groups i.e., six treatment conditions in total, therefore two-way AVOVA with multiple comparisons tests help pinpoint exactly which groups (e.g., the 6 different levels of the 'stimulation' and 'cytokine' treatments) are significantly different from each other. This was important for understanding the specific effects of our treatments. The reader can therefore also deduce how these six treatment conditions compare to each other.

In contrast, performing multiple single comparisons independently of the rest of the dataset (e.g. t tests), increases the risk of false positives (type 1 error). Multiple comparisons ANOVA with post-tests adjust for this, helping to reduce the likelihood of a type 1 error. These stats are more stringent, and it is therefore harder to get P values <0.05. Hence, if we compared all six treatment groups without adjustment, you increase the chance of finding false positives due to the sheer number of comparisons, leading to biased and incorrect conclusions.

In our case, multiple comparisons tests were essential after the two-way ANOVA because they helped to objectively identify specific treatment group differences and control the overall error rate when we were extracting our conclusions, thereby reducing any risk of biases in our conclusions.

A one-way ANOVA is used to test the effect of a single variable with more than two groups contained in the dataset. For example, in our case if you only want to test how different 'stimulation' groups affect ECAR or OCR, only in unprimed macrophages, a one-way ANOVA would be used.

The current study used two-way ANOVA to test the effects of two variables (priming and stimulation, or in some cases priming and inhibition) each containing 3 groups, and see if there is any interaction between the two factors. For example, in our case this allowed us to examine how the 'stimulation' and the 'cytokine' priming affect ECAR/OCR levels and to determine if the effect of 'stimulation' depends on the 'cytokine' priming.

• More justification could be given for the dose of IFNγ used for priming. Inflammatory priming is typically performed with a "low-dose" treatment (e.g., ~1 ng/ml), whereas the authors use 10 ng/ml, which would be considered a high dose. It would be useful to repeat select experiments with a more standard low-dose treatment of IFNg to demonstrate that this is also sufficient to induce the observed metabolic changes.

Previous work has identified little difference in the response of AM and peripheral monocytes to low versus high doses of IFN-γ (6). We have inserted the following into the study limitations (lines 479-481).  

“Furthermore, only one dose of IFN-γ was utilised due to limitations in AM yield, however, recently both low and high doses of IFN-γ have been shown to have similar effects on AM in vitro (6).”

• Check for accuracy of the Fig.4 legend. Also check that 4G and 4B math is consistent.

The legend for Figure 4 has been amended for incorrect A,B to state G,H. The math has been double checked for accuracy and is correct. 3 out of 10 MDM donors produced IL-1β in the absence of IFN-γ in Figure 4B, therefore the average used to calculate the data represented in Figure 4G was brought down markedly by donors who produced little or no IL-1β.  

• Functional plasticity is a vague term and difficult to interpret in this context. It is stated that AM have greater functional plasticity, but MDMs appear to have greater capacity to secrete IL-1β and respond more robustly to IL-4 in terms of T cell stimulation. On that note, the claims regarding antigen presentation would be more impactful if a direct comparison of antigen presentation capacity was made between AM and MDM.

Our data suggests that AM have a greater ability to alter cytokine production, such as IL1β. To consider different culture and plating requirements of MDM v AM cytokine concentration was normalised and expressed in terms of fold change.  This gives a more controlled and accurate comparison of the ability of IFN-γ or IL-4 to modulate cytokine production in AM compared with MDM.  

The terms ‘functional plasticity’ and phenotypic plasticity’ have now been defined in the manuscript in lines 60-63.  

We have therefore added the following sentence into the conclusion of the manuscript (lines 490-493). “Cumulatively, the data presented herein suggests that the MDM maybe more phenotypically plastic than the AM, while the AM have enhanced functional plasticity in their ability to produce cytokine after exposure Th1 and Th2 cytokines.”

However, we acknowledge that the MDM may be regarded as more plastic because of their ability to respond robustly to IL-4, whereas the phenotypic and functional changes in the AM in response to IL4 are more limited. Whilst the focus of our work was to determine if AM are a tractable target to promote immunity in the lungs through upregulation of pro-inflammatory effector function, their ability to downregulated inflammation in response to IL-4 is comparatively less profound compared with MDM.  

We acknowledge the shortcomings of our work which did not allow us to directly measure antigen processing in the AM, due to limitations in the cellular yield from BALF. We have edited the text (lines 251-252 and 286) to clarify this for the reader.  

• Inconsistent normalization complicates interpretation of metabolic data. For example, it is unclear, for example, whether changes in glycolysis and oxidative phosphorylation in primed cells are due to priming or subsequent treatments. Check harmony of methods for analysis of "metabolic assays" with Fig.1 data, axis, and legend.

We have addressed this comment, which is similar to points made by the other reviewers and amended the manuscript to increase clarity. These changes are outlined in the response to reviewer 1, point 5 (recommendations for the author). In addition, we have amended the metabolic assay method (lines 111-112) to state that “Post stimulation the ECAR and OCR were continually sampled at 20-minute intervals for times indicated.”

• A direct comparison of cytokine production after priming and stimulation with Mtb or LPS is limited by inconsistent axes. The data may not support a claim that AM has greater "functional plasticity" without a direct comparison of antigen presentation. Moreover, MDM secrete more IL-1β than AM. The claim that that AM "have increased ability to produce all cytokines assayed in response to Mtb stimulation" does not appear to be supported by the data.

We have amended the text to clarify this issue (lines 313-315). “These data suggest that the AM have greater functional plasticity in terms of their ability to upregulate cytokine production in response to IFN-γ, compared with the MDM. IFN-γ primed AM have enhanced IL-10 and TNF production in response to Mtb and LPS, respectively.”  

We have amended the manuscript and have replaced “IFN-γ primed AM have increased ability to produce all cytokines assayed in response to Mtb stimulation” with the following (lines 421-423) “IFNγ primed AM stimulated with Mtb have significantly more production of IL-1β, TNF and IL-10 compared with unprimed controls. This is in contrast with IFN-γ primed MDM which only upregulate TNF compared to their unprimed controls.”

• AM populations could be defined experimentally.

Airway macrophages were adherence purified from bronchoalveolar lavage fluid defined as CD68+CD14+ as per rebuttal figure 1. The purpose of this study was to examine if human peripherally derived or lung resident macrophages were plastic in response to the classical polarising cytokines IFNγ and IL-4. We have identified that the AM and MDM do indeed have different functional and metabolic responses to these cytokines. However, determining functional differences within the AM subpopulations is beyond the scope of the current study and hampered by low cell numbers in human BALF.  

References

(1) Conzelmann M, Wagner AH, Hildebrandt A, Rodionova E, Hess M, Zota A, Giese T, Falk CS, Ho AD, Dreger P, Hecker M, Luft T. IFN-γ activated JAK1 shifts CD40-induced cytokine profiles in human antigen-presenting cells toward high IL-12p70 and low IL-10 production. Biochemical pharmacology 2010; 80: 2074-2086.

(2) Fries KM, Sempowski GD, Gaspari AA, Blieden T, Looney RJ, Phipps RP. CD40 Expression by human fibroblasts. Clinical Immunology and Immunopathology 1995; 77: 42-51.

(3) Gu W, Chen J, Yang L, Zhao KN. TNF-α promotes IFN-γ-induced CD40 expression and antigen process in Myb-transformed hematological cells. TheScientificWorldJournal 2012; 2012: 621969.

(4) Hershman MJ, Appel SH, Wellhausen SR, Sonnenfeld G, Polk HC, Jr. Interferon-gamma treatment increases HLA-DR expression on monocytes in severely injured patients. Clinical and experimental immunology 1989; 77: 67-70.

(5) Maenaka A, Kenta I, Ota A, Miwa Y, Ohashi W, Horimi K, Matsuoka Y, Ohnishi M, Uchida K, Kobayashi T. Interferon-γ-induced HLA Class II expression on endothelial cells is decreased by inhibition of mTOR and HMG-CoA reductase. FEBS open bio 2020; 10: 927-936.

(6) Thiel BA, Lundberg KC, Schlatzer D, Jarvela J, Li Q, Shaw R, Reba SM, Fletcher S, Beckloff SE, Chance MR, Boom WH, Silver RF, Bebek G. Human alveolar macrophages display marked hyporesponsiveness to IFN-γ in both proteomic and gene expression analysis. PLoS One 2024; 19: e0295312.

Author Response

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1 (Public Review):

(1) The authors' primary research question revolves around the inquiry of "how far in advance semantic information might become available from parafoveal preview." In contrast to prior studies, the current research seeks to achieve a breakthrough in terms of timing by employing innovative technology. They mention in the manuscript that "most of these studies have been limited to measuring parafoveal preview from fixations to an immediately adjacent word... We tackle these core issues using a new technique that combines the use of frequency tagging and the measurement of magnetoencephalography (MEG)-based signals." However, the argumentation for how this new technology constitutes a breakthrough is not sufficiently substantiated. Specifically, there are two aspects that require further clarification. Firstly, the authors should clarify the importance of investigating the timing of semantic integration in their research question. They need to justify why previous studies focusing on the preview effect during fixations to an immediately adjacent word cannot address their specific inquiry about "how far in advance semantic information might become available from parafoveal preview," which requires examining parafoveal processing (POF). Secondly, in terms of the research methodology, the authors should provide a more comprehensive explanation of the advantages offered by MEG technology in the observation of the timing of semantic integration compared to the techniques employed in prior research. Indeed, the authors have overlooked some rather significant studies in this area. For instance, the research conducted by Antúnez, Milligan, Hernández-Cabrera, Barber, & Schotter in 2022 addresses the same research question mentioned in the current study and employs a similar experimental design. Importantly, they utilize a natural reading paradigm with synchronized ERP and eye-tracking recordings. Collectively, these studies, along with the series of prior research studies employing ERP techniques and RSVP paradigms discussed by the authors in their manuscript, provide ample evidence that semantic information becomes available and integrated from words before fixation occurs. Therefore, the authors should provide a more comprehensive citation of relevant research and delve deeper into explaining the potential contributions of their chosen technology to this field.

We express our gratitude to the reviewer for providing insightful comments. Firstly, we clarify the advantages of the RIFT technique. The revised paragraph is on Page 4 with tracked changes and is copied as follows:

“…… The RIFT technique provides a notable advantage by generating a signal — the tagging response signal — specifically yoked to just the tagged word. This ensures a clear separation in processing the tagged word from the ongoing processing of other words, addressing a challenge faced by eye tracking and ERP/FRP approaches. Moreover, RIFT enables us to monitor the entire dynamics of attentional engagement with the tagged word, which may begin a few words before the tagged word is fixated.”

We also rephase our research questions in the introduction section on Page 5 with tracked changes:

“This paradigm allows us to address three questions. First, we aimed to measure when in the course of reading people begin to direct attention to parafoveal words. Second, we sought to ascertain when semantic information obtained through parafoveal preview is integrated into the sentence context. Modulations of pre-target RIFT responses by the contextual congruity of target words would serve as evidence that parafoveal semantic information has not only been extracted and integrated into the sentence context but that it is affecting how readers allocate attention across the text. Third, we explored whether these parafoveal semantic attention effects have any relationship to reading speed.”

Secondly, we would like to elucidate the significance of investigating the timing of semantic integration and why this complements existing findings of parafoveal processing (POF) during reading. Our manuscript has been revised accordingly, with specific modifications highlighted on Page 2. The revised passage reads as follows:

“…… eye tracking-based evidence for the extraction of parafoveal semantic information …… was eventually extended into English …… For example, Schotter and Jia (2016) showed preview benefits on early gaze measures for plausible compared to implausible words, even for plausible words that were unrelated to the target. These results demonstrate that semantic information can indeed be extracted from parafoveal words. However, due to the limitations of the boundary paradigm, which only assesses effects after target words have been fixated, it is challenging to precisely determine when and how parafoveal semantic processing takes place. Furthermore, it is generally hard to distinguish between the effects of cross-saccade integration (e.g., mismatch between the preview and the word fixated) and the effects of how differing words fit into the context itself (Veldre and Andrews, 2016a, 2016b).”

Thirdly, we now better highlight the contributions of Antúnez et al. paper as they have provided important evidence for parafoveal semantic processing during natural reading. The relevant modifications are highlighted on Page 3. The revised passage is as follows: “Although many of these effects have been measured in the context of unnatural reading paradigms (e.g., the “RSVP flanker paradigm”), similar effects obtain during natural reading. Using the stimuli and procedures from Schotter and Jia (2016), Antúnez et al. (2022) showed that N400 responses, measured relative to the fixation before the target words (i.e., before the boundary change while the manipulated words were in parafoveal preview), were sensitive to the contextual plausibility of these previewed words. These studies suggest that semantic information is available from words before they are fixated, even if that information does not always have an impact on eye fixation patterns.”

References:

Schotter ER, Jia A. 2016. Semantic and plausibility preview benefit effects in English: Evidence from eye movements. J Exp Psychol Learn Mem Cogn 42:1839–1866. doi:10.1037/xlm0000281

Veldre A, Andrews S. 2016a. Is Semantic Preview Benefit Due to Relatedness or Plausibility? J Exp Psychol Hum Percept Perform 42:939–952. doi:10.1037/xhp0000200

Veldre A, Andrews S. 2016b. Semantic preview benefit in English: Individual differences in the extraction and use of parafoveal semantic information. J Exp Psychol Learn Mem Cogn 42:837–854. doi:10.1037/xlm0000212

Antúnez M, Milligan S, Andrés Hernández-Cabrera J, Barber HA, Schotter ER. 2022. Semantic parafoveal processing in natural reading: Insight from fixation-related potentials & eye movements. Psychophysiology 59:e13986. doi:10.1111/PSYP.13986

(2) Further, the authors emphasize semantic integration in their observed results but overlook the intricate relationship between access, priming, and integration. This assertion appears overly confident. Despite using low-constraint sentences and low-predicted targets (lines 439-441), differences between congruent and incongruent conditions may be influenced by word-level factors. For instance, in the first coherent sentence, such as "Last night, my lazy brother came to the party one minute before it was over" (line 1049), replacing the keyword "brother" with an incongruent word could create an incoherent sentence, possibly due to semantic violation, relation mismatch with "lazy," or prediction error related to animate objects. A similar consideration applies to the second example sentence, "Lily says this blue jacket will be a big fashion trend this fall" (line 1050), where the effect might result from a discrepancy between "blue" and an incongruent word. However, the authors do not provide incongruent sentences to substantiate their claims. I recommend that the authors discuss alternative explanations and potentially control for confounding factors before asserting that their results unequivocally reflect semantic integration. My intention is not to dispute the semantic integration interpretation but to stress the necessity for stronger evidence to support this assertion.

We agree with the reviewer that stimulus control is very critical for this kind of work and apologize for the lack of clarity in the original manuscript.

(1) We fully agree that word-level factors can be an important confound, which is why we carefully controlled word-level factors in the experimental design. As detailed in the Appendix of the original manuscript, each pair of target words has been strategically embedded into two sentences, allowing for the creation of both congruent and incongruent sentence pairs through the interchange of these words. We now have explicitly specified this design in all sentences, as reflected in the edited manuscript on Page 38. For example, considering the exemplar pair of “brother/jacket”,

“Last night, my lazy brother/jacket came to the party one minute before it was over.

Lily says this blue jacket/brother will be a big fashion trend this fall.”

In this design, the pair of target words is presented in both congruent and incongruent sentences. Participant A reads “lazy brother” and “blue jacket”, while Participant B reads “lazy jacket” and “blue brother”. This approach ensures that the same target words appear in both congruent and incongruent conditions across participants, serving as an effective control for word-level factors.

(2) We acknowledge that the consideration of word-level information is crucial when making claims about contextual integration in the current study. However, we don’t think there are many cases in the stimulus set where a single feature like animacy is enough to create the mismatch. Instead, the stimuli were written so that it is not possible to strongly predict any word or even a specific semantic feature, so that appreciating the mismatch requires the comprehender to integrate the word into the context (and especially to integrate the word with the immediately preceding one). However, this more local modifier/noun plausibility may behave differently from a more global contextual plausibility, which is a limitation of the stimulus set and has been discussed in the revised manuscript, as indicated by the tracked changes on Page 16, as copied below:

“Two noteworthy limitations exist in the current study. Firstly, the construction of pretarget–target word pairs consistently follows an adjective–noun phrase structure, potentially leading to semantic violations arising from immediate local incongruence rather than a broader incongruence derived from the entire sentential context. While the context preceding target words was deliberately minimized to ensure a pure effect of bottom-up parafoveal processing rather than the confounding impact of top-down prediction, it is essential to recognize that information from both local and global contexts can exert distinct effects on word processing during natural reading (Wong et al., 2022). Future investigations should incorporate more information-rich contexts to explore the extent to which the parafoveal semantic integration effect observed in this study can be generalized.”

References:

Wong R, Veldre A, Andrews S. 2022. Are There Independent Effects of Constraint and Predictability on Eye Movements During Reading? J Exp Psychol Learn Mem Cogn. doi:10.1037/XLM0001206

Reviewer #2 (Public Review):

This MEG study used co-registered eye-tracking and Rapid Invisible Frequency Tagging (RIFT) to track the effects of semantic parafoveal preview during natural sentence reading. Unpredictable target words could either be congruent or incongruent with sentence context. This modulated the RIFT response already while participants were fixating on the preceding word. This indicates that the semantic congruency of the upcoming word modulates visual attention demands already in parafoveal preview.

The quest for semantic parafoveal preview in natural reading has attracted a lot of attention in recent years, especially with the development of co-registered EEG and MEG. Evidence from dynamic neuroimaging methods using innovative paradigms as in this study is important for this debate.

We express our gratitude to the reviewer for recognizing the significance of our research question in the domain of natural reading.

Major points:

(1) The authors frame their study in terms of "congruency with sentence context". However, it is the congruency between adjective-noun pairs that determines congruency (e.g. "blue brother" vs "blue jacket", and examples p. 16 and appendix). This is confirmed by Suppl Figure 1, which shows a significantly larger likelihood of refixations to the pre-target word for incongruent sentences, probably because the pre-target word is most diagnostic for the congruency of the target word. The authors discuss some possibilities as to why there is variability in parafoveal preview effects in the literature. It is more likely to see effects for this simple and local congruency, rather than congruency that requires an integration and comprehension of the full sentence. I'm not sure whether the authors really needed to present their stimuli in a full-sentence context to obtain these effects. This should be explicitly discussed and also mentioned in the introduction (or even the abstract).

We have addressed this limitation of the study explicitly in the revised manuscript. The modifications can be found in the tracked changes on Page 16, and is copied as follows:

“Two noteworthy limitations exist in the current study. Firstly, the construction of pretarget–target word pairs consistently follows an adjective–noun phrase structure, potentially leading to semantic violations arising from immediate local incongruence rather than a broader incongruence derived from the entire sentential context. While the context preceding target words was deliberately minimized to ensure a pure effect of bottom-up parafoveal processing rather than the confounding impact of top-down prediction, it is essential to recognize that information from both local and global contexts can exert distinct effects on word processing during natural reading (Wong et al., 2022). Future investigations should incorporate more information-rich contexts to explore the extent to which the parafoveal semantic integration effect observed in this study can be generalized.”

References:

Wong R, Veldre A, Andrews S. 2022. Are There Independent Effects of Constraint and Predictability on Eye Movements During Reading? J Exp Psychol Learn Mem Cogn. doi:10.1037/XLM0001206

(2) The authors used MEG and provided a source estimate for the tagging response (Figure 2), which unsurprisingly is in the visual cortex. The most important results are presented at the sensor level. This does not add information about the brain sources of the congruency effect, as the RIFT response probably reflects top-down effects on visual attention etc. Was it necessary to use MEG? Would EEG have produced the same results? In terms of sensitivity, EEG is better than MEG as it is more sensitive to radial and deeper sources. This should be mentioned in the discussion and/or methods section.

Source estimation was exclusively provided for the tagging response rather than the congruency effect because we posit that this conditional contrast would emanate from the same brain regions exhibiting the tagging responses in general. As depicted in the following figure, source localization for the congruency effect was identified in the left association cortex (Brodmann area 18), the same area as the source localization for the tagging response (the negative cluster observed here is due to the incongruent minus congruent contrast). While we agree with the Reviewer that the RIFT result might indicate a top-down effect on visual attention, it is important to note that, due to the low-pass filter property of synapses, observing a tagging response at a high frequency beyond the visual cortex is challenging.

Author response image 1.

We discussed the necessity of using MEG in the edited manuscript with tracked changes on Page 20, and is copied as follows:

“While the current study was conducted using MEG, these procedures might also work with EEG. If so, this would make our approach accessible to more laboratories as EEG is less expensive. However, there are currently no studies directly comparing the RIFT response in EEG versus MEG. Therefore, it would be of great interest to investigate if the current findings can be replicated using EEG.”

(3) The earliest semantic preview effects occurred around 100ms after fixating the pre-target word (discussed around l. 323). This means that at this stage the brain must have processed the pre-target and the target word and integrated their meanings (at some level). Even in the single-word literature, semantic effects at 100 ms are provocatively early. Even studies that tried to determine the earliest semantic effects arrived at around 200 ms (e.g. (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3382728/, https://psycnet.apa.org/record/2013-17451-002). The present results need to be discussed in a bit more detail in the context of the visual word recognition literature.

We have incorporated this valuable suggestion into the discussion section to enhance the clarity of our key result regarding the timing of parafoveal semantic integration. The revised manuscript with tracked changes can be found on Page 14, and the relevant passage is provided below:

“Our results also provide information about the time course of semantic integration …… by as early as within 100 ms after fixating on the pre-target word. The timing of this parafoveal semantic effect appears remarkably early, considering that typical semantic access for a single word occurs no earlier than around 200 ms, as demonstrated in the visual word recognition literature (Carreiras et al., 2014). For instance, in a Go/NoGo paradigm, the earliest distinguishable brain activity related to category-related semantic information of a word occurs at 160 ms (Amsel et al., 2013; Hauk et al., 2012). Therefore, the RIFT results presented here suggest that natural reading involves parallel processing that spans multiple words. The level of (covert) attention allocated to the target word, as indexed by the significant difference in RIFT responses compared to the baseline interval, was observed even three words in advance (see Figure 2C). This initial increase in RIFT coincided with the target entering the perceptual span (McConkie and Rayner, 1975; Rayner, 1975; Underwood and McConkie, 1985), likely aligning with the initial extraction of lower-level perceptual information about the target. The emerging sensitivity of the RIFT signal to target plausibility, detected around 100 ms after the fixation on the pre-target word, suggests that readers at that time had accumulated sufficient semantic information about the target words and integrated that information with the evolving sentence context. Therefore, it is plausible that the initial semantic processing of the target word commenced even before the pre-target fixation and was distributed across multiple words. This parallel processing of multiple words facilitates rapid and fluent reading.”

References:

Carreiras M, Armstrong BC, Perea M, Frost R. 2014. The what, when, where, and how of visual word recognition. Trends Cogn Sci 18:90–98. doi:10.1016/j.tics.2013.11.005

Amsel BD, Urbach TP, Kutas M. 2013. Alive and grasping: Stable and rapid semantic access to an object category but not object graspability. Neuroimage 77:1–13. doi:10.1016/J.NEUROIMAGE.2013.03.058

Hauk O, Coutout C, Holden A, Chen Y. 2012. The time-course of single-word reading: Evidence from fast behavioral and brain responses. Neuroimage 60:1462. doi:10.1016/J.NEUROIMAGE.2012.01.061

McConkie GW, Rayner K. 1975. The span of the effective stimulus during a fixation in reading. Percept Psychophys 17:578–586. doi:10.3758/BF03203972

Rayner K. 1975. The perceptual span and peripheral cues in reading. Cogn Psychol 7:65–81.

Underwood NR, McConkie GW. 1985. Perceptual Span for Letter Distinctions during Reading. Read Res Q 20:153. doi:10.2307/747752

(4) As in previous EEG/MEG studies, the authors found a neural but no behavioural preview effect. As before, this raises the question of whether the observed effect is really "critical" for sentence comprehension. The authors provide a correlation analysis with reading speed, but this does not allow causal conclusions: Some people may simply read slowly and therefore pay more attention and get a larger preview response. Some readers may hurry and therefore not pay attention and not get a preview response. In order to address this, one would have to control for reading speed and show an effect of RIFT response on comprehension performance (or vice versa, with a task that is not close to ceiling performance). The last sentence of the discussion is currently not justified by the results.

We acknowledge that the correlation analysis between the RIFT effect and reading speed on the group level lacks causality, making it less ideal for addressing this question. We have incorporated this acknowledgment as one of the limitations of the current study in the revised manuscript on Page 16, as indicated by the tracked changes, and the relevant passage is provided below:

“Two noteworthy limitations exist in the current study. …… Secondly, the correlation analysis between the pre-target RIFT effect and individual reading speed (Figure 5) does not establish a causal relationship between parafoveal semantic integration and reading performance. Given that the comprehension questions in the current study were designed primarily to maintain readers’ attention and the behavioural performance reached a ceiling level, employing more intricate comprehension questions in future studies would be ideal to accurately measure reading comprehension and reveal the impact of semantic parafoveal processing on it.”

We reformulated the last sentence:

“These results support the idea that words are processed in parallel and suggest that early and deep parafoveal processing may be important for fluent reading.”

(5) L. 577f.: ICA components were selected by visual inspection. I would strongly recommend including EOG in future recordings when the control of eye movements is critical.

We appreciate the reviewer for providing this valuable suggestion. We acknowledge that EOG recordings were not included in the current study due to restrictions on MEG data collection from the University of Birmingham during the COVID-19 pandemic. In our future studies, we will follow the reviewer's suggestion to incorporate EOG recordings in data collection. This addition will facilitate optimal eye movement-related artifact rejection through ICA, as recommended by Dimigen in his methodological paper:

Dimigen, O. (2020). Optimizing the ICA-based removal of ocular EEG artifacts from free viewing experiments. NeuroImage, 207, 116117.

(6) The authors mention "saccade planning" a few times. I would suggest looking at the SWIFT model of eye movement control, which is less mechanistic than the dominant EZ-Reader model (https://psycnet.apa.org/record/2005-13637-003). It may be useful for the framing of the study and interpretation of the results (e.g. second paragraph of discussion).

In the revised manuscript, we have provided a more comprehensive explanation eye movements/saccade planning, aligning it with the SWIFT model. Please refer to Page 15 with tracked changes, and the updated passage is provided below:

“The results of the present study are aligned with the SWIFT model of eye movement control in natural reading (Engbert et al., 2005), wherein the activation field linked to a given word is hypothesized to be both temporally and spatially distributed. Indeed, we found that the initial increase in covert attention to the target word occurred as early as three words before, as measured by RIFT responses (Figure 2C). These covert processes enable the detection of semantic incongruity (Figure 3B and Figure 3C). However, it may occur at the non-labile stage of saccade programming, preventing its manifestation in fixation measures of the currently fixated pre-target word (Figure 1B). Therefore, the RIFT technique’s capacity to yoke patterns to a specific word offers a unique opportunity to track the activation field of word processing during natural reading.”

References:

Engbert R, Nuthmann A, Richter EM, Kliegl R. 2005. Swift: A dynamical model of saccade generation during reading. Psychol Rev 112:777–813. doi:10.1037/0033-295X.112.4.777

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

While the manuscript is well-written and presents a structured analysis of the data, it requires further clarification and substantiation regarding the originality of the research questions, the advantages of the proposed methodology, and the interpretation of the results related to semantic integration. Additional references and a more thorough discussion of related research are needed to strengthen the manuscript's contribution to the field.

We appreciate the reviewer's kind words about this manuscript and the insightful comments and suggestions provided. In the revised manuscript, we have now placed additional emphasis on the importance of investigating semantic integration within the realm of parafoveal processing in natural reading. We have clarified the advantages of employing MEG and RIFT and expanded upon our results in the context of Antúnez et al.'s 2022 paper, as suggested by the reviewer.

Reviewer #2 (Recommendations For The Authors):

(1) L. 59: The "N400" has been linked to much more than "semantic access". I think it is widely accepted that "access" happens (or at least begins) earlier, and that the N400 reflects high-level integration processes etc.

Earlier debates about whether the N400 is more linked to access or integration have resolved in favour of an access account, but with a growing appreciation of the blurred boundaries between constructions like access, priming, and integration, as Reviewer 1 also pointed out in comment #2.

(2) L. 177: I wasn't sure about the selection of sensors. Were the same sensors used for all participants (whether they had a tagging response or not)?

We appreciate the reviewer for highlighting the confusion regarding the sensor selection procedure in the study. In response, we have added further clarifications about this procedure in the Method section of the revised manuscript. The relevant changes can be found on Page 25 with tracked changes, and the modified passage is reproduced below:

"Please note that the tagging response sensors may vary in number across participants (7.9 ± 4.5 sensors per participant, M ± SD). Additionally, they may have a different but overlapping spatial layout, primarily over the visual cortex. For the topography of all tagging response sensors, please refer to Figure 2A."

(3) Ll. 247ff.: I don't understand the idea of a "spill-over effect". The future cannot spill into the past. Or does this refer to possible artefacts or technical problems?

In the revised manuscript, we have rephrased this passage with tracked changes on Page 11, and the updated version is provided below:

“We conducted a similar analysis of the coherence measured when participants fixated the target word and found no significant modulations related to the contextual congruity of that target word. …… Thus, the parafoveal semantic integration effect identified during the pre-target intervals cannot be attributed to signal contamination from fixations on the target word induced by the temporal smoothing of filters.”

(4) I struggled to follow the "internal attention" explanation for the paradoxical RIFT effect (p. 11/12).

We appreciate the reviewer for pointing out the confusion, and we have rephrased the passage in the revised manuscript with tracked changes on Page 13. The revised version is provided below:

"Previous work has demonstrated that tagging responses decrease as attention shifts from an external task (e.g., counting visual targets) to an internal task (e.g., counting heartbeats) (Kritzman et al., 2022). Similarly, in a reading scenario, visually perceiving the flickering word constitutes an external task, while the internal task involves the semantic integration of previewed information into the context. If more attentional resources are internally directed when faced with the challenge of integrating a contextually incongruent word, fewer attentional resources would remain for processing the flickering word. This may be the kind of shift reflected in the reduction in RIFT responses."

References:

Kritzman L, Eidelman-Rothman M, Keil A, Freche D, Sheppes G, Levit-Binnun N. 2022. Steady-state visual evoked potentials differentiate between internally and externally directed attention. Neuroimage 254:119133.

(5) L. 572: Why was detrending necessary on top of a 0.5 Hz high-pass filter? Was detrending applied to the continuous raw data, or to epochs? Was it just the linear trend or other polynomial terms?

We agree with the Reviewer that, given the prior application of a 0.5Hz high-pass filter to the data, the detrending does not alter the data. Nonetheless, we included this procedure in the manuscript for the sake of completeness. In the revised manuscript, we have provided additional clarification on this point, as indicated by the tracked changes on Page 23. The modified passage is presented below:

"Subsequently, detrending was applied individually to each channel of the continuous raw data to factor out the linear trend."

(6) Source analysis, p. 25f.: How was the beamformer regularized?

This information was already included in the original manuscript on Page 26. The original text is provided below for reference:

“No regularisation was performed to the CSD matrices (lambda = 0).”

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

This study focuses on characterizing a previously identified gene, encoding the secreted protein Ppe1, that may play a role in rice infection by the blast fungus Magnaporthe oryzae. Magnaporthe oryzae is a hemibiotrophic fungus that infects living host cells before causing disease. Infection begins with the development of a specialized infection cell, the appressorium, on the host leaf surface. The appressorium generates enormous internal turgor that acts on a thin penetration peg at the appressorial base, forcing it through the leaf cuticle. Once through this barrier, the peg elaborates into bulbous invasive hyphae that colonizes the first infected cell before moving to neighboring cells via plasmodesmata. During this initial biotrophic growth stage, invasive hyphae invaginate the host plasma membrane, which surrounds growing hyphae as the extra-invasive hyphae membrane (EIHM). To avoid detection, the fungus secretes apoplastic effectors into the EIHM matrix via the conventional ER-Golgi secretion pathway. The fungus also forms a plant-derived structure called the biotrophic interfacial complex (BIC) that receives cytoplasmic effectors through an unconventional secretion route before they are delivered into the host cell. Together, these secreted effector proteins act to evade or suppress host innate immune responses. Here the authors contribute to our understanding of M. oryzae infection biology by showing how Ppe1, which localizes to both the appressorial penetration peg and to the appressorial-like transpressoria associated with invasive hyphal movements into adjacent cells, maximizes host cell penetration and disease development and is thus a novel contributor to rice blast disease.

We sincerely appreciate the reviewer’s thoughtful evaluation of our work. We are grateful for your recognition of Ppe1 as a novel contributor to M. oryzae infection biology and your insightful summary of its spatio-temporal localization and functional importance in host penetration. We also appreciate devoting your time to provide us with constructive feedback, which greatly strengthens our manuscript.

Strengths:

A major goal of M. oryzae research is to understand how the fungus causes disease, either by determining the physiological underpinnings of the fungal infection cycle or by identifying effectors and their host targets. Such new knowledge may point the way to novel mitigation strategies. Here, the authors make an interesting discovery that bridges both fungal physiology and effector biology research by showing how a secreted protein Ppe1, initially considered an effector with potential host targets, associates with its own penetration peg (and transpressoria) to facilitate host invasion. In a previous study, the authors had identified a small family of small secreted proteins that may function as effectors. Here they suggest Ppe1 (and, later in the manuscript, Ppe2/3/5) localizes outside the penetration peg when appressoria develops on surfaces that permit penetration, but not on artificial hard surfaces that prevent peg penetration. Deleting the PPE1 gene reduced (although did not abolish) penetration, and a fraction of those that penetrated developed invasive hyphae that were reduced in growth compared to WT. Using fluorescent markers, the authors show that Ppe1 forms a ring underneath appressoria, likely where the peg emerges, which remained after invasive hyphae had developed. The ring structure is smaller than the width of the appressorium and also lies within the septin ring known to form during peg development. This so-called penetration ring also formed at the transpressorial penetration point as invasive hyphae moved to adjacent cells. This structure is novel, and required for optimum penetration during infection. Furthermore, Ppe1, which carries a functional signal peptide, may form on the periphery of the peg, together suggesting it is secreted and associated with the peg to facilitate penetration. Staining with aniline blue also suggests Ppe1 is outside the peg. Together, the strength of the work lies in identifying a novel appressorial penetration ring structure required for full virulence.

We are deeply grateful to the reviewer for the clear understanding and insightful evaluation of our work. Your recognition of the novel contribution and scientific merit of our study is both encouraging and motivating. We sincerely appreciate the time, expertise and constructive feedback dedicated to reviewing our manuscript, as the comments have been instrumental in enhancing the quality of this work.

Weaknesses:

The main weakness of the paper is that, although Ppe1 is associated with the peg and optimizes penetration, the function of Ppe1 is not known. The work starts off considering Ppe1 a secreted effector, then a facilitator of penetration by associating with the peg, but what role it plays here is only often speculated about. For example, the authors consider at various times that it may have a structural role, a signaling role orchestrating invasive hyphae development, or a tethering role between the peg and the invaginated host plasma membrane (called throughout the host cytoplasmic membrane, a novel term that is not explained). However, more effort should be expended to determine which of these alternative roles is the most likely. Otherwise, as it stands, the paper describes an interesting phenomenon (the appressorial ring) but provides no understanding of its function.

We sincerely appreciate the reviewer’s comments. We have revised "host cytoplasmic membrane" to "host plasma membrane" throughout the manuscript for consistency. To further investigate the role of the Ppe1 in the interaction between M. oryzae and rice, we overexpressed PPE1 in rice ZH11. A pCXUN-SP-GFP-Ppe1 vector containing a signal peptide and an N-terminal GFP tag was constructed (pCXUN-SP-GFP-Ppe1), and 35 GFP-PPE1-OX plants (T0) were subsequently obtained through Agrobacterium-mediated rice transformation. Subsequently, PCR and qRT-PCR validation were performed on the T0 transgenic plants. The PCR results showed that the inserted plasmid could be amplified from the genomic DNA extracted from the leaves of all the resulting T0 plants (Author response image 1A). qRT-PCR results indicated that most T0 transgenic plants could transcriptionally express PPE1 (Author response image 1B). T0 plants with higher expression levels were selected for western blot analysis, which confirmed the presence of GFP-Ppe1 bands of the expected size (Author response image 1C). To further explore the targets of Ppe1 in rice, the leaf sheaths of T0 plants were inoculated with M. oryzae strain Guy11. Total proteins were extracted at 24 hours post-inoculation (hpi) and subjected to immunoprecipitation using GFP magnetic beads. Silver staining revealed more interacting protein bands in T0 plants compared to ZH11 and GFP-OX controls (Author response image 1D). These samples were then analyzed by mass spectrometry in which 331 rice proteins that potentially interact with Ppe1 were identified (Author response image 1E). Subsequently, yeast two-hybrid assays were performed on 13 putative interacting proteins with higher coverage, but no interaction was detected between Ppe1 and these proteins (Author response image 1F-G). Considering that the identification and functional validation of interacting proteins is a labor-intensive and time-consuming endeavor, we will focus our future efforts on in-depth studies of Ppe1's function in rice.

