Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

The authors investigate the role of HSPA2 during mouse preimplantation development. Knocking down HSPA2 in zygotes, the authors describe lower chances of developing into blastocysts, which show a reduced number of inner cell mass cells. They find that HSPA2 mRNA and protein levels show some heterogeneity among blastomeres at the 4-cell stage and propose that HSPA2 could contribute to skewing their relative contribution to embryonic lineages. To test this, the authors try to reduce HSPA2 expression in one of the 2-cell stage blastomere and propose that it biases their contribution to towards extra-embryonic lineages. To explain this, the authors propose that HSPA2 would interact with CARM1, which controls chromatin accessibility around genes regulating differentiation into embryonic lineage.

Strengths:

(1) The study offers simple and straightforward experiments with large sample sizes.

Thanks for your kind recognition.

(2) Unlike most studies in the field, this research often relies on both mRNA and protein levels to analyses gene expression and differentiation.

Thanks for your kind recognition.

Weaknesses:

(1) Image and statistical analyses are not well described.

Thanks for your advisable comment. We redescribe the image and statistical analyses in our revised version (line 255-257).

(2) The functionality of the overexpression construct is not validated.

Thanks for your kind suggestion. We validate the functionality of the overexpression construct in our revised version (Figure S3).

(3) Tracking of KD cells in embryos injected at the 2-cell stage with GFP is unclear.

Thanks for your kind suggestion. We randomly co-injected green fluorescent protein (Gfp) mRNA as a linage tracer with either Hspa2-siRNA or NC-FAM into one of the 2 -cell, and then monitored embryo development to the blastocyst stage (line 342-344).

(4) A key rationale of the study relies on measuring small differences in the levels of mRNA and proteins using semi-quantitative methods to compare blastomeres. As such, it is not possible to know whether those subtle differences are biologically meaningful. For example, the lowest HSPA2 level of the embryo with the highest level is much higher than the top cell from the embryo with the lowest level. What does this level mean then? Does this mean that some blastomeres grafted from strong embryos would systematically outcompete all other blastomeres from weaker embryos? That would be very surprising. I think the authors should be more careful and consider the lack of quantitative power of their approach before reaching firm conclusions. Although to be fair, the authors only follow a long trend of studies with the same intrinsic flaw of this approach.

Thanks for your advisable comment. Indeed, despite the approach drew on previous research (Zhou Cell 2018), we were clearly aware that this approach can only reflect relative comparisons. This means that the relative difference among the blastomeres from the same embryo were detected and compared. We did not compare the absolute levels of mRNA between different embryos. We also offered simple and straightforward experiments with large sample sizes to confirm this conclusion.

(5) Some of the analyses on immunostaining do not take into account that this technique only allows for semi-quantitative measurements and comparisons.

a) Some of the microscopy images are shown with an incorrect look-up table.

b) Some of the schematics are incorrect and misleading.

Thanks for your advisable comment. We revised microscopy images and schematics in our revised version.

Reviewer #2 (Public review):

Summary:

In this study, Gao et al. use RNA-seq to identify Hspa2 as one of the earliest transcripts heterogeneously distributed between blastomeres. Functional studies are performed using siRNA knockdown showing Hspa2 may bias cells toward the ICM lineage via interaction with the known methyltransferase CARM1.

Strengths:

This study tackles an important question regarding the origins of the first cell fate decision in the preimplantation embryo. It provides novelty in its identification of Hspa2 as a heterogeneous transcript in the early embryo and proposes a plausible mechanism showing interactions with Carm1. Multiple approaches are used to validate their functional studies (FISH, WB, development rates, proteomics). Given only 4 other transcripts/RNA have been identified at or before the 4-cell stage (LincGET, CARM1, PRDM14, HMGA1), this would be an important addition to our understanding of how TE vs ICM fate is established.

Thanks for your kind recognition.

The RNA-seq results leading the authors to focus on Hspa2 are not included in the manuscript. This dataset would serve as an important resource but is neither included nor discussed. Nor is it mentioned whether Hspa2 was identified in prior RNA-seq embryos studies (for example Deng Science 2014).