Author response image 1.

Screening of Ppe1 candidate targets in rice. (A) The determination of GFP-PPE1 construct in transgenic rice. (B) The expression of PPE1 transgenic rice (T0) was verified by qRT-PCR. (C) Western blot analysis of Ppe1 expression in transgenic rice. (D) Rapid silver staining for detection of the purified proteins captured by the GFP-beads. (E) Venn diagram comparing the number of proteins captured in the different samples. (F) Identity of the potential targets of Ppe1 in rice. (G) Yeast two-hybrid assay showing negative interaction of Ppe1 with rice candidate proteins.

The inability to nail down the function of Ppe1 likely stems from two underlying assumptions with weak support. Firstly, the authors assume that Ppe1 is secreted and associated with the peg to form a penetration ring between the plant cell wall and cytoplasm membrane. However, the authors do not demonstrate it is secreted (for instance by blocking Ppe1 secretion and its association with the peg using brefeldin A).

To investigate the secretion pathway of Ppe1 in M. oryzae, we determined the inhibitory effects of Brefeldin A (BFA) on conventional ER-to-Golgi secretion in fungi as suggested by the reviewer. We inoculated rice leaf sheaths with conidia suspensions from the Ppe1-mCherry and PBV591 strains (containing a Pwl2-mCherry-NLS and Bas4-GFP co-expressing constructs) and treated them with BFA. We found that, even after exposure to BFA for 5 to 11 hours, the Ppe1-mCherry still formed its characteristic ring conformation (Author response image 2). Similarly, in the BFA-treated samples, the cytoplasmic effector Pwl2-mCherry accumulated at the BIC, while the apoplastic effector Bas4-GFP was retained in the invasive hyphae (Author response image 2). These results indicate that Ppe1 is not secreted through the conventional ER-Golgi secretion pathway.

Author response image 2.

The secretion of Ppe1 is not affected by BFA treatment. (A) and (B) The Ppe1-mCherry fluorescent signal was still observed both in the presence and absence of BFA. (C) Following BFA treatment, the secretion of the apoplastic effector Bas4-GFP was blocked while that of the cytoplasmic effector Pwl2-mCherry was not affected. The rice leaf sheath tissue was inoculated with 50 μg/mL BFA (0.1% DMSO) at 17 hpi. Images were captured at 22 hpi for A and 28 hpi for B and C. Scale bars = 10 µm.

Also, they do not sufficiently show that Ppe1 localizes on the periphery of the peg. This is because confocal microscopy is not powerful enough to see the peg. The association they are seeing (for example in Figure 4) shows localization to the bottom of the appressorium and around the primary hyphae, but the peg cannot be seen. Here, the authors will need to use SEM, perhaps in conjunction with gold labeling of Ppe1, to show it is associating with the peg and, indeed, is external to the peg (rather than internal, as a structural role in peg rigidity might predict). It would also be interesting to repeat the microscopy in Figure 4C but at much earlier time points, just as the peg is penetrating but before invasive hyphae have developed - Where is Ppe1 then? Finally, the authors speculate, but do not show, that Ppe1 anchors penetration pegs on the plant cytoplasm membrane. Doing so may require FM4-64 staining, as used in Figure 2 of Kankanala et al, 2007 (DOI: 10.1105/tpc.106.046300), to show connections between Ppe1 and host membranes. Note that the authors also do not show that the penetration ring is a platform for effector delivery, as speculated in the Discussion.

We sincerely appreciate the reviewer's valuable suggestion regarding SEM with immunogold labeling to precisely visualize Ppe1's association with penetration peg. While we fully acknowledge this would be an excellent approach, after consulting several experts in the field, we realized that the specialized equipment and technical expertise required for fungal immunogold-SEM are currently unavailable to us. We sincerely hope that the reviewer will understand this technical limitation.

To further strengthen our evidence for the role of Ppe1's in anchoring penetration peg to the plant plasma membrane, we provided new co-localization images of Ppe1 and penetration peg (Fig. S7). At 16 hours post-inoculation (hpi), when the penetration peg was just forming and prior to the development of invasive hyphae, the Ppe1-mCherry fluorescence forms a tight ring-like structure closely associated with the base of the appressorium. As at 23 hpi, the circular Ppe1-mCherry signal was still detectable beneath the appressorium, and around the penetration peg which differentiated into the primary invasive hyphae. Furthermore, we obtained 3D images of the strain expressing both Ppe1-mCherry and Lifeact-GFP during primary invasive hyphal development. The results revealed that Ppe1 forms a ring-like structure that remains anchored to the penetration peg during fungal invasion (Fig. S6).

We also conducted FM4-64 staining experiment as recommended by the reviewer. Although the experiment provided valuable insights, we found that the resolution was insufficient to precisely delineate the spatial relationship between Ppe1 and host membranes at the penetration peg (Author response image 3). To optimize this colocalization, we tested the localization between Ppe1-mCherry ring and rice plasma membrane marker GFP-OsPIP2 (Fig. S8). These new results provide compelling complementary evidence supporting our conclusion that Ppe1 functions extracellularly at the host-pathogen interface. We hope these additional data will help address the reviewer's concerns regarding Ppe1's localization.

Author response image 3.

FM4-64-stained rice leaf sheath inoculated with M. oryzae strain expressing Ppe1-GFP. Ppe1-GFP ring was positioned above the primary invasive hyphae. Scale bar = 5 µm.

Secondly, the authors assume Ppe1 is required for host infection due to its association with the peg. However, its role in infection is minor. The majority of appressoria produced by the mutant strain penetrate host cells and elaborate invasive hyphae, and lesion sizes are only marginally reduced compared to WT (in fact, the lesion density of the 70-15 WT strain itself seems reduced compared to what would be expected from this strain). The authors did not analyze the lesions for spores to confirm that the mutant strains were non-pathogenic (non-pathogenic mutants sometimes form small pinprick-like lesions that do not sporulate). Thus, the pathogenicity phenotype of the knockout mutant is weak, which could contribute to the inability to accurately define the molecular and cellular function of Ppe1.

We appreciate the reviewer’s comments. To ensure the reliability of our findings, we conducted spray inoculation experiments with multiple independent repeats. Our results consistently demonstrated that deletion of the PPE1 gene significantly attenuates the virulence of M. oryzae. Further analysis of lesion development and sporulation in the Δ_ppe1_ mutant revealed that it retains the ability to produce conidia. To validate these observations, we generated a PPE1 knockout in the wild-type reference strain Guy11. Similarly, we observed a significant decrease in the pathogenicity of the Δ_ppe1_ mutants generated from the wild-type Guy11 strain compared to Guy11 in the spray assay (Fig S2). These results collectively indicate the importance of Ppe1 in the pathogenicity of M. oryzae to rice.

In summary, it is important that the role of Ppe1 in infection be determined.

Reviewer #2 (Public review):

The article focuses on the study of Magnaporthe oryzae, the fungal pathogen responsible for rice blast disease, which poses a significant threat to global food security. The research delves into the infection mechanisms of the pathogen, particularly the role of penetration pegs and the formation of a penetration ring in the invasion process. The study highlights the persistent localization of Ppe1 and its homologs to the penetration ring, suggesting its function as a structural feature that facilitates the transition of penetration pegs into invasive hyphae. The article provides a thorough examination of the infection process of M. oryzae, from the attachment of conidia to the development of appressoria and the formation of invasive hyphae. The discovery of the penetration ring as a structural element that aids in the invasion process is a significant contribution to the understanding of plant-pathogen interactions. The experimental methods are well-documented, allowing for reproducibility and validation of the results.

We sincerely appreciate the thoughtful and insightful evaluation of our work. Thank you for recognizing the significance of our findings regarding the penetration ring and the functional role of Ppe1 during host invasion.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Line 48: "after appressorium- or transpressorium-mediated penetration of plant cell wall" - transpressoria do not penetrate the plant cell wall.

Thank you for your valuable suggestion. For improved clarity, we have rephrased the sentence as follows: In this study, we showed that a penetration ring is formed by penetration pegs after appressorium-mediated penetration of plant cell wall.

Line 143: "approximately 25% of the 143 appressoria formed by the Δppe1 mutant had no penetration peg" - It is not possible to see the penetration peg by confocal microscopy.

Thank you for your valuable suggestion. We have revised the sentence as follows: In contrast, approximately 25% of the appressoria formed by the Δ_ppe1_ mutant had no penetration.

Line 159: "inner cycle" -should be inner circle?

We gratefully acknowledge the reviewer's careful reading. The typographical error has been corrected throughout the revised manuscript.

Line 255: "These results indicate that initiation of penetration peg formation is necessary for the formation of the penetration ring." Actually, more precisely, they indicate that penetration is necessary.

We appreciate this suggestion and have revised the text to be more concise: These results indicate that penetration is necessary for the formation of the penetration ring.

Line 282: "unlike subcellular localizations of other effectors"- is this an effector if no plant targets are known?

We appreciate this suggestion and have revised the text as follows: unlike subcellular localizations of Bas4, Slp1, Pwl2, and AvrPiz-t.

Line 299: "it may function as a novel physical structure for anchoring penetration pegs on the surface of plant cytoplasm membrane after cell wall penetration" - an interaction with the plant plasma membrane was not shown and this is speculative.

We have provided new evidence to show the spatial positioning of Ppe1-mCherry ring with the rice plasma membrane (see figure S8)

Line 301: "It is also possible that this penetration ring functions as a collar or landmark that is associated with the differentiation of penetration pegs (on the surface of cytoplasm membrane) into primary invasive hyphae enveloped in the EIHM cytoplasm membrane (Figure 7)." The alternative conclusions for Ppe1 function, either interacting with host membranes or acting as a developmental landmark, need to be resolved here.

We appreciate this suggestion and have revised the text as follows: It is also possible that this penetration ring functions as a collar that is associated with the differentiation of penetration pegs into primary invasive hyphae enveloped in the EIHM (Figure 7).

Line 317: "is likely a structural feature or component for signaling the transition of penetration pegs to invasive hyphae",- if the authors think Ppe1 has these roles, why do they refer to Ppe1 as an effector?

Many thanks for these comments. We have revised this and refer to Ppe1 as a secreted protein throughout the revised manuscript.

Line 337: "After the penetration of plant cell wall, the penetration ring may not only function as a physical structure but also serve as an initial effector secretion site for the release of specific effectors to overcome plant immunity in early infection stages"- which is it? Also, no evidence is provided to suggest it is a platform for effector secretion.

We sincerely appreciate your valuable suggestion. We have revised this sentence as follows: After the penetration of plant cell wall, the penetration ring may not only function as a physical structure but also serve as a secretion site for the release of specific proteins to overcome plant immunity during the early infection stages.

Reviewer #2 (Recommendations for the authors):

(1) While the study suggests the penetration ring as a structural feature, it remains unclear whether it also serves as a secretion site for effectors. Further exploration of this aspect would strengthen the conclusions.

We thank the reviewer for this useful suggestion. In this study, we demonstrated that Ppe1 proteins form a distinct penetration ring structure at the site where the penetration peg contacts the plant plasma membrane prior to differentiation into primary invasive hyphae (Figs. 2 and 7). Thus, we reasoned that penetration ring may function as a novel physical structure. Notably, additional Ppe family members (Ppe2, Ppe3, and Ppe5) were also found to localize to this penetration ring (Fig. 6B), suggesting that it also serves as a secretion site for releasing proteins. To test whether Ppe1 and Ppe2 label to the same site, we analyzed the colocalization between Ppe1-GFP and Ppe2-mCherry. The results showed that Ppe1-GFP and Ppe2-mCherry are well colocalized (Author response image 4). This study primarily focuses on the discovery and characterization of the penetration ring. The potential role of this structure in effector translocation will be investigated in future studies.

Author response image 4.

Ppe1 co-localizes with Ppe2 at the penetration ring in M. oryzae. Line graphs were generated at the directions pointed by the white arrows. Scale bar = 2μm.

(2) The article could benefit from a discussion on the broader implications of these findings for developing resistant crop varieties or new fungicidal strategies.

We have incorporated this discussion as suggested (lines 358-360).

(3) What is the significance of the formation of the penetration ring in the pathogenicity of the rice blast fungus? Or, how does it assist the fungus in its infection process?

Our findings have several significant implications. First, we believe that the discovery of the penetration ring as a novel physical structure associated with the differentiation of invasive hyphae represents a breakthrough in plant-pathogen interactions that will be of interest to fungal biologists, pathologists and plant biologists. Secondly, our study presents new role of the peg as a specialized platform for secretory protein deployment, in addition to its commonly known role as a physical penetration tool for the pathogen. Thirdly, we identify Ppe1 as a potential molecular target for controlling the devastating rice blast disease, as Ppe homologs are absent in plants and mammals. We have incorporated this discussion in the revised manuscript (lines 354-362).

Author response:

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1 (Public Review):

Strengths:

Overall there are some very interesting results that make an important contribution to the field. Notably, the results seem to point to differential recruitment of the PL-DMS pathway in goal-tracking vs sign-tracking behaviors.

Thank you.

Weaknesses:

There is a lot of missing information and data that should be reported/presented to allow a complete understanding of the findings and what was done. The writing of the manuscript was mostly quite clear, however, there are some specific leaps in logic that require more elaboration, and the focus at the start and end on cholinergic neurons and Parkinson's disease are, at the moment, confusing and require more justification.

In the revised paper, we provide additional graphs and information in support of results, and we further clarify procedures and findings. Furthermore, we expanded the description of the proposed interpretational framework that suggests that the contrasts between the cortical-striatal processing of movement cues in sign- versus goal trackers are related to previously established contrasts between the capacity for the  cortical cholinergic detection of attention-demanding cues.

Reviewer #2 (Public review):

Strengths:

The power of the sign- and goal-tracking model to account for neurobiological and behavioral variability is critically important to the field's understanding of the heterogeneity of the brain in health and disease. The approach and methodology are sound in their contribution to this important effort.

The authors establish behavioral differences, measure a neurobiological correlate of relevance, and then manipulate that correlate in a broader circuitry and show a causal role in behavior that is consistent with neurobiological measurements and phenotypic differences.

Sophisticated analyses provide a compelling description of the authors' observations.

Thank you.

Weaknesses:

It is challenging to assess what is considered the "n" in each analysis (trial, session, rat, trace (averaged across a session or single trial)). Representative glutamate traces (n = 5 traces (out of hundreds of recorded traces)) are used to illustrate a central finding, while more conventional trial-averaged population activity traces are not presented or analyzed. The latter would provide much-needed support for the reported findings and conclusions. Digging deeper into the methods, results, and figure legends, provides some answers to the reader, but much can be done to clarify what each data point represents and, in particular, how each rat contributes to a reported finding (ie. single trial-averaged trace per session for multiple sessions, or dozens of single traces across multiple sessions).

Representative traces should in theory be consistent with population averages within phenotype, and if not, discussion of such inconsistencies would enrich the conclusions drawn from the study. In particular, population traces of the phasic cue response in GT may resemble the representative peak examples, while smaller irregular peaks of ST may be missed in a population average (averaged prolonged elevation) and could serve as a rationale for more sophisticated analyses of peak probability presented subsequently.

We have added two new Tables to clarify the number of rats per phenotype and sex used for each experiment described in the paper (Table 1), and the number of glutamate traces (range, median and total number) extracted for each analysis of performance-associated glutamate levels and the impact of CNO-mediated inhibition of fronto-striatal glutamate (Table 3).

As the timing of glutamate peaks varies between individual traces and subjects, relative to turn and stop cue onset or reward delivery, subject-and trial-averaged glutamate traces would “wash-out” the essential findings of phenotype- and task event-dependent patterns of glutamate peaks. In the detailed responses to the reviewers, we illustrate the results of an analysis of averaged traces to substantiate this view. Furthermore, as detailed in the section on statistical methods, and as mentioned by the reviewer under Strengths, we used advanced statistical methods to assure that data from individual animals contribute equally to the overall result, and to minimize the possibility that an inordinate number of trials obtained from just one or a couple of rats biased the overall analysis.

Reviewer #3 (Public review):

Strengths:

Overall these studies are interesting and are of general relevance to a number of research questions in neurology and psychiatry. The assessment of the intersection of individual differences in cue-related learning strategies with movement-related questions - in this case, cued turning behavior - is an interesting and understudied question. The link between this work and growing notions of corticostriatal control of action selection makes it timely.

Thank you.

Weaknesses:

The clarity of the manuscript could be improved in several places, including in the graphical visualization of data. It is sometimes difficult to interpret the glutamate results, as presented, in the context of specific behavior, for example.

We appreciate the reviewer’s concerns about the complexity of some of the graphics, particularly the results from the arguably innovative analysis illustrated in Figure 6. Figure 6 illustrates that the likelihood of a cued turn can be predicted based on single and combined glutamate peak characteristics. The revised legend for this figure provides additional information and examples to ease the readers’ access to this figure. In addition, as already mentioned above, we have added several graphs to further illustrate our findings.

(Recommendations for the authors)

Reviewer #1 (Recommendations for the authors):

(1) The differences in behavioral phenotype according to vendor (Figure 1c) are slightly concerning, could the authors please elaborate on why they believe this difference is? Are there any other differences in these stocks- i.e. weight, appearance, other types of behaviors?

Differences in PCA behavior across vendors or specific breeding colonies were documented previously and may reflect the impact of environmental, developmental and genetic factors (references added in the revised manuscript). We included animals from both vendors to increase phenotypic variability and due to animal procurement constraints during COVID-related restrictions.

(2) Possibly related to the above, the rats in Figure 1a and Figure 2 are different strains. Please clarify.

In the revised legend of Figure 2 we clarify that the rat shown in the photographs is a Long-Evans rat that was not part of the experiments described in this paper. This rat was used to generate these photos as the black-spotted fur provided better contrast against the white treadmill belt.

(3) Figure 3c, the pairwise comparison showing a significant increase from Day 1 to Day 3 is hard to understand unless this is a lasting change. Is this increase preserved at Day 4? Examination of either a linear trend across days or a simple comparison of either Day 1 & 2 against Day 3 & 4 or, minimally Day 1 against Day 4 would communicate this message. Otherwise, there doesn't seem to be much of a case for improvement across test sessions, which would also be fine in my view.

As the analysis of post-criterion performance also revealed an effect of DAY, we felt compelled to report and illustrate the results of pairwise comparisons in Fig. 3c. In agreement with the reviewer’s point, we did not further comment on this finding in the manuscript.

(4) Figure 4e. I find it extremely unlikely that every included electrode was located exactly at anterior 0.5mm. Please indicate the range - most anterior and most posterior of the included electrodes in the study.

The schematic section shown in Fig. 4e depicted that AP level of that section and collapsed all placements onto that level. As detailed in Methods, electrode placements needed to be within the following stereotaxic space: AP: -0.3 to 0.6 mm, ML: 2 to 2.5 mm, and DV: -4.2 to -5 mm (see Methods). To clarify this issue, the text in Results and the legend was modified and the 0.5 mm label was removed from Fig. 4e.

(5) The paper generally is quite data light and there are a lot of extra results reported that aren't shown in the figures. There are 17 instances of the phrase "not shown", some are certainly justified, but a lot of results are missing…

We followed the reviewer’s suggestion and added several graphs. The revised Figure 5 includes the new graph 5d that shows the number of glutamate traces with just 1, 2 or 3 peaks occurring during cue presentation period. Likewise, the revised Figure 7 includes the new graph 7h that shows the number of glutamate traces with just 1, 2 or 3 peaks following the administration of CNO or its vehicle. In both cases, we also revised the analysis of peak number data, by counting the number of cases (or traces) with just 1, 2 or 3 peaks and using Chi-squared tests to determine the impact of phenotype and, in the latter case, of CNO. In addition, the revised Figure 7 now includes a graph showing the main effects of phenotype and CNO in reward delivery-locked glutamate maximum peak concentrations (Fig. 7k). In revising these sections, we also removed the prior statement about glutamate current rise times as this isolated observation had no impact on subsequent analyses or the discussion.

Concerning the reviewer’s point 5d (DMS eGFP transfection correlations Figure 8), the manuscript clarifies that the absence of such a correlation was expected given that eGFP expression in the DMS does not accurately reproduce the prelimbic-DMS projection space that was inhibited by CNO. In contrast, the correlations between the efficacy of CNO and DREADD expression measures in prelimbic cortex were significant and are graphed (Figs. 8g and 8j).

(6) Please clarify the exact number of animals in each experiment. The caption of Figure 3 seems to suggest there are 29 GTs and 22 STs in the initial experiment, but the caption of Figure 5b seems to suggest there are N=30 total rats being analyzed (leaving 21 un-accounted for), or is this just the number of GTs (meaning there is one extra)?

We have added Table 1 to clarify the number of animals used across different experiments and stages. Additionally, we have included a new Table 3 that identifies, for each graph showing results from the analyses of glutamate concentrations, the number of rats from which recordings were obtained and the number of traces per rat (range, median, and total).

(7) Relatedly, in Figures 5c-f and Figures 7g-i, the data seem to be analyzed by trial rather than subject-averaged, please clarify and what is the justification for this?

As detailed Experimental design and statistical analyses, we employed linear mixed-effects modeling to analyze the amperometric data that generated figures 5 and 7 to minimize the risk of bias due to an excessive number of trials obtained from specific rats. LMMs were chosen to analyze these repeated (non-independent) data to address issues that may be present with subject-averaged data. For clarity, throughout the results for these figures, the numerator in the F-ratio reflects the degrees of freedom from the fixed effects (phenotype/sex) and the denominator reflects the error term influenced by the number of subjects and the within-subject variance.

Concerning the illustration and analysis of trial- or subject-averaged glutamate traces please see reviewer 2, point 1 and the graph in that section. Within a response bin, such as the 2-s period following turn cues, glutamate peaks – as defined in Methods - occur at variable times relative to cue onset. Averaging traces over a population of rats or trials would “wash-out” the phenotype- and task event-dependent patterns of glutamate concentration peaks, yielding, for example, a single, nearly 2-s long plateau for cue-locked glutamate recordings from STs (see Figure 5b versus the graph shown in response to reviewer 2, point 1).

(8) Likewise on page 22, the number of animals from which these trials were taken should be stated "The characteristics of glutamate traces (maximum peak concentration, number of peaks, and time to peak) were extracted from 548 recordings of turn cue trials, 364 of which yielded a turn (GTs: 206, STs: 158) and 184 a miss (GTs: 112, STs: 72).".

The number of animals is now included in the text and listed in Table 3.

(9) The control group for Figure 7 given the mCherry fluorophore - given the known off-target effects of CNO, this is a very important control. Minimally, this data should be shown, but it is troubling that the ST group has n=2, I don't really understand how any sort of sensible stats can be conducted with a group this size, and obviously it's too small to find any significant differences if they were there.

As discussed on p. 14-15 in the manuscript under the section Clozapine N-Oxide, the conversion rate of CNO to clozapine suggests that approximately 50-100 times the dose of clozapine (compared to our 5.0 mg/kg CNO dosage) would be required to produce effects on rodent behavior (references on p. 14-15).

Regarding evidence from control rats expressing the empty construct, the revised manuscript clarifies that no effects of CNO on cued turns were found in 5 GTs expressing the empty control vector. Although CNO had no effects in STs expressing the DREADD, we also tested the effects of CNO in 2 STs expressing the empty control vector (individual turn rates following vehicle and CNO are reported for these 2 STs). Moreover, we extracted turn cue-locked glutamate traces (vehicle: 18 traces; 16 CNO traces) from an empty vector-expressing GT and found that administration of CNO neither reduced maximum glutamate peak concentrations nor the proportion of traces with just one peak. The absence of effects of CNO on cued turning performance and on turn-cue locked glutamate dynamics are consistent with prior studies showing no effects of 5.0 mg/kg CNO in rats not expressing the DREADD vector (references in manuscript).

(10) Figure 8b - the green circle indicated by 1 is definitely not the DMS, this is the DLS, and animals with virus placement in this region should be excluded.

The reviewer of course is correct and that exactly was the point of that illustration, as such a transfection space would have received the lowest possible rating (as indicated by the “1” in the green space). Fig. 8b was intended to illustrate expression efficacy ratings and does not indicate actual viral transfection spaces. Because the results described in the manuscript did not include data from a brain with a striatal transfection space as was illustrated in green in the original Fig. 8b, we removed that illustration of an off-target transfection space.  

(11) Figure 8j, the correlation specifically counts double-labeled PL hM4Di + eGFP neurons. Separating dual-labeled cells from all mCherry-labeled cells seems very strange given the nature of the viral approach. There seems to be an assumption that there are some neurons that express the mCherry-hM4Di that don't also have the AAV-Cre (eGFP). Obviously, if that were true this poses a huge problem for your viral approach and would mean that you're inhibiting a non-selective population of neurons. More likely, the AAV-Cre (eGFP) is present in all of your mCherry-hM4Di cells, just not at levels visible without GFP antibody amplification. Ideally, staining should be done to show that all cells with mCherry also have eGFP, but minimally this correlation should include all cells expressing mCherry with the assumption that they must also have the AAV-Cre.

As noted on page 15 in the Visualization and Quantification of eGFP/mCherry-Expressing Neurons section, eGFP expression in our viral approach was notably bright and did not necessitate signal enhancement. Furthermore, given the topographic organization of prelimbic-DMS projections on the on hand, and the variable transfection spaces in cortex and striatum on the other hand, the speculation that AAV-Cre may have been present in all mCherry cells is without basis. Second, there certainly are mCherry-positive cells that do not also express the retrogradely transported AAV-Cre, and that therefore were not affected by CNO. Third, the entire point of this dual vector strategy was to selectively inhibit prelimbic-striatal projections, and the strong correlation between double-labeled neuron numbers and cued turn scores substantiates the usefulness of this approach.

(12) Discussion, a bit more interpretation of the results would be good. Specifically - does the PL-DMS inhibition convert GTs to STs? There were several instances where the behavior and glutamate signals seemed to be pushed to look like STs but also a lot of missing data so it is hard to say. One would assume this kind of thing if, as I think is being said (please clarify), the ST phenotype is being driven by glutamatergic drive either locally or from sources other than PL cell bodies, presumably silencing the PL cell body inputs in GTs also leaves other glutamatergic inputs as the primary sources?

We agree with the reviewer that one could say, perhaps somewhat colloquially, that PL-DMS inhibition turns GTs to STs, in terms of turning performance and associated glutamate peak dynamics. The newly added data graphs are consistent with this notion. However, there are of course numerous other neurobiological characteristics which differ between GTs and STs and are revealed in the context of other behavioral or physiological functions.  In the Discussion, and as noted by the reviewer, we discuss alternative sources of glutamatergic control in STs and the functional implications of bottom-up mechanisms. In the revised manuscript, we have updated references and made minor revisions to improve this perspective.

(13) I found the abstract really detailed and very dense, it is pretty hard to understand in its current form for someone who hasn't yet read the paper. At this level, I would recommend more emphasis on what the results mean rather than listing the specific findings, given that the task is still quite opaque to the reader.

We revised the abstract, in part by deleting two rather dense but non-essential statements of results and by adding a more accessible conclusion statement.

(14) There are a lot of abbreviations: CTTT, PD, PCA, GT, ST, MEA, GO, LMM, EMMs, PL, DMS. Some of these are only mentioned a few times: MEA, LMM, and EMMs are all mentioned less than 5 times. To reduce mental load for the reader, you could spell these ones out, or include a table somewhere with all of the abbreviations.

We added a list of Abbreviations and Acronyms and eliminated abbreviations that were used infrequently.

(15) Generally, the logic that cortico-striatal connections contribute to GT vs ST seems easy to justify, however, the provided justification is missing a line of connection: "As such biases of GTs and STs were previously shown to be mediated in part via contrasting cholinergic capacities for the detection of cues (Paolone et al., 2013; Koshy Cherian et al., 2017; Pitchers et al., 2017a; Pitchers et al., 2017b), we hypothesized that contrasts in the cortico-striatal processing of movement cues contribute to the expression of these opponent biases." Please elaborate on why specifically cholinergic involvement suggests corticostriatal involvement. I think there are probably more direct reasons for the current hypothesis.

Done – see p. 4-5.

(16) Along the same line, paragraph 3 of the intro about Parkinson's disease and cholinergics seems slightly out of place. This is because the specific or hypothesized link between these things and corticostriatal glutamate has not been made clear. Consider streamlining the message specifically to corticostriatal projections in the context of the function you are investigating.

Done – see p. 4-5.

(17) Page 8, paragraph 2. There is a heading or preceding sentence missing from the start of this paragraph: "Contrary to the acclimation training phase, during which experimenters manually controlled the treadmill, this phase was controlled entirely by custom scripts using Med-PC software and interface (MedAssociates).".

Revised and clarified.

(18) Page 13 "We utilized a pathway-specific dual-vector chemogenetic strategy (e.g., Sherafat et al., 2020) to selectively inhibit the activity of fronto-cortical projections to the DMS". The Hart et al (2018) reference seems more appropriate being both the same pathway and viral combination approach.

Yes, thank you, we’ve updated the citation.

(19) Pages 20-21: "Maximum glutamate peak concentrations recorded during the cue period were significantly higher in GTs than in STs (phenotype: F(1,28.85)= 8.85, P=0.006, ηp 2=0.23; Fig. 5c). In contrast, maximum peak amplitudes locked to other task events all were significantly higher in STs." The wording here is misleading, both Figures 5c and 5d report glutamate peaks during the turn cue, the difference is what the animal does. So, it should be something like "Maximum glutamate peak concentrations recorded during the cue period were significantly higher in GTs than in STs when the animal correctly made a turn (stats) but this pattern reversed on missed trials when the animal failed to turn (stats)..." or something similar.

Yes, thank you. We have revised this section accordingly.  

(20) Same paragraph: "Contingency tables were used to compare phenotype and outcome-specific proportions and to compute the probability for turns in GTs relative to STs." What is an outcome-specific proportion?

This has been clarified.

.

(21) Page 22 typo: "GTs were only 0.74 times as likely as GTs to turn".

Fixed.

(22) The hypothesis for the DREADDs experiment isn't made clear enough. Page 23 "In contrast, in STs, more slowly rising, multiple glutamate release events, as well as the presence of relatively greater reward delivery-locked glutamate release, may have reflected the impact of intra-striatal circuitry and ascending, including dopaminergic, inputs on the excitability of glutamatergic terminals of corticostriatal projections" As far as I can understand, the claim seems to be that glutamate release might be locally modulated in the case of ST, on account of the profile of glutamate release- more slowly rising, multiple events, and reward-locked. Please clarify why these properties would preferentially suggest local modulation.

We have revised and expanded this section to clarify the basis for this hypothesis.

(23) The subheadings for the section related to Figure 7 "CNO disrupts..." "CNO attenuates..." presumably you mean fronto-striatal inhibition disrupts/attenuates. As it stands, it reads like the CNO per se is having these effects, off-target.

Fixed.

(24) The comparison of the results in the discussion against a "hypothetical" results section had the animals not been phenotyped behaviorally is unnecessary and overly speculative, given that 30-40% of rats don't fall into either of these two categories. I think the point here is to emphasize the importance of taking phenotype into account. This point can surely be made directly in its own sentence, probably somewhere towards the end of the discussion).

We have partly followed the reviewer’s advice and separated the discussion of the hypothetical results from the summary of main findings. However, we did not move this discussion toward the end of the Discussion section as we believe that it justifies the guiding focus of the discussion on the impact of phenotype.

(25) The discussion, like the introduction, talks a lot about cholinergic activity. As noted, this link is unclear - particularly how it links with the present results, please clarify or remove. Likewise high-frequency oscillations.

We have revised relevant sections in the Introduction (see above) and Discussion sections. However, given the considerable literature indicating contrasts between the cortical cholinergic-attentional capacities of GTs and STs, the interpretation of the current findings in that larger context is justified.

(26) Typo DSM in the discussion x 2.

Thanks, fixed.

Reviewer #2 (Recommendations for the authors):

(1) As mentioned in the Public Review, it is challenging to assess what is considered the "n" in each analysis, particularly for the glutamate signal analysis (trial, session, rat, trace (averaged across session or single trial)). Representative glutamate traces are used to illustrate a central finding, while more conventional trial-averaged population activity traces are not presented or analyzed. For example, n = 5 traces, out of hundreds of recorded traces, with each rat contributing 1-27 traces across multiple sessions suggests ~1-2% of the data are shown as time-resolved traces. Representative traces should in theory be consistent with population averages within phenotype, and if not, discussion of such inconsistencies would enrich the conclusions drawn from the study. In particular, population traces of the phasic cue response in GT may resemble the representative peak examples, while smaller irregular peaks of ST may be missed in a population average (averaged prolonged elevation in signal) and could serve as rationale for more sophisticated analyses of peak probability presented subsequently (and relevant to opening paragraph of discussion where hypothetical data rationale is presented).

We have added the new Table 1 to provide a complete account of the number of rats, per phenotype and sex, for each component of the experiments. In addition, the new Table 3 provides the range, median and total number of glutamate traces that were analyzed and formed the foundation of the individual data graphs depicting the results of glutamate concentration analyses.

We chose not to present trial- or subject-averaged traces, as glutamate peaks occur at variable times relative to the onset of turn and stop cues and reward delivery, and therefore averaging across a population of rats or trials would obscure phenotype- and task event-dependent patterns of glutamate peaks. The attached graph serves to illustrate this issue. The graph shows turn cue-locked glutamate concentrations (M, SD) from trials that yielded turns, averaged over all traces used for the analysis of the data shown in Fig. 5d (see also Table 3, top row). Because of the variability of peak times, trial- and subject-averaging of traces from STs yielded a nearly 2-s long elevated plateau of glutamate concentrations (red triangles), contrasting with the presence single and multiple peaks in STs as illustrated in Figs. 5b and 5e. Furthermore, averaging of traces from GTs obscured the presence of primarily single turn cue-locked peaks. Because of the relatively large variances of averaged data points, again reflecting the variability of peak times, analysis of glutamate levels during the cue period did not indicate an effect of phenotype (F(1,190)=1.65, P\=0.16). Together, subject- or trial-averaged traces would not convey the glutamate dynamics that form the essence of the amperometric findings obtained from our study. We recognize, as inferred by the reviewer, that smaller irregular peaks in STs may have been missed given the definition of a glutamate peak (see Methods). It is in part for that reason that we conducted a prospective analysis of the probability for turns given a combination of peak characteristics (maximum peak concentration and peak numbers; Fig. 6).

(2)To this latter point, the relationship between the likelihood to turn and the size of glutamate peak is focused on the GT phenotype, which limits understanding of how smaller multiple peaks relate to variables of interest in ST (missed turns, stops, reward). If it were possible to determine the likelihood for each phenotype, without a direct contrast of one phenotype relative to the other, this would be a more straightforward description of how signal frequency and amplitude relate to relevant behaviors in each group. Depending on the results, this could be done in addition to or instead of the current analysis in Figure 6.

We considered the reviewer’s suggestion but could not see how attempts to analyze the role of maximum glutamate concentrations and number of peaks within a single phenotype would provide any significant insights beyond the current description of results. Moreover, as stressed in the 2nd paragraph of the Discussion (see Reviewer 1, point 24), the removal of the phenotype comparison would nearly completely abolish the relationships between glutamate dynamics and behavior from the current data set.

Author response image 1.

(3) If Figure 6 is kept, a point made in the text is that GT is 1.002x more likely than ST to turn at a given magnitude of Glu signal. 1.002 x more likely is easily (perhaps mistakenly) interpreted as nearly identical likelihood. Looking closely at the data, perhaps what is meant is @ >4uM the difference between top-line labeled {b} and bottom-line labeled {d,e} is 1.002? If not, there may be a better way to describe the difference as 1x could be interpreted as the same/similar.

Concerning the potential for misinterpretation, the original manuscript stated (key phrase marked here in red font): Comparing the relative turn probabilities at maximum peak concentrations >4 µM, GTs were 1.002 times more likely (or nearly exactly twice as likely) as STs to turn if the number of cue-evoked glutamate peaks was limited to one (rhombi in Fig. 6a)  when compared to the presence of 2 or 3 peaks (triangles in Fig. 6a). However, we appreciate the reviewer’s concern about the complexity of this statement and, as it merely re-emphasized a result already described, it was deleted.

(4) For Figure 7e, the phenotype x day interaction is reported, but posthocs are looking within phenotype (GT) at treatment effects. Is there a phenotype x day x treatment, or simply phenotype x treatment (day collapsed) to justify within-group treatment posthocs?