Thanks for your advisable comment. To identify genes that show a significantly high variability across blastomeres in the same embryo, we regressed out the embryo effect by established a new method, which will be published and uploaded to the database in the future. Thus, the RNA-seq results leading the we focus on Hspa2 are not included in the manuscript.   

In addition, the functional studies are centered on Hspa2 knockdown at the zygote (1-cell) stage, which would largely target maternal transcript. Given the proposed mechanism relies on Hspa2 heterogeneity post-ZGA (late 2-cell stage), the knockdown studies don't necessarily test this and thus don't provide direct support to the authors' conclusions. The relevance of the study would be improved if the authors could show that zygotic knockdown leads to symmetric Hspa2 levels at the late 2-cell and/or 4-cell stage. It may be possible that zygotic knockdown leads to lower global Hspa2 levels, but that asymmetry is still generated at the 4-cell stage.

Thanks for your advisable comment. We showed that the Hspa2 levels at the late 2-cell and 4cell stage after zygotic knockdown in our revised version (Figure S1 G-H, line 450-452).

Furthermore, the authors show that Hspa2 knockdown at the 1-cell stage lowers total Carm1 levels at the 4-cell stage. However, it is unclear how total abundance within the embryo alters lineage specification within blastomeres. The authors go on to propose a plausible mechanism involving Hspa2 and Carm1 interaction, but do not discuss how expression levels may be involved.

Thanks for your advisable comment. Previous research suggests that heterogeneous activity of the methyltransferase CARM1 results in differential methylation of histone H3R26 to modulate establishment of lineage specification (Zernicka-Goetz Cell 2018). Thus, we didn't discuss the total abundance within the embryo alters lineage specification.

Recommendations for the authors:  

Reviewer #1 (Recommendations for the authors):

(1) Major issue with analyses:

Image analysis needs to be much better explained than simply saying that ImageJ was used. Where are cells measured (at their equatorial plane? What is the size of the ROI?)? Ideally, the ROI and/or raw measurements should be provided.

Thanks for your advisable comment. We redescribe the Image analysis in our revised version (line 187-194). 

What are the objective criteria determining whether a cell is counted as GFP positive, CDX2 positive, or OCT4 positive? This is very unclear and key to the interpretation of many experiments.

Thanks for your advisable comment. We think that the cell containing fluorescence signals above background noise were counted positive.

Statistical analyses mention ANOVA in the methods but the student's t-test in the figure legend. Which is which? Most data are heavily normalized, which would unlikely fit the description for Student's t-test analyses.

Thanks for your advisable comment. We redescribe the statistical analyses in our materials and methods (line 253-260).

Figure 5H describes a relative fluorescence intensity with control at 1. The legend describes a normalization to "DNA" (I guess the authors meant DAPI), which is unlikely to give 1. This suggests that additional normalization was done and is not described. Is that the case? Also, since the authors propose that HSPA2 would control Histone modification and chromatin packing, I do not think that using DAPI is an appropriate way of normalizing the fluorescence signal.

Thanks for your advisable comment. We replaced DNA with DAPI in our revised version. Based on previous studies, we adopted DAPI as a normalized fluorescence signal (Zhou Cell 2018, Zernicka-Goetz Cell 2018).

Figure 1E shows data normalized to the lowest level while Figure 1H is normalized to the highest level. A consistent representation would be welcome.

Thanks for your advisable comment. We revised the Figure 1H in our revised version.

Is Figure 1C showing a t-test between correlations?

Yes, Figure 1C shows the t-test between correlation.

(2) Major issue with the interpretation of semi-quantitative methods and measurements:

qPCR, WB, immunostaining are all semi-quantitative methods that require some kind of normalization due to non-linear bias in the way the molecules are picked up. Such normalization makes it difficult to know whether a detectable difference is meaningful biologically speaking i.e. if a difference of 1 CT between blastomeres can be detected after qPCR, is it meaningful? If that were the case, then embryos with lower CT than others (Figure 1D) would not be able to develop into blastocyst, like siRNA injected embryos, or grafting a blastomere with a high CT onto an embryo with low CT would lead to the systematic differentiation of these strong blastomeres into ICM.