We have revised the analysis and illustration of the data shown in Figs 7e and 7f, by averaging the test scores from the two tests, per animal, of the effects of vehicle and CNO, to be able to conduct a simpler 2-way analysis of the effects of phenotype and treatment.

(5) Ideally, viral control is included as a factor in this analysis as well. The separate analysis for viral controls was likely done due to low n, however negative findings from an ANOVA in which an n=2 (ST) should be interpreted with extreme caution. The authors already have treatment control (veh, CNO) and may consider dropping the viral controls completely due to the lack of power to perform appropriate analyses.

This issue has been clarified – see reviewer 1, point 9.

Minor:

(1) In the task description, it could be clearer how reward delivery relates to turns and stops. For example, does the turn cue indicate the rat will be rewarded at the port behind it? Does the stop cue indicate that the rat will be rewarded at the port in front of it? This makes logical sense, but the current text does not describe the task in this way, instead focusing on what is the correct action (seemingly but unlikely independent of reinforcement).

We have updated the task description in Methods and the legend of Figure 2 to indicate the location of reward delivery following turns and stops.

(2) For the peak analysis, what is the bin size for determining peaks? It is indicated that the value before and after the peak is >1 SD below the peak value, so it is helpful to know the temporal bin resolution for this definition.

As detailed on p 11-12 under Amperometry Data Processing and Analysis of Glutamate Peaks, we analyzed glutamate concentrations recorded at a frequency of 5 Hz (200 ms bins) throughout the 2-second-long presentation of turn and stop cues and for a 2-second period following reward delivery.

(3) Long Evans rats are pictured in Figure 2 (presumably contrast with a white background is better here), while SD rats are pictured in Figure 1. Perhaps stating why LE rats are pictured would help clear up any ambiguity about the strains used, as a quick look gives the impression two strains are used in two different tasks.

Yes, see reviewer 1, point 2.

(4) In Figure 7e, the ST and GT difference in turns/turn cue does not seem to replicate prior findings for tracking differences for this measure (Figure 3b). ST from the chemogenetic cohort seems to perform better than rats whose behavior was examined prior to glutamate sensor insertion. What accounts for this difference? Training and testing conditions/parameters?

The reviewer is correct. The absence of a significant difference between vehicle-treated GTs and vehicle-treated STs in Fig. 7e reflects a relatively lower turn rate in GTs than was seen in the analysis of baseline behavior (Fig. 3b; note the different ordinates of the two figures, needed to show the impact of CNO in Fig. 7e). Notably, the data in Fig. 7e are based on fewer rats (12 versus 29 GTs and 10 versus 22 STs; Table 1) and on rats which at this point had undergone additional surgeries to infuse the DREADD construct and implant electrode arrays. We can only speculate that these surgeries had greater detrimental effects in GTs, perhaps consistent with evidence suggesting that immune challenges trigger a relatively greater activation of their innate immune system (Carmen et al., 2023). We acknowledged this issue in the revised Results.

(5) The authors are encouraged to revise for grammar (are vs. is, sentence ending with a preposition, "not only" clause standing alone) and word choice (i.e. in introduction: insert, import, auditorily). Consider revising the opening sentence on page 5 for clarity.

We have revised the entire text to improve grammar and word choice.

(6) Do PD fallers refer to rats or humans? if the latter, this may be a somewhat stigmatizing word choice.

We have replaced such phrases using more neutral descriptions, such as referring to people with PD who frequently experience falls.

(7) Page 27 What does "non-instrumental" behavior mean?

We have re-phrased this statement without using this term.

(8) The opening paragraph of the discussion is focused on comparing reported results (with phenotype as a factor) to a hypothetical description of results (without phenotype as a factor) that were not presented in the results section. There is one reference to a correlation analysis on collapsed data, but otherwise, no reporting of data overall rats without phenotype as a factor. If this is a main focus, including these analyses in the results would be warranted. If this is only a minor point leading to discussion, authors could consider omitting the hypothetical comparison.

We have revised this section - see reviewer 1 point 24.

Reviewer #3 (Recommendations for the authors):

(1) These are really interesting studies. I think there are issues in data presentation/analysis that make it difficult to parse what exactly is happening in the glutamate signals, and when. Overall the paper is just a bit of a difficult read. A generally standard approach for showing neural recording data of many kinds, including, for example, subject-averaged traces, peri-event histograms, heatmaps, etc summarizing and quantifying the results - would be helpful. Beyond the examples in Figure 5, I would suggest including averaged traces of the glutamate signals and quantification of those traces.

We have addressed these issues in multiple ways, see the response to several points of reviewers 1 and 2, particularly reviewer 2, point 1.

(2) Figure 6 (and the description in the response letter) is also very non-intuitive. It's unclear how the examples shown relate to the reported significance indicators/labels/colors etc in the figure. I would suggest rethinking this figure overall, and if there is a more direct quantitative way to connect signal features with behavior. Again, drawing from standard visualization approaches for neural data could be one approach.

See also reviewer 2 points 1 and 3. Furthermore, we have revised the text in Results and the legend to improve the accessibility of Fig. 6.

(3) As far as I can tell, all of the glutamate sensor conclusions reflect analysis collapsed across 100s of trials. Do any of the patterns hold for a subjects-wise analysis? How variable are individual subjects?

We employed linear mixed-effect model analyses and added a random subject intercept to account for subject variability outside fixed effects (phenotype and treatment). The variance of the intercept ranged 0.01-1.71 SEM across outcome (cued turns/cued stops/misses). See also reviewer 1, point 7 and reviewer 2, point 1.

Author Response

The following is the authors’ response to the original reviews.

eLife assessment

This important study reports a novel measurement for the chemotactic response to potassium by Escherichia coli. The authors convincingly demonstrate that these bacteria exhibit an attractant response to potassium and connect this to changes in intracellular pH level. However, some experimental results are incomplete, with additional controls/alternate measurements required to support the conclusions. The work will be of interest to those studying bacterial signalling and response to environmental cues.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

This paper shows that E. coli exhibits a chemotactic response to potassium by measuring both the motor response (using a bead assay) and the intracellular signaling response (CheY phosporylation level via FRET) to step changes in potassium concentration. They find increase in potassium concentration induces a considerable attractant response, with an amplitude larger than aspartate, and cells can quickly adapt (but possibly imperfectly). The authors propose that the mechanism for potassium response is through modifying intracellular pH; they find both that potassium modifies pH and other pH modifiers induce similar attractant responses. It is also shown, using Tar- and Tsr-only mutants, that these two chemoreceptors respond to potassium differently. Tsr has a standard attractant response, while Tar has a biphasic response (repellent-like then attractant-like). Finally, the authors use computer simulations to study the swimming response of cells to a periodic potassium signal secreted from a biofilm and find a phase delay that depends on the period of oscillation.

Strengths:

The finding that E. coli can sense and adapt to potassium signals and the connection to intracellular pH is quite interesting and this work should stimulate future experimental and theoretical studies regarding the microscopic mechanisms governing this response. The evidence (from both the bead assay and FRET) that potassium induces an attractant response is convincing, as is the proposed mechanism involving modification of intracellular pH.

Weaknesses:

The authors show that changes in pH impact fluorescent protein brightness and modify the FRET signal; this measurement explains the apparent imprecise adaptation they measured. However, this effect reduces confidence in the quantitative accuracy of the FRET measurements. For example, part of the potassium response curve (Fig. 4B) can be attributed to chemotactic response and part comes from the pH modifying the FRET signal. Measuring the full potassium response curve of the no-receptor mutants as a control would help quantify the true magnitude of the chemotactic response and the adaptation precision to potassium.

Response: We thank the reviewer for the suggestion. We have now measured the full potassium response curve for the no-receptor mutant (HCB1414-pVS88), as shown in Fig. S4. We characterized the pH effects on CFP and YFP channels at different concentrations of KCl, and the relationship between the ratio of the signal post- to pre-KCl addition and the KCl concentration was established for both channels, as shown in Fig. S4C. The pH-corrected signal after KCl addition for strains with receptors was obtained by dividing the original signal after KCl addition by this ratio at the specific KCl concentration. This was done for both CFP and YFP channels. The pH-corrected responses for the Tar-only and Tsr-only strains are represented by red dots in Fig. 5BC. The recalculated response curve and adaptation curve for the wild-type strain are shown in Fig. S5. The same correction was applied to Fig. 3 as well. We also re-performed the simulations using the corrected dose-response curve and replotted Fig. 6, though the simulation results did not change much.

We have now added a subsection “Revised FRET responses by correcting the pH effects on the brightness of eCFP and eYFP” at line 296 in “Results” to describe this.

The measured response may also be impacted by adaptation. For other strong attractant stimuli, the response typically shows a low plateau before it recovers (adapts). However, in the case of Potassium, the FRET signal does not have an obvious plateau following the stimuli. Do the authors have an explanation for that? One possibility is that the cells may have already partially adapted when the response reaches its minimum, which could indicate a different response and/or adaptation dynamics from that of a regular chemo-attractant? In any case, directly measuring the response to potassium in mutants without adaptation enzymes (CheR, CheB) and with the receptors in different methylation levels would shed more light on the problem.

Response: We appreciate the reviewer’s insightful questions. To observe the low plateau before adaptation, a saturating amount of attractant should be added in a stepwise manner. According to the dose-response curve we measured for potassium, a saturating amount of potassium would be close to 100 mM. In fact, there is a small segment of the low plateau in the step response to 30 mM KCl (Fig. 4C or Fig. S5A). To observe more of this low plateau, we could have used a higher concentration of KCl. However, a stimulation higher than 30 mM KCl will induce substantial physiological changes in the cell, resulting in a significant decrease in fluorescence for both channels (Fig. S7). Therefore, the range of KCl concentration that can be reliably applied in FRET measurements is limited.

The half-time of adaptation at 30 mM KCl was measured to be approximately 80 s, demonstrating a faster adaptation than 0.1 mM MeAsp, which induced a similar magnitude of response. Nevertheless, this is still significantly slower than the time required for medium exchange in the flow chamber, which takes less than 10 s to replace 99% of the medium. Thus, the effect on the measured response magnitude due to adaptation should be small (less than 10%).

We thank the reviewer for the suggestion of measuring the response to potassium in mutants without adaptation enzymes (CheR, CheB) and with the receptors in different methylation levels. However, these mutants are typically less sensitive than the wild-type, exhibiting higher values of K0.5 (Sourjik & Berg, PNAS 99:123, 2002), and thus require an even higher KCl concentration to see the low plateau. Consistent with this, we attempted to measure the response to potassium in a cheRcheB mutant (HCB1382-pVS88). As shown in Fig. R1 below, there is no response to up to 30 mM KCl, suggesting that the sensitive region of the mutant is beyond 30 mM KCl.

The relevant text was added at line 413-424.

Author response image 1.

The response of the cheRcheB mutant (HCB1382-pVS88) to different concentrations of KCl. The blue solid line denotes the original signal, while the red dots represent the pH-corrected signal. The vertical purple (green) dashed lines indicate the moment of adding (removing) 0.01 mM, 0.1 mM, 0.3 mM, 1 mM, 3 mM, 10 mM and 30 mM KCl, in chronological order.

There seems to be an inconsistency between the FRET and bead assay measurements, the CW bias shows over-adaptation, while the FRET measurement does not.

Response: We thank the reviewer for pointing this out. We have now demonstrated that the imprecise adaptation shown in the FRET assay primarily resulted from the pH-induced intensity change of the fluorescent proteins. As shown in Fig. S5A&C, the FRET signal also shows over-adaptation, similar to the bead assay, when we recalculated the response by correcting the CFP and YFP channels.

Now we clarified it at line 315.

The small hill coefficient of the potassium response curve and the biphasic response of the Tar-only strain, while both very interesting, require further explanation since these are quite different than responses to more conventional chemoattractants.

Response: We thank the reviewer for pointing this out. We have now recalculated the pH-corrected results for the dose-response curve (Fig. S5) and the biphasic response of the Tar-only strain (Fig. 5C). The new Hill coefficient is 0.880.14 (meanSD), which is close to the response to MeAsp (1.2) (ref. 46). We suspected that this Hill coefficient of slightly less than 1 resulted from the different responses of Tar and Tsr receptors to potassium.

The Tar-only strain exhibits a repellent response to stepwise addition of low concentrations of potassium less than 10 mM, and a biphasic response above (Fig. 5C). This biphasic response might result from additional pH-effects on the activity of intracellular enzymes such as CheRB and CheA, which may have a different timescale and response from the Tar receptor. We have now added the penultimate paragraph in “Discussion” to talk about the response of the Tar-only strain.

Reviewer #2 (Public Review):

Summary:

Zhang et al investigated the biophysical mechanism of potassium-mediated chemotactic behavior in E coli. Previously, it was reported by Humphries et al that the potassium waves from oscillating B subtilis biofilm attract P aeruginosa through chemotactic behavior of motile P aeruginosa cells. It was proposed that K+ waves alter PMF of P aeruginosa. However, the mechanism was this behaviour was not elusive. In this study, Zhang et al demonstrated that motile E coli cells accumulate in regions of high potassium levels. They found that this behavior is likely resulting from the chemotaxis signalling pathway, mediated by an elevation of intracellular pH. Overall, a solid body of evidence is provided to support the claims. However, the impacts of pH on the fluorescence proteins need to be better evaluated. In its current form, the evidence is insufficient to say that the fluoresce intensity ratio results from FRET. It may well be an artefact of pH change. Nevertheless, this is an important piece of work. The text is well written, with a good balance of background information to help the reader follow the questions investigated in this research work.

In my view, the effect of pH on the FRET between CheY-eYFP and CheZ-eCFP is not fully examined. The authors demonstrated in Fig. S3 that CFP intensity itself changes by KCl, likely due to pH. They showed that CFP itself is affected by pH. This result raises a question of whether the FRET data in Fig3-5 could result from the intensity changes of FPs, but not FRET. The measured dynamics may have nothing to do with the interaction between CheY and CheZ. It should be noted that CFP and YFP have different sensitivities to pH. So, the measurement is likely confounded by the change in intracellular pH. Without further experiments to evaluate the effect of pH on CFP and YFP, the data using this FRET pair is inconclusive.

Response: We thank the reviewer for pointing this out. We have now measured the full potassium response curve for the no-receptor mutant (HCB1414-pVS88), as shown in Fig. S4. We characterized the pH effects on CFP and YFP channels at different concentrations of KCl, and the relationship between the ratio of the signal post- to pre-KCl addition and the KCl concentration was established for both channels, as shown in Fig. S4C. The pH-corrected signal after KCl addition for strains with receptors was obtained by dividing the original signal after KCl addition by this ratio at the specific KCl concentration. This was done for both CFP and YFP channels. The pH-corrected responses for the Tar-only and Tsr-only strains are represented by red dots in Fig. 5BC. The recalculated response curve and adaptation curve for the wild-type strain are shown in Fig. S5. The same correction was applied to Fig. 3 as well. We also re-performed the simulations using the corrected dose-response curve and replotted Fig. 6, though the simulation results did not change much.

We have now added a subsection “Revised FRET responses by correcting the pH effects on the brightness of eCFP and eYFP” at line 296 in “Results” to describe this.

The data in Figure 1 is convincing. It would be helpful to include example videos. There is also ambiguity in the method section for this experiment. It states 100mM KCl was flown to the source channel. However, it is not clear if 100 mM KCl was prepared in water or in the potassium-depleted motility buffer. If KCl was prepared with water, there would be a gradient of other chemicals in the buffer, which confound the data.

Response: We apologize for the ambiguity. The KCl solution used in this work was prepared in the potassium-depleted motility buffer. We have now clarified this at both lines 116 and 497. We now provided an example video, Movie S1, with the relevant text added at line 123.

The authors show that the FRET data with both KCl and K2SO4, and concluded that the chemotactic response mainly resulted from potassium ions. However, this was only measured by FRET. It would be more convincing if the motility assay in Fig1 is also performed with K2SO4.

Response: We thank the reviewer for the suggestion. The aim of comparing the responses to KCl and K2SO4 was to determine the role of chloride ions in the response and to prove that the chemotactic response of E. coli to KCl comes primarily from its response to potassium ions. It is more sensitive to compare the responses to KCl and K2SO4 by using the FRET assay. In contrast, the microfluidic motility assay is less sensitive in revealing the difference in the chemotactic responses, making it difficult to determine the potential role of chloride ions.

Methods:

• Please clarify the promotes used for the constitutive expression of FliCsticky and LacI.

Response: The promoters used for the constitutive expression of LacIq and FliCsticky were the Iq promoter and the native promoter of fliC, respectively (ref. 57).

Now these have been clarified at line 471.

• Fluorescence filters and imaging conditions (exposure time, light intensity) are missing.

Response: Thank you for the suggestion. We have now added more descriptions at lines 535-546: The FRET setup was based on a Nikon Ti-E microscope equipped with a 40× 0.60 NA objective. The illumination light was provided by a 130-W mercury lamp, attenuated by a factor of 1024 with neutral density filters, and passed through an excitation bandpass filter (FF02-438/24-25, Semrock) and a dichroic mirror (FF458-Di02-25x36, Semrock). The epifluorescent emission was split into cyan and yellow channels by a second dichroic mirror (FF509-FDi01-25x36, Semrock). The signals in the two channels were then filtered by two emission bandpass filters (FF01-483/32-25 and FF01-542/32-25, Semrock) and collected by two photon-counting photomultipliers (H7421-40, Hamamatsu, Hamamatsu City, Japan), respectively. Signals from the two photomultipliers were recorded at a sampling rate of 1 Hz using a data-acquisition card installed in a computer (USB-1901(G)-1020, ADlink, New Taipei, Taiwan).

• Please clarify if the temperature was controlled in motility assays.

Response: All measurements in our work were performed at 23 ℃. It was clarified at line 496.

• L513. It is not clear how theta was selected. Was theta set to be between 0 and pi? If not, P(theta) can be negative?

Response: The θ was set to be between 0 and π. This has now been added at line 581.

• Typo in L442 (and) and L519 (Koff)

Response: Thank you. Corrected.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

(1) From the motor measurements the authors find that the CW bias over-adapts to a level larger than prestimulus, but this is not seen in the FRET measurements. What causes this inconsistency? Fig. 2D seems to rule out any change in CheY binding to the motor.

Response: We thank the reviewer for pointing this out. We have now demonstrated that the imprecise adaptation shown in the FRET assay primarily resulted from the pH-induced intensity change of the fluorescent proteins. As shown in Fig. S5A&C, the FRET signal also shows over-adaptation, similar to the bead assay, when we recalculated the response by correcting the CFP and YFP channels.

We now clarified it at line 315.

(2) It would be useful to compare the response amplitude for potassium (Fig. 3C) to a large concentration of both MeAsp and serine. This is a fairer comparison since your work shows potassium acts on both Tar and Tsr. Alternatively, testing a much larger concentration (~10^6 micromolar) at which MeAsp also binds to Tsr would also be useful.

Response: We thank the reviewer for pointing this out. We have now recalculated the response to potassium by correcting the pH-induced effects on fluorescence intensity of CFP and YFP. The response to 30 mM KCl was 1.060.10 times as large as that to 100 μM MeAsp. The aim of the comparison between the responses to potassium and MeAsp was to provide an idea of the magnitude of the chemotactic response to potassium. The stimulus of 100 μM MeAsp is already a saturating amount of attractant and induces zero-kinase activity, thus using a higher stimulus (adding serine or a larger concentration of MeAsp) is probably not needed. Moreover, a larger concentration (~10^6 micromolar) of MeAsp would also induce an osmotactic response.

(3) The fitted Hill coefficient (~0.5) to the FRET response curve is quite small and the authors suggest this indicates negative cooperativity. Do they have a proposed mechanism for negative cooperativity? Have similar coefficients been measured for other responses?

Response: We thank the reviewer for pointing this out. We have now recalculated the pH-corrected results for the dose-response curve (Fig. S5). The new Hill coefficient is 0.880.14 (meanSD), which is close to the response to MeAsp (1.2) (ref. 46). We suspect that this Hill coefficient of slightly less than 1 results from the differing responses of Tar and Tsr receptors to potassium.

(3a) The authors state a few times that the response to potassium is "very sensitive", but the low Hill coefficient indicates that the response is not very sensitive (at least compared to aspartate and serine responses).

Response: We apologize for the confusion. We described the response to potassium as “very sensitive” due to the small value of K0.5. This has now been clarified at line 236.

(3b) Since the measurements are performed in wild-type cells the response amplitude following the addition of potassium may be biased if the cell has already partially adapted. This seems to be the case since the FRET time series does not plateau after the addition of the stimulus. The accuracy of the response curve and hill coefficient would be more convincing if the experiment was repeated with a cheR cheB deficient mutant.

Response: We thank the reviewer for raising these questions. To observe the low plateau before adaptation, a saturating amount of attractant should be added in a stepwise manner. According to the dose-response curve we measured for potassium, a saturating amount of potassium would be close to 100 mM. In fact, there is a small segment of the low plateau in the step response to 30 mM KCl (Fig. 4C or Fig. S5A). To observe more of this low plateau, we could have used a higher concentration of KCl. However, a stimulation higher than 30 mM KCl will induce substantial physiological changes in the cell, resulting in a significant decrease in fluorescence for both channels (Fig. S7). Therefore, the range of KCl concentration that can be reliably applied in FRET measurements is limited.

The half-time of adaptation at 30 mM KCl was measured to be approximately 80 s, demonstrating a faster adaptation than 0.1 mM MeAsp, which induced a similar magnitude of response. Nevertheless, this is still significantly slower than the time required for medium exchange in the flow chamber, which takes less than 10 s to replace 99% of the medium. Thus, the effect on the measured response magnitude due to adaptation should be small (less than 10%).

We thank the reviewer for the suggestion of measuring the response to potassium in mutants without adaptation enzymes (CheR, CheB) and with the receptors in different methylation levels. However, these mutants are typically less sensitive than the wild-type, exhibiting higher values of K0.5 (ref. 46), and thus require an even higher KCl concentration to see the low plateau. Consistent with this, we attempted to measure the response to potassium in a cheRcheB mutant (HCB1382-pVS88). As shown in Fig. R1, there is no response to up to 30 mM KCl, suggesting that the sensitive region of the mutant is beyond 30 mM KCl.

The relevant text was added at line 413-424.

(4) The authors show that the measured imprecise adaptation can be (at least partially) attributed to pH impacting the FRET signal by changing eCFP and eYFP brightness.

(4a) Comparing Fig. 5C and D, the chemosensing and pH response time scales look similar. Therefore, does the pH effect bias the measured response amplitude (just as it biases the adapted FRET level)?

Response: We agree with the reviewer that the pH effect on CFP and YFP biases the measured response amplitude. We have now performed the measurement of dose-response curve to potassium for the no-receptor mutant (HCB1414-pVS88), as shown in Fig. S4. The pH effects on CFP and YFP were corrected. The dose-response curve and adaptation curve were recalculated and plotted in Fig. S5.

(4b) It would help to measure a full response curve (at many concentrations) for the no-receptor strain as a control. This would help distinguish, as a function of concentration, how much response can be attributed to pH impacting the FRET signal versus the true chemotactic response.

Response: We thank the reviewer for the suggestion. We have now performed the measurements for the no-receptor strain. The impact of pH on CFP and YFP has been corrected. The pH-corrected results, previously in Fig.3-5, are now presented in Fig. 3, Fig. S5 and Fig. 5, respectively.

(5) The biphasic response of Tar is strange and warrants further discussion. Do the authors have any proposed mechanisms that lead to this behavior? For the 10mM and 30mM KCl measurements there is a repellent response followed by an attractant response for both adding and removing the stimuli, why is this?

Response: We thank the reviewer for pointing this out. The Tar-only strain exhibits a repellent response to stepwise addition of low concentrations of potassium less than 10 mM, and a biphasic response above (Fig. 5C). This biphasic response might result from additional pH-effects on the activity of intracellular enzymes such as CheRB and CheA, which may have a different timescale and response from the Tar receptor. We have now added the penultimate paragraph in “Discussion” to talk about the response of the Tar-only strain.

(5a) The fact that Tar and Tsr are both attractant (after the initial repellant response in Tar) appears to be inconsistent with previous work on pH response (Ref 52, Yang and Sourjik Molecular Microbiology (2012) 86(6), 1482-1489). This study also didn't see any biphasic response.

Response: We thank the reviewer for pointing this out. The Tar-only strain shows a repellent response to stepwise addition of low concentrations of potassium, specifically less than 10 mM. This is consistent with previous observations of the response of Tar to changes in intracellular pH (refs. 44,45) and also with the work of Yang and Sourjik (new ref. 53), although the work in ref. 53 dealt with the response to external pH change, and bacteria were known to maintain a relatively stable intracellular pH when external pH changes (Chen & Berg, Biophysical Journal (2000) 78:2280-2284). Interestingly, the Tar-only strain exhibits a biphasic response to high potassium concentrations of 10 mM and above. This biphasic response might result from additional pH-effects on the activity of intracellular enzymes such as CheRB and CheA (ref. 56), which may have a different timescale and response from the Tar receptor. We have now added the penultimate paragraph in “Discussion” to talk about the response of the Tar-only strain.

(5b) The response of Tar to the removal of sodium benzoate (Fig. S2) seems to be triphasic, is there any explanation for this?

Response: We thank the reviewer for pointing this out. We have now acknowledged in the legend of Fig. S2 that this response is interesting and warrants further exploration: “The response to the removal of sodium benzoate seems to be a superposition of an attractant and a repellent response, the reason for which deserves to be further explored.”

(6) Fitting the MWC model leads to N=0.35<1. It is fine to use this as a phenomenological parameter, but can the authors comment on what might be causing such a small effective cluster size for potassium response?

Response: We thank the reviewer for pointing this out. We have now recalculated the pH-corrected results for the dose-response curve (Fig. S5). The new Hill coefficient is 0.880.14 (meanSD), which is close to the response to MeAsp (1.2) (ref. 46). We now refit the MWC model to the pH-corrected dose-response curve, obtaining N of 0.85. We think the small N is due partly to the fact that we are fitting the curve with four parameters: N, Kon, Koff, and fm, while only three features of the sigmoid does-response curve are relevant (the vertical scale, the midpoint concentration, and the slope of the sigmoid). Future experiments may determine these parameters more accurately, but they should not significantly affect the simulation results as long as the wild-type dose-response curve is accurate.

(7) The results of the modeling are closely related to Zhu et. al. Phys. Rev. Lett. 108, 128101. Is the lag time for large T related to the adaptation time?

Response: We thank the reviewer for pointing this out. We used a similar framework of modeling as Zhu et. al. The potassium response was also analogous to the chemotactic response to MeAsp. Thus, the results are closely related to Zhu et al. We have now cited Zhu et al. (Ref. 52) and noted this at line 366.

The lag time for large T is related to the adaptation time. We have now simulated the chemotaxis to potassium for large T with different adaptation time by varying the methylation rate kR. The results are shown in Fig. S8. The simulated lag time decreases with the methylation rate kR, but levels off at high values of kR. Now this has been added at line 603.

Minor issues:

• Fig. 1C: should the axis label be y?

Response: Yes, thank you. Now corrected.

• Line 519: Koff given twice, the second should be Kon.

Response: Thank you. Corrected.

• When fitting the MWC model (Eq. 3 and Fig. 6B) did you fix a particular value for m?

Response: m was treated as a fitting parameter, grouped in the parameter fm.

Reviewer #2 (Recommendations For The Authors):

Minor points: - I suggest explaining the acronyms when they first appear in the text (eg CMC, CW, CCW).

Response: Thank you. Now they have been added.

• L144. L242. "decrease" is ambiguous since membrane potential is negative. I understand the authors meant less negative (which is an increase). I suggest to avoid this expression.

Response: Thank you for the suggestion. Now they have been replaced by “The absolute value of the transmembrane electrical potential will decrease”.

• For Fig 1b - it says the shaded area is SEM in the text, but SD in the legend. Please clarify.

Response: Thank you. The annotation in the legend has now been revised as SEM.

• Fig 1C label of x axis should be "y" instead of "x" to be consistent with Fig 1A.

Response: Thank you. It has now been revised.

• In Figure 2, the number of independent experiments as well as the number of samples should be included.

Response: Thank you. The response in Fig. 2C is the average of 83 motors from 5 samples for wild-type strain (JY26-pKAF131). The response in Fig. 2D is the average of 22 motors from 4 samples for the chemotaxis-defective strain (HCB901-pBES38). They have now been added to the legend.

• Regarding the attractant or repelling action of potassium and sucrose, it would be important to have a move showing the cells' behaviours.

Response: We thank the reviewer for the suggestion. We have now provided Movie S1 to show the cells’ behavior to potassium. As shown in Fig. 3B, the chemotactic response to 60 mM sucrose is very small compared to the response to 30 mM KCl. This implies that a noticeable response to sucrose necessitates higher concentrations of stimulation. However, Jerko et al. [Rosko, J., Martinez, V. A., Poon, W. C. K. & Pilizota, T. Proc. Natl Acad. Sci. USA 114, E7969-E7976 (2017).] have shown that high concentrations of sucrose lead to a significant reduction in the speed of the flagella motor. Thus, in a motility assay for sucrose, the osmolarity-induced motility effect may overwhelm the minor repellent-like response.

Author response:

The following is the authors’ response to the original reviews.

eLife Assessment

This well-written report uses functional neuroimaging in human observers to provide convincing evidence that activity in the early visual cortex is suppressed at locations that are frequently occupied by a task-irrelevant but salient item. This suppression appears to be general to any kind of stimulus, and also occurs in advance of any item actually appearing. The work in its present form will be valuable to those examining attention, perception, learning and prediction, but with a few additional analyses could more informatively rule out potential alternative hypotheses. Further discussion of the mechanistic implications could clarify further the broad extent of its significance. 

We thank the editor and the reviewers for the positive evaluation of our manuscript and the thoughtful comments. Below we provide a detailed point-by-point reply to the reviewers’ comments.

In addition to addressing the reviewers' comments, we have improved the figure legends by explicitly describing the type of error bars depicted in the figures, information which was previously only listed in the Materials and Methods section. Specifically, the statement: “Error bars denote within-subject SEM” was added to several figures, as applicable. We believe that briefly reiterating this information in the figure legends enhances clarity and enables readers to interpret the results more accurately and efficiently. We also updated our code and data sharing statement, as well as opened the repository for the public: “Analysis and experiment code, as well as data required to replicate the results reported in this manuscript are available here: https://doi.org/10.17605/OSF.IO/G4RXV. Raw MRI data is available upon request.”

Public Reviews

Reviewer #1 (Public review): 

Summary: 

The authors investigated if/how distractor suppression derived from statistical learning may be implemented in early visual cortex. While in a scanner, participants conducted a standard additional singleton task in which one location more frequently contained a salient distractor. The results showed that activity in EVC was suppressed for the location of the salient distractor as well as for neighbouring neutral locations. This suppression was not stimulus specific - meaning it occurred equally for distractors, targets and neutral items - and it was even present in trials in which the search display was omitted. Generally, the paper was clear, the experiment was well-designed, and the data are interesting. Nevertheless, I do have several concerns mostly regarding the interpretation of the results. 

(1) My biggest concern with the study is regarding the interpretation of some of the results. Specifically, regarding the dynamics of the suppression. I appreciate that there are some limitations with what you might be able to say here given the method but I do feel as if you have committed to a single interpretation where others might still be at play. Below I've listed a few alternatives to consider. 

We agree with the reviewer that there are important alternatives to consider. Adequately addressing these alternatives will substantially increase the inferences we can draw from our data. Therefore, we address each alternative interpretation in detail below.

(a) Sustained Suppression. I was wondering if there is anything in your results that would speak for or against the suppression being task specific. That is, is it possible that people are just suppressing the HPDL throughout the entire experiment (i.e., also through ITI, breaks, etc., rather than just before and during the search). Since the suppression does not seem volitional, I wonder if participants might apply a blanket suppression to HPDL un l they learn otherwise. Since your localiser comes a er the task you might be able to see hints of sustained suppression in the HPDL during these trials.  

It is indeed possible that participants suppressed the HPDL throughout the entire experiment, instead of proactively instantiating suppression on each trial. While possible, we believe that this account is less likely to explain the present results, given the utilized analysis approach, a voxel-wise GLM fit to the BOLD data per run (see Materials and Methods for details). Specifically, we derived parameter estimates from this GLM per location to estimate the relative suppression. Sustained suppression would modulate BOLD responses throughout the run, i.e. presumably also during the implicit baseline period used to estimate the contrast parameter estimates per location. Hence, sustained suppression should not result in a differential modulation between locations, as the BOLD response at the HPDL during the baseline period would be equally suppressed as during the trial. Inspired by the reviewer’s comment, we now clarify this critical point in the manuscript’s Discussion section:

“Third, participants might have suppressed the HPDL consistently throughout the experiment. This sustained suppression account differs from the proactive suppression proposed here. While this alternative is plausible, we believe that it is less likely to account for the present results, given the analysis conducted. Specifically, we computed voxel-wise parameter estimates and contrasted the obtained betas between locations. Under a sustained suppression account, the HPDL would show suppression even during the implicit baseline period, which would obscure the observed BOLD suppression at and near the HPDL.” 

(b) Enhancement followed by suppression. Another alternative that wasn't discussed would be an initial transient enhancement of the HPDL which might be brought on by the placeholders followed by more sustained suppression through the search task. Of course, on the whole this would look like suppression, but this still seems like it would hold different implications compared to simply "proactive suppression". This would be something like search and destroy however could be on the location level before the actual onset of the search display.  

R1 correctly points out that BOLD data, given the poor temporal resolution, do not allow for the detection of potential transient enhancements at the HPDL followed by a later and more pronounced suppression (akin to “search and destroy”). We fully agree with this assessment. However, we also argue that a transient enhancement followed by sustained suppression before search display onset constitutes proactive suppression in line with our interpretation, because suppression would still arise proactively (i.e., before search, and hence distractor, onset). Whether transient enhancement precedes suppression cannot be elucidated by our data, but we believe that it constitutes an interesting avenue for future studies using me-resolved and spatially specific recording methods. We now clarify this important implementational variation in the updated manuscript.

“Finally, due to the limited temporal resolution of BOLD data, the present data do not elucidate whether the present suppression is preceded by a brief attentional enhancement of the HPDL, as implied by some prior work (Huang et al., 2024). On this account the HPDL would see transient enhancement, followed by sustained suppression, akin to a ‘search and destroy’ mechanism. Critically, we believe that this variation would nonetheless constitute proactive distractor suppression as the suppression would still arise before search onset. Using temporally and spatially resolved methods to explore potential transient enhancements preceding suppression is a promising avenue for future research charting the neural mechanisms underlying distractor suppression.”

(2) I was also considering whether your effects might be at least partially attributable to priming type effects. This would be on the spatial (not feature) level as it is clear that the distractors are switching colours. Basically, is it possible that on trial n participants see the HPDL with the distractor in it and then on trial n+1 they suppress that location. This would be something distinct from the statistical learning framework and from the repetition suppression discussion you have already included. To test for this, you could look at the trials that follow omission or trials. If there is no suppression or less suppression on these trials it would seem fair to conclude that the suppression is at least in part due to the previous trial. 

We agree with the reviewer that it is plausible that participants particularly suppress locations which on previous trials contained a distractor. To address this possibility, we conducted a new analysis and adjusted the manuscript accordingly:

“Second, participants may have suppressed locations that contained the distractor on the previous trial, reflecting a spatial priming effect. This account constitutes a complementary but different perspective than statistical learning, which integrates implicit prior knowledge across many trials. We ruled out that spatial priming explains the present results by contrasting BOLD suppression magnitudes on trials with the distractor at the HPDL and trials where the distractor was not at the HPDL on the previous trial. Results, depicted in Supplementary Figure 4 showed that distractor suppression was statistically significant across both trial types, including trials without a distractor at the HPDL on the preceding trial. This indicates that the observed BOLD suppression is unlikely to be driven by priming and is instead more consistent with statistical learning. Moreover, results did not yield a statistically significant difference between trial types based on the distractor location in the preceding trial. However, these results should not be taken to suggest that spatial priming cannot contribute to distractor suppression; for details see: Supplementary Figure 4.” (p. 13).

We note that this analysis approach slightly differs from the reviewer’s suggestion, which considered omission trials. However, we decided to exclude trials immediately following an omission to ensure that both conditions were matched as closely as possible. In particular, omission trials represent extended rest periods, which could alter participants’ state and especially modulate the visually evoked BOLD responses (e.g., potentially increasing the dynamic range) compared to trials that did not follow omissions. Our analysis approach avoids this difference while still addressing the hypothesis put forward by the reviewer. We now provide the full explanation and results figure of this priming analysis in the figure text of Supplementary Figure 4: 

Reviewer #2 (Public review): 

The authors of this work set out to test ideas about how observers learn to ignore irrelevant visual information. Specifically, they used fMRI to scan participants who performed a visual search task. The task was designed in such a way that highly salient but irrelevant search items were more likely to appear at a given spatial location. With a region-of-interest approach, the authors found that activity in visual cortex that selectively responds to that location was generally suppressed, in response to all stimuli (search targets, salient distractors, or neutral items), as well as in the absence of an anticipated stimulus. 

Strengths of the study include: A well-written and well-argued manuscript; clever application of a region of interest approach to fMRI design, which allows articulating clear tests of different hypotheses; careful application of follow-up analyses to rule out alternative, strategy-based accounts of the findings; tests of the robustness of the findings to detailed analysis parameters such as ROI size; and exclusion of the role of regional baseline differences in BOLD responses. 

We thank the reviewer for the positive evaluation of our manuscript.

The report might be enhanced by analyses (perhaps in a surface space) that distinguish amongst the multiple "early" retinotopic visual areas that are analysed in the aggregate here. 

We agree with the reviewer that an exploratory analysis separating early visual cortex (EVC) into its retinotopic areas could be an interesting addition. Our reasoning to combine early visual areas into one mask in the original analyses was two-fold: First, we did not have an a priori reason to expected distinct neural suppression between these early ROIs. Therefore, we did not acquire retinotopy data to reliably separate early visual areas (e.g. V1, V2 and V3), instead opting to increase the number of search task trials. The lack of retinotopy data inherently limits the reliability of the resulting cortical segmentation. However, we now performed an analysis separating early visual cortex into V1 and V2 and report the details as Supplementary Text 1:

“In an exploratory analysis we investigated whether subdivisions of EVC exhibit different representations of priority signals. In brief, we used FreeSurfer to reconstruct brain surfaces (recon-all) from each subject’s anatomical scan. From these reconstructions we derived V1_exvivo and V2_exvivo labels, which were transformed into volume space using ‘mri_label2vol’ and merged into a bilateral mask for each ROI. We then selected the voxels within each ROI that were most responsive to the four stimulus locations, based on independent localizer data. This voxel selection followed the procedure outlined in the Materials and Methods: Region of Interest (ROI) Definition. To accommodate the subdivision into two ROIs (V1 and V2) compared to the single EVC ROI in the main analysis, we halved the number of voxels selected per location. Finally, we applied the same ROI analysis to investigate distractor suppression during search and omission trials, following the procedure described in Materials and Methods: Statistical Analysis. 

Results of this more fine-grained ROI analyses are depicted in Supplementary Figure 1. First, the results from V2 qualitatively mirrored our primary ROI analysis. BOLD responses in V2 differed significantly between stimulus types (main effect of stimulus type: F_(2,54) = 31.11, p < 0.001, 𝜂 = 0.54). Targets elicited larger BOLD responses compared to distractors (t₍₂₇₎ = 3.05, p_holm = 0.004, d = 0.06) and neutral stimuli (t₍₂₇₎ = 7.82, p_holm < 0.001, d = 0.14). Distractors also evoked larger responses than neutral stimuli (t₍₂₇₎ = 4.78, p_holm < 0.001, d = 0.09). These results likely reflect top-down modulation due to target relevance and bo om-up effects of distractor salience. Consistent with the primary ROI analysis, the manipula on of distractor predictability showed a distinct pattern of location specific BOLD suppression in V2 (main effect of location: F_(1.1,52.8) = 5.01, p = 0.030, 𝜂 = 0.16). Neural populations with receptive fields at the HPDL showed significantly reduced BOLD responses compared to the diagonally opposite neutral location (NL-far; post hoc test HPDL vs NL-far: t₍₂₇₎ = 2.69, p_holm = 0.022, d = 0.62). Again, this suppression was not confined to the HPDL but also extended to close by neutral locations (NL-near vs NL-far: t₍₂₇₎ = 2.79, p_holm = 0.022, d = 0.65). BOLD responses did not differ between HPDL and NL-near locations (HPDL vs NL-near: t₍₂₇₎ = 0.11, p_holm = 0.915, d = 0.03; BF₁₀ = 0.13). As in the EVC ROI analysis, this suppression pattern was consistent across distractor, target, and neutral stimuli presented at the HPDL and NL-near locations compared to NL-far. In sum, neural responses in V2 were significantly modulated by the distractor contingencies, evident as reduced BOLD responses in neural populations with receptive fields at the HPDL and neutral locations near the location of the frequent distractor (NL-near), relative to the neutral location diagonally across the HPDL (NL-far). 

In V1, BOLD responses also differed significantly between stimulus types (main effect of stimulus type: F_(1.3,35.6) = 6.69, p = 0.009, 𝜂 = 0.20). Targets elicited larger BOLD responses compared neutral stimuli (t₍₂₇₎ = 3.52, p_holm = 0.003, d = 0.12) and distractors evoked larger responses than neutral stimuli (t₍₂₇₎ = 2.62, p_holm = 0.023, d = 0.09). However, no difference between targets and distractors was observed (t₍₂₇₎ = 0.90, p_holm = 0.375, d = 0.03; BF₁₀ = 0.17), suggesting reduced sensitivity to task-related effects in V1. Indeed, analyzing the effect of distractor predictability for BOLD responses in V1 showed a different result than in V2 and the combined EVC ROI. There was no significant main effect of location (F_(2,54) = 2.20, p = 0.120, 𝜂 = 0.08; BF₁₀ = 0.77). BOLD responses at NL-near and NL-far were similar (BF₁₀ = 0.171), with the only reliable difference found between target stimuli at the HPDL and NL-far locations (W = 94, p_holm = 0.012, r = 0.54).”  

We include the new result figure as Supplementary Figure 5

We now include reference to these results in the manuscript’s Discussion section:

“Are representations of priority signals uniform across EVC? A priori we did not have any hypotheses regarding distinct neural suppression profiles across different early visual areas, hence our primary analyses focused stimulus responses neural populations in EVC, irrespective of subdivision. However, an exploratory analysis suggests that distractor suppression may show different patterns in V1 compared to V2 (Supplementary Figure 5 and Supplementary Text 1). In brief, results in V2 mirrored those reported for the combined EVC ROI (Figure 4). In contrast, results in V1 appeared to be only partially modulated by distractor contingencies, and if so, the modulation was less robust and not as spatially broad as in V2. This suggests the possibility of different effects of distractor predictability across subdivisions of early visual areas. However, these results should be interpreted with caution. First, our design did not optimize the delineation of early visual areas (e.g., no functional retinotopy), limiting the accuracy of V1 and V2 segmentation. Additionally, analyses were conducted in volumetric space, which further reduces spatial precision. Future studies could improve this by including retinotopy runs to accurately delineate V1, V2, and V3, and by performing analyses in surface space. Higher-resolution functional and anatomical MRI sequences would also help elucidate how distractor suppression is implemented across EVC with greater precision.”

Furthermore, the study could benefit from an analysis that tests the correlation over observers between the magnitude of their behavioural effects and their neural responses. 

R2 highlights that behavioral facilitation and neural suppression could be correlated across participants. The rationale is that if neural suppression in EVC is related to the facilitation of behavioral responses, we should expect a positive relationship between neural suppression at the HPDL and RTs across participants. In this analysis we focused on the contrast between HPDL and NL-far, as this contrast was statistically significant in both the RT (Figure 2) and the neural suppression analysis (Figure 4). First, we computed for each participant the behavioural benefit of distractor suppression as: RT_facilitation = RT_NL-far – RT_HPDL. Thereby RT facilitation reflects the response speeding due to a distractor appearing at the high probability distractor location compared to the far neutral location. Next, we computed neural suppression as: BOLD_suppression = BOLD_NL-far – BOLD_HPDL Thus, positive values reflect the suppression of BOLD responses at the HPDL comparted to the NL-far location. The BOLD suppression index was computed for each stimulus type separately, as in the main ROI analysis (i.e. for Targets, Neutrals and Distractors). Finally, we correlated RT_facilitation with BOLD_suppression across participants using Pearson correlation. Results showed a small, but not statistically significant correlation between RT facilitation and BOLD suppression for distractor (r₍₂₆₎ = 0.22, p = 0.257), target (r₍₂₆₎ = 0.10, p = 0.598) and neutral (r₍₂₆₎ = 0.13, p = 0.519) stimuli. Thus, while the direc on of the correlation was in line with the specula on by the reviewer in the “ Recommendations for the authors”, results were not statistically reliable and therefore inconclusive. As also noted in our preliminary reply to the reviewer comments, it was a priori unlikely that this analysis would yield a statistically significant correlation. An a priori power analysis suggested that, to reach a power of 0.8 at a standard alpha of 0.05, given the present sample size of n=28, the effect size would need to exceed r > 0.75, which seemed unlikely for the correlation of behavioural and neural difference scores. Given the inconclusive nature of the results, we prefer to not include this additional analysis in the manuscript, as we believe that it does not add to the main message of the paper but have it accessible to the interested reader in the public “peer review process”.

The study provides an advance over previous studies, which iden fied enhancement or suppression in visual cortex as a function of search target/distractor predictability, but in less spatially-specific way. It also speaks to open questions about whether such suppression/enhancement is observed only in response to the arrival of visual information, or instead is preparatory, favouring the la er view. The theoretical advance is moderate, in that it is largely congruent with previous frameworks, rather than strongly excluding an opposing view or providing a major step change in our understanding of how distractor suppression unfolds. 

We agree with the reviewer that our results are an advancement of prior work, particularly with respect to narrowing down the role of sensory areas and the proactive nature of distractor suppression. However, we argue that this represents a significant step forward for several reasons. First, to our knowledge, the literature on distractor suppression, and visual search in general, is by no means unanimous with respect to the conclusion that distractor suppression is instantiated proactively (Huang et al., 2021, 2022). Indeed, there are several studies suggesting the opposite account; reactive suppression (Chang et al., 2023) or contributions by both proactive and reactive mechanisms (Sauter et al., 2021; Wang et al., 2019). Moreover, studies in support of proactive distractor suppression did not investigate the involvement of (early) sensory areas during suppression. Conversely, to our knowledge most studies investigating the involvement of sensory cortex during distractor suppression did not address the question whether suppression arises proactive or reactively.

Recommendations for the authors: 

Reviewer #1 ( Recommendations for the authors): 

Minor Points: 

(1) There are several disconnects between the behaviour and the MR results - i.e. not stimulus specific yet there are no deficits for targets appearing the HPDL, also no behavioural suppression for the NLNear but neural suppression found. Nevertheless, the behaviour is used as a way to rule out potential attentional strategies when considering whether there is enhancement in the NL-Far condition. I realise you have a few other points here, but I think it's worth addressing what could be seen as a double standard.

The reviewer points out an important concern, which we feel could have better been addressed in the manuscript. From our point of view a partial dissociation between neural modulations in EVC and eventual behavioural facilitation is not surprising, given the extensive neural processing beyond EVC required for behaviour. However, this assessment may differ, if one stresses an explicit volitional attentional strategy over an implicit statistical learning account. That said, we clearly do not want to create the impression of using a double standard. The lack of behavioural facilitation for targets at NLfar is not a critical part of our argument against explicit attentional strategies. Therefore, we rephrased the relevant paragraph in the Discussion section to now emphasize the importance of the control analysis excluding participants who reported the correct HPDL in the questionnaire (Figure 5), but nonetheless yielded qualitatively identical results to the main ROI analysis (Figure 4). In our opinion, this control analysis provides more compelling evidence against a volitional attentional strategy account without the risk of crea ng the impression of applying a double standard in the interpretation of behavioural data. Additionally, we now acknowledge the limitation of relying on behavioral data in ruling out volitional attentional strategies in the updated manuscript:

“It is well established that attention enhances BOLD responses in visual cortex (Maunsell, 2015; Reynolds & Chelazzi, 2004; Williford & Maunsell, 2006). If participants learned the underlying distractor contingencies, they could deploy an explicit strategy by directing their attention away from the HPDL, for example by focusing attention on the diagonally opposite neutral location. This account provides an alternative explanation for the observed EVC modulations. However, while credible, the current findings are not consistent with such an interpretation. First, there was no behavioral facilitation for target stimuli presented at the far neutral location, contrary to what one might expect if participants employed an explicit strategy. However, given the partial dissociation between neural suppression in EVC and behavioral facilitation, additional neural data analyses are required to rule out volitional attention strategies. Thus, we performed a control analysis that excluded all participants that indicated the correct HPDL location in the questionnaire, thereby possibly expressing explicit awareness of the contingencies. This control analysis yielded qualitatively identical results to the full sample, showing significant distractor suppression in EVC. Therefore, it is unlikely that explicit attentional strategies, and the enhancement of locations far from the HPDL, drive the results observed here. Instead the current finding are consistent with an account emphasizing the automa c deployment of spatial priors (He et al., 2022) based on implicitly learned statistical regularities.”

(2) Does the level of suppression change in any way through the experiment? I.e., does it get stronger in the second vs. first half of the experiment? 

The reviewer askes an interesting question, whether BOLD suppression may change across the experiment. To address this question, we performed an additional analysis testing BOLD suppression in EVC during the first compared to second half of the MRI experiment. Here we defined BOLD suppression as: BOLD_suppression = ((BOLD_NL-far – BOLD_HPDL) + (BOLD_NL-far – BOLD_NL-near)) / 2. Thus, in this formula on of BOLD suppression we summarize the two primary BOLD suppression effects observed in our main results (Figure 4). Additionally, as we previously did not observe any significant differences in BOLD suppression magnitudes between different stimulus types (i.e. suppression was similar for target, distractor and neutral stimuli), we collapsed across stimulus types in this analysis.

Results, depicted below, showed that during both the initial (Run 1+2) and later part (Run 4+5) of the MRI experiment BOLD suppression was statistically significant (BOLD suppression Run 1+2: W = 331, p = 0.003, r = 0.63; BOLD suppression Run 4+5: W = 320, p = 0.007, r= 0.58) , confirming our main results of reliable distractor suppression even in this subset of trials. However, we did not observe any statistically significant differences between early and late runs of the experiment (t₍₂₇₎ = -0.21, p = 0.835, d = -0.04). In fact, a Bayesian paired t-test provided evidence for the absence of a difference in BOLD suppression between early compared to later runs (BF₁₀ = 0.205), suggesting that distractor suppression in EVC was stable throughout the experiment. A qualitatively similar, pattern was evident during omission trials, with significant distractor suppression during early runs (t₍₂₇₎ = 2.70, p = 0.012, d = 0.51), but not quite a statistically significant modulation for later runs (t₍₂₇₎ = 1.97, p = 0.059, d = 0.37). Again, there was no evidence for a difference in suppression magnitudes across the experiment (W = 198, p = 0.920, d = -0.025) and support for the absence of a difference in BOLD suppression between early and late runs (BF₁₀ = 0.278).

Author response image 1.

Analysis of BOLD suppression magnitudes in EVC across the MRI experiment phases. BOLD suppression was comparable between early (Run 1+2) and late (Run 4+5) phases of the MRI experiment, suggesting consistent suppression in EVC following statistical learning. Error-bars denote within-subject SEM. * p < 0.05, ** p < 0.01, = BF₁₀ < 1/3.

In sum, results suggest that distractor suppression in EVC was stable across runs and did not change significantly throughout the experiment. This result was a priori likely, given that participants already underwent behavioral training before entering the MRI. This enabled them to establish modified spatial priority maps, containing the high probability distractor location contingencies, already before the first MRI run. While specula ve, it is possible that participants may still have consolidated the spatial priority maps during the initial runs, but that this additional consolation is not evident in the data, as later runs may see less engagement by participants due to increasing fa gue towards the end of the MRI experiment. Indeed, rapid learning and stable suppression throughout the remainder of the experiment is also reported by prior work (Lin et al., 2021). We believe that it is highly interesting for future studies to investigate the development of distractor suppression across learning, with initial exposure to the contingencies inside the MRI. However, as the present results are inconclusive, we prefer to not include this analysis in the main manuscript, as it may not provide significant additional insight into the neural mechanisms underlying distractor suppression. 

(3) In the methods vs. results you have reported the probabili es slightly differently. In the methods you say the HPDL was 6x more likely to contain a distractor whereas in the results you say 4x. Based on the reported trial numbers I think it should be 4, but probably you want to double check that this is consistent and correct throughout. 

We thank the reviewer for bringing this inconsistency to our attention. We have corrected this oversight in the adjusted manuscript: 

“One of the four locations of interest was designated the high probability distractor location (HPDL), which contained distractor stimuli (unique color) four mes more o en than any of the remaining three locations of interest. In other words, if a distractor was present on a given trial (42 trials per run), the distractor appeared 57% (24 trials per run) at the HPDL and at one of the other three locations with equal probability (i.e., 14% or 6 trials per run per location).” 

Reviewer #2 ( Recommendations for the authors): 

The authors have performed their analyses in the volume rather than the surface, and have grouped together V1, V2, and V3 as "early visual cortex". As the authors' claims lean heavily on the idea that they are measuring "early" visual responses, the study would be improved by delinea ng the ROIS within these different retinotopic regions. Such an approach might be facilitated by analysing data on the reconstructed surface. 

Please refer to our reply to this analysis suggested in the Public review.

The authors rightly tread carefully on the causal link between their neural findings and the behavioural outcomes. The picture might be clarified somewhat further by testing for a positive relationship between behavioural effect sizes and neural effect sizes across participants. e.g. to what extent is the search advantage when distractors are presented at the "HPDL" linked to greater suppression of BOLD at the HDPL region of early visual cortex? 

Please refer to our reply to this analysis suggested in the Public review.

Some of the claims based on null hypotheses would be better supported by Bayesian tests e.g. page 6 "This pattern of results was the same regardless whether the distractor, target, or a neutral stimulus presented at the HPDL and NL-near locations compared to NL-far ..." and "BOLD responses between HPDL and NL-near locations did not reliably differ ..." This is similar to the approach that the authors adopted later in the section "Ruling out attentional modulation".

We agree with the reviewer that our ROI analyses would benefit from providing evidence for the absence of a modulation. Accordingly, we updated our results by adding equivalent Bayesian tests. Bayes Factors were computed using JASP 0.18.2 (JASP Team, 2024; RRID:SCR_015823) with default settings; i.e. for Bayesian paired t-tests with a Cauchy prior width of 0.707. Qualitative interpretations of BFs were based on Lee and Wagenmakers (2014). We now report the obtained BF in the Results section. 

“BOLD responses between HPDL and NL-near locations did not reliably differ (HPDL vs NL-near: t₍₂₇₎ = 0.47, p_holm = 0.643, d = 0.08; BF₁₀ = 0.19).”

And:

“Neural responses at HPDL and NL-near did not reliably differ (t₍₂₇₎ = 0.21, p_holm = 0.835 d = 0.04; BF₁₀ = 0.21).”

Moreover, we now denote any equivalent results (defined as BF₁₀<1/3) in Fig. 4 and Fig. 5, and included the descrip on of the associated symbol in the figure text (“ = BF₁₀ < 1/3”).

Additionally, we now also report the BF for all paired t-tests reported in Supplementary Table 1.

Finally, we addressed the statement: “This pattern of results was the same regardless whether the distractor, target, or a neutral stimulus presented at the HPDL and NL-near locations compared to NLfar”. Our inten on was to emphasize that the pattern of results reported in the sentence preceding it was evident for distractor, target, or neutral stimulus, and not to suggest that the magnitude of the effect is the same. Hence, to more accurate reflect the results, we changed this sentence to:  “This pattern of results was present regardless whether the distractor, target, or a neutral stimulus presented at the HPDL and NL-near locations compared to NL-far”

Author response:

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1 (Public Review):

Based on previous publications suggesting a potential role for miR-26b in the pathogenesis of metabolic dysfunction-associated steatohepatitis (MASH), the researchers aim to clarify its function in hepatic health and explore the therapeutical potential of lipid nanoparticles (LNPs) to treat this condition. First, they employed both whole-body and myeloid cell-specific miR-26b KO mice and observed elevated hepatic steatosis features in these mice compared to WT controls when subjected to WTD. Moreover, livers from whole-body miR-26b KO mice also displayed increased levels of inflammation and fibrosis markers. Kinase activity profiling analyses revealed distinct alterations, particularly in kinases associated with inflammatory pathways, in these samples. Treatment with LNPs containing miR-26b mimics restored lipid metabolism and kinase activity in these animals. Finally, similar anti-inflammatory effects were observed in the livers of individuals with cirrhosis, whereas elevated miR-26b levels were found in the plasma of these patients in comparison with healthy control. Overall, the authors conclude that miR-26b plays a protective role in MASH and that its delivery via LNPs efficiently mitigates MASH development.

The study has some strengths, most notably, its employ of a combination of animal models, analyses of potential underlying mechanisms, as well as innovative treatment delivery methods with significant promise. However, it also presents numerous weaknesses that leave the research work somewhat incomplete. The precise role of miR-26b in a human context remains elusive, hindering direct translation to clinical practice. Additionally, the evaluation of the kinase activity, although innovative, does not provide a clear molecular mechanisms-based explanation behind the protective role of this miRNA.

Therefore, to fortify the solidity of their conclusions, these concerns require careful attention and resolution. Once these issues are comprehensively addressed, the study stands to make a significant impact on the field.

We would like the reviewer for his/her careful evaluation of our manuscript and appreciate his/her appraisal for the strengths of our study. Regarding the weaknesses, we have addressed these as good as possible during the revision of our manuscript.

We can already state that miR-26b has clear anti-inflammatory effects on human liver slices, which is in line with our results demonstrating that miR-26b plays a protective role in MASH development in mice. The notion that patients with liver cirrhosis have increasing plasma levels of miR-26b, seems contradictory at first glance. However, we believe that this increased miR-26b expression is a compensatory mechanism to counteract the MASH/cirrhotic effects. However, the exact source of this miR-26b remains to be elucidated in future studies.

The performed kinase activity analysis revealed that miR-26b affects kinases that particularly play an important role in inflammation and angiogenesis. Strikingly and supporting these data, these effects could be inverted again by LNP treatment. Combined, these results already provide strong mechanistic insights on molecular and intracellular signalling level. Although the exact target of miR-26b remains elusive and its identification is probably beyond the scope of the current manuscript due to its complexity, we believe that the kinase activity results already provide a solid mechanistic basis.

Reviewer #1 (Recommendations For The Authors):

A list of recommendations for the authors is presented below:

(1) The title should emphasize that the majority of experiments were conducted in mice to accurately reflect the scope of the study.

As suggested we have updated our title to include the statement that we primarily used a murine model:

“MicroRNA-26b protects against MASH development in mice and can be efficiently targeted with lipid nanoparticles.”

(2) It would be useful to know more about miR-26b function, including its target genes, tissue-specific expression, and tissue vs. circulating levels. Is it expected that the two strains of the miRNA (i.e., -3p and -5p) act this similarly? Also, miR-26b expression in the liver of individuals with cirrhosis should be determined.

The function of miR-26b is still rather elusive, making functional studies using this miR very interesting. In a previous study, describing our used mouse model (Van der Vorst et al. BMC Genom Data, 2021) we have eluded several functions of miR-26b that are already investigated. This was particularly already described in carcinogenesis and the neurological field.

Target gene wise, there are already several targets described in miRbase. However, for our experiments we feel that determination of the specific target genes is beyond the scope of the current manuscript and rather a focus of follow-up projects.

Regarding the expression of miR-26b, the liver and blood have rather high and similar expressions of both miR-26b-3p and miR-26b-5p as shown in Author response image 1.

Author response image 1.

Expression of miR-26b-3p and -5p. Expression of miR-26b-3p (left) and miR-26b-5p (right), generated by using the miRNATissueAtlas 2025 (Rishik et al. Nucleic Acids Research, 2024). Unfortunately, due to restrictions in tissue availability and the lack of stored RNA samples, we are unable to measure miR-26b expression in the human livers. However, based on the potency of the miR-26b mimic loaded LNPs in the mice (Revised Supplemental Figure 2A-B), we are confident that these LNPs also resulted in a overexpression of miR-26b in the human livers.

(3) Please explain the rationale behind primarily using whole-body miR-26b KO mice rather than the myeloid cell-specific KO model for the studies.

The main goal of our study is the elucidation of the general role of miR-26b in MASH formation. Therefore, we decided to primarily focus on the whole-body KO model. While we used the myeloid cell-specific KO model to highlight that myeloid cells play an important role in the observed phenotypes, we believe the whole-body KO model is more appropriate as main focus, particularly also in light of the used LNP targeting that also provides a whole-body approach. Furthermore, this focus on the whole-body model also reflects a more therapeutically relevant approach.

(4) The authors claim that treatment with LNPs containing miR-26b "replenish the miR-26b level in the whole-body deficient mouse" but the results of this observation are not presented.

This is indeed a valid point that we have now addressed. We have measured the mir26b-3p and mir26b-5p expression levels in livers from mice after 4-week WTD with simultaneous injection with either empty LNPs as vehicle control (eLNP) or LNPs containing miR-26b mimics (mLNP) every 3 days. As shown in Revised Supplemental Figure 2A-B, mLNP treatment clearly results in an overexpression of the mir26b in the livers of these mice. We have rephrased the text accordingly by stating that mLNP results in an “overexpression” rather than “replenishment”.

(5) The number of 3 human donors for the precision-cut liver slices is clearly insufficient and clinical parameters need to be shown. Additionally, inconsistencies in individual values in Figures 8B-E need clarification.

Unfortunately, due to restrictions in tissue availability, we are unable to increase our n-number for these experiments. Clinical parameters are not available, but the liver slices were from healthy tissue.

We have performed these experiments in duplicates for each individual donor. We have now specified this also in the figure legend to explain the individual values in the graphs:

“…(3 individual donors, cultured in duplicates).”

(6) Figure 2D: Please include representative images.

As suggested we have included representative images in our revised manuscript.

(7) Address discrepancies in the findings across different experimental settings. For example, the expression levels of the lipid metabolism-related genes are not significantly modulated in whole-body miR-26b KO mice (except for Sra), but they are in the myeloid cell-specific model (but not Sra), and none of them are restored after LNPs injections.

Although Cd36 is not significantly increased in the whole-body miR-26b KO mice, there is a clear tendency towards increased expression, which is now also validated on protein level (Revised Figure 1K-L). In the myeloid cell-specific model we see a similar tendency, although the gene expression difference of Sra is not significantly changed. This could be due to the difference in the model, since only myeloid cells are affected, suggesting that the effects on Sra are to a large extend driven by non-myeloid cells. This would also fit to the tendency to decreased Sra expression in the mimic-LNP treated mice. Due to the larger variation, this difference did not reach significance, which is rather a statistical issue due to relatively small n-numbers. At this moment, we cannot exclude that these receptors are differentially regulated by different cell-types. For this, future studies are needed focussing on cell-specific targeting of miR-26b in somatic cells, like hepatocytes.

(8) Figure 4A the images are not representative of the quantification.

We have selected another representative image that is exactly reflecting the average Sirius red positive area, to reflect the quantification appropriately.

(9) Figures 5 and 7: Are there not significantly decreased/increased kinases? A deeper analysis of these kinase alterations is necessary to understand how miR-26b exerts its role. A comparison analysis of these two datasets might clarify this regard.

We indeed very often see in these kinome analysis that the general tendency of kinase activity is unidirectional. We believe that this is caused by the highly interconnected nature of kinases. Activation of one signalling cascade will also results in the activation of many other cascades. However, it is interesting to see which pathways are affected in our study and we find it particularly interesting to see that the tendencies is exactly opposite between both comparisons as KO vs. WT shows increase kinase activities, while KO-LNP vs. KO shows a decrease again. Further showing that the method is reflecting a true biological effect that is mediated by miR26b.

(10) Determinations of the effect of LNPs containing miR-26b in the KO mice are limited to only a few observations (that are not only significant). More extensive findings are needed to conclusively demonstrate the effectiveness of this treatment method. Similar to the experiments with human liver samples (Figures 8A-E).

We have now elaborated our observations in the mouse model using LNPs by also analysing the effects on inflammation and fibrosis. However, it seems that the treatment time was not long enough to see pronounced changes on these later stages of disease development. Interestingly, the expression of Tgfb was significantly reduced, suggesting at least that the LNPs on genetic levels have an effect already on fibrotic processes. Thereby, it can be suggested that longer mLNP treatment may result in more effects on protein level as well, which remains to be determined in future studies.

Unfortunately, due to restrictions in tissue availability, we are unable to increase our n-number or read-outs for these experiments at this moment.

(11) In Figures 8F-H, the observed increase in circulating miR-26b levels in the plasma of cirrhotic individuals seems contradictory to its proposed protective role. This discrepancy requires clarification.

In the revised discussion (second to last paragraph), we have now elaborated more on the findings in the plasma of cirrhotic individuals in comparison to our murine in-vivo results, to highlight and discuss this discrepancy.

(12) Figures 8F-H legend mentions using 8-11 patients per group, but the methods section lacks corresponding information about these individuals.

These patients, together with inclusion/exclusion criteria and definition of cirrhosis are described in the method section 2.14.

(13) Figure 8G has 7 data points in the cirrhosis group, instead of 8. Any data exclusion should be justified in the methods section.

As defined in method section 2.15, we have identified outliers using the ROUT = 1 method, which is the reason why Figure 8G only has 7 data points instead of 8.

Reviewer #2 (Public Review):

Summary:

This manuscript by Peters, Rakateli, et al. aims to characterize the contribution of miR-26b in a mouse model of metabolic dysfunction-associated steatohepatitis (MASH) generated by a Western-type diet on the background of Apoe knock-out. In addition, the authors provide a rescue of the miR-26b using lipid nanoparticles (LNPs), with potential therapeutic implications. In addition, the authors provide useful insights into the role of macrophages and some validation of the effect of miR-26b LNPs on human liver samples.

Strengths:

The authors provide a well-designed mouse model, that aims to characterize the role of miR-26b in a mouse model of metabolic dysfunction-associated steatohepatitis (MASH) generated by a Western-type diet on the background of Apoe knock-out. The rescue of the phenotypes associated with the model used using miR-26b using lipid nanoparticles (LNPs) provides an interesting avenue to novel potential therapeutic avenues.

Weaknesses:

Although the authors provide a new and interesting avenue to understand the role of miR-26b in MASH, the study needs some additional validations and mechanistic insights in order to strengthen the author's conclusions.

(1) Analysis of the expression of miRNAs based on miRNA-seq of human samples (see https://ccb-compute.cs.uni-saarland.de/isomirdb/mirnas) suggests that miR-26b-5p is highly abundant both on liver and blood. It seems hard to reconcile that despite miRNA abundance being similar in both tissues, the physiological effects claimed by the authors in Figure 2 come exclusively from the myeloid (macrophages).

We agree with the reviewer that the effects observed in the whole-body KO model are most likely a combination of cellular effects, particularly since miR-26b is also highly expressed in the liver. However, with the LysM-model we merely want to demonstrate that the myeloid cells at least play an important, though not exclusive, role in the phenotype. In the discussion, we also further elaborate on the fact that the observed changes in the liver can me mediated by hepatic changes.

To stress this, we have adjusted the conclusion of Figure 2:

“Interestingly, mice that have a myeloid-specific lack of miR-26b also show increased hepatic cholesterol levels and lipid accumulation demonstrated by Oil-red-O staining, coinciding with an increased hepatic Cd36 expression (Figure 2), demonstrating that myeloid miR-26b plays a major, but not exclusive, role in the observed steatosis.”

(2) Similarly, the miRNA-seq expression from isomirdb suggests also that expression of miR-26a-5p is indeed 4-fold higher than miR-26b-5p both in the liver and blood. Since both miRNAs share the same seed sequence, and most of the supplemental regions (only 2 nt difference), their endogenous targets must be highly overlapped. It would be interesting to know whether deletion of miR-26b is somehow compensated by increased expression of miR-26a-5p loci. That would suggest that the model is rather a depletion of miR-26.

UUCAAGUAAUUCAGGAUAGGU mmu-miR-26b-5p mature miRNA

UUCAAGUAAUCCAGGAUAGGCU mmu-miR-26a-5p mature miRNA

This is a very valid point raised by the reviewer, which we actually already explored in a previous study, describing our used mouse model (Van der Vorst et al. BMC Genom Data, 2021). In this manuscript, we could show that miR-26a is not affected by the deficiency of miR-26b (Figure 1G in: Van der Vorst et al. BMC Genom Data, 2021).

(3) Similarly, the miRNA-seq expression from isomirdb suggests also that expression of miR-26b-5p is indeed 50-fold higher than miR-26b-3p in the liver and blood. This difference in abundance of the two strands is usually regarded as one of them being the guide strand (in this case the 5p) and the other being the passenger (in this case the 3p). In some cases, passenger strands can be a byproduct of miRNA biogenesis, thus the rescue experiments using LNPs with both strands in equimolar amounts would not reflect the physiological abundance miR-26b-3p. The non-physiological overabundance of miR-26b-3p would constitute a source of undesired off-targets.

We agree with the reviewer on this aspect and this is something we had to consider while generating the mimic LNPs. However, we believe that we do not observe and undesired off-target effects, as the effects of the mimic LNPs at least on functional outcomes are relatively mild and only restricted to the expected effects on lipids. Furthermore, the effects on the kinase profile due to the mimic LNP treatment are in line with our expectations. Combined these results suggest at least that potential off-target effects are minor.

(4) It would also be valuable to check the miRNA levels on the liver upon LNP treatment, or at least the signatures of miR-26b-3p and miR-26b-5p activity using RNA-seq on the RNA samples already collected.

This is indeed a valid point that we have now addressed. We have measured the mir26b-3p and mir26b-5p expression levels in livers from mice after 4-week WTD with simultaneous injection with either empty LNPs as vehicle control (eLNP) or LNPs containing miR-26b mimics (mLNP) every 3 days. As shown in Supplemental Figure 2A-B, mLNP treatment clearly results in an overexpression of the mir26b in the livers of these mice. We have rephrased the text accordingly by stating that mLNP results in an “overexpression” rather than “replenishment”.

(5) Some of the phenotypes described, such as the increase in cholesterol, overlap with the previous publication by van der Vorst et al. BMC Genom Data (2021), despite in this case the authors are doing their model in Apoe knock-out and Western-type diet. I would encourage the authors to investigate more or discuss why the initial phenotypes don't become more obvious despite the stressors added in the current manuscript.

In our previous publication (BMC Genom Data; 2021), we actually did not see any changes in circulating lipid levels. However, in that study we did not evaluate the livers of the mice, so we do not have any information about the hepatic lipid levels.

As mentioned by the reviewer, we believe that we see much more pronounced phenotypes in the current model because we use the combined stressor of Apoe-/- and Western-type diet, which cannot be compared to the wildtype and chow-fed mice used in the BMC Genom Data manuscript.

(6) The authors have focused part of their analysis on a few gene makers that show relatively modest changes. Deeper characterization using RNA-seq might reveal other genes that are more profoundly impacted by miR-26 depletion. It would strengthen the conclusions proposed if the authors validated that changes in mRNA abundance (Sra, Cd36) do impact the protein abundance. These relatively small changes or trends in mRNA expression, might not translate into changes in protein abundance.

As suggested by the reviewer we have now also confirmed that the protein expression of CD36 and SRA is significantly increased upon miR-26b depletion, visualized as Figure 1K-L in the revised manuscript. Unfortunately, we do not have enough material left to perform similar analysis for the LysM-model or the LNP-model, although based on the whole-body effects we are confident that at least for CD36/SRA in this case the gene expression matches effects observed on protein level.

(7) In Figures 5 and 7, the authors run a phosphorylation array (STK) to analyze the changes in the activity of the kinome. It seems that a relatively large number of signaling pathways are being altered, I think that should be strengthened by further validations by Western blot on the collected tissue samples. For quite a few of the kinases, there might be antibodies that recognise phosphorylation. The two figures lack a mechanistic connection to the rest of the manuscript.<br /> On this point we respectfully have to disagree with the reviewer. We have used a kinase activity profiling approach (PamGene) to analyse the real-time activity of kinases in our lysates. This approach is different than the classical Western blot approach in which only the presence or absence of a specific phosphorylation is detected. Thereby, Western blot analysis does not analyse phosphorylation in real-time, but rather determines whether there has been phosphorylation in the past. Our approach actually determines the real-time, current activity of the kinases, which we believe is a different and perhaps even more reliable read-out measurement. Therefore, validation by Western blot would not strengthen these observations.

We have particularly tried to connect these observations to the rest of the manuscript by highlighting the observed signalling cascades that are affected, highlighting a role in inflammation and angiogenesis, thereby providing some mechanistic insights.

Reviewer #2 (Recommendations For The Authors):

I would encourage the authors to follow-up on some of the more miRNA focused comments made above, which would strengthen the mechanistic part of the work presented.

I suggest the authors tone down some of some of the claims made (eg. "clearly increased expression", "exacerbated hepatic fibrosis"), given that some of it might need further validation.