Thanks for your advisable comment. The CT values represent the relative mRNA levels of Hspa2 between blastomeres, and the higher CT value represents the lower expression of Hspa2 at mRNA level. Figure 1D shows the Hspa2 mRNA levels between blastomeres. The blastomere with lowlevel expression of the Hspa2 mRNA is not bias an ICM fates.  

The same goes for fluorescence analyses (Figure 1F). Can the authors also provide the measurements for DAPI as they did for HSPA2? I am sure that with enough measurements, DAPI is variable enough to give a statistical difference among blastomeres with questionable biological meaning.

I think the reasoning used here (unfortunately following the reasoning that has been used in a series of studies by other groups) of ranking blastomeres after semi-quantitative measurement is fundamentally flawed.

Thanks for your advisable comment. The DAPI was determined by the maximal area using a custom Python script. Based on previous studies, we adopted DAPI as a normalized fluorescence signal (Zhou Cell 2018). This approach is to normalize embryo-to-embryo variance from the technical reason.

(3) Major issue with overexpression experiment:

While the siRNA experiment is partially validated by qPCR and WB measurements of HSPA2 after KD, the overexpression experiment is not. Do the authors have any evidence that the construct they use is produced into protein and functional? Can the authors check by WB? Can the authors rescue the siRNA with their overexpression?

Thanks for your advisable comment. We verified the overexpression experiment by WB in in our revised version (Figure S3, line 360-361). Considering that siRNA degrades mRNA and prevents the mRNA translation process, we did not co-inject the siRNA with their overexpression.

The lack of effect of HSPA2 overexpression on blastocyst formation is difficult to reconcile with the interpretation from the authors that levels of HSPA2 bias lineages.

Have the authors tried lower concentrations? Have the authors tried FISH on their half-injected 2cell embryos? Of course, if the antibody against HSPA2 would work with immunostaining, that would be ideal.

Thanks for your advisable comment. We chose the concentrations for our study based on previous research (Zernicka-Goetz Cell 2016). To verified Hspa2 was successfully inject into one blastomere at the 2-cell stage, we observed green fluorescence after co-injected GFP mRNA with either siRNA or NC-FAM into one blastomere of the two-cell embryos. Thus, we didn't try FISH on half-injected 2-cell embryos. We tried to perform immunostaining experiments with various HSPA2 antibodies (Proteintech: 12797-1-AP, Abcam: ab108416) and no good results were achieved.

Author response image 1.

(4) Major issue with tracking of injected cells:

It is unclear what counts as a GFP-positive cell. In Figure 3D, most cells appear to have the same level of GFP.

Thanks for your advisable comment. The cell containing green fluorescence signals above background noise were counted GFP-positive in Figure 3D. Most cells seem to have the same level of GFP because they are daughter cells of the blastomeres injected with GFP.

In the images of GFP-expressing cells used to track the control of KD cells shown in Figure 3A, it seems that the control embryos have mostly GFP cells in the ICM. Is that the case, or just a bad example?

Thanks for your advisable comment. The green fluorescent signals in Figure 3A represented OCT4 protein, an ICM marker.

Can the authors do FISH against HSPA2 and visualize their GFP cells to validate the heterogeneous expression in situ?

Thanks for your advisable comment. We have verified the heterogeneous expression of HSPA2 in Figure1.

(5) Issue with fluorescent images:

Many images are shown with inappropriate look-up tables with saturated DAPI, OCT4, CDX2, and FISH. This raises the doubt that analyses were made on saturated images, which would be incorrect.

The LUT of Figure 5H should be adjusted similarly between the control and siRNA.

Thanks for your advisable comment. We revised some images which showed inappropriate lookup tables in our revised version. The LUT of Figure 5H had been adjusted between the control and siRNA. 