Wherever needed we have tuned down the tone of some claims, although we believe that most claims are already written carefully enough and in line with the observed results.

Some of the panels that are supposed to have the same amount of animals have variable N, despite they come from the same exact number of RNA samples or tissue lysates (eg. 1G and 1H, vs 1I and 1J).

This is indeed correct and caused by the fact that some analysis resulted in statistical outliers as identified using the ROUT = 1 method, as also specified in section 2.15 of the method section.

It would be nice to have representative images of oil-red-o in all the figures where it is quantified (or at least in the supplementary figures).

As suggested by the reviewer, we have now included representative images for the LysM-model (Revised Figure 2D) and the LNP-model (Revised Figure 6D) as well.

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public Review):

In this study, the authors aim to understand why decision formation during behavioural tasks is distributed across multiple brain areas. They hypothesize that multiple areas are used in order to implement an information bottleneck (IB). Using neural activity recorded from monkey DLPFC and PMd performing a 2-AFC task, they show that DLPFC represents various task variables (decision, color, target configuration), while downstream PMd primarily represents decision information. Since decision information is the only information needed to make a decision, the authors point out that PMd has a minimal sufficient representation (as expected from an IB). They then train 3-area RNNs on the same task and show that activity in the first and third areas resemble the neural representations of DLPFC and PMd, respectively. In order to propose a mechanism, they analyse the RNN and find that area 3 ends up with primarily decision information because feedforward connections between areas primarily propagate decision information.

The paper addresses a deep, normative question, namely why task information is distributed across several areas.

Overall, it reads well and the analysis is well done and mostly correct (see below for some comments). My major problem with the paper is that I do not see that it actually provides an answer to the question posed (why is information distributed across areas?). I find that the core problem is that the information bottleneck method, which is evoked throughout the paper, is simply a generic compression method.

Being a generic compressor, the IB does not make any statements about how a particular compression should be distributed across brain areas - see major points (1) and (2).

If I ignore the reference to the information bottleneck and the question of why pieces of information are distributed, I still see a more mechanistic study that proposes a neural mechanism of how decisions are formed, in the tradition of RNN-modelling of neural activity as in Mante et al 2013. Seen through this more limited sense, the present study succeeds at pointing out a good model-data match, and I could support a publication along those lines. I point out some suggestions for improvement below.

We thank the reviewer for their comments, feedback and suggestions. We are glad to hear you support the good model-data match for this manuscript.  With your helpful comments, we have clarified the connections to the information bottleneck principle and also contrasted it against the information maximization principle (the InfoMax principle), an alternative hypothesis. We elaborate on these issues in response to your points below, particularly major points (1) and (2). We also address all your other comments below.

Major points

(1) It seems to me that the author's use of the IB is based on the reasoning that deep neural networks form decisions by passing task information through a series of transformations/layers/areas and that these deep nets have been shown to implement an IB. Furthermore, these transformations are also loosely motivated by the data processing inequality.

On Major Point 1 and these following subpoints, we first want to make a high-level statement before delving into a detailed response to your points as it relates to the information bottleneck (IB). We hope this high-level statement will provide helpful context for the rest of our point-by-point responses. 

We want to be clear that we draw on the information bottleneck (IB) principle as a general principle to explain why cortical representations differ by brain area. The IB principle, as applied to cortex, is only stating that a minimal sufficient representation to perform the task is formed in cortex, not how it is formed. The alternative hypothesis to the IB is that brain areas do not form minimal sufficient representations. For example, the InfoMax principle states that each brain area stores information about all inputs (even if they’re not necessary to perform the task). InfoMax isn’t unreasonable: it’s possible that storing as much information about the inputs, even in downstream areas, can support flexible computation and InfoMax also supports redundancy in cortical areas. Indeed, many studies claim that action choice related signals are in many cortical areas, which may reflect evidence of an InfoMax principle in action for areas upstream of PMd.

While we observe an IB in deep neural networks and cortex in our perceptual decision-making task, we stress that its emergence across multiple areas is an empirical result. At the same time, multiple areas producing an IB makes intuitive sense: due to the data processing inequality, successive transformations typically decrease the information in a representation (especially when, e.g., in neural networks, every activation passes through the Relu function, which is not bijective). Multiple areas are therefore a sufficient and even ‘natural’ way to implement an IB, but multiple areas are not necessary for an IB. That we observe an IB in deep neural networks and cortex emerge through multi-area computation is empirical, and, contrasting InfoMax, we believe it is an important result of this paper. 

Nevertheless, your incisive comments have helped us to update the manuscript that when we talk about the IB, we should be clear that the alternative hypothesis is non-minimal representations, a prominent example of which is the InfoMax principle. We have now significantly revised our introduction to avoid this confusion. We hope this provides helpful context for our point-by-point replies, below.

However, assuming as a given that deep neural networks implement an IB does not mean that an IB can only be implemented through a deep neural network. In fact, IBs could be performed with a single transformation just as well. More formally, a task associates stimuli (X) with required responses (Y), and the IB principle states that X should be mapped to a representation Z, such that I(X;Z) is minimal and I(Y,Z) is maximal. Importantly, the form of the map Z=f(X) is not constrained by the IB. In other words, the IB does not impose that there needs to be a series of transformations. I therefore do not see how the IB by itself makes any statement about the distribution of information across various brain areas.

We agree with you that an IB can be implemented in a single transformation. We wish to be clear that we do not intend to argue necessity: that multiple areas are the only way to form minimal sufficient representations. Rather, multiple areas are sufficient to induce minimal sufficient representations, and moreover, they are a natural and reasonably simple way to do so. By ‘natural,’ we mean that minimal sufficient representations empirically arise in systems with multiple areas (more than 2), including deep neural networks and the cortex at least for our task and simulations. For example, we did not see minimal sufficient representations in 1- or 2-area RNNs, but we did see them emerge in RNNs with 3 areas or more. One potential reason for this result is that sequential transformations through multiple areas can never increase information about the input; it can only maintain or reduce information due to the data processing inequality.

Our finding that multiple areas facilitate IBs in the brain is therefore an empirical result: like in deep neural networks, we observe the brain has minimal sufficient representations that emerge in output areas (PMd), even as an area upstream (DLPFC) is not minimal. While the IB makes a statement that this minimal sufficient representation emerges, to your point, the fact that it emerges over multiple areas is not a part of the IB – as you have pointed out, the IB doesn’t state where or how the information is discarded, only that it is discarded. Our RNN modeling later proposes one potential mechanism for how it is discarded. We updated the manuscript introduction to make these points:

“An empirical observation from Machine Learning is that deep neural networks tend to form minimal sufficient representations in the last layers. Although multi-layer computation is not necessary for an IB, they provide a sufficient and even “natural” way to form an IB. A representation z = f(x) cannot contain more information than the input x itself due to the data processing inequality[19]. Thus, adding additional layers typically results in representations that contain less information about the input.”

And later in the introduction:

“Consistent with these predictions of the IB principle, we found that DLPFC has information about the color, target configuration, and direction. In contrast, PMd had a minimal sufficient representation of the direction choice. Our recordings therefore identified a cortical IB. However, we emphasize the IB does not tell us where or how the minimal sufficient representation is formed. Instead, only our empirical results implicate DLPFC-PMd in an IB computation. Further, to propose a mechanism for how this IB is formed, we trained a multi-area RNN to perform this task. We found that the RNN faithfully reproduced DLPFC and PMd activity, enabling us to propose a mechanism for how cortex uses multiple areas to compute a minimal sufficient representation.”

In the context of our work, we want to be clear the IB makes these predictions:

Prediction 1: There exists a downstream area of cortex that has a minimal and sufficient representation to perform a task (i.e.,. I(X;Z) is minimal while preserving task information so that I(Z;Y) is approximately equal to  I(X;Y)). We identify PMd as an area with a minimal sufficient representation in our perceptual-decision-making task. 

Prediction 2 (corollary if Prediction 1 is true): There exists an upstream brain area that contains more input information than the minimal sufficient area. We identify DLPFC as an upstream area relative to PMd, which indeed has more input information than downstream PMd in our perceptual decision-making task. 

Note: as you raise in other points, it could have been possible that the IB is implemented early on, e.g., in either the parietal cortex (dorsal stream) or inferotemporal cortex (ventral stream), so that DLPFC and PMd both contained minimal sufficient representations. The fact that it doesn’t is entirely an empirical result from our data. If DLPFC had minimal sufficient representations for the perceptual decision making task, we would have needed to record in other regions to identify brain areas that are consistent with Prediction 2. But, empirically, we found that DLPFC has more input information relative to PMd, and therefore the DLPFC-PMd connection is implicated in the IB process.

What is the alternative hypothesis to the IB? We want to emphasize: it isn’t single-area computation. It’s that the cortex does not form minimal sufficient representations. For example, an alternative hypothesis (“InfoMax”) would be for all engaged brain areas to form representations that retain all input information. One reason this could be beneficial is because each brain area could support a variety of downstream tasks. In this scenario, PMd would not be minimal, invalidating Prediction 1. However, this is not supported by our empirical observations of the representations in PMd, which has a minimal sufficient representation of the task. We updated our introduction to make this clear:

“But cortex may not necessarily implement an IB. The alternative hypothesis to IB is that the cortex does not form minimal sufficient representations. One manifestation of this alternative hypothesis is the “InfoMax” principle, where downstream representations are not minimal but rather contain maximal input information22. This means information about task inputs not required to perform the task are present in downstream output areas. Two potential benefits of an InfoMax principle are (1) to increase redundancy in cortical areas and thereby provide fault tolerance, and (2) for each area to support a wide variety of tasks and thereby improve the ability of brain areas to guide many different behaviors. In contrast to InfoMax, the IB principle makes two testable predictions about cortical representations. Prediction 1: there exists a downstream area of cortex that has a minimal and sufficient representation to perform a task (i.e., I(X; Z) is minimal while preserving task information so that I(Z; Y) ≈ I(X; Y)). Prediction 2 (corollary if Prediction 1 is true): there exists an upstream area of cortex that has more task information than the minimal sufficient area.”

Your review helped us realize we should have been clearer in explaining that these are the key predictions of the IB principle tested in our paper. We also realized we should be much clearer that these predictions aren’t trivial or expected, and there is an alternative hypothesis. We have re-written the introduction of our paper to highlight that the key prediction of the IB is minimal sufficient representations for the task, in contrast to the alternative hypothesis of InfoMax.

A related problem is that the authors really only evoke the IB to explain the representation in PMd: Fig 2 shows that PMd is almost only showing decision information, and thus one can call this a minimal sufficient representation of the decision (although ignoring substantial condition independent activity).

However, there is no IB prediction about what the representation of DLPFC should look like.

Consequently, there is no IB prediction about how information should be distributed across DLPFC and PMd.

We agree: the IB doesn’t tell us how information is distributed, only that there is a transformation that eventually makes PMd minimal. The fact that we find input information in DLPFC reflects that this computation occurs across areas, and is an empirical characterization of this IB in that DLPFC has direction, color and context information while PMd has primarily direction information. To be clear: only our empirical recordings verified that the DLPFC-PMd circuit is involved in the IB. As described above, if not, we would have recorded even further upstream to identify an inter-areal connection implicated in the IB.

We updated the text to clearly state that the IB predicts that an upstream area’s activity should contain more information about the task inputs. We now explicitly describe this in the introduction, copy and pasted again here for convenience.

“In contrast to InfoMax, the IB principle makes two testable predictions about cortical representations. Prediction 1: there exists a downstream area of cortex that has a minimal and sufficient representation to perform a task (i.e., I(X; Z) is minimal while preserving task information so that I(Z; Y) ≈ I(X; Y)). Prediction 2 (corollary if Prediction 1 is true): there exists an upstream area of cortex that has more task information than the minimal sufficient area.

Consistent with the predictions of the IB principle, we found that DLPFC has information about the color, target configuration, and direction. In contrast, PMd had a minimal sufficient representation of the direction choice. Our recordings therefore identified a cortical IB. However, we emphasize the IB does not tell us where or how the minimal sufficient representation is formed. Instead, only our empirical results implicate DLPFC-PMd in an IB computation Further, to propose a mechanism for how this IB is formed, we trained a multi-area RNN to perform this task.”  

The only way we knew DLPFC was not minimal was through our experiments. Please also note that the IB principle does not describe how information could be lost between areas or layers, whereas our RNN simulations show that this may occur through preferential propagation of task-relevant information with respect to the inter-area connections.  

(2) Now the authors could change their argument and state that what is really needed is an IB with the additional assumption that transformations go through a feedforward network. However, even in this case, I am not sure I understand the need for distributing information in this task. In fact, in both the data and the network model, there is a nice linear readout of the decision information in dPFC (data) or area 1 (network model). Accordingly, the decision readout could occur at this stage already, and there is absolutely no need to tag on another area (PMd, area 2+3).

Similarly, I noticed that the authors consider 2,3, and 4-area models, but they do not consider a 1-area model. It is not clear why the 1-area model is not considered. Given that e.g. Mante et al, 2013, manage to fit a 1-area model to a task of similar complexity, I would a priori assume that a 1-area RNN would do just as well in solving this task.

While decision information could indeed be read out in Area 1 in our multi-area model, we were interested in understanding how the network converged to a PMd-like representation (minimal sufficient) for solving this task. Empirically, we only observed a match between our model representations and animal cortical representations during this task when considering multiple areas. Given that we empirically observed that our downstream area had a minimal sufficient representation, our multi-area model allowed how this minimal sufficient representation emerged (through preferential propagation of task-relevant information).

We also analyzed single-area networks in our initial manuscript, though we could have highlighted these analyses more clearly to be sure they were not overlooked. We are clearer in this revision that we did consider a 1-area network (results in our Fig 5). While a single-area RNN can indeed solve this task, the single area model had all task information present in the representation, and did not match the representations in DLPFC or PMd. It would therefore not allow us to understand how the network converged to a PMd-like representation (minimal sufficient) for solving this task. We updated the schematic in Fig 5 to add in the single-area network (which may have caused the confusion).

We have added an additional paragraph commenting on this in the discussion. We also added an additional supplementary figure with the PCs of the single area RNN (Fig S15). We highlight that single area RNNs do not resemble PMd activity because they contain strong color and context information. 

In the discussion:

“We also found it was possible to solve this task with single area RNNs, although they did not resemble PMd (Figure S15) since it did not form a minimal sufficient representation. Rather, for our RNN simulations, we found that the following components were sufficient to induce minimal sufficient representations: (1) RNNs with at least 3 areas, following Dale’s law (independent of the ratio of feedforward to feedback connections).”

I think there are two more general problems with the author's approach. First, transformations or hierarchical representations are usually evoked to get information into the right format in a pure feedforward network. An RNN can be seen as an infinitely deep feedforward network, so even a single RNN has, at least in theory, and in contrast to feedforward layers, the power to do arbitrarily complex transformations. Second, the information coming into the network here (color + target) is a classical xor-task. While this task cannot be solved by a perceptron (=single neuron), it also is not that complex either, at least compared to, e.g., the task of distinguishing cats from dogs based on an incoming image in pixel format.

An RNN can be viewed as an infinitely deep feedforward network in time. However, we wish to clarify two things. First, our task runs for a fixed amount of time, and therefore this RNN in practice is not infinitely deep in time. Second, if it were to perform an IB operation in time, we would expect to see color discriminability decrease as a function of time. Indeed, we considered this as a mechanism (recurrent attenuation, Figure 4a), but as we show in Supplementary Figure S9, we do not observe it to be the case that discriminability decreases through time. This is equivalent to a dynamical mechanism that removes color through successive transformations in time, which our analyses reject (Fig 4). We therefore rule out that an IB is implemented through time via an RNN’s recurrent computation (viewed as feedforward in time). Rather, as we show, the IB comes primarily through inter-areal connections between RNN areas. We clarified that our dynamical hypothesis is equivalent to rejecting the feedforward-in-time filtering hypothesis in the Results: 

“We first tested the hypothesis that the RNN IB is implemented primarily by recurrent dynamics (left side of Fig. 4a). These recurrent dynamics can be equivalently interpreted as the RNN implementing a feedforward neural network in time.”  

The reviewer is correct that the task is a classical XOR task and not as complex as e.g., computer vision classification. That said, our related work has looked at IBs for computer vision tasks and found them in deep feedforward networks (Kleinman et al., ICLR 2021). Even though the task is relatively straightforward, we believe it is appropriate for our conclusions because it does not have a trivial minimal sufficient representation: a minimal sufficient representation for XOR must contain only target, but not color or target configuration information. This can only be solved via a nonlinear computation. In this manner, we favor this task because it is relatively simple, and the minimal sufficient representations are interpretable, while at the same time not being so trivially simple (the minimal sufficient representations require nonlinearity to compute).  

Finally, we want to note that this decision-making task is a logical and straightforward way to add complexity to classical animal decision-making tasks, where stimulus evidence and the behavioral report are frequently correlated. In tasks such as these, it may be challenging to untangle stimulus and behavioral variables, making it impossible to determine if an area like premotor cortex represents only behavior rather than stimulus. However, our task decorrelates both the stimulus and the behaviors. 

(3) I am convinced of the author's argument that the RNN reproduces key features of the neural data. However, there are some points where the analysis should be improved.

(a) It seems that dPCA was applied without regularization. Since dPCA can overfit the data, proper regularization is important, so that one can judge, e.g., whether the components of Fig.2g,h are significant, or whether the differences between DLPFC and PMd are significant.

We note that the dPCA codebase optimizes the regularization hyperparameter through cross-validation and requires single-trial firing rates for all neurons, i.e., data matrices of the form (n_Neurons x Color x Choice x Time x n_Trials), which are unavailable for our data. We recognized that you are fundamentally asking whether differences are significant or not. We therefore believe it is possible to address this through a statistical test, described further below. 

In order to test whether the differences of variance explained by task variables between DLPFC and PMd are significant, we performed a shuffle test. For this test, we randomly sampled 500 units from the DLPFC dataset and 500 units from the PMd dataset. We then used dPCA to measure the variance explained by target configuration, color choice, and reach direction (e.g., Var^True_DLPFC,Color, Var^True_PMd,Color).

To test if this variance was significant, we performed the following shuffle test. We combined the PMd and DLPFC dataset into a pool of 1000 units and then randomly selected 500 units from this pool to create a surrogate PMd dataset and used the remaining 500 units as a surrogate DLPFC dataset. We then again performed dPCA on these surrogate datasets and estimated the variance for the various task variables (e.g., Var_{ShuffledDLPFC,Color}  ,Var_{ShuffledPMd,Color}).

We repeated this process for 100 times and estimated a sampling distribution for the true difference in variance between DLPFC and PMd for various task variables (e.g., Var^True_DLPFC,Color - Var^True_PMd,Color). At the same time, we estimated the distribution of the variance difference between surrogate PMd and DLPFC dataset for various task variables (e.g., Var_{ShuffleDLPFC,Color} - Var_{ShufflePMd,Color}). 

We defined a p-value as the number of shuffles in which the difference in variance was higher than the median of the true difference and divided it by 100. Note, for resampling and shuffle tests with n shuffles/bootstraps, the lowest theoretical p-value is given as 2/n, even in the case that no shuffle was higher than the median of the true distribution. Thus, the differences were statistically significant (p < 0.02) for color and target configuration but not for direction (p=0.72). These results are reported in Figure S6 and show both the true sampling distribution and the shuffled sampling distributions.

(b) I would have assumed that the analyses performed on the neural data were identical to the ones performed on the RNN data. However, it looked to me like that was not the case. For instance, dPCA of the neural data is done by restretching randomly timed trials to a median trial. It seemed that this restretching was not performed on the RNN. Maybe that is just an oversight, but it should be clarified. Moreover, the decoding analyses used SVC for the neural data, but a neural-net-based approach for the RNN data. Why the differences?

Thanks for bringing up these points. We want to clarify that we did include SVM decoding for the multi-area network in the appendix (Fig. S4), and the conclusions are the same. Moreover, in previous work, we also found that training with a linear decoder led to analogous conclusions (Fig. 11 of Kleinman et al, NeurIPS 2021).  As we had a larger amount of trials for the RNN than the monkey, we wanted to allow a more expressive decoder for the RNN, though this choice does not affect our conclusions. We clarified the text to reflect that we did use an SVM decoder.

“We also found analogous conclusions when using an SVM decoder (Fig. S4).”

dPCA analysis requires trials of equal length. For the RNN, this is straightforward to generate because we can set the delay lengths to be equal during inference (although the RNN was trained on various length trials and can perform various length trials). Animals must have varying delay periods, or else they will learn the timing of the task and anticipate epoch changes. Because animal trial lengths were therefore different, their trials had to be restretched. We clarified this in the Methods.

“For analyses of the RNN, we fixed the timing of trials, obviating the need to to restretch trial lengths. Note that while at inference, we generated RNN trials with equal length, the RNN was trained with varying delay periods.” 

(4) The RNN seems to fit the data quite nicely, so that is interesting. At the same time, the fit seems somewhat serendipitous, or at least, I did not get a good sense of what was needed to make the RNN fit the data. The authors did go to great lengths to fit various network models and turn several knobs on the fit. However, at least to me, there are a few (obvious) knobs that were not tested.

First, as already mentioned above, why not try to fit a single-area model? I would expect that a single area model could also learn the task - after all, that is what Mante et al did in their 2013 paper and the author's task does not seem any more complex than the task by Mante and colleagues.

Thank you for bringing up this point. As mentioned in response to your prior point, we did analyze a single-area RNN (Fig. 5d). We updated the schematic to clarify that we analyzed a single area network. Moreover, we also added a supplementary figure to qualitatively visualize the PCs of the single area network (Fig. S15). While a single area network can solve the task, it does not allow us to study how representations change across areas, nor did it empirically resemble our neural recordings. Single-area networks contain significant color, context, and direction information. They therefore do not form minimal representations and do not resemble PMd activity.

Second, I noticed that the networks fitted are always feedforward-dominated. What happens when feedforward and feedback connections are on an equal footing? Do we still find that only the decision information propagates to the next area? Quite generally, when it comes to attenuating information that is fed into the network (e.g. color), then that is much easier done through feedforward connections (where it can be done in a single pass, through proper alignment or misalignment of the feedforward synapses) than through recurrent connections (where you need to actively cancel the incoming information). So it seems to me that the reason the attenuation occurs in the inter-area connections could simply be because the odds are a priori stacked against recurrent connections. In the real brain, of course, there is no clear evidence that feedforward connections dominate over feedback connections anatomically.

We want to clarify that we did pick feedforward and feedback connections based on the following macaque atlas, reference 27 in our manuscript: 

Markov, N. T., Ercsey-Ravasz, M. M., Ribeiro Gomes, A. R., Lamy, C., Magrou, L., Vezoli, J., Misery, P., Falchier, A., Quilodran, R., Gariel, M. A., Sallet, J., Gamanut, R., Huissoud, C., Clavagnier, S., Giroud, P., Sappey-Marinier, D., Barone, P., Dehay, C., Toroczkai, Z., … Kennedy, H. (2014). A weighted and directed interareal connectivity matrix for macaque cerebral cortex. Cerebral Cortex , 24(1), 17–36.

We therefore believe there is evidence for more feedforward than feedback connections. Nevertheless, as stated in response to your next point below, we ran a simulation where feedback and feedforward connectivity were matched.

More generally, it would be useful to clarify what exactly is sufficient:

(a) the information distribution occurs in any RNN, i.e., also in one-area RNNs

(b) the information distribution occurs when there are several, sparsely connected areas

(c) the information distribution occurs when there are feedforward-dominated connections between areas

We better clarify what exactly is sufficient. 

- We trained single-area RNNs and found that these RNNs contained color information; additionally two area RNNs also contained color information in the last area (Fig 5d). 

- We indeed found that the minimal sufficient representations emerged when we had several areas, with Dale’s law constraint on the connectivity. When we had even sparser connections, without Dale’s law, there was significantly more color information, even at 1% feedforward connections; Fig 5a.

- When we matched the percentage of feedforward and feedback connections with Dale’s law constraint on the connectivity (10% feedforward and 10% feedback), we also observed minimal sufficient representations (Fig S9). 

Together, we found that minimal sufficient representations emerged when we had several areas (3 or greater), with Dale’s law constraint on the connectivity, independent of the ratio of feedforward/feedback connections. We thank the reviewer for raising this point about the space of constraints leading to minimal sufficient representations in the late area. We clarified this in the Discussion.

“We also found it was possible to solve this task with single area RNNs, although they did not resemble PMd (Figure S15) since it did not form a minimal sufficient representation. Rather, for our RNN simulations, we found that the following components were sufficient to induce minimal sufficient representations: RNNs with at least 3 areas, following Dale’s law (independent of the ratio of feedforward to feedback connections).”

Thank you for your helpful and constructive comments!

Reviewer #2 (Public Review):

Kleinman and colleagues conducted an analysis of two datasets, one recorded from DLPFC in one monkey and the other from PMD in two monkeys. They also performed similar analyses on trained RNNs with various architectures.

The study revealed four main findings. (1) All task variables (color coherence, target configuration, and choice direction) were found to be encoded in DLPFC. (2) PMD, an area downstream of PFC, only encoded choice direction. (3) These empirical findings align with the celebrated 'information bottleneck principle,' which suggests that FF networks progressively filter out task-irrelevant information. (4) Moreover, similar results were observed in RNNs with three modules.

We thank the reviewer for their comments, feedback and suggestions, which we address below.

While the analyses supporting results 1 and 2 were convincing and robust, I have some concerns and recommendations regarding findings 3 and 4, which I will elaborate on below. It is important to note that findings 2 and 4 had already been reported in a previous publication by the same authors (ref. 43).

Note the NeurIPS paper only had PMd data and did not contain any DLPFC data. That manuscript made predictions about representations and dynamics upstream of PMd, and subsequent experiments reported in this manuscript validated these predictions. Importantly, this manuscript observes an information bottleneck between DLPFC and PMd.

Major recommendation/comments:

The interpretation of the empirical findings regarding the communication subspace in relation to the information bottleneck theory is very interesting and novel. However, it may be a stretch to apply this interpretation directly to PFC-PMd, as was done with early vs. late areas of a FF neural network.

In the RNN simulations, the main finding indicates that a network with three or more modules lacks information about the stimulus in the third or subsequent modules. The authors draw a direct analogy between monkey PFC and PMd and Modules 1 and 3 of the RNNs, respectively. However, considering the model's architecture, it seems more appropriate to map Area 1 to regions upstream of PFC, such as the visual cortex, since Area 1 receives visual stimuli. Moreover, both PFC and PMd are deep within the brain hierarchy, suggesting a more natural mapping to later areas. This contradicts the CCA analysis in Figure 3e. It is recommended to either remap the areas or provide further support for the current mapping choice.

We updated the Introduction to better clarify the predictions of the information bottleneck (IB) principle. In particular, the IB principle predicts that later areas should have minimal sufficient representations of task information, whereas upstream areas should have more information. In PMd, we observed a minimal sufficient representation of task information during the decision-making task. In DLPFC, we observed more task information, particularly more information about the target colors and the target configuration.

In terms of the exact map between areas, we do not believe or intend to claim the DLPFC is the first area implicated in the sensorimotor transformation during our perceptual decision-making task. Rather, DLPFC best matches Area 1 of our model. It is important to note that we abstracted our task so that the first area of our model received checkerboard coherence and target configuration as input (and hence did not need to transform task visual inputs). Indeed, in Figure 1d we hypothesize that the early visual areas should contain additional information, which we do not model directly in this work. Future work could model RNNs to take in an image or video input of the task stimulus. In this case, it would be interesting to assess if earlier areas resemble visual cortical areas. We updated the results, where we first present the RNN, to state the inputs explicitly and be clear the inputs are not images or videos of the checkerboard task.

“The RNN input was 4D representing the target configuration and checkerboard signed coherence, while the RNN output was 2D, representing decision variables for a left and right reach (see Methods).”

Another reason that we mapped Area 1 to DLPFC is because anatomical, physiological and lesion studies suggest that DLPFC receives inputs from both the dorsal and ventral stream (Romanski, et, al, 2007; Hoshi, et al, 2006; Wilson, at al, 1993). The dorsal stream originates from the occipital lobe, passes through the posterior parietal cortex, to DLPFC, which carries visuospatial information of the object. The ventral stream originates from the occipital lobe, passes through the inferior temporal cortex, ventrolateral prefrontal cortex to DLPFC, which encodes the identity of the object, including color and texture. In our RNN simulation, Area 1 receives processed inputs of the task: target configuration and the evidence for each color in the checkerboard. Target configuration contains information of the spatial location of the targets, which represents the inputs from the dorsal stream, while evidence for each color by analogy is the input from the ventral stream. Purely visual areas would not fit this dual input from both the dorsal and ventral stream. A potential alternative candidate would be the parietal cortex which is largely part of the dorsal stream and is thought to have modest color inputs (although there is some shape and color selectivity in areas such as LIP, e.g., work from Sereno et al.). On balance given the strong inputs from both the dorsal and ventral stream, we believe Area 1 maps better on to DLPFC than earlier visual areas.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

(1) Line 35/36: Please specify the type of nuisance that the representation is robust to. I guess this refers to small changes in the inputs, not to changes in the representation itself.

Indeed it refers to input variability unrelated to the task. We clarified the text.

(2) For reference, it would be nice to have a tick for the event "Targets on" in Fig.2c.

In this plot, the PSTHs are aligned to the checkerboard onset. Because there is a variable time between target and checkerboard onset, there is a trial-by-trial difference of when the target was turned on, so there is no single place on the x-axis where we could place a “Targets on” tick. In response to this point, we generated a plot with both targets on and check on alignment, with a break in the middle, shown in Supplementary Figure S5. 

(3) It would strengthen the comparison between neural data and RNN if the DPCA components of the RNN areas were shown, as they are shown in Fig.2g,h for the neural data.

We include the PSTHs plotted onto the dPCA components here for Area 1 of the exemplar network. Dashed lines indicate a left reach, while solid lines indicate a right reach, and the color corresponds to the color of the selected target. As expected, we find that the dPCA components capture the separation between components. We emphasize that the trajectory paths along the decoder axes are not particularly meaningful to interpret, except to demonstrate whether variables can be decoded or not (as in Fig 2g,h, comparing DLPFC and PMd). The decoder axes of dPCA are not constrained in any way, in contrast to the readout (encoder) axis (see Methods). This is why our manuscript focuses on analyzing the readout axes. However, if the reviewer strongly prefers these plots to be put in the manuscript, we will add them.   

Author response image 1.

(4) The session-by-session decode analysis presented in Fig.2i suggests that DLPFC has mostly direction information while in Area 1 target information is on top, as suggested by Fig.3g. An additional decoding analysis on trial averaged neural data, i.e. a figure for neural data analogous to Fig.3g,h, would allow for a more straightforward and direct comparison between RNN and neural data. 

We first clarify that we did not decode trial-averaged neural data for either recorded neural data or RNNs. In Fig 3g, h (for the RNN) all decoding was performed on single trial data and then averaged. We have revised the main manuscript to make this clear. Because of this, the mean accuracies we reported for DLPFC and PMd in the text are therefore computed in the same way as the mean accuracies presented in Fig 3g, h. We believe this likely addresses your concern: i.e., the mean decode accuracies presented for both neural data and the RNN were computed the same way. 

If the above paragraph did not address your concern, we also wish to be clear that we presented the neural data as histograms, rather than a mean with standard error, because we found that accuracies were highly variable depending on electrode insertion location. For example, some insertions in DLPFC achieved chance-levels of decoding performance for color and target configuration. For this reason, we prefer to keep the histogram as it shows more information than reporting the mean, which we report in the main text. However, if the reviewer strongly prefers us to make a bar plot of these means, we will add them.

(5) Line 129 mentions an analysis of single trials. But in Fig.2i,j sessions are analyzed. Please clarify.

For each session, we decode from single trials and then average these decoding accuracies, leading to a per-session average decoding accuracy. Note that for each session, we record from different neurons. In the text, we also report the average over the sessions. We clarified this in the text and Methods.

(6) Fig.4c,f show how color and direction axes align with the potent subspaces. We assume that the target axis was omitted here because it highly aligns with the color axis, yet we note that this was not pointed out explicitly.

You are correct, and we revised the text to point this out explicitly.

“We quantified how the color and direction axis were aligned with these potent and null spaces of the intra-areal recurrent dynamics matrix of Area 1 ($\W^1_{rec}$). We did not include the target configuration axis for simplicity, since it highly aligns with the color axis for this network.”

(7) The caption of Fig.4c reads: "Projections onto the potent space of the intra-areal dynamics for each area." Yet, they only show area 1 in Fig.4c, and the rest in a supplement figure. Please refer properly.

Thank you for pointing this out. We updated the text to reference the supplementary figure.

(8) Line 300: "We found the direction axis was more aligned with the potent space and the color axis was more aligned with the null space." They rather show that the color axis is as aligned to the potent space as a random vector, but nothing about the alignments with the null space. Contrarily, on line 379 they write "...with the important difference that color information isn't preferentially projected to a nullspace...". Please clarify.

Thank you for pointing this out. We clarified the text to read: “We found the direction axis was more aligned with the potent space”. The text then describes that the color axis is aligned like a random vector: “In contrast, the color axis was aligned to a random vector.”

(9) Line 313: 'unconstrained' networks are mentioned. What constraints are implied there, Dale's law? Please define and clarify.

Indeed, the constraint refers to Dale’s law constraints. We clarified the text: “Further, we found that W₂₁ in unconstrained 3 area networks (i.e., without Dale's law constraints) had significantly reduced…”

(10) Line 355 mentions a 'feedforward bottleneck'. What does this exactly mean? No E-I feedforward connections, or...? Please define and clarify.

This refers to sparser connections between areas than within an area, as well as a smaller fraction of E-I connections. We clarified the text to read:

“Together, these results suggest  that a connection bottleneck in the form of neurophysiological architecture constraints (i.e., sparser connections between areas than within an area, as well as a smaller fraction of E-I connections) was the key design choice leading to RNNs with minimal color representations and consistent with the information bottleneck principle.”

(11) Fig.5c is supposedly without feedforward connections, but it looks like the plot depicts these connections (i.e. identical to Fig.5b).

In Figure 5, we are varying the E to I connectivity in panel B, and the E-E connectivity in panel C. We vary the feedback connections in Supp Fig. S12. We updated the caption accordingly. 

(12) For reference, it would be nice to have the parameters of the exemplar network indicated in the panels of Fig.5.

We updated the caption to reference the parameter configuration in Table 1 of the Appendix.

(13) Line 659: incomplete sentence

Thank you for pointing this out. We removed this incomplete sentence.

(14) In the methods section "Decoding and Mutual information for RNNs" a linear neural net decoder as well as a nonlinear neural net decoder are described, yet it was unclear which one was used in the end.

We used the nonlinear network, and clarified the text accordingly. We obtained consistent conclusions using a linear network, but did not include these results in the text. (These are reported in Fig. 11 of Kleinman et al, 2021). Moreover, we also obtain consistent results by using an SVM decoder in Fig. S4 for our exemplar parameter configuration.

(15) In the discussion, the paragraph starting from line 410 introduces a new set of results along with the benefits of minimal representations. This should go to the results section.

We prefer to leave this as a discussion, since the task was potentially too simplistic to generate a clear conclusion on this matter. We believe this remains a discussion point for further investigation.

(16) Fig S5: hard to parse. Show some arrows for trajectories (a) (d) is pretty mysterious: where do I see the slow dynamics?

Slow points are denoted by crosses, which forms an approximate line attractor. We clarified this in the caption.

Reviewer #2 (Recommendations For The Authors):

Minor recommendations (not ordered by importance)

(1) Be more explicit that the recordings come from different monkeys and are not simultaneously recorded. For instance, say 'recordings from PFC or PMD'. Say early on that PMD recordings come from two monkeys and that PFC recordings come from 1 of those monkeys. Furthermore, I would highlight which datasets are novel and which are not. For instance, I believe the PFC dataset is a previously unpublished dataset and should be highlighted as such.

We added: “The PMd data was previously described in a study by Chandrasekaran and colleagues” to the main text which clarifies that the PMd data was previously recorded and has been analyzed in other studies.

(2) I personally feel that talking about 'optimal', as is done in the abstract, is a bit of a stretch for this simple task.

In using the terminology “optimal,” we are following the convention of IB literature that optimal representations are sufficient and minimal. The term “optimal” therefore is task-specific; every task will have its own optimal representation. We clarify in the text that this definition comes from Machine Learning and Information Theory, stating:

“The IB principle defines an optimal representation as a representation that is minimal and sufficient for a task or set of tasks.”

In this way, we take an information-theoretic view for describing multi-area representations. This view was satisfactory for explaining and reconciling the multi-area recordings and simulations for this task, and we think it is helpful to provide a normative perspective for explaining the differences in cortical representations by brain area. Even though the task is simple, it still allows us to study how sensory/perceptual information is represented, and well as how choice-related information is being represented.