(6) Issue with schematics:

Schematics of blastomere isolation grown into blastocyst-like structures are misleading since the final blastocyst-like structure should not have a zona pellucida and should have fewer cells than regular blastocysts.

Thanks for your advisable comment. We revised schematics of blastomere grown into morula in our revised version (Figure 1A and Figure S1A).

The summary schematics in the final figure should not state HSPA2 -/- since experiments in the study did not use KO but KD.

Thanks for your advisable comment. We revised the summary schematics in our revised version.

The blastocysts are the same sizes as the cleavage stage or morula embryos which implies that cells lose volume to the lumen, which is not the case.

Thanks for your advisable comment. We revised the schematics in our revised version.

(7) Issue with data presentation:

In the tables within the figures, the number of decimals given should be the same for the mean and SE (one decimal should be more than enough).

Thanks for your advisable comment. We revised the figure 2H in our revised version.

The comparison of cell number and distribution within embryos (e.g. Figure 2B) would be best represented by a correlation analysis of TE vs ICM cells.

Thanks for your advisable comment. We add the figure of a correlation analysis of TE vs ICM cells in our revised version (Figure 3B).

The docking simulations are described in the main text as "experiments".

Thanks for your advisable comment. We redescribed the docking simulations in our revised version.

(8) Issue with data interpretation:

The reduced number of ICM cells is interpreted as a slowed-down cell cycle. This could also be explained by failed cytokinesis and the generation of binucleated or polyploid cells. Have the authors checked for that? For example, by looking at their DAPI staining. 

Thanks for your advisable comment. Our RNA-seq results revealed that the differentially expressed genes (DEGs) at blastocyst stage with HSPA2 knocking down are closely related to negative regulation of cell cycle, G1/S transition of mitotic cell cycle, mitotic cell cycle phase transition and regulation of mitotic cell cycle phase transition. Additionally, the previous study demonstrated that knockdown of HSPA2 reduced cell proliferation and led to G1/S phase cell cycle arrest (Hu Ann Transl Med 2019). Additionally, the lower cell number in ICM may also associated with failed cytokinesis and the generation of binucleated or polyploid cells. Thus, we guessed that HSPA2 has a role in ICM lineage establishment, although half of the ICM cells were able to survive with HSPA2 deficiency (line 463-472).

It is unclear to me why reduced ICM should lead to fewer blastocysts. Blastocysts should be able to form as long as their TE is fine. In Figure 2G, embryos seem to be cultured in close proximity, which is fine if they are healthy but not if some of the embryos start dying and releasing toxic compounds (e.g. ROS). Have the authors tried removing the dying KD embryos to see if the development of the remaining embryos would improve?

Thanks for your advisable comment. We think HSPA2 may affect blastocyst development by affecting other signaling pathways. And, the GO enriched terms was closely related to blastocyst development (Figure 2E). There was no significant difference in morula formation rate between Hspa2-KD group and NC group, thus the assumption that the toxic compounds released by some of the embryos that lead to downregulation of blastocyst rate may not be correct. Indeed, the rate of blastocyst formation in Hspa2-KD embryos was reduced significantly lower when few embryos was cultured separately. In addition, we discussed the possibility that the lower cell number in ICM may also associated with failed cytokinesis and the generation of binucleated or polyploid cells.

Author response image 2.

Reviewer #2 (Recommendations for the authors):

One of the significant findings in the paper is the discovery portion where Hspa2 is identified as a heterogeneous transcript. To improve the logic and impact of the manuscript, it may benefit from reorganizing some of the figures and text. For example:

(1) The paragraph in the introduction (Lines 56-68) should be moved to the discussion as the Hspa2 reveal should be in section 3.1, not prior to the RNA-seq results presented in Figure 1.

Thanks for your advisable comment. We think it is more logical that HSPA2 needs to be introduced in the introduction.

(2) Add text at the beginning of Section 3.1 to describe the rationale and results for the RNAseq. It would help the readers if the authors clearly stated why they chose the 4-cell stage.

Thanks for your advisable comment. We explain why we chose the 4-cell stage in our revised version (line 272-273).