(3) It is mentioned (and even highlighted) in the abstract that we don't know why the brain distributes computations. I agree with that statement, but I don't think this manuscript answers that question. Relatedly, the introduction mentions robustness as one reason why the brain would distribute computations, but then raises the question of whether there is 'also a computational benefit for distributing computations across multiple areas'. Isn't the latter (robustness) a clear 'computational benefit'?

We decided to keep the word “why” in the abstract, because this is a generally true statement (it is unclear why the brain distributes computation) that we wish to convey succinctly, pointing to the importance of studying this relatively grand question (which could only be fully answered by many studies over decades). We consider this the setting of our work. However, to avoid confusion that we are trying to give a full answer to this question, we are now more precise in the first paragraph of our introduction as to the particular questions we ask that will take a step towards this question. In particular, the first paragraph now asks these questions, which we answer in our study.

“For example, is all stimuli and decision-related information present in all brain areas, or do the cortical representations differ depending on their processing stage? If the representations differ, are there general principles that can explain why the cortical representations differ by brain area?”

We also removed the language on robustness, as we agree it was confusing. Thank you for these suggestions. 

(4) Figure 2e and Fig. 3d, left, do not look very similar. I suggest zooming in or rotating Figure 2 to highlight the similarities. Consider generating a baseline CCA correlation using some sort of data shuffle to highlight the differences.

The main point of the trajectories is to demonstrate that both Area 1 and DLPFC represent both color and direction. We now clarify this in the manuscript. However, we do not intend for these two plots to be a rigorous comparison of similarity. Rather, we quantify similarity using CCA and our decoding analysis. We also better emphasize the relative values of the CCA, rather than the absolute values.

(5) Line 152: 'For this analysis, we restricted it to sessions with significant decode accuracy with a session considered to have a significant decodability for a variable if the true accuracy was above the 99th percentile of the shuffled accuracy for a session.' Why? Sounds fishy, especially if one is building a case on 'non-decodability'. I would either not do it or better justify it.

The reason to choose only sessions with significant decoding accuracy is that we consider those sessions to be the sessions containing information of task variables. In response to this comment, we also now generate a plot with all recording sessions in Supplementary Figure S7. We modified the manuscript accordingly.

“For this analysis, we restricted it to sessions with significant decode accuracy with a session considered to have a significant decodability for a variable if the true accuracy was above the 99th percentile of the shuffled accuracy for a session. This is because these sessions contain information about task variables. However, we also present the same analyses using all sessions in Fig. S7.”

(6) Line 232: 'The RNN therefore models many aspects of our physiological data and is therefore'. Many seems a stretch?

We changed “many” to “key.”

(7) The illustration in Fig. 4B is very hard to understand, I recommend removing it.

We are unsure what this refers to, as Figure 4B represents data of axis overlaps and is not an illustration. 

(8) At some point the authors use IB instead of information bottleneck (eg line 288), I would not do it.

We now clearly write that IB is an abbreviation of Information Bottleneck the first time it is introduced.  

(9) Fig. 5 caption is insufficient to understand it. Text in the main document does not help. I would move most part of this figure, or at least F, to supplementary. Instead, I would move the results in S11 and S10 to the main document.

We clarified the caption to summarize the key points. It now reads: 

“Overall, neurophysiological architecture constraints in the form of multiple areas, sparser connections between areas than within an area, as well as a smaller fraction of E-I connections lead to a minimal color representation in the last area.”

(10) Line 355: 'Together, these results suggest that a connection bottleneck in the form of neurophysiological architecture constraints was the key design choice leading to RNNs with minimal color representations and consistent with the information bottleneck principle.' The authors show convincingly that increased sparsity leads to the removal of irrelevant information. There is an alternative model of the communication subspace hypothesis that uses low-rank matrices, instead of sparse, to implement said bottlenecks (https://www.biorxiv.org/content/10.1101/2022.07.21.500962v2)

We thank the reviewer for pointing us to this very nice paper. Indeed, a low-rank connectivity matrix is another mechanism to limit the amount of information that is passed to subsequent areas. In fact, the low-rank matrix forms a hard-version of our observations as we found that task-relevant information was preferentially propagated along the top singular mode of the inter-areal connectivity matrix. In our paper we observed this tendency naturally emerges through training with neurophysiological architecture constraints. In the paper, for the multi-area RNN, they hand-engineered the multi-area network, whereas our network is trained. We added this reference to our discussion. 

Thank you for your helpful and constructive comments.

Author response:

The following is the authors’ response to the original reviews.

Response to Public Reviewer Comments:

Reviewer 1:

In this work, Veseli et al. present a computational framework to infer the functional diversity of microbiomes in relation to microbial diversity directly from metagenomic data. The framework reconstructs metabolic modules from metagenomes and calculates the per-population copy number of each module, resulting in the proportion of microbes in the sample carrying certain genes. They applied this framework to a dataset of gut microbiomes from 109 inflammatory bowel disease (IBD) patients, 78 patients with other gastrointestinal conditions, and 229 healthy controls. They found that the microbiomes of IBD patients were enriched in a high fraction of metabolic pathways, including biosynthesis pathways such as those for amino acids, vitamins, nucleotides, and lipids. Hence, they had higher metabolic independence compared with healthy controls. To an extent, the authors also found a pathway enrichment suggesting higher metabolic independence in patients with gastrointestinal conditions other than IBD indicating this could be a signal for a general loss in host health. Finally, a machine learning classifier using high metabolic independence in microbiomes could predict IBD with good accuracy. Overall, this is an interesting and well-written article and presents a novel workflow that enables a comprehensive characterization of microbiome cohorts.

We thank the reviewer for their interest in our study, their summary of its findings, and their kind words about the manuscript quality.

Reviewer 2:

This study builds upon the team's recent discovery that antibiotic treatment and other disturbances favour the persistence of bacteria with genomes that encode complete modules for the synthesis of essential metabolites (Watson et al. 2023). Veseli and collaborators now provide an in-depth analysis of metabolic pathway completeness within microbiomes, finding strong evidence for an enrichment of bacteria with high metabolic independence in the microbiomes associated with IBD and other gastrointestinal disorders. Importantly, this study provides new open-source software to facilitate the reconstruction of metabolic pathways, estimate their completeness and normalize their results according to species diversity. Finally, this study also shows that the metabolic independence of microbial communities can be used as a marker of dysbiosis. The function-based health index proposed here is more robust to individuals' lifestyles and geographic origin than previously proposed methods based on bacterial taxonomy.

The implications of this study have the potential to spur a paradigm shift in the field. It shows that certain bacterial taxa that have been consistently associated with disease might not be harmful to their host as previously thought. These bacteria seem to be the only species that are able to survive in a stressed gut environment. They might even be important to rebuild a healthy microbiome (although the authors are careful not to make this speculation).

This paper provides an in-depth discussion of the results, and limitations are clearly addressed throughout the manuscript. Some of the potential limitations relate to the use of large publicly available datasets, where sample processing and the definition of healthy status varies between studies. The authors have recognised these issues and their results were robust to analyses performed on a per-cohort basis. These potential limitations, therefore, are unlikely to have affected the conclusions of this study.

Overall, this manuscript is a magnificent contribution to the field, likely to inspire many other studies to come.

We thank the reviewer for their endorsement of our study and their precision regarding the evaluation of its strengths. We also appreciate their high expectations for its impact in the field.

Reviewer 3:

The major strength of this manuscript is the "anvi-estimate-metabolism' tool, which is already accessible online, extensively documented, and potentially broadly useful to microbial ecologists.

We thank the reviewer for their recognition of the computational advances in this study. We also thank the reviewer for their suggestions that we have addressed below, which allowed us to strengthen our manuscript.

However, the context for this tool and its validation is lacking in the current version of the manuscript. It is unclear whether similar tools exist; if so, it would help to benchmark this new tool against prior methods.

The reviewer brings up a very good point about the lack of context for the `anvi-estimate-metabolism` program. While our efforts that led to the emergence of this software included detailed benchmarking efforts, a formal assessment of its performance and accuracy was indeed lacking. We are thankful for our reviewer to point this out, which motivated us to perform additional analyses to address such concerns. Our revision contains a new, 34-page long supplementary information file (Supplementary File 2) that includes a section titled “Comparison of anvi-estimate-metabolism to existing tools for metabolism reconstruction”. The text therein describes the landscape of currently available software for metabolism reconstruction and describes the features that make `anvi-estimate-metabolism` unique – namely, (1) its implementation of metrics that make it suitable for metagenome-level analyses (i.e., pathway copy number and stepwise interpretation of pathway definitions) and (2) its ability to process user-defined metabolic pathways rather than exclusively relying on KEGG. As described in that section, there is currently no other tool that can compute copy numbers of metabolic pathways from metagenomic data. Hence, it is not quite possible to benchmark the copy number methodology used in our study against prior methods; however, our benchmarking of this functionality with synthetic genomes and metagenomes (described later in this document) does provide necessary quantitative insights into its accuracy and efficiency.

While comparison of the copy number calculations to other tools was not possible due to the unique nature of this functionality, it was possible to benchmark our gene function annotation methodology against existing tools that also annotate genes with KEGG KOfams, which is a step commonly used by various tools that aim to estimate metabolic potential in genomes and metagenomes. In the anvi’o software ecosystem the annotation of genes for metabolic reconstruction is implemented in `anvi-run-kegg-kofams`, and represents a step that is required by `anvi-estimate-metabolism`. As our comparisons were quite extensive and involved additional researchers, we described them in another study which we titled “Adaptive adjustment of significance thresholds produces large gains in microbial gene annotations and metabolic insights” (doi:10.1101/2024.07.03.601779) that is now cited from within our revision in the appropriate context. Briefly, our comparison of anvi’o, Kofamscan, and MicrobeAnnotator using 396 publicly-available bacterial genomes from 11 families demonstrated that `anvi-run-kegg-kofams` is able to identify an average of 12.8% more KO annotations per genome than the other tools, especially in families commonly found in the gut environment (Figure 1). Furthermore, anvi’o recovered the highest proportion of annotations that were independently validated using eggNOG-mapper. Our comparisons also showed that annotations from anvi’o yield at least 11.6% more complete metabolic modules than Kofamscan or MicrobeAnnotator, including the identification of butyrate biosynthesis in Lachnospiraceae genomes at rates similar to manual identification of this pathway in this clade (Figure 2a). Overall, our findings that are now described extensively in DOI:10.1101/2024.07.03.601779 show that our method captures high-quality annotations for accurate downstream metabolism estimates.

We hope these new data help increase the reviewer’s confidence in our results.

Simulated datasets could be used to validate the approach and test its robustness to different levels of bacterial richness, genome sizes, and annotation level.

We thank the reviewer for this suggestion. It was an extremely useful exercise that not only helped us elucidate the nuances of our approach, but also enabled us to further highlight its strengths in our manuscript. We created simulated datasets including a total of 409 synthetic metagenomes that we used to test the robustness of our approach to different genome sizes, community sizes, and levels of diversity. Overall, our tests with these synthetic metagenomes demonstrated that our approach of computing PPCN values to summarize the metabolic capacity within a metagenomic community is accurate and robust to differences in all three critical variables. Most of these variables were weakly correlated between PPCN or PPCN accuracy, and the few correlations that were stronger in fact further supported our original hypothesis that we generated from our comparisons of healthy and IBD gut metagenomes. The methods and results of our validation efforts are explained in detail in our new Supplementary File 2 (see the section titled “Validation of per-population copy number (PPCN) approach on simulated metagenomic data”), but we copy here the subsection that summarizes our findings for the reviewer’s convenience:

Overall impact on the comparison between healthy and IBD gut metagenomes

“In summary, our validation strategy revealed good accuracy at estimating metagenome-level metabolic capacity relative to our genome-level knowledge in the simulated data. While it often underestimated average genomic completeness by ignoring partial copies of metabolic pathways and often overestimated average genomic copy number due to the effect of pathway complementarity between different community members, the magnitude of error was overall limited in range and the error distributions were centered at or near 0. Furthermore, we observed these broad error trends in all cases we tested, and therefore we expect that they would also apply to both sample groups in our comparative analysis. Thus, we next considered how the PPCN approach might have influenced our analyses that considered metagenomes from healthy individuals and from those who have IBD – two groups that differed from one another with respect to some of the variables considered in our tests.

Most of the correlations between PPCN or PPCN accuracy and sample parameters were weak, yet significant (Table 1). They showed that community size and diversity level have limited influence on the PPCN calculation, while genome size does not influence its accuracy. The only exception was the moderate correlation between PPCN and genome size, particularly for the subset of IBD-enriched pathways. It was a negative correlation with the proportion of small genomes in a metagenome, indicating that PPCN values for these pathways are larger when there are more large genomes in the community and suggesting that these pathways tend to occur frequently in larger genomes. This is in line with our observation that IBD communities contain more large genomes and therefore confirms our interpretation that the populations surviving in the IBD gut microbiome are those with the genomic space to encode more metabolic capacities.

If we consider even the weak correlations, two of those relationships indicate that our approach would be more accurate for IBD metagenomes than for healthy metagenomes. For instance, PPCN accuracy was slightly higher for smaller communities (as in IBD samples), with a weakly positive correlation between PPCN error and community size. It was also slightly more accurate for less diverse communities (as in IBD samples), with a weakly positive correlation between PPCN error and number of phyla. The only opposing trend was the weakly positive correlation between PPCN error and proportion of smaller genomes, which favors higher accuracy in communities with smaller genomes (as in healthy samples). Given that our analysis focuses on the pathways enriched in IBD samples, an overall higher accuracy in IBD samples would increase the confidence in our enrichment results.

We also examined the accuracy of our method to predict the number of populations within a metagenome based on the distribution and frequency of single-copy core genes (i.e., the denominator in the calculation of PPCN). Our benchmarks show that the estimates are overall accurate, where most errors reflect a negligible amount of underestimations of the actual number of populations. Errors occurred more frequently for the realistic synthetic assemblies generated from simulated short read data than for the ideal synthetic assemblies generated from the combination of genomic contigs. The correlations between estimation accuracy and sample parameters indicated that the population estimates are more accurate for smaller communities and communities with more large genomes, as in IBD samples (Table 2). Thus, this method is more likely to underestimate the community size in healthy samples, and these errors could lead to overestimation of PPCN in healthy samples relative to IBD samples. Thus, the enrichment of a given pathway in the IBD samples would have to overcome its relative overestimation in the healthy sample group, making it more likely that we identified pathways that were truly enriched in the IBD communities.

Overall, the consideration of our simulations in the context of healthy vs IBD metagenomes suggest that slight biases in our estimates as a function of unequal diversity with sample groups should have driven PPCN calculations towards a conclusion that is opposite of our observations under neutral conditions. Thus, clear differences between healthy vs IBD metagenomes that overcome these biases suggest that    biology, and not potential bioinformatics artifacts, is the primary driver of our observations.”

Accordingly, we have added the following sentence summarizing the validation results to our paper:

“Our validation of this method on simulated metagenomic data demonstrated that it is accurate in capturing metagenome-level metabolic capacity relative to genome-level metabolic capacity estimated from the same data (Supplementary File 2, Supplementary Table 6).”

Early in this process of validation, we identified and fixed two minor bugs in our codebase. The bugs did not affect the results of our paper and therefore did not warrant a re-analysis of our data. The first bug, which is detailed in the Github issue https://github.com/merenlab/anvio/issues/2231 and fixed in the pull request https://github.com/merenlab/anvio/pull/2235, led to the overestimation of the number of microbial populations in a metagenome when the metagenome contains both Bacteria and Archaea. None of the gut metagenomes analyzed in our paper contained archaeal populations, so this bug did not affect our community size estimates.

The second bug, which is detailed in the Github issue https://github.com/merenlab/anvio/issues/2217 and fixed in the pull request https://github.com/merenlab/anvio/pull/2218, caused inflation of stepwise copy numbers for a specific type of metabolic pathway in which the definition contained an inner parenthetical clause. This bug affected only 3 pathways in the KEGG MODULE database we used for our analysis, M00083, M00144, and M00149. It is worth noting that one of those pathways, M00083, was identified as an IBD-enriched module in our analysis. However, the copy number inflation resulting from this bug would have occurred equivalently in both the healthy and IBD sample groups and thus should not have impacted our comparative analysis.

Regardless, we are grateful for the suggestion to validate our approach since it enabled us to identify and eliminate these minor issues.

The concept of metabolic independence was intriguing, although it also raises some concerns about the overinterpretation of metagenomic data. As mentioned by the authors, IBD is associated with taxonomic shifts that could confound the copy number estimates that are the primary focus of this analysis. It is unclear if the current results can be explained by IBD-associated shifts in taxonomic composition and/or average genome size. The level of prior knowledge varies a lot between taxa; especially for the IBD-associated gamma-Proteobacteria.

The reviewer brings up an important point, and we are thankful for the opportunity to clarify the impact of taxonomy on our analysis. Though IBD has been associated with taxonomic shifts in the gut microbiome, a major problem with such associations is that the taxonomic signal is extremely variable, leading to inconsistency in the observed shifts across different studies (doi:https://doi.org/10.3390/pathogens8030126). Indeed, one of the most comprehensive prior studies into this topic demonstrated that inter-individual variation is the largest contributor to all multi-omic measurements aiming to differentiate between the gut microbiome of individuals with IBD from that of healthy individuals, including taxonomy (doi:10.1038/s41586-019-1237-9). We therefore took a different approach to study this question that is independent of taxonomy, by focusing on metabolic potential estimated directly from metagenomes to elucidate an ecological explanation behind the reduced diversity of the IBD gut microbiome, which studies of taxonomic composition alone are not able to provide. Furthermore, the variability inherent to taxonomic profiles of the gut microbiome makes it unlikely that taxonomic shifts could confound our analysis, especially given our large sample set encompassing a variety of individuals with different origins, ages, and genders.

We agree with the reviewer that our level of prior knowledge varies substantially across taxa. Regardless, the only prior knowledge with any bearing on our ability to estimate metabolic capacity in a taxonomy-independent manner is the extent of sequence diversity captured by our annotation models for the enzymes used in metabolic pathways. During our analysis, we had observed that metagenomes in the healthy group had fewer gene annotations than those in the IBD group and we therefore shared the reviewer’s concern about potential annotation bias, whereby less-studied genomes are not always incorporated into the Hidden Markov Models for annotating KEGG Orthologs, perhaps making it more likely for us to miss annotations in these genomes (and leading to lower completeness scores for metabolic pathways in the healthy samples). Our annotation method partially addresses this limitation by taking a second look at any unannotated genes and mindfully relaxing the bit score similarity thresholds to capture annotations for any genes that are slightly too different from reference sequences for annotation with default thresholds. As mentioned previously, our recent preprint demonstrates the efficacy of this strategy (doi:10.1101/2024.07.03.601779). To further address this concern, we also investigated the extent of distant homology in these metagenomes using AGNOSTOS (doi:https://doi.org/10.7554/eLife.67667), which showed a higher proportion of unknown genes in the healthy metagenomes and suggested that a substantial portion of the unannotated genes are not distant homologs of known enzymes that we failed to annotate due to lack of prior knowledge about them, but rather are completely novel functions. To describe these results, we added the following paragraph and two accompanying figures (Supplementary Figure 4g-h) to the section “Differential annotation efficiency between IBD and Healthy samples” in Supplementary File 1:

“To understand the potential origins of the reduced annotation rate in healthy metagenomes, we ran AGNOSTOS (Vanni et al. 2022) to classify known and unknown genes within the healthy and IBD sample groups. AGNOSTOS clusters genes to contextualize them within an extensive reference dataset and then categorizes each gene as ‘known’ (has homology to genes annotated with Pfam domains of known function), ‘genomic unknown’ (has homology to genes in genomic reference databases that do not have known functional domains), or ‘environmental unknown’ (has homology to genes from metagenomes or MAGs that do not have known functional domains). The resulting classifications confirm that healthy metagenomes contain fewer ‘known’ genes than metagenomes in the IBD sample group – the proportion of ‘known’ genes classified by AGNOSTOS is about 3.0% less in the healthy metagenomes than in the IBD sample group, which is similar to the ~3.5% decrease in the proportion of ‘unannotated’ genes observed by simply counting the number of genes with at least one functional annotation (Supplementary Figure 4g-h, Supplementary Table 1e). Furthermore, the majority of the unannotated genes in either sample group were categorized by AGNOSTOS as ‘genomic unknown’ (Supplementary Figure 4g), suggesting that the unannotated sequences are genes without biochemically-characterized functions currently associated with them and are thus legitimately lacking a functional annotation in our analysis, rather than representing distant homologs of known protein families that we failed to annotate. Based upon the classifications, a systematic technical bias is unlikely driving the annotation discrepancy between the sample groups.”

Furthermore, we have already discussed this limitation and its implications in our manuscript (see section “Key biosynthetic pathways are enriched in microbial populations from IBD samples”). To further clarify that our approach is independent of taxonomy, we have now also amended the following statement in our introduction:

“Here we implemented a high-throughput, taxonomy-independent strategy to estimate metabolic capabilities of microbial communities directly from metagenomes and investigate whether the enrichment of populations with high metabolic independence predicts IBD in the human gut.”

Finally, the reviewer is also correct that genome size is a part of the equation, as genome size and level of metabolic capacity are inextricable. In fact, we observed this in our analysis, as already stated in our paper:

“HMI genomes were on average substantially larger (3.8 Mbp) than non-HMI genomes (2.9 Mbp) and encoded more genes (3,634 vs. 2,683 genes, respectively)”

Since larger genomes have the space to encode more functional capacity, it follows that having higher metabolic independence would require a microbe to have a larger genome. The validation of our method on simulated metagenomic data supported this idea by demonstrating that the IBD-enriched metabolic pathways are commonly identified in large genomes. The validation also proved that genome size does not influence the accuracy of our approach (Supplementary File 2).

It can be difficult to distinguish genes for biosynthesis and catabolism just from the KEGG module names and the new normalization tool proposed herein markedly affects the results relative to more traditional analyses.

We agree with the reviewer that KEGG module names do not clearly indicate the presence of biosynthetic genes of interest. That said, KEGG is a commonly-used and extensively-curated resource, and many biologists (including ourselves) trust their categorization of genes into pathways. We hope that readers who are interested in specific genes within our results would make use of our publicly-available datasets (which include gene annotations) to conduct a targeted analysis based on their expertise and research question.

However, we would like to respectfully note that the ability to distinguish the genes within each KEGG module may not be very useful to most readers, and is unlikely to have a meaningful impact in our findings. As the reviewer most likely appreciates, the presence of individual genes in isolation can be insufficient to indicate biosynthetic capacity, considering that 1) most biosynthetic pathways involve several biochemical conversions requiring a series of enzymes, 2) enzymes are often multi-functional rather than exclusive to one pathway, and 3) different organisms in a community may utilize enzymes encoded by different genes to perform the same or similar biochemical reaction in a pathway. We therefore made the choice to analyze metabolic capacity at the pathway level, because this would better reflect the biosynthetic abilities encoded by the multiple microbial populations within each metagenome.

The reviewer also suggests that our novel normalization method affects our results, yet we believe that this normalization strategy is one of the strengths of our study in comparison to ‘more traditional analyses’ as it enables an appropriate comparison between metagenomes describing microbial communities of dramatically different degrees of richness. Indeed, we suspect that the lack of normalization in more traditional analyses may be one reason why prior analyses have so far failed to uncover any mechanistic explanation for the loss of diversity in the IBD gut microbiome. We hope that our validation efforts were sufficiently convincing in demonstrating the suitability of our approach, and copy here a particularly illuminating section of the validation results that we have added to Supplementary Information File 2:

“As expected, we observed a significant positive correlation between metagenomic copy number (the numerator of PPCN) and community size in each group, likely driven by the increase in the copy number of core metabolic pathways in larger communities (Supplementary Figure 18). Interestingly, this correlation was much stronger for the subset of IBD-enriched pathways (0.49 <= R <= 0.67) than for all modules (0.12 <= R <=0.13).

“However, the correlation was much weaker and often nonsignificant for the normalized PPCN data in both groups of modules (all modules: 0.01 < R < 0.04, enriched modules: 0.04 < R < 0.09, Supplementary Table 6b, Supplementary Figure 19), which demonstrates the suitability of our normalization method to remove the effect of community size in comparisons of metagenome-level metabolic capacity.”

As such, it seems safer to view the current analysis as hypothesis-generating, requiring additional data to assess the degree to which metabolic dependencies are linked to IBD.

We certainly agree with the reviewer that our study, similar to the vast majority of studies published every year, is a hypothesis-generating work. Any idea proposed in any scientific study in life sciences will certainly benefit from additional data analyses, and therefore we respectfully do not accept this as a valid criticism of our work. The inception of this study is linked to an earlier work that hypothesized high metabolic independence as a determinant of microbial fitness in stressed gut communities (doi:10.1186/s13059-023-02924-x), which lacked validation on larger sets of data. Our study tests this original hypothesis using a large number of metagenomes, and lends further support for it with approaches that are now better validated. Furthermore, there are other studies that agree with our interpretation of the data (doi:10.1101/2023.02.17.528570, doi:10.1038/s41540-021-00178-6), and we look forward to more computational and/or experimental work in the future to generate more evidence to evaluate these insights further.

Response to Recommendations for the Authors

Reviewer 1:

My main comments include:

- From the results reported in lines 178-185, it seems that metabolic pathways in general were enriched in IBD microbiomes, not specifically biosynthetic pathways. Can we really say then that the signal is specific for biosynthesis capabilities?

We apologize for the confusion here. When we read the text again, we ourselves were confused with our phrasing.

The reviewer is correct that a similar proportion of both biosynthetic and non-biosynthetic pathways had elevated per-population copy number (PPCN) values in the IBD samples. However, the low microbial diversity associated with IBD and the on average larger genome size of individual populations contributes to this relative enrichment of the majority of metabolic modules. To remove this bias and identify specific modules whose enrichment was highly conserved across microbial populations associated with IBD, we implemented two criteria: 1) we selected modules that passed a high statistical significance threshold in our enrichment test (Wilcoxon Rank Sum Test, FDR-adjusted p-value < 2e-10), and 2) we accounted for effect size by ranking these modules according to the difference between their median PPCN in IBD samples and their median PPCN in healthy samples, and keeping only those in the top 50% (which translated to an effect size threshold of > 0.12).

This analysis revealed a set of metabolic modules that were consistently and highly significantly enriched in microbial communities associated with IBD. The majority of these metabolic modules encode biosynthesis pathways. Our use of the terms “elevated”, “enriched”, and “significantly enriched” in the previous version of the text was confusing to the reader. We thank the reviewer for pointing this out, and we hope that our revision of the text clarifies the analysis strategy and observations:

“To gain insight into potential metabolic determinants of microbial survival in the IBD gut environment, we assessed the distribution of metabolic modules within samples from each group (IBD and healthy) with and without using PPCN normalization. Without normalizing, module copy numbers were overall higher in healthy samples (Figure 2a) and modules exhibited weak differential occurrence between cohorts (Figure 2b, 2c, Supplementary Figure 3). The application of PPCN reversed this trend, and most metabolic modules were elevated in IBD (Supplementary Figure 5). This observation is influenced by two independent aspects of the healthy and IBD microbiota. The first one is the increased representation of microbial organisms with smaller genomes in healthy individuals (Watson et al. 2023), which increases the likelihood that the overall copy number of a given metabolic module is below the actual number of populations. In contrast, one of the hallmarks of the IBD microbiota is the generally increased representation of organisms with larger genomes (Watson et al. 2023). The second aspect is that the generally higher diversity of microbes in healthy individuals increases the denominator of the PPCN. This results in a greater reduction in the PPCN of metabolic modules that are not shared across all members of the diverse gut microbial populations in health.

To go beyond this general trend and identify modules that were highly conserved in the IBD group, we first selected those that passed a relatively high statistical significance threshold in our enrichment test (Wilcoxon Rank Sum Test, FDR-adjusted p-value < 2e-10). We then accounted for effect size by ranking these modules according to the difference between their median PPCN in IBD samples and their median PPCN in healthy samples, and keeping only those in the top 50% (which translated to an effect size threshold of > 0.12). This stringent filtering revealed a set of 33 metabolic modules that were significantly enriched in metagenomes obtained from individuals diagnosed with IBD (Figure 2d, 2e), 17 of which matched the modules that were associated with high metabolic independence previously (Watson et al. 2023) (Figure 2f). This result suggests that the PPCN normalization is an important step in comparative analyses of metabolisms between samples with different levels of microbial diversity.”

Lines 178-185 from our original submission have been removed to avoid further confusion. These results can be found in Supplementary File 1 (section “Module enrichment without consideration of effect size leads to nonspecific results”).

It is not entirely clear to me what is meant by PPCN normalization. Normalize the number of copy numbers to the overall number of genes?

The idea behind using per-population copy number (PPCN) is to normalize the prevalence of each metabolic module found in an environment with the number of microbial populations within the same sample. PPCN achieves this by dividing the pathway copy numbers by the number of microbial populations in a given metagenome, which we estimate from the frequency of bacterial single-copy core genes. We have updated the description of the per-population copy number (PPCN) calculation to clarify its use:

“Briefly, the PPCN estimates the proportion of microbes in a community with a particular metabolic capacity (Figure 1, Supplementary Figure 2) by normalizing observed metabolic module copy numbers with the ‘number of microbial populations in a given metagenome’, which we estimate using the single-copy core genes (SCGs) without relying on the reconstruction of individual genomes.”

We also note that the equation for PPCN is shown in Figure 1.

It is also not clear to me how the classifier predicts stress on microbiomes rather than dysbiosis.

The reviewer asks an interesting question since it is true that we could also use the term “dysbiosis” rather than “stress”. Yet we refrained from the use of dysbiosis as it is considered a poorly-defined term to describe an altered microbiome often associated with a specific disease (doi:https://doi.org/10.3390/microorganisms10030578), such as IBD, relative to another poorly-defined state, “healthy microbiome” (doi:https://doi.org/10.1002/phar.2731). We do consider that stress is not necessarily a term that is less vague than dysbiosis, yet it has the advantage of being more common in studies of ecology compared to dysbiosis. Our relatively neutral stance towards which term to use has shifted dramatically due to one critical observation in our study: the identical patterns of enrichment of HMI microbes in individuals diagnosed with IBD as well as in healthy individuals treated with antibiotics. We appreciate that the observed changes in the antibiotics case can also fulfill the definition of “dysbiosis”, but the term “stress response” more accurately describes what the classifier identifies in our opinion.

What is the advantage of using the estimate-metabolism pipeline presented in this article over workflows such as those using genome-scale models, which are repeatedly cited and discussed?

Genome-scale models are often appropriate for a big-picture view of metabolism, and especially when the capability to perform quantitative simulations like flux-balance analysis is needed. For our investigation, we wanted a more specific and descriptive summary of metabolic capacity, so we focused on individual KEGG modules, which qualitatively describe subsets of the vast metabolic network with pathway names that all readers can understand, rather than working with an abstract model of the entire network. Furthermore, genome-scale models would have prevented us from assessing the redundancy (copy number) of metabolic pathways, as these networks usually focus on the presence-absence of gene annotations for enzymes in the network rather than the copy number of these annotations. The copy number metric has been critical for our analyses, considering that we are focusing on metabolic capacity at the community level and require the ability to normalize this metabolic capacity by the size of the community described by each metagenome. Finally, assessing a discrete set of metabolic pathways yielded a corresponding set of features that we used to create the machine learning classifier, whereas data from genome-scale models would not be as easily transferable into classifier features.

Minor comments:

Figure 2d and e are mentioned in the text before Figure 2a.

We thank the reviewer for catching this. We have rewritten the section as follows to put the figure references in numerical order:

!To gain insight into potential metabolic determinants of microbial survival in the IBD gut environment, we assessed the distribution of metabolic modules within samples from each group (IBD and healthy) with and without using PPCN normalization. Without normalizing, module copy numbers were overall higher in healthy samples (Figure 2a) and modules exhibited weak differential occurrence between cohorts (Figure 2b, 2c, Supplementary Figure 3). After the application of PPCN, most metabolic modules were elevated in IBD (Supplementary Figure 5). This observation is a product of two independent aspects of the healthy and IBD microbiota. The first one is the increased representation of microbial organisms with smaller genomes in healthy individuals (Watson et al. 2023), which increases the likelihood that the overall copy number of a given metabolic module is below the actual number of populations. In contrast, one of the hallmarks of the IBD microbiota is the generally increased representation of organisms with larger genomes (Watson et al. 2023). The second aspect is that the generally higher diversity of microbes in healthy individuals increases the denominator of the PPCN due to the higher number of populations detected in these samples. This results in a greater reduction in the PPCN of metabolic modules that are not shared across all members of the diverse gut microbial populations in health. To go beyond this general trend and identify modules that were highly conserved in the IBD group, we first selected those that passed a relatively high statistical significance threshold in our enrichment test (Wilcoxon Rank Sum Test, FDR-adjusted p-value <2e-10). We then accounted for effect size by ranking these modules according to the difference between their median PPCN in IBD samples and their median PPCN in healthy samples, and keeping only those in the top 50% (which translated to an effect size threshold of > 0.12). This stringent filtering revealed a set of 33 metabolic modules that were significantly enriched in metagenomes obtained from individuals diagnosed with IBD (Figure 2d, 2e), 17 of which matched the modules that were associated with high metabolic independence previously (Watson et al. 2023) (Figure 2f). This result suggests that the PPCN normalization is an important step in comparative analyses of metabolisms between samples with different levels of microbial diversity.!

How much preparation is needed for users that want to apply the estimate-metabolism pipeline to their own datasets? From the documentation at anvi'o, it still seems like a significant effort.

We thank the reviewer for this important question. The use of anvi-estimate-metabolism is simple, but the concept it makes available and the means it offers its users to interact with their data are not basic, thus its use requires some effort. Anvi’o provides users with the ability to directly interact with their data at each step of the analysis to have full control over the analysis and to make informed decisions on the way. In comparison to pre-defined analysis pipelines that often require no additional input from the user, this approach requires some level of involvement of the user throughout the process – namely, they must run a few programs in series rather than running just one pipeline command that quietly handles everything on their behalf. The most basic workflow for using `anvi-estimate-metabolism` is quite straightforward and requires four simple steps following the installation of anvi’o: 1. Run the program `anvi-setup-kegg-data` to download the KEGG data. 2. Convert the assembly FASTA file into an anvi’o-compatible database format with gene calls by running `anvi-gen-contigs-database`. 3. Annotate genes with KOs with the program `anvi-run-kegg-kofams`. 4. Get module completeness scores and copy numbers by running `anvi-estimate-metabolism`. In addition, we provide simple tutorials (such as the one at https://anvio.org/tutorials/fmt-mag-metabolism/) and reproducible bioinformatics workflows online (including for this study at https://merenlab.org/data/ibd-gut-metabolism/) which helps early career researchers to apply similar strategies to their own datasets. We are happy to report that we have been using this tool in our undergraduate education, and observed that students with no background in computation were able to apply it to their questions without any trouble.

Reviewer 2:

Congratulations on this great work, the manuscript is a pleasure to read. Minor questions that the authors might want to clarify:

L 275: Why use reference genomes from the GTDB (for only 3 phyla) instead of using MAGs reconstructed from the data? I understand that assemblies based on individual samples would probably not yield enough complete MAGs, but I would expect that co-binning the assemblies for the entire dataset would.

We thank the reviewer for their kind words. We certainly agree that metagenome assembled genomes (MAGs) reconstructed directly from the assemblies would by nature represent the populations in these communities better than reference genomes. However, one of our aims in this study was to avoid the often error-prone and time-consuming step of reconstructing MAGs. Most automatic binning algorithms inevitably make mistakes, and especially for metabolism estimation, low quality MAGs can introduce a bias in the analysis. At the same time the manual curation of each bin to remove any contamination would require a substantial effort and make the workflow less accessible for others to use. As an example, in our previous work (doi:10.1186/s13059-023-02924-x), careful refinement of MAGs from just two co-assemblies took two months. Here, we developed the PPCN workflow as a more scalable, assembly-level analysis to avoid the need for binning in the first place.

To supplement and confirm the metagenome-level results, we decided to run a genome-level analysis. We used the GTDB since it represents the most comprehensive, dereplicated collection of reference genomes across the tree of life. We chose those 3 phyla in particular because of their ecological relevance in the human gut environment. Bacteroidetes and

Firmicutes together represent the majority (up to ~90%) of the populations in healthy individuals (doi:10.1038/nature07540), and Proteobacteria represent the next most abundant phylum on average (2% ± 10%) (doi:10.1371/journal.pone.0206484).

L 403: Should the Franzosa and Papa papers be referenced as numbers?

Thanks for pointing this out. The rogue numerical citation was actually an artifact of the submission and was corrected to a long-format citation in the online version of the manuscript on the eLife website.

Reviewer 3:

The lack of any experimental validation contributes to the tentative nature of the conclusions that can be drawn at this time. Numerous studies have looked at the metabolism of gut bacterial species during in vitro growth, which could be mined to test if the in silico predictions of metabolism can be supported. Alternatively, the authors could isolate key strains of interest and study them in culture or in mouse models of IBD.

We appreciate these suggestions and agree with the reviewer that experimental validation is important. However, we do not agree that either the use of mouse models or the isolation of individual microbial strains would be an appropriate experimental test in this case. The use of humanized gnotobiotic mice has critical limitations (see doi:10.1016/j.cell.2019.12.025 and references within the section on “human microbiota-associated murine models”). As it is not possible to establish a mouse model whose gut microbiota fully reflect the human gut microbiome, such an approach would neither be appropriate to validate our findings, nor would it have been possible to produce the insights we have gained based on environmental data. We are not sure how exactly a mouse model, even when ignoring the well established limitations, could improve or validate a comprehensive analysis of a large “environmental” datasets that resulted in highly significant signals.

We are also not sure that we understand how the reviewer believes that the isolation of individual strains would aid in validating our findings. While we appreciate that not all relevant genes are captured by the available annotation routines and that some genes may be misannotated, the large dataset used here renders these concerns negligible. Isolating a small subset of bacterial populations would hardly lead to a representative sample and testing their metabolic capacities in vitro would not improve the reliability of our analysis.

Boilerplate suggestions as vague as “isolate key strains of interest” or “experiment in mouse models of IBD” do not add or retract anything from our findings. Our findings and hypotheses are well supported by our data and extensive analyses.

Line 9 - not sure this approach is hypothesis testing in the traditional sense, you might reword.

Hypothesis testing occurs when one makes an observation, develops an hypothesis that explains the observation, and then gathers and analyzes data to investigate whether additional data support or disprove the hypothesis. We are not convinced a reword is necessary.

Line 40 - the lack of consistent differences in IBD and healthy individuals does not mean that the microbiome doesn't impact disease. It's important to consider all the mechanistic studies in animal models and other systems.

Our study does not claim that microbiome has no impact on the course of disease.

Line 50 - this seemed out of place and undercuts the current findings. Upon checking Ref. 31, the analysis seems distinct enough to not mention in the introduction.

We disagree. Ref 31 uses genome-scale metabolic models to identify the loss of cross-feeding interactions in the gut microbiome of individuals with IBD, which is another way of saying that the microbes in IBD no longer rely on their community for metabolic exchange – in other words, they are metabolically independent. This is an independent observation that is parallel to our results and confirms our analysis; hence, it is important to keep in our introduction.

Line 55 - Ref. 32 looked at FMT, which should be explicitly stated here.

The reviewer’s suggestion is not helpful. Ref 32 has a significant focus on IBD as it compares a total of 300 MAGs generated from individuals with IBD to 264 MAGs from healthy individuals and shows differences in metabolic enrichment between healthy and IBD samples independent of taxonomy, thus setting the stage for our current work. What model has been used to generate the initial insights that led to the IBD-related conclusion in Ref 32 has no significance in this context.

Lines 92-107 - this text is out of place in the Results section and reads more like a review article. Please trim it down and move it to the introduction.

We would like to draw the reviewer’s attention to the fact that this is a “Result and Discussion” section. In this specific case it is important for readers to appreciate the context for our new tool, as the reviewer commented in the public review. We kindly disagree with the reviewer’s suggestion to remove this text as that would diminish the context.

Line 107 - is "selection" the word you meant to use?

If the frequency of a given metabolic module remains the same or increases despite the decreasing diversity of the microbial community, it is conceivable to assume that its enrichment indicates the presence of a selective process to which the module responds. It is indeed the word we meant to use.

Line 110 - this is the first mention of this new method, need to add it to the abstract and introduction.

The reviewer must have overlooked the text passages in which we mention the strategy we developed within the abstract:

“Here, we tested this hypothesis on a large scale, by developing a software framework to quantify the enrichment of microbial metabolisms in complex metagenomes as a function of microbial diversity.”

And in the last paragraph of the introduction:

“Here we implemented a high-throughput, taxonomy-independent strategy to estimate metabolic capabilities of microbial communities directly from metagenomes…”

Figure 1 - a nice summary, but no data is shown to support the validity of this model. Consider shrinking the cartoon and adding validation with simulated datasets.

We hope we have addressed this recommendation with the extensive validation efforts summarized above.

Line 134 - need to state the FDR and effect size cutoffs used.

We have reworded this sentence as follows to clarify which thresholds were used:

“We identified significantly enriched modules using an FDR-adjusted p-value threshold of p < 2e-10 and an effect size threshold of > 0.12 from a Wilcoxon Rank Sum Test comparing IBD and healthy samples.”

I'm also concerned about the simple comparison of IBD to healthy without adjusting for confounders like study, geographical location, age, sex, drug use, diet, etc. More text is needed to explain the nature of these data, how much metadata is available, and which other variables distinguish IBD from healthy.

The reviewer is correct that there is a large amount of interindividual variation between samples due to host and environmental factors. However, the lack of adjusting for confounders was intentional, and in fact one of the critical strengths of our study. We observe a clear signal between healthy individuals and individuals diagnosed with IBD, despite the amount of interindividual variation in our diverse set of samples from 13 different studies (details of which are summarized in Supplementary Table 1). The clear increase in predicted metabolic capacity that we consistently observe in IBD patients using both metagenomes and genomes across diverse cohorts points to metabolic independence as a high-level trend that is predictive of microbial prevalence in stressed gut environments irrespective of host factors.

Line 145 - calling PPCN normalization an "essential step" is a huge claim and requires a lot more data to back it up. Might be best to qualify this statement.

We hope we have addressed this recommendation with our validation efforts. Supplementary Figures 18 and 19 in particular show evidence for the necessity of the normalization step. It is indeed an essential step if the purpose is to compare metabolic enrichment between cohorts of highly different microbial diversity.

Figure 2a - the use of a 1:1 trend line seems potentially misleading. I would replace it with a best-fit line.

Our purpose here was not to show the best fit. Instead, the 1:1 trend line separates the modules based on their relative abundance distribution between healthy individuals and individuals diagnosed with IBD. If the module is to the left of the line, it has a higher median copy number in healthy individuals and if the module is to the right, it has a higher median copy number in individuals with IBD. The line also helps to demonstrate the shift that occurs between the unnormalized data in Figure 2a. Without the normalization, more modules occur to the left of the

1/1 line as a result of the higher raw copy numbers in healthy metagenomes which simply contain more microbial populations. With the normalization (Figure 2d), more modules fall on the right side of the 1/1 line due to higher PPCN values. A best-fit line would not serve well for these purposes.

The text should be revised to state that this analysis actually did find many significant differences and to discuss whether they were the same modules identified in Figure 2d.

We apologize for the confusion and thank the reviewer for bringing this issue to our attention. As mentioned above, the disparate levels of microbial diversity between healthy individuals and individuals with IBD resulted in much larger copy numbers of metabolic modules in healthy samples reflecting the often much larger communities. Hence, we ran statistical tests only on normalized (PPCN) data. The p-values associated with each module in Figure 2a, as well as the colors of each point, are based on the PPCN data in Figure 2d. We aimed to improve the clarity of the visual comparison between normalized and unnormalized results by identifying the same set of IBD-enriched modules in plots a-c and plots d-f.

That being said, the reviewer’s comment made us realize the potential for confusion when using the normalized data’s statistical results in Figure 2a that otherwise shows results from unnormalized data. We have now run the same statistical test on the unnormalized (raw copy number) data and re-generated Figure 2a with the new FDR-adjusted p-values and points colored based on the statistical tests using unnormalized data. We’ve also removed the arrow connecting to Figure 2b (since we no longer show the same set of IBD-enriched modules in Figures 2a and 2b), and added a dashed line to indicate the effect size threshold (similar to the one in Figure 2d). We have updated the legend for Figure 2a-d to reflect these changes:

When we used the same p-value threshold (p < 2e-10) as before and also filtered for an effect size larger than the mean (the same strategy used to set our effect size threshold for the normalized data), there are 10 modules that are significantly enriched based on the unnormalized data. Of course, it is difficult to gauge the relevance of these 10 modules to microbial fitness in the IBD gut environment since their raw copy numbers do not tell us anything about the relative proportion of community members that harbor these modules. Therefore, we are reluctant to add these modules to the results text. For the record, only 3 of those modules were also significantly enriched based on the normalized PPCN values: M00010 (Citrate cycle, first carbon oxidation), M00053 (Pyrimidine deoxyribonucleotide biosynthesis), and M00121 (Heme biosynthesis).

Figure 2c,f - these panels raise a lot of concerns given that the choice of method inverts the trend. Without additional data/validation, it's hard to know which method is right.

We hope we have addressed this recommendation with the extensive validation efforts summarized above. Inversion of the trend is an expected outcome, because the raw copy numbers of most metabolic modules are much lower in the IBD sample group due to lower community sizes.

Line 167 - Need to take the KEGG names with a grain of salt, just because it says "biosynthesis" doesn't mean that the pathway goes in that direction in your bacterium of interest.

We believe the reviewer is under a misapprehension regarding the general reversibility of KEGG metabolic modules, or indeed of metabolic pathways. Most metabolic pathways have one or several (practically) irreversible reactions. To demonstrate this for the 33 IBD-enriched modules, we evaluated their reversibility based upon their corresponding KEGG Pathway Maps, which indicate reaction reversibility via double-sided arrows. Aside from the signature modules M00705 and M00627, in 26 out of 31 pathway modules one or more irreversible reactions render these pathways one-directional. Indeed, on average the majority (54%) of the reactions in a given module are irreversible. When focusing on the 23 “biosynthesis” modules, 22 out of 23 (96%) modules have at least one irreversible reaction, and on average 64% of a given module’s reactions are irreversible. These data (which can be accessed at doi:10.6084/m9.figshare.27203226 for the reviewer’s convenience) challenge the reviewer’s notion that pathway directionality is free to change arbitrarily, since the presence of even one irreversible reaction effectively blocks the flux in the opposing direction. Thus, “biosynthesis” is indeed a meaningful term in KEGG module names.

That said, KEGG Pathway Maps, though highly curated, are likely not the final word on whether a given reaction in a metabolic pathway can be considered reversible or irreversible in each microbial population and under all conditions. And our analysis, like many others that rely on metagenomic data, does not consider the environmental conditions in the gut such as temperature or metabolite concentrations that might influence the Gibbs free energy and thus the directionality of these reactions in vivo. However, even assuming general reversibility of metabolic pathways, this would not invalidate the fact that these microbes have the metabolic capacity to synthesize the respective molecules. In other words, the potential reversibility of pathways is irrelevant to our analysis since we are describing metabolic potential. The lac operon in E. coli might only be expressed in the absence of glucose, but E. coli always has the capability to degrade lactose regardless of whether that pathway is active. Thus, our overall conclusion that gut microbes associated with IBD are metabolically self-sufficient (encoding the enzymatic capability to synthesize certain key metabolites) remains valid irrespective of fixed or flexible pathway directionality.

It's also important to be careful not to conflate KEGG modules (small subsets of a pathway) with the actual metabolic pathway. It's possible to have a module change in abundance while not altering the full pathway. Inspection of the individual genes could help in this respect - are they rate-limiting steps for biosynthesis or catabolism?

The reviewer is absolutely correct that KEGG modules do not necessarily represent full pathways. We have updated the language in our manuscript to explicitly refer to “modules” rather than “pathways” whenever appropriate, to restrict the scope of the analysis to metabolic modules rather than full pathways.

That said, we do not see how “inspection of individual genes” would improve our analysis. The strength of looking at complete modules rather than individual genes is that we can gain conclusive insights into a certain metabolic capacity. Of course, no pathway or module stands alone. However, the enrichment of metabolic modules does conclusively indicate that these modules are beneficial under the given conditions, such as stress caused by inflammation or antibiotic use. Whether a certain step in a module or pathway is rate limiting is completely irrelevant for this analysis.

Line 177 - I'm not a big fan of the HMI acronym. Is there a LMI group? It seems simplistic to lump all of metabolism into dependent or independent, which in reality will differ depending on the specific substrate, the growth condition, and the strain.

While we are sorry that our study failed to provide the reviewer with a term they could be a fan of, their input did not change our view that HMI, an acronym we have adapted from a previously peer-reviewed study (doi:10.1186/s13059-023-02924-x), is a powerfully simplistic means to describe a phenomenon we observe and demonstrate in multiple different ways with our extensive analyses. The argument that HMI or LMI status will differ given the growth condition, substrate availability, or strain differences is not helping this case either: our analyses cut across a large number of humans and naturally occurring microbial systems in their guts that are exposed to largely variable ‘growth conditions’ and ‘substrates’ and composed of many strain variants of similar populations. Yet, we observe a clear role for HMI despite all these differences. Perhaps it is because HMI simply describes a higher metabolic capacity based on a defined subset of largely biosynthetic pathways that we observe to be consistently enriched in a large dataset covering a large variety of host, environmental and diet factors and indicates that a population has a higher metabolic capacity to not rely on ecosystem services. We show in our analysis that in the inflamed gut these capacities are indeed required, which is why HMI populations are enriched in IBD samples. HMI has no relation to any of the constraints mentioned by the reviewer, which is one of the major strengths of this metric.

Line 198 - It seems like a big assumption to state that efflux and drug resistance are unrelated to biosynthesis, as they could be genetically or even phenotypically linked.

We agree with the reviewer and are thankful for their input. We have weakened the assertion in this statement.

“These capacities may provide an advantage since antibiotics are a common treatment for IBDs (Nitzan et al. 2016), but are not necessarily related to the systematic enrichment of biosynthesis modules that likely provide resilience to general environmental stress rather than to a specific stressor such as antibiotics.”

Lines 202-218 - I'd suggest removing this paragraph. The "non-IBD" data introduces even more complications to the meta-analysis and seems irrelevant to the current study.

We thank the reviewer for this suggestion. Non-IBD data is important, but its relevance to the primary aims of the study is indeed negligible. We now have moved this paragraph to Supplementary File 1 (under the section “‘Non-IBD’ samples are intermediate to IBD and healthy samples”).

The health gradient is particularly problematic, putting cancer closer to healthy than IBD.

We took the reviewer’s advice and have swapped the order of the studies in Supplementary Figure 6 to place the cancer samples from Feng et al. closer to the IBD samples, on the other side of the non-IBD samples from the IBD studies.

Lines 235-257 - should trim this down and move to the discussion.

As mentioned above, we have opted for a “Results and Discussion format” for our manuscript, so we believe this discussion is in the correct place. We find it important to clearly highlight the limitations and potential biases of our work and trimming this text would take away from that goal.

Figure 3 - panels are out of order. Need to put the current panel D below current panel C. Also, relabel panel letters to go top to bottom (the bottom panel should be D). Could change current panel 3D to a violin plot to match current 3C.

We have updated Figure 3 by converting panel A into a new supplementary figure (Supplementary Figure 8), moving panels C and D below panel B, and relabeling the panels accordingly.

Figure 3B - this panel was incredibly useful and quite surprising to me in many respects. I would have assumed that the Bacteroides would be in the "HMI" bin. Is this a function of the specific strains included here? Was B. theta or B. fragilis included?

The reviewer makes an excellent observation that has been keeping us awake at night, yet somehow was not appropriately discussed in the text until their input. We are very thankful for their attention to detail here.

It is indeed true that Bacteroides genomes are often detected with increased abundance in individuals with IBD and likely have a survival advantage in the IBD gut environment, Bacteroides fragilis and Bacteroides thetaiotaomicron being some of the most dominant residents of the IBD gut. Their non-HMI status is not a function of which strains were included, since all taxa here are represented by the representative genomes available in the publicly available Genome Taxonomy Database. Their non-HMI status comes from the fact that they have HMI scores of around 24 to 26, which fall slightly below the threshold score of 26.4 that we used to classify genomes as HMI. This threshold is back-calculated from the metabolic completion requirement of at least 80% average completion of all 33 metabolic modules that are significantly enriched in IBD. So these genomes are right there at the edge, but not quite over it.

Thanks to this comment by our reviewer, we started wondering whether we should follow a more ‘literature-driven’ approach to set the threshold for HMI, rather than the 80% cutoff, and in fact attempted to lower the HMI score threshold to see if we could include more of the IBD-associated Bacteroides in the HMI bin. Author response table 1 below shows the relevant subset of our new Supplementary Table 3h, which describes the data from our tests on different thresholds.

Author response table 1.

Number and proportion of Bacteroides genomes classified as HMI at each HMI score threshold. There were 20 total Bacteroides genomes in the set of 338 gut microbes identified from the GTDB. The HMI score is computed by adding the percent completeness of all 33 IBD-enriched KEGG modules. The full table can be viewed in Supplementary Table 3h.

Lowering the threshold to 24.75, which corresponds to an average of 75% completeness in the 33 IBD-enriched modules, enabled the classification of 6 Bacteroides genomes as HMI, including B. fragilis, B. intestinalis, B. theta, and B. faecis. However, it also identified several microbes that are not IBD-associated as HMI, including 75 genomes from the Lachnospiraceae family and 18 genomes from the Ruminococcaceae family. In the latter family, several Faecalibacterium genomes, including 10 representatives of Faecalibacterium prausnitzii, were considered HMI using this threshold. These microbes are empirically known to decrease in abundance during inflammatory gastrointestinal conditions (doi:10.3390/microorganisms8040573, doi:10.1093/femsre/fuad039), and therefore these genomes should not be considered HMI – at least not under the working definition of HMI used in our study. To avoid including such a large number of obvious false positives in the HMI bin, we decided to maintain a higher threshold despite the exclusion of Bacteroides genomes.

This outcome demonstrates that our reductionist approach does not successfully capture every microbial population that is associated with IBD. Nevertheless, and in our opinion very surprisingly, the metric does capture a very large proportion of genomes with increased detection and abundance in IBD samples, as demonstrated by the peaks of detection/abundance that match to HMI status Author response image 1.

Author response image 1.

Screenshots of Figure 3 that demonstrate the overlapping signal between HMI status and genome detection/abundance in IBD.

Furthermore, the violin plots in Figure 3B (formerly Figure 3C) clearly reflect the increased representation of HMI populations in IBD metagenomes. Although our classification method is imperfect, it still demonstrates the predictive power of metabolic competencies in identifying which microbes will survive in stressful gut environments. To ensure that readers recognize the crude nature of this classification strategy and the possibility that high metabolic independence can be achieved in different ways, we have added the following sentences to the relevant section of our manuscript:

“Given the number of ways a genome can pass or fail this threshold, this arbitrary cut-off has significant shortcomings, which was demonstrated by the fact that several species in the Bacteroides group were not classified as HMI despite their frequent dominance of the gut microbiome of individuals with IBD (Saitoh et al. 2002; Wexler 2007; Vineis et al. 2016) (Supplementary File 1). That said, the genomes that were classified as HMI by this approach were consistently higher in their detection and abundance in IBD samples (Figure 3a). It is likely that there are multiple ways to have high metabolic independence which are not fully captured by the 33 IBD-enriched metabolic modules identified in this study.”

We have also included a discussion of these findings in Supplementary Information File 1 (see section “Examining the impact of different HMI score thresholds on genome-level results”).

This panel also makes it clear that many of these modules are widespread in all genomes and thus unlikely to meaningfully differ in the microbiome. It would be interesting to use this type of analysis to identify a subset of KEGG modules with high variability between strains.

The figure makes it ‘look like’ many of these modules are widespread in all genomes and thus unlikely to meaningfully differ in the microbiome, but our quantitative analyses clearly demonstrate that these modules indeed differ meaningfully between microbiomes of healthy individuals and those diagnosed with IBD. For instance, the classifier that we built relying exclusively upon these modules’ PPCN values was able to reliably distinguish between the healthy and IBD sample groups in our dataset. The fact that the differentiating signal does not rely on rare metabolic or signature modules is what makes the classifier powerful enough to differentiate between “healthy” and “stressed” microbiomes in 86% of cases. Modules that are by nature less common could not serve this purpose. That said, we do agree with the reviewer that it might be interesting to study variability of KEGG modules as a function of variability between strains. This does not fall into the scope of this work, but we hope to assist others with the technical aspects of such work.

Considering the entirety of the exchange in this section, perhaps there is a broader discussion to be had around this topic. In retrospect, not being able to perfectly split microbes into two groups that completely recapitulate their enrichment in healthy or IBD samples by a crude metric and an arbitrary threshold is not surprising at all. What is surprising is that such a crude metric in fact works for the vast majority of microbes and predicts their increased presence in the IBD gut by only considering their genetic make up. In some respects, we believe that the inability of this cutoff to propose a perfect classifier is similar to the limited power of metabolic independence concept and the classes of HMI or LMI to capture and fully explain microbial fitness in health and disease. What is again surprising here is that these almost offensively simple classes do capture more than what one would expect. We can envision a few ways to implement a more sophisticated HMI/LMI classifier, and it is certainly an important task that is achievable. However, we are hopeful that this technical work can also be done better by others in our field, and that step forward, along with further scrutinizing the relevance of HMI/LMI classes to understand metabolic factors that contribute to the biodiversity of stressful environments, will have to remain as future work.

We thank the reviewer again for their comment here and pushing us to think more carefully and address the oddity regarding the poor representation of Bacteroides as HMI by our cutoff.

Given that a lot of the gaps are in the Firmicutes, this panel also makes me more concerned about annotation bias. How many of these gaps are real?

Analyses relying on gene annotations all suffer equally from the potential for missannotation or missing annotations, which primarily result from limitations in our reference databases for functional data. For instance, the Hidden Markov models for microbial genes in the KEGG Ortholog database are generated from a curated set of gene sequences primarily originating from cultivable microorganisms and particularly from commonly-used model organisms; hence, they do not capture the full extent of sequence diversity observed in populations that are less well-represented in reference databases – a category which includes several Firmicutes, as the reviewer points out. For KEGG KOfams in particular, the precomputed bit score thresholds for distinguishing between ‘good’ and ‘bad’ matches to a given model are often too stringent to enable annotation of genes that are just slightly too divergent from the set of known sequences, thus resulting in missing annotations. Based on our experience with these sorts of issues, we implemented a heuristic that reduces the number of missing annotations for KOs and captures significantly more homologs than other state-of-the-art approaches, as described in doi:10.1101/2024.07.03.601779. We refer the reviewer to our response to the related public comment about annotation bias above, which includes additional details about our investigations of annotation bias in our data. In comparison to the current standard, the heuristic we implemented improves functional annotation results. However, neither our nor any other bioinformatic study that relies on functional gene annotation can exclude the potential for annotation bias.

Figure 3B plotting issues - need to use the full names of the modules; for example, M00844 is "arginine biosynthesis, ornithine => arginine", which changes the interpretation. Need a key for the heatmap on the figure. The tree is difficult to see, needs a darker font.

We have darkened the lines of the tree and dendrogram, and added a legend for the heatmap gradient (see new version of Figure 3 above). Unfortunately, we could not fit the full names of the modules into the figure due to space constraints. However, the full module name and other relevant information can be found in Supplementary Table 2a, and the matrix of pathway completeness scores in these genomes (e.g., the values plotted in the heatmap) can be found in Supplementary Table 3b. We are not sure what the reviewer refers to when stating that “for example, M00844 is "arginine biosynthesis, ornithine => arginine", which changes the interpretation”. There is no ambiguity regarding the identity of KEGG module M00844, which is arginine biosynthesis from ornithine.

Line 321 - more justification for the 80% cutoff is needed along with a sensitivity analysis to see if this choice matters for the key results.

Inspired by this comment, and the one above regarding the classification of Bacteroides genomes, we tested several HMI score thresholds ranging from 75% to 85% average completeness of the 33 IBD-enriched modules. For each threshold, we computed all the key statistics reported in this section of our paper, including the statistical tests. We found that the choice of HMI score threshold does not influence the overall conclusions drawn in this section of our manuscript. Author response table 2 below shows the relevant subset of our new Supplementary Table 3h, which describes the results for each threshold:

Author response table 2.

Key genome-level results at each HMI score threshold. The HMI score is computed by adding the percent completeness of all 33 IBD-enriched KEGG modules. WRS – Wilcoxon Rank Sum test; KW – Kruskal-Wallis test. The full table can be viewed in Supplementary Table 3h

We’ve summarized these findings in a new section of Supplementary File 1 entitled “Examining the impact of different HMI score thresholds on genome-level results”. We copy below the relevant text for the reviewer’s convenience:

“Determining the HMI status of a given genome required us to set a threshold for the HMI score above which a genome would be considered to have high metabolic independence. We tested several different thresholds by varying the average percent completeness of the 33 IBD-enriched metabolic modules that we expected from the

‘HMI’ genomes from ≥ 75% (corresponding to an HMI score of ≥ 24.75) to ≥ 85% (corresponding to an HMI score of ≥ 28.05). For each threshold, we computed the same statistics and ran the same statistical tests as those reported in our main manuscript to assess the impact of these thresholds on the results (Supplementary Table 3h). At the highest threshold we tested (HMI score ≥ 28.05), a small proportion of the reference genomes (7%, or n = 24) were classified as HMI, so we did not test higher thresholds.

We found that the results from comparing HMI genomes to non-HMI genomes are similar regardless of which HMI score threshold is used to classify genomes into either group. No matter which HMI score threshold was used, the mean genome size and mean number of genes were higher for HMI genomes than for non-HMI genomes. On average, the HMI genomes were about 1 Mb larger and had 1,032 more gene calls than non-HMI genomes. We ran two Wilcoxon Rank Sum statistical tests to assess the following null hypotheses: (1) HMI genomes do not have higher detection in IBD samples than non-HMI genomes, and (2) HMI genomes do not have higher detection in healthy samples than non-HMI genomes. For both tests, the p-values decreased (grew more significant) as the HMI score threshold decreased due to the inclusion of more genomes in the HMI bin. The first test for higher detection of HMI genomes than non-HMI genomes in IBD samples yielded p-values less than α = 0.05 at all HMI score thresholds. The second test for higher detection of HMI genomes than non-HMI genomes in healthy samples yielded p-values less than α = 0.05 for the three lowest HMI score thresholds (HMI score ≥ 24.75, ≥ 25.08, or ≥ 25.41). However, irrespective of significance threshold and HMI score threshold, there was always far stronger evidence to reject the first null hypothesis than the second, given that the p-value for the first test in IBD samples was 1 to 5 orders of magnitude lower (more significant) than the p-value for the second test in healthy samples.

IBD samples harbored a significantly higher fraction of genomes classified as HMI than healthy or non-IBD samples, regardless of HMI score threshold (p < 1e-15, Kruskal-Wallis Rank Sum test). The p-values for this test increased (grew less significant) as the HMI score threshold decreased. This suggests that, at higher thresholds, relatively more genomes drop out of the HMI fraction in healthy/non-IBD samples than in IBD samples, thereby leading to larger differences and more significant p-values. Consequently, the HMI scores of genomes detected in IBD samples must be higher than the HMI scores of genomes detected in the other sample groups – indeed, the average HMI score of genomes detected within at least one IBD sample is 24.75, while the average score of genomes detected within at least one healthy sample is 22.78. Within a given sample, the mean HMI score of genomes detected within that sample is higher for the IBD group than in the healthy group: the average per-sample mean HMI score is 25.14 across IBD samples compared to the average of 23.00 across healthy samples.”

Lines 357 and 454 - I would remove the discussion of the "gut environment" which isn't really addressed here. The observed trends could just as easily relate to microbial interactions or the effects of diet and pharmaceuticals. Perhaps the issue is the vague nature of this term, which I read to imply changes in the mammalian host. Given the level of evidence, I'd opt to keep the options open and discuss what additional data would help resolve these questions.

We are in complete agreement with the reviewer that microbial interactions are likely an important driver of our observations. In healthy communities, microbial cross-feeding enables microbes with lower metabolic independence to establish and increase microbial diversity. Which is exactly why we are stating that “Community-level signal translates to individual microbial populations and provides insights into the microbial ecology of stressed gut environments”.

Diet or usage of prescription drugs on the other hand, as discussed previously, likely varies substantially over the various cohorts investigated, and is thus not a driver of the observed trends. Instead, HMI works as a high level indicator that is not influenced by these variable host habits.

Lines 354-394 - Could remove or dramatically trim down this text. Too much discussion for a results section.

We kindly remind the reviewer that our manuscript is written following a “Results and Discussion” format. This section provides necessary context and justification for our classifier implementation, so we have left it as-is.

Lines 395-441 - This section raised a lot of issues and could be qualified or even removed. The model was trained on modules that were IBD-associated in the same dataset, so it's not surprising that it worked. An independent test set would be required to see if this model has any broader utility.

The point that we selected the IBD-enriched modules as features should not raise any concerns, as these modules would have emerged as the most important (ie, most highly weighted) features in our model even if we had included all modules in our training data. This is because machine learning classifiers by design pick out the features that best distinguish between classes, and the 33 IBD-associated modules are a selective subset of these (if they were not, they would not have been significantly enriched in the IBD sample group). That said, a carefully conducted feature selection process prior to model training is a standard best-practice in machine learning; thus, if anything, this should be interpreted as a point of confidence rather than a concern. Furthermore, we evaluated our model using cross-validation, a standard practice in the machine learning field that assesses the stability of model performance by training and testing the model on different subsets of the data. This effort established that the model is robust across different inputs as demonstrated by the per-fold confusion matrix and the ROC curve. These are all standard approaches in machine learning to quantify the model tradeoff between bias and variance. As for the independent test set, we went far and beyond, and applied our model to the antibiotic time-series dataset described later in this section, which, in our opinion, and likely also in the opinion of many experts, serves as one of the most convincing ways to test the utility of any model. Classification results here show that our hypothesis concerning the relevance of metabolic independence to microbial survival in stressed gut environments applies beyond the IBD case and includes antibiotic use, which is indeed a stronger validation for this hypothesis than any test we could have done on other IBD-related datasets. Regardless, we agree that any ‘broader’ utility of our model, such as its applications in clinical settings for diagnostic purposes, is something we certainly can not make strong claims about without more data. We have therefore qualified this section by adding the following sentence:

“Determining whether such a model has broader utility as a diagnostic tool requires further research and validation; however, these results demonstrate the potential of HMI as an accessible diagnostic marker of IBD.”

The application to the antibiotic intervention data raises additional concerns, as the model will predict IBD (labeled "stress" in Figure 5) where none exists.

We apologize for this misunderstanding. The label “stress” actually means stress, not IBD. The figure the reviewer is referring to demonstrates that metabolic modules enriched in the gut microbiome of IBD patients are also temporarily enriched in the gut microbiome of healthy individuals treated with antibiotics for the duration of the treatment. While the classifier uses PPCN values for 33 metabolic modules enriched in microbiomes of IBD patients, it does not mean that this enrichment is exclusive to IBD. The classifier will distinguish between metagenomes in which the PPCN values for those 33 metabolic modules is higher and metagenomes in which the PPCN values are lower. Hence, our analysis demonstrates that during antibiotic usage in healthy individuals, the PPCN values of these 33 metabolic modules spike in a similar fashion to how they would in the gut community of a person with IBD. This points to a more general trend of high metabolic independence as a factor supporting microbial survival in conditions of stress; that is, the increase in metabolic independence is not specific to the IBD condition but rather a more generic ecological response to perturbations in the gut microbial community. We have clarified this point with the following addition to the paragraph summarizing these results:

“All pre-treatment samples were classified as ‘healthy’ followed by a decline in the proportion of ‘healthy’ samples to a minimum 8 days post-treatment, and a gradual increase until 180 days post treatment, when over 90% of samples were classified as ‘healthy’ (Figure 5, Supplementary Table 4b). In other words, the increase in the HMI metric serves as an indicator of stress in the gut microbiome, regardless of whether that stress arises from the IBD condition or the application of antibiotics. These observations support the role of HMI as an ecological driver of microbial resilience during gut stress caused by a variety of environmental perturbations and demonstrate its diagnostic power in reflecting gut microbiome state.”

We’ve also added the following sentence to the end of the legend for Figure 5:

“Samples classified as ‘healthy’ by the model were considered to have ‘no stress’ (blue), while samples classified as ‘IBD’ were considered to be under ‘stress’ (red).”

Figure S5A - should probably split this into 2 graphs since different data is analyzed.

It is true that different sets of modules are used in either half of the figure; however, there is a significant amount of overlap between the sets (17 modules), which is why there are lines connecting the points for the same module as described in the figure legend. We are using this figure to make the point that the median PPCN value of each module increases, in both sets of modules, from the healthy sample group to the IBD sample group. Therefore, we believe the current presentation is appropriate.

Figure S6A – this shows a substantial study effect and raises concerns about reproducibility.

We examined potential batch effects in Supplementary Information File 1 (see section “Considerations of Batch Effect”), and found that any study effect was minor and overcome by the signal between groups:

“The similar distribution of the median normalized copy number for each of the 33 IBD-enriched metabolic modules (summarized across all samples within a given study), across all studies within a given sample group (Supplementary Figure 6b), confirms that the sample group explains more of the trend than the study of origin.”

Furthermore, within Supplementary Figure 6a, there is a clear increase between the non-IBD controls from Franzosa et al. 2018 and the IBD samples from the same study, as well as between the non-IBD controls from Schirmir et al. 2018 and the IBD samples from that study. As there is no study effect influencing those two comparisons, this reinforces the evidence that there is a true increase in the normalized copy numbers of these modules when comparing samples from more healthy individuals to those from less healthy individuals.

Figure S7B - check numbers, which I think should sum to 33.

The numbers should not sum to 33. In this test to determine whether the two largest studies had excessive influence on the identity of the IBD-enriched modules, we repeated our strategy to obtain 33 IBD-enriched modules (those with the 33 smallest p-values from the statistical test) from each set of samples – either (1) samples from Le Chatelier et al. 2013 and Vineis et al. 2016, or (2) samples that are not from those two studies. The 2 sets, containing 33 modules each, gives us a total of 66 IBD-enriched modules. By comparing those two sets, we found that 20 modules were present in both sets – hence the value of 20 in the center of the Venn Diagram. In each set, 13 modules were unique – hence the value of 13 on either side. 13 + 13 + 2*20 = 66 total modules.

We again thank our reviewers for their time and interest, and invaluable input.

Author response:

The following is the authors’ response to the original reviews.

Reveiwer#1 (Public Review):

Weaknesses:

While the novel compound showed a promising potency to the HER2-positive gastric cancer cells and xenograft model, it would be great to also to be evaluated with the HER2-positive breast cancer cell models. The author did not compare the current compounds with other therapeutic strategies targeting HER2 expression at the genetic level. It is unclear whether the EGFR inhibitors gefitinib and canertinib but not HER2-specific inhibitors (i.e. tucatinib) were used as a control in the manuscript.

We appreciate the reviewer’s insightful comments. Evaluating compound 10 on HER2-positive breast cancer cells is indeed crucial, especially given the established HER2-targeting therapies for breast cancer. In response to this concern, we conducted additional experiments to investigate the impact of compound 10 on HER2-positive breast cancer cell lines AU565 and BT474, specifically assessing its HER2 downregulating activity (Author response image 1).

Author response image 1.

HER2 downregulatory effect of compound 10 in HER2-positive breast cancer cell lines, AU565 and BT474.

The selection of gefitinib (an EGFR tyrosine kinase inhibitor) and canertinib (a pan-HER inhibitor) as positive controls in our manuscript is based on their demonstrated ability to inhibit the protein-protein interaction (PPI) between ELF3 and MED23, as previously reported (J Adv Res. 47, (2023) 173-87. 10.1016/j.jare.2022.08.003; Cancer letters. 325, (2012) 72-9. 10.1016/j.canlet.2012.06.004). In referenced studies, SEAP reporter gene assay was utilized to screen compounds for their capacity to disrupt the ELF3-MED23 PPI. This assay involves GAL4-ELF3 binding to a GAL4 binding site in the SEAP reporter gene, followed by interaction with MED23, leading to RNA polymerase II recruitment and SEAP expression in cells (J Am Chem Soc. 2004, 126(49), 15940. doi: 10.1021/ja0445140). Canertinib exhibited stronger inhibitory activity against ELF3-MED23 PPI compared to gefitinib, but also showed non-specific cytotoxicity. YK1 was subsequently developed based on structural analysis of the interfaces between gefitinib and MED23, and between ELF3 and MED23. Considering the previously validated inhibitory activities of gefitinib and canertinib, these drugs were selected as positive controls in the current study to compare the ELF3-MED23 inhibitory efficacy of novel compounds.

Reveiwer#1 (Recommendations For the Authors):

(1) It is unclear how compound 5 did not inhibit HER2 overexpression at mRNA but at protein levels as compounds 3 and 10. Could the author further explain the potential mechanism for compound 5?