(3) As this is the first time Hspa2 is identified, consider moving Figure S1C to the main figure to show expression throughout development.

Thanks for your advisable comment. We moved Figure S1C to the main figure in our revised version (line 286-291).

(4) Figure 1C: the correlation between Hspa2 and ICM markers would be strengthened if additional transcripts were used (Oct4, Sox2, Sox21). The graph in 1C would also be more informative if represented as a scatter plot with correlation coefficients (Nanog log2TPM vs Hspa2 log2TPM), rather than bar graphs.

Thanks for your advisable comment. We chose Nanog as the correlation between Hspa2 and Nanog, a ICM markers, was showing the strongest correlation in result. And, the figure 1C shows the stronger positive correlation between Nanog and Hspa2 in gene expression than random gene pairs (n=100, n means the number of random gene pairs). Thus, the figure 1C with bar graphs is easier to understand.

(5) Figure 1D: how were individual blastomeres grouped into B1-4? Individually run and then pooled based on relative expression?

Thanks for your advisable comment. Blastomeres are named B1 to B4 according to increasing Hspa2 concentration in figure 1E.

(6) Figures 1F, 1I, 5H: the DAPI channel appears to be saturated, but is used to normalize fluorescence intensity and may incorrectly account for light scattering within the embryo. Please clarify by adding more details regarding image analysis. Were partial stacks through the nucleus used for analysis, or max projections? Graph axes should be "relative fluorescence intensity."

Thanks for your advisable comment. We added the details of fluorescence images analysis. The graph axes had revised in our revised version.

(7) Line 278: the results in Figure S1C would benefit from more text regarding expression patterns throughout development. The maternal transcript appears to have a sharp downregulation by the early 2-cell stage, and is then upregulated coinciding with ZGA.

Thanks for your advisable comment. We added more describe of the Figure in main text (LINE 285-290).

(8) For the analyses in Figure 2 I-J and 2K-L, were arrested embryos excluded from analysis? This is an important detail as including arrested embryos would significantly bias the RNA-seq results. 

Thanks for your advisable comment. The arrested embryos were excluded in Figure 2 I-J and 2K-L.

(9) Figures 2G-H would be aided by converting the table in 2H to a bar graph and adding development rates for all stages (2-, 4-, 8-, morula, and blast). This would also show when an arrest occurs.

Thanks for your advisable comment. We converted the table in 2H to a bar graph.

(10) Blast rates are represented with too many significant digits (Figures 2H, 4B). They should only be reported to the closest ones given the unit of measure (number of blasts divided by number of zygotes). For instance, a blast rate of 81.63 {plus minus} 2.000 reflects excessive precision that is not measured in the data, it should rather read 82 {plus minus} 2%. This is also true for % cells (Figures 3E, 4H).

Thanks for your advisable comment. Values were rounded down to the one decimal place (rounded down).

(11) The clarity and impact of Figure 3A and 3D would benefit from 2D slices through the ICM. 

Thanks for your advisable comment. In order to get more comprehensive understanding of the 3D structure of blastocyst of Figure 3A and 3D, we did not choose 2D slices.

(12) To improve clarity and logic, separate the 1-cell and 2-cell knockdown experiments in the text and figures:

a) 1-cell knockdown with RNA-seq results (Fig 2A-F).

b) 1-cell knockdown showing less ICM/pluripotency markers in (combine Figures 2G-M and Figures 3A-B; "new Fig 3").

c) 2-cell knockdown tracing lineage (Figures 2D-E; "new Fig 4").

The new Figures 3 and 4 should mirror one another (i.e. for each knockdown experiment, development rates and cell counts should be included). For the 2-cell knockdown (Figures 2 D-E), what were the developmental rates (8-cell, morula, blast)?

Thanks for your advisable comment. However, in order to the overall logical of the article, we do not separate the 1-cell and 2-cell knockdown experiments in the text and figures. And, we added the developmental rates (8-cell, morula, blast) of 2-cell knockdown group in our revised version (Figure S2).