While the exact mechanism remains unclear, the results indicated that compound 5 likely affects the protein level of HER2 through somewhat non-specific mechanisms rather than by inhibiting the ELF3-MED23 PPI. Based on this assessment, compound 5 was excluded from further investigation.

(2) The HER2 expression and its downstream signaling pathway assay are unclear about the approach. It needs to be included in the methods or supplementary.

We investigated the ELF3-MED23 PPI inhibitory activity and its subsequent effect on HER2 downregulation using a comprehensive approach involving multiple techniques to ensure precise and unbiased experimental results.

To assess PPI inhibition, we employed the following assays:

· SEAP reporter gene assay

· Fluorescence polarization (FP)

· Split-luciferase complementation assay

· GST-pulldown

· Immunoprecipiation (IP)

HER2 expression levels were evaluated through:

· SEAP reporter gene assay

· Luciferase promoter assay

· Quantification of HER2 mRNA using qPCR

· Measurement of HER2 protein levels via western blot analysis

To evaluate downstream signaling of HER2, we analyzed:

· Phosphorylation levels of MAPK (pMAPK) and AKT (pAKT)

These methods were systematically applied to elucidate the mechanism of action of compound 10 in inhibiting ELF3-MED23 interaction and subsequently downregulating HER2.

For clarity, we have revised the manuscript to provide a detailed description of the experimental methods to assess PPI, as described below.

“SEAP assay was performed as previously described to measure ELF3-MED23 PPI-dependent HER2 transcription [29]. In this assay, the GAL4-ELF3 fusion protein binds to one of the five GAL4 binding sites on the reporter gene (pG4IL2SX). The interaction between the GAL4-ELF3 fusion protein and endogenous MED23 induces the expression of the SEAP. Once expressed, SEAP acts as a phosphatase on the substrate 4-MUP (4-methyl umbelliferyl phosphate), resulting in increased fluorescence. The mammalian expression vector, …”

“FP assay was conducted following a previously described method to evaluate the molecular interaction between ELF3 and MED23 [29]. The FP assay operates on the principle of the molecular rotation dynamics. When a fluorescently labeled small molecule is excited by polarized light, the emitted fluorescence can be polarized or depolarized depending on the molecular status. Free small molecules rotate rapidly, altering the orientation of their fluorescence dipole and emitting depolarized light. However, when these small molecules bind to large molecules, such as proteins, the resulting complex rotates more slowly, and the emitted light retains much of its original polarization. In this study, different concentrations of (His)6-MED23391–582, as the large molecule, and 10 nM of FITC-labeled ELF3129–145 peptide, as the fluorescence-labeled small molecule, were combined in …”

(3) It is confusing to me about the order of the experiments, in which the SAR work came after the synthesis and a series of biochemical studies for the characterization of the synthetic compounds. What is the specific reason for this order?

We concluded that the current approach is appropriate because the analysis was not intended for structural modification and optimization through SAR (Structure-Activity Relationship) analysis. Instead, the primary objective was to elucidate the structural basis underlying the efficacy of PPI inhibition among compounds sharing the same scaffold. We believe this will provide valuable insights for future design and synthesis of new compounds.

(4) The yield for each step of the general synthesis needs to be included in the scheme 1.

Scheme 1 has been updated to include the yield of each step of the synthesis process.

(5) In line 532, the authors stated 28 compounds, should it be 26?

‘Twenty-eight compounds’ includes 26 newly synthesized compounds and 2 positive controls, gefitinib and canertinib.

(6) Introduction part, lines 74 to 75, "While HER2 gene amplification is the primary mechanism responsible for HER2 overexpression" may not be confirmed in lung cancers.

HER2 overexpression is usually a direct consequence of gene amplification, although overexpression can occur by other mechanisms [Nat Rev Cancer. 2009;9:463–475. doi: 10.1038/nrc2656.; Cell. 2007;129:1275–1286. doi: 10.1016/j.cell.2007.04.034.]. The levels of HER2 protein expression and gene amplification are linearly associated and highly concordant in breast cancer, colorectal cancer, ovarian cancer, and esophageal adenocarcinoma [World J Gastrointest Oncol. 2019, 11(4): 335–347. doi: 10.4251/wjgo.v11.i4.335; J Clin Oncol. 2002;20:719–26. doi.org/10.1200/JCO.2002.20.3.71; Oncology. 2001;61(Suppl 2):14–21. doi.org/10.1159/000055397; Science. 1989, 244(4905):707-12. doi: 10.1126/science.2470152; Cancer. 2014 Feb 1; 120(3): 415–424. doi: 10.1002/cncr.28435]. As reviewer mentioned, the linear association between of HER2 protein expression and gene amplification has not been fully established for NSCLC [ESMO Open. 2022, 100395. doi: 10.1016/j.esmoop.2022.100395].

Therefore, we change the sentence as describe below.

“While HER2 gene amplification is the primary mechanism responsible for HER2 overexpression in most HER2-positive cancers, except in lung cancer [16], high transcription rates of HER2 per gene copy have also been observed to contribute.”

(7) The abstract part, lines 31 and 32, the detailed experimental data for SEAP needs to be expressed in another way.

SEAP is a type of reporter gene assay. We revised the manuscript as follows and we additionally described it method part.

“Upon systematic analysis, candidate compound 10 was selected due to its potency in downregulating reporter gene activity of HER2 promoter confirmed by SEAP activity and its effect on HER2 protein and mRNA levels.”

(8) The author should combine the box for Chalcone, pyrazoline, Licochalcone E, and YK-1, Figures 1 and 2 into a new single Figure.

We revised the manuscript following the reviewer's comments.

(9) Provide the list of antibodies and sources for the cell-based and western blot assays.

Table S1 presents detailed information about the antibodies and dilution ratios used in the cell-based and western blot assays.

Reveiwer#2 (Public Reviews):

Weaknesses:

The rationale behind the proposed structural modifications for the three groups of compounds is not clear.

Reveiwer#2 (Recommendations For the Authors):

(1) Based on previous work experience, it would be interesting to evaluate the in silico mode of interaction of compound 10.

As suggested by the reviewers, we additionally performed in silico docking study to identify the mode of interaction of compound 10 (Author response image 2). As shown below, the results indicate that compound 10 shares a similar binding orientation with YK1, forming an H-bond with the H449 residue. Although it does not interact with the D400 residue, it was predicted to create an additional H-bond with S450, which is right next to H449, thereby reinforcing the overall binding of compound 10 to MED23. Moreover, compound 10 was additionally predicted to form a pi-pi interaction with F399, which has been previously identified as an important interaction for compounds to demonstrate outstanding PPI inhibitory effect against ELF3 and MED23.

Author response image 2.

Docking analysis of compound 10.

(2) The chalcones presented in this study are structurally similar to those previously presented by the group (ref 29). In said work, most of the compounds exhibited activities with IC50 values between 1.3 and 3 μM, with inhibition values at 10 μM ranging between 80 and 90% in the SEAP assay. These results are similar to those observed in this paper for the same assay. Can an explanation be found?

Chalcones are inherently flexible molecules, giving them a high chance of occupying critical hotspot residues within the binding interface of ELF3-MED23, irrespective of the side chains introduced to this moiety. However, depending on the type of side chains introduced, the overall drug-like properties of compounds can be significantly altered, while still maintaining their PPI inhibitory effect. The significance of this study lies in our effort to enhance metabolic stability through extensive introduction of methoxy groups and other hydrophobic side chains to the chalcone skeleton, while preserving high PPI inhibitory activity.

(3) Is the replacement of H and OH by OMe necessary? Does it improve any property (activity, selectivity, bioavailability, solubility, etc.)? Regarding the derivatives of group 2, why did they decide to replace the O-H, which in silico demonstrated favorable hydrogen bond interactions with Asp400? How do these molecules look in the binding site? Perhaps this is a point to discuss since the substitution of OH led to the obtaining of inactive molecules, or is the effect due to substitution with the terminal aromatic ring with 3 OMe?

We modified the hydroxyl group moiety of YK-1 into a methoxy group to reduce the polarity of the compound, thereby enhancing its cell membrane permeability (Author response image 3) and reducing the likelihood of rapid elimination through phase II metabolic pathways in vivo. Additionally, we considered the potential conversion of the methoxy group back to a hydroxyl group via phase I metabolism in vivo.

Author response image 3.

Impact of methoxy group introduction on TPSA (total polar surface area) of each molecule. TPSA of each molecule containing chalcone structure were calculated using the Molinspiration webserver.

(4) Lines 134 and 134: "Only compounds are in red."

We revised the manuscript following the reviewer's comments.

(5) Line 171: "Chalcone skeleton, shown in red."

We revised the manuscript following the reviewer's comments.

(6) Line 350: "N-1-acetyl-4,5-dihydropyrazoline."

We revised the manuscript following the reviewer's comments.

(7) Scheme 1. Replace "h" with "hr".

We revised the manuscript following the reviewer's comments. Scheme 1 has been replaced by a new version.

(8) Where is "Table S1" in SI?

Tables S1 and S2 are supposed to be included in SI. We will ensure that Tables S1 and S2 are properly uploaded to the SI section.

(9) In Figure 6, Graph D, to enhance comprehension, please incorporate red arrows indicating drug administration.

We revised Figure 6 (D) following the reviewer's comments. Red arrows indicating drug administration have been incorporated, along with a descriptive comment "Drug administration" next to each arrow. Additionally, the figure legend now includes a clear description of these additions.

Reveiwer#3 (Public review):

Weaknesses:

Compound 10 potency as PPI inhibitor has been shown in only one cell line NCI-N87.

Reveiwer#3 (Recommendations For the Authors):

(1) The authors should show this compound 10 is effective in other gastric cancer cells like KATOIII, SNU1.

We evaluated the HER2 downregulating activity of compound 10 in the gastric cancer cell line, SNU216, which is confirmed to express high level of HER2 protein (Author response image 4).

Author response image 4.

HER2 downregulatory effect of compound 10 in HER2-positive gastric cancer cell line, SNU216. (A) Expression levels of HER2 and ELF3 in various gastric cancer cell lines. (B) HER2 downregulation in the SNU216 cell line following treatment with compound 10.

Author response:

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

In this work by Wang et al., the authors use single-molecule super-resolution microscopy together with biochemical assays to quantify the organization of Nipah virus fusion protein F (NiV-F) on cell and viral membranes. They find that these proteins form nanoscale clusters which favors membrane fusion activation, and that the physical parameters of these clusters are unaffected by protein expression level and endosomal cleavage. Furthermore, they find that the cluster organization is affected by mutations in the trimer interface on the NiV-F ectodomain and the putative oligomerization motif on the transmembrane domain, and that the clusters are stabilized by interactions among NiV-F, the AP2-complex, and the clathrin coat assembly. This work improves our understanding of the NiV fusion machinery, which may have implications also for our understanding of the function of other viruses.

Strengths:

The conclusions of this paper are well-supported by the presented data. This study sheds light on the activation mechanisms underlying the NiV fusion machinery.

Weaknesses:

The authors provide limited details of the convolutional neural network they developed in this work. Even though custom-codes are made available, a description of the network and specifications of how it was used in this work would aid the readers in assessing its performance and applicability. The same holds for the custom-written OPTICS algorithm. Furthermore, limited details are provided for the imaging setup, oxygen scavenging buffer, and analysis for the single-molecule data, which limits reproducibility in other laboratories. The claim of 10 nm resolution is not backed up by data and seems low given the imaging conditions and fluorophores used. Fourier Ring Correlation analysis would have validated this claim. If the authors refer to localization precision rather than resolution, then this should be specified and appropriate data provided to support this claim.

We thank reviewer 1 for these suggestions. We described key steps in imaging setup, singlemolecule data reconstruction, the OPTICS algorithm in cluster identification, and 1D CNN in

classification of the OPTICS data in the Materials and Methods section. We also provided a recipe for the imaging buffer. We refer to 10 nm localization precision rather than resolution. The localization precision achieved by our SMLM system is shown in the Author response image 1.

Author response image 1.

The localization precision of the custom-built SMLM. Shows the distribution of localization error at the x (dX), y (dY), and z (dZ) direction in nanometer of blinks generated from Alexa Flour 647 labeled to NiV-F expressed on the plasma membrane of PK13 cells. The lateral precision is <10 nm and the axial precision is < 20 nm. 

Reviewer #2 (Public Review): 

Summary:

In this manuscript, Wang and co-workers employ single molecule light microscopy (SMLM) to detect NiV fusion protein (NiV-F) in the surface of cells. They corroborate that these glycoproteins form microclusters (previously seen and characterized together with the NiVG and Nipah Matrix protein by Liu and co-workers (2018) also with super-resolution light microscopy). Also seen by Liu and coworkers the authors show that the level of expression of NiV-F does not alter the identity of these microclusters nor endosomal cleavage. Moreover, mutations and the transmembrane domain or the hexamer-of-trimer interface seem to have a mild effect on the size of the clusters that the authors quantified.

Importantly, it has also been shown that these particles tend to cluster in Nipah VLPs.

We thank reviewer #2 for the comments and suggestions. This paper is built on Liu et al 1 to further characterize the nanoclusters formed by NiV-F and their role in membrane fusion activation. While Liu et al. studied the NiV glycoprotein distribution at the NiV assembly sites to inform mechanisms in NiV assembly and release, Wang et al. analyzed the nanoorganization and distribution of NiV-F at the prefusion conformation, providing insights into the membrane fusion activation mechanisms.  

Strengths:

The authors have tried to perform SMLM in single VLPs and have shown partially the importance of NiV-F clustering.

Weaknesses:

The labelling strategy for the NiV-F is not sufficiently explained. The use of a FLAG tag in the extracellular domain should be validated and compared with the unlabelled WT NiV-F when expressed in functional pseudoviruses (for example HIV-1 based particles decorated with NiV-F). This experiment should also be carried out for both infection and fusion (including BlaM-Vpr as a readout for fusion). I would also suggest to run a time-of-addition BlaM experiment to understand how this particular labelling strategy affects single virion fusion as compared to the the WT.  

We thank reviewer #2 for this suggestion. We have made various efforts to validate the expression and function of FLAG-tagged NiV-F. The NiV-F-FLAG shows comparable cell surface expression levels and induces similar cell-cell fusion levels in 293T cells as that of untagged NiV-F 1. The NiV-F-FLAG also showed similar levels of virus entry as untagged NiV-F when both were pseudotyped on a recombinant Vesicular Stomatitis Virus (VSV) with the VSV glycoprotein replaced by a Renilla luciferase reporter gene (VSV-ΔG-rLuc; Fig. S1D). We also performed a virus entry kinetics assay using NiV VLPs expressing NiV-M-βlactamase (NiV-M-Bla), NiV-G-HA, and NiV-F-FLAG, NiV-F-AU1 or untagged NiV-F. The intracellular AU1 tag is located at the C-terminus of NiV-F (Genbank accession no. AY816748.1). However, we detected different levels of NiV-M-Bla in equal volume of VLPs, suggesting that the tags in NiV-F affect the budding of the VLPs (Author response image 2A). Therefore, we performed fusion kinetics assay by using VLPs expressing the same levels of NiV-M-Bla. Among them, the NiV-F-FLAG on VLPs shows the most efficient fusion between VLP and HEK293T cell membranes (Author response image 2B), significantly more efficient than that of untagged NiV-F and NiV-FAU1. However, we cannot attribute the enhanced fusion activity to the FLAG tag, because the readout of this assay relies on both the levels of β-lactamase (introduced by NiV-M-Bla in VLPs) and the NiV-F constructs. The tags in NiV-F could affect both the budding of VLPs and the stoichiometry of F and M in individual VLPs. We did not use the HIV-based pseudovirus system because the incorporation of NiV-F into HIV pseudoviruses requires a C-terminal deletion 2,3.

In summary, the FLAG tag does not affect cell-cell fusion 1 and virus entry when pseudotyped to the recombinant VSV-ΔG-rLuc viruses (Fig. S1D). Given that we do not observe any difference in clustering between an HA- and FLAG-tagged NiV-F constructs on PK13 cell surface (Fig. S1A-C), we conclude that the FLAG tag has minimal effect on both the fusion activity and the nanoscale distribution of NiV-F. 

Author response image 2.

Viral entry is not affected by labeling of NiV-F. A) Western blot analysis of NiV-M-Bla in NiV-VLPs generated by HEK293T cells expressing NiV-M-Bla, NiV-G-HA and NiV-F-FLAG, untagged NiV-F, or NiV-F-AU1. Equal volume of VLPs were separated by a denaturing 10% SDS–PAGE and probed against β-lactamase (SANTA CRUZ, sc-66062). B) NiV-VLPs expressing NiV-M-BLa, NiV-G-HA, and NiV-F-FLAG, untagged NiV-F or NiV-F-AU1 expression plasmids were bond to the target HEK293T cells loaded with CCF2-AM dye at 4°C. The Blue/Green (B/G) ratio was measured at 37°C for 4 hrs at a 3-min interval. Results were normalized to the maximal B/G ratio of NiV-F-FLAG-NiV VLPs. Results from one representative experiment out of three independent experiments are shown. 

It would also be very important to compare the FLAG labelling approach with recent advances in the field (for instance incorporating noncanonical amino acids (ncAAs) into NiVF by amber stop-codon suppression, followed by click chemistry). 

We are greatly thankful for this comment from reviewer #2. Labeling noncanonical amino acids (ncAAs) with biorthogonal click chemistry is indeed a more precise labeling strategy compared to the traditional epitope labeling approach used in this paper. We will explore the applications of ncAAs labeling in single-molecule localization imaging and virus-host interactions in future projects. 

In this paper, the FLAG tag inserted in NiV-F protein seems to have minimal effect on the NiV-F-induced virus entry and cell-cell fusion 1 (Fig. S1). Although the FLAG tag labeling approach may increase the detectable size of NiV-F nanoclusters due to the use of the antibody complex, it should not affect our conclusions drawn from the relative comparisons between wt and mutant NiV-F or control and drug-treated cells. 

The correlation between the existence of microclusters of a particular size and their functionality is missing. Only cell-cell fusion assays are shown in supplementary figures and clearly, single virus entry and fusion cannot be compared with the biophysics of cell-cell fusion. Not only the environment is completely different, membrane curvature and the number of NiV-F drastically varies also. Therefore, specific fusion assays (either single virus tracking and/or time-of-addition BlaM kinetics with functional pseudoviruses) are needed to substantiate this claim.  

We thank Reviewer 2 for the suggestion. To support the link between F clustering and viruscell membrane fusion, we conducted pseudotyped virus entry and VLP fusion kinetics assays, as shown in revised Figure S4. The viral entry results (Fig. S4 E and F) corroborate that of the cell-cell fusion assay (Fig. S4A and B) and previously published data 4. The fusion kinetics confirmed that the real-time fusion kinetics was affected by mutations at the hexameric interface, with the hypo-fusogenic mutants L53D and V108D exhibited reduced entry efficiency while the hyper-fusogenic mutant Q393L showed increased efficiency (Fig. S4G and H). The results were described in detail in the revised manuscript. 

Additionally, we performed a pseudotyped virus entry assay on the LI4A (Fig. S6F and G) and YA (Fig. S7F and G) mutants to verify the function of these mutants on viruses in revised Supplemental Figures. Neither LI4A nor YA incorporated into the VSV/NiV pseudotyped viruses as shown by the Western blot analyses of the pseudovirions (Fig. S6F and S7F), and thus did not induce virus entry, consisting with the cell-cell fusion results (Fig. S6C, D and Fig. S7C, D). We did not perform the entry kinetic assay of these two mutants as they do not incorporate into VLPs or pseudovirions. 

The authors also claim they could not characterize the number of NiV-F particles per cluster. Another technique such as number and brightness (Digman et al., 2008) could support current SMLM data and identify the number of single molecules per cluster. Also, this technology does not require complex microscopy apparatus. I suggest they perform either confocal fluorescence fluctuation spectroscopy or TIRF-based nandb to validate the clusters and identify how many molecule are present in these clusters.  

We thank reviewer 2 for this suggestion. Determining the true copy number of NiV-F in individual clusters could verify whether the F clusters on the plasma membrane are hexamer-of-trimer assemblies. Regardless, it does not affect our conclusion that the organization of NiV-F into nanoclusters affects the membrane fusion triggering ability. The confocal fluorescence fluctuation spectroscopy (FFS) and TIRF-based analyses are accessible tools for quantifying fluorophore copy numbers and/or stoichiometry based on fluorescence fluctuation or photobleaching. However, these methods are unable to quantify the number of proteins in individual clusters because they analyze fluorophores either in the entire cell (as in wide-field epifluorescence microscopy coupled with FFS and TIRF-coupled photobleaching) 5–7 or within a large excitation volume (confocal laser scanning microscopycoupled FFS) 8. Both of these volumes are significantly larger than a single NiV-F cluster, which has an average diameter of 24-26 nm (Fig. 1F). 

The current SMLM setup is useful for characterizing the protein distribution and organization. However, quantifying the true protein copy number within a nanocluster is challenging because of the stochasticity of fluorophore blinking and the unknown labeling stoichiometry 9–11. To address the challenge in fluorophore blinking, quantitative DNA-PAINT (qDNA-PAINT) may be used because the on-off frequency of the fluorophores is tied to the well-defined kinetic constants of DNA binding and the influx rate of the imager strands, rather than the stochasticity of fluorophore blinking. Thus, the frequency of blinks can be translated to protein counting 12. To address the challenge in unknown labeling stoichiometry, DNA origami can be used as a calibration standard 11. DNA origami supports handles at a regular space with several to tens of nanometers apart, and the handles can be conjugated with a certain number of proteins of interest. The copy number of protein interest in the experimental group can be determined by comparing the SMLM localization distribution of the sample to that of the DNA origami calibration standard. Given the requirement of a more sophisticated SMLM setup and a high-precision calibration tool, we will explore the quantification of NiV-F copy numbers in nanoclusters in a future project. 

Also, it is not clear how many cells the authors employ for their statistics (at least 30-50 cells should be employed and not consider the number of events blinking events. I hope the authors are not considering only a single cell to run their stats... The differences between the mutants and the NiV-F is minor even if their statistical analyses give a difference (they should average the number and size of the clusters per cell for a total of 30-50 cells with experiments performed at least in three different cells following the same protocol). Overall, it seems that the authors have only evaluated a very low number of cells.

We disagree with this comment from Reviewer #2. The sample size for cluster analysis in SMLM images was chosen by considering the target of the study (cells and VLPs) and the data acquisition and analysis standards in the SMLM imaging field. We also noted the sample size (# of ROI and cells) in the figure legend. 

Below, we compared the sample sizes in our study to those in similar studies that used comparable imaging and cluster analysis methods from 2015 to 2024. The classical clustering analysis methods are categorized into global clustering (e.g. nearest neighbor analysis, Ripley’s K function, and pair correlation function) and complete clustering, such as density-based analysis (e.g. DBSCAN, Superstructure, FOCAL, ToMATo) and Tessellationbased analysis (e.g. Delaunay triangulation, Voronoii Tessellation). The global clustering analysis method provides spatial statistics for global protein clustering or organization (e.g. clustering extent), while the complete clustering approach extracts information from a single-cluster level, such as the morphology and localization density of individual clusters. We used the density-based analyses, DBSCAN and OPTICS, for cluster analysis on cell plasma membranes and VLP membranes. 

Author response table 1.

The comparison of imaging methods, analysis methods, and sample size in the current study to other studies conducted from 2015 to 2024.

They should also compare the level of expression (with the number of molecules per cell provided by number and brightness) with the total number of clusters. 

We thank reviewer 2 for this suggestion. We compared the level of expression with the total number of clusters for F-WT in Figure 1I in the main text.  

The same applies to the VLP assay. I assume the authors have only taken VLPs expressing both NiV-M and NiV-F (and NiV-G). But even if this is not clearly stated I would urge the authors to show how many viruses were compared per condition (normally I would expect 300 particles per condition coming from three independent experiments. As a negative control to evaluate the cluster effect I would mix the different conditions. Clearly you have clusters with all conditions and the differences in clustering depending on each condition are minimal. Therefore you need to increase the n for all experiments.

We thank reviewer 2 for this comment. We acquired and analyzed more images of NiV VLPs bearing F-WT, Q393L, L53D, and V108D. Results are shown in the revised Figure 4 and the number of VLPs (>300) used for analysis is specified in the figure legend. An increased number of VLP images does not affect the classification result in Figure 4C. 

As for the suggestion on “evaluating the cluster effect at different mixed conditions”, I assume that reviewer 2 would like to see how the presence of different viral structural proteins (F, M, and G) on VLPs could affect F clustering.  We showed that the organization of NiV envelope proteins on the VLP membrane is similar in the presence or absence of NiV-M by direct visualization 27, suggesting that the effect of NiV-M on F-WT clustering on VLPs is minimal. We also show comparable incorporation of NiV-F among the NiV-F hexamer-oftrimer mutants (Fig. 4A). Therefore, we did not test the F clustering at different F, M, and G combinations in this paper. However, this could be an interesting question to pursue in a paper focusing on NiV VLP production. 

Reviewer #3 (Public Review):

Summary:

The manuscript by Wang and colleagues describes single molecule localization microscopy to quantify the distribution and organization of Nipah virus F expressed on cells and on virus-like particles. Notably the crystal structure of F indicated hexameric assemblies of F trimers. The authors propose that F clustering favors membrane fusion.

Strengths:

The manuscript provides solid data on imaging of F clustering with the main findings of:

-  F clusters are independent of expression levels

-  Proteolytic cleavage does not affect F clustering

-  Mutations that have been reported to affect the hexamer interface reduce clustering on cells and its distribution on VLPs - - F nanoclusters are stabilized by AP

Weaknesses:

The relationship between F clustering and fusion is per se interesting, but looking at F clusters on the plasma membrane does not exclude that F clustering occurs for budding. Many viral glycoproteins cluster at the plasma membrane to generate micro domains for budding. 

This does not exclude that these clusters include hexamer assemblies or clustering requires hexamer assemblies. 

We thank reviewer #3 for this question. We did not focus on the role of NiV-F clusters for budding in the current manuscript, although this is an interesting topic to pursue. In this manuscript, we observed that NiV VLP budding is decreased for some cluster-disrupting mutants, such as F-YA, and F-LI4A. however, F-V108D showed increased budding compared to F-WT (Fig. 4A). We also observed that VLPs and VSV/NiV pseudoviruses expressing L53D have little NiV-G (Fig. 4A, Fig. S4F and S4H), although the incorporation level of L53D is comparable to that of wt F in both VLPs and pseudovirions (Fig. 4A and Fig. S4F). L53D is a hypofusogenic mutant with decreased clustering ability. Therefore, our current data do not show a clear link between F clustering and NiV VLP budding or glycoprotein incorporation. 

We reported that both NiV-F and -M form clusters at the plasma membrane although NiV-F clusters are not enriched at the NiV-M positive membrane domains 1. This result indicates that NiV-M is the major driving force for assembly and budding, while NiV-F is passively incorporated into the assembly sites. The central role of NiV-M in budding is also supported by a recent study showing that NiV-M induces membrane curvature by binding to PI(4,5)P2 in the inner leaflet of the plasma membrane 28. However, the expression of NiV-F alone induces the production of vesicles bearing NiV-F 29 and NiV-F recruits vesicular trafficking and actin cytoskeleton factors to VLPs either alone or in combination with NiV-G and -M, indicating a potential autonomous role in budding 30. Additionally, several electron microscopy studies show that the paramyxovirus F forms 2D lattice interspersed above the M lattice, suggesting the participation of F in virus assembly and budding. Nonetheless, the evidence above suggests that NiV-F may play a role in budding, but our data cannot correlate NiV-F clustering to budding. 

Assuming that the clusters are important for entry, hexameric clusters are not unique to Nipah virus F. Similar hexameric clusters have been described for the HEF on influenza virus C particles (Halldorsson et al 2021) and env organization on Foamy virus particles (Effantin et al 2016), both with specific interactions between trimers. What is the organization of F on Nipah virus particles? If F requires to be hexameric for entry, this should be easily imaged by EM on infectious or inactivated virus particles. 

We thank reviewer #3 for this suggestion. The hexamer-of-trimer NiV-F is observed on the VLP surface by electron tomography 4. The NiV-F hexamer-of-trimers are arranged into a soccer ball-like structure, with one trimer being part of multiple hexamer-of-trimers. The implication of NiV-F clusters in virus entry and the potential mechanism for NiV-F higherorder structure formation are discussed in the revised manuscripts. 

AP stabilization of the F clusters is curious if the clusters are solely required for entry? Virus entry does not recruit the clathrin machinery. Is it possible that F clusters are endocytosed in the absence of budding? 

We thank reviewer #3 for this question. The evidence from the current study does not exclude the role of NiV-F clustering in virus budding. NiV-F is known to be endocytosed in the virus-producing cells for cleavage by Cathepsin B or L at endocytic compartments at a pH-dependent manner31–33 in the absence of budding. However, given that all cleaved and uncleaved NiV-F have an endocytosis signal sequence at the cytoplasmic tail and are able to interact with AP-2 for endosome assembly and the cleaved and uncleaved F may have similar clustering patterns (Fig. 2), we do not think NiV-F clustering is specifically regulated for the cleavage of NiV-F. A plausible hypothesis is that NiV-F clusters are stabilized by multiple intrinsic factors (e.g. trimer interface) and host factors (e.g. AP-2) on cell membrane for cell-cell fusion and virus budding. We linked the clustering to the fusion ability of NiV-F in this study, but the NiV-F clustering may also be important in facilitating virus budding. Once in the viruses, the higher-order assembly of the clusters (e.g. lattice) may form due to protein enrichment, and the cell factors may not be the major maintenance force. 

Clusters are required for budding. 

Other points:

Fig. 3: Some of the V108D and L53D clusters look similar in size than wt clusters. It seems that the interaction is important but not absolutely essential. Would a double mutant abrogate clustering completely?

We thank Reviewer #3 for the suggestion. We generated a double mutant of NIV-F with L53D and V108D (NiV-F-LV) and assessed its expression and processing. Although the mutant retained processing capability, it exhibited minimal surface expression, making it unfeasible to analyze its nano-organization on the cell or viral membrane.

Author response image 4.

The expression and fusion activity of Flag-tagged NiV-F and NiV-F L53D-V108D (LV). (A) Representative western blot analysis of NiV-F-WT, LV in the cell lysate of 293T cells. 293T cells were transfected by NiV-F-WT or the LV mutant. The empty vector was used as a negative control. The cell lysates were analyzed on SDS-PAGE followed by western blotting after 28hrs post-transfection. F0 and F2 were probed by the M2 monoclonal mouse antiFLAG antibody. GAPDH was probed by monoclonal mouse anti-GAPDH. (B) Representative images of 293T cell-cell fusion induced by NiV-G and NiV-F-WT or NiV-F-LV. 293T cells were co-transfected with plasmids coding for NiV-G and empty vector (NC) or NiV-F constructs. Cells were fixed at 18 hrs post-transfection. Arrows point to syncytia. Scale bar: 10um. (C) Relative cell-cell fusion levels in 293T cells in (B). Five fields per experiment were counted from three independent experiments. Data are presented as mean ± SEM. (D) The cell surface expression levels of NiV-F-WT, NiV-F-LV in 293T cells measured by flow cytometry. Mean fluorescence Intensity (MFI) values were calculated by FlowJo and normalized to that of F-WT. Data are presented as mean ± SEM of three independent experiments. Statistical significance was determined by the unpaired t-test with Welch’s correction (*P<0.05, **P<0.01, ***P<0.001, ****P<0.0001). Values were compared to that of the NiV-F-WT.

Fig. 4: The distribution of F on VLPs should be confirmed by cryoEM analyses. This would also confirm the symmetry of the clusters. The manuscript by Chernomordik et al. JBC 2004 showed that influenza HA outside the direct contact zone affects fusion, which could be further elaborated in the context of F clusters and the fusion mechanism.

We thank reviewer 3 for this suggestion. The distribution of F on VLPs was resolved by electron tomogram which showed that the NiV-F hexamer-of-trimers are arranged into a soccer ball-like structure 4. The role of influenza HA outside of the contact zone in fusion activation is an interesting phenomenon. It may address the energy transmission within and among clusters. We will pursue this topic in a future project.  

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

•  Please define all used abbreviations throughout the manuscript and in the SI.

We defined the abbreviations at their first usage. 

•  The sentence starting with "Additionally, ..." on line 155 appears to be incomplete.

We corrected this sentence.  

•  The statement starting with "As reported, ..." on line 181 should be supported by a reference.

We added a reference. 

•  In Fig. 4C, it is unclear what the x and y axes represent.  

Fig. 4C is a t-SNE plot for visualizing high-dimensional data in a low-dimensional space. It maintains the local data structure but does not represent exact quantitative relationships. In other words, points that are close together in Fig. 4C are also close in the high-dimensional space, meaning the OPTICS plots, which reflect the clustering patterns, are similar for two points that are positioned near each other in Fig. 4C. Therefore, the x and y axes do not represent the original, quantitative data, and thus the axis titles are meaningless.  

•  The reference on line 306 appears to be unformatted.

We reformatted the reference.  

Reviewer #2 (Recommendations For The Authors):

The authors need to include the overall statistics for each experiment (at least 30 to 50 cells with three independent experiments are needed). 

We highlighted the sample size (number of ROI and number of cells) used for analysis in the figure legend. The determination of the sample size is justified in Table 1 in the response letter. 

The authors need to generate a functional pseudovirus system (for example HIVpp/NiV F) to run both infectivity and fusion experiments (including Apr-BlaM assay). 

We tested viral entry using a VSV/NiV pseudovirus system and the viral entry kinetics using VLPs expressing NiV-M-β-lactamase. The results are presented in Fig. S1, S4, S6, and S7.  

Reviewer #3 (Recommendations For The Authors):

Even low resolution EM data on VLPs or viruses would strengthen the conclusions.

We thank this reviewer for the suggestion. We cited the NiV VLP images acquired by electron tomography 4, but we currently have limited resources to perform cryoEM on NiV VLPs.  

References.

(1) Liu, Q., Chen, L., Aguilar, H. C. & Chou, K. C. A stochastic assembly model for Nipah virus revealed by super-resolution microscopy. Nature Communications 9, 3050 (2018).

(2) Khetawat, D. & Broder, C. C. A Functional Henipavirus Envelope Glycoprotein Pseudotyped Lentivirus Assay System. Virology Journal 7, 312 (2010).

(3) Palomares, K. et al. Nipah Virus Envelope-Pseudotyped Lentiviruses Efficiently Target ephrinB2Positive Stem Cell Populations In Vitro and Bypass the Liver Sink When Administered In Vivo. J Virol 87, 2094–2108 (2013).

(4) Xu, K. et al. Crystal Structure of the Pre-fusion Nipah Virus Fusion Glycoprotein Reveals a Novel Hexamer-of-Trimers Assembly. PLoS Pathog 11, e1005322 (2015).

(5)    Bakker, E. & Swain, P. S. Estimating numbers of intracellular molecules through analysing fluctuations in photobleaching. Sci Rep 9, 15238 (2019).

(6) Nayak, C. R. & Rutenberg, A. D. Quantification of Fluorophore Copy Number from Intrinsic

Fluctuations during Fluorescence Photobleaching. Biophys J 101, 2284–2293 (2011).

(7) Salavessa, L. & Sauvonnet, N. Stoichiometry of ReceptorsReceptors at the Plasma MembranePlasma membrane During Their EndocytosisEndocytosis Using Total Internal Reflection Fluorescent (TIRF) MicroscopyMicroscopy Live Imaging and Single-Molecule Tracking. in Exocytosis and Endocytosis: Methods and Protocols (eds. Niedergang, F., Vitale, N. & Gasman, S.) 3–17 (Springer US, New York, NY, 2021). doi:10.1007/978-1-0716-1044-2_1.

(8) Slenders, E. et al. Confocal-based fluorescence fluctuation spectroscopy with a SPAD array detector. Light Sci Appl 10, 31 (2021).

(9) Annibale, P., Vanni, S., Scarselli, M., Rothlisberger, U. & Radenovic, A. Identification of clustering artifacts in photoactivated localization microscopy. Nat Methods 8, 527–528 (2011).

(10) Baumgart, F. et al. Varying label density allows artifact-free analysis of membrane-protein nanoclusters. Nat Methods 13, 661–664 (2016).

(11) Zanacchi, F. C. et al. A DNA origami platform for quantifying protein copy number in super-resolution. Nat Methods 14, 789–792 (2017).

(12) Jungmann, R. et al. Multiplexed 3D cellular super-resolution imaging with DNA-PAINT and Exchange-PAINT. Nature Methods 11, 313–318 (2014).

(13) Rubin-Delanchy, P. et al. Bayesian cluster identification in single-molecule localization microscopy data. Nat Methods 12, 1072–1076 (2015).

(14) Griffié, J. et al. 3D Bayesian cluster analysis of super-resolution data reveals LAT recruitment to the T cell synapse. Sci Rep 7, 4077 (2017).