For the overexpression experiment (Figure 4), why were injections performed at the zygote stage versus the 2-cell stage? Given the significant downregulation of maternal transcript demonstrated in Figure S1C, it seems plausible that the injected RNA was also downregulated.

Thanks for your advisable comment. For the overexpression experiment, we first chose to inject Hspa2 mRNA at the zygote stage and found that the overexpression of Hspa2 does not induce blastomere cells to bias an ICM fate. The qRT-PCR results indicated that the expression level of Hspa2 in overexpression group was significantly increased compared with normal group at 4cell and blastocyst stage (Figure 4C, 4D).  In addition, there is no guarantee that an equal amount of Hspa2 mRNA be injected into each blastomere in 2-cell stage. Thus, we did not microinject Hspa2 mRNA into the 2-cell stage.

The 3.5 subheading overstates the results as the Hspa2-Carm1 interaction is not linked to lineage segregation. For example, a more specific subtitle might be, "Hspa2 interacts with Carm1 and alters H3R26me2 levels."

Thanks for your advisable comment. We revised the subtitle in our revised version (line 376).

Figures 5B-C and 5D-E. The qRT-PCR and WB analysis of knockdown blasts shows a correlation between Hspa2 downregulation and Carm1 downregulation. However, if the proposed mechanism is Hspa2 binding to Carm1 to mediate downstream methylation, why would it be expected to alter transcript levels at the 4-cell or blast stage? Please add further details and discussion in the results and discussion sections.

Thanks for your advisable comment. The reason we chose to work at the 4-cell stage is because previous studies on CARM1 have focused on the 4-cell stage (Zernicka-Goetz Cell 2018,2016). 

In the discussion, the statement in Lines 430-431 is an overinterpretation: "the heterogeneity of HSPA2... acts as an upstream factor to drive [the] first cell-fate decision." The knockdown experiments don't alter heterogeneity per se, but total abundance. Furthermore, the results do not show that heterogeneity drives heterogeneity in H3R26me2 patterns, for example.

Thanks for your advisable comment. We redescribe the relevant statement in the discussion.

More needs to be said regarding the ICM cells that persisted in the 1-cell KD experiment (Fig 3B). Lines 449-450 point out this result, but do not propose any plausible explanations. For instance, ICM cells may still form due to the incomplete knockdown achieved or the possibility that redundant pathways exist.

Thanks for your advisable comment. We redescribe the relevant statement in our revised version (line 468-473).

The 5th paragraph of the discussion seems incomplete. The authors point out a possible link between Hspa2 and Hippo and Wnt signaling pathways, but need to expand their discussion on how this may act as an additional mechanism incorporating Hspa2 with lineage segregation.

Thanks for your advisable comment. We redescribe the 5th paragraph of the discussion (line 483-494).

Statistics: all comparisons with greater than 2 groups should be performed with a one-way ANOVA and multiple comparisons, rather than Student's t-test (Figures 1B, 1D, 1E, 1F).

All figure legends lack statistical test details.

Thanks for your advisable comment. All figure legends added statistical test details in statistical analysis.

Minor comments:

In all graphs, individual blastomere expression levels should be represented as boxwhisker/bar/scatter/violin plots since the comparison is groups rather than time points (i.e. symbols should not be connected with a line in Figures 1B, 1D, 1F-G, 1I, S1D, S1F).

Thanks for your advisable comment. Each colored line represents a single cell, and the dots of the same color represent the blastomere of the same cell. Thus, we use a line representation individual blastomere.

For all fluorescent images, having two representative images may be confusing for the reader. Figures may be improved by just including one representative image for each stage/treatment (Figures 1F, 1I, S1F, 3A, 3D, 4E, 4G).

Thanks for your advisable comment. The figures just including one representative image for each stage in our revised version. In addition, two representative images from each group were shown for each treatment (Figures 3A, 3D, 4E, 4G).

The manuscript would be improved with thorough grammar and typo editing.

For example:

(1) Lines 18, 73, the wording is confusing, consider: "knockdown of Hspa2 in one of the two-cell blastomeres biased its progeny towards the trophectoderm lineage.".

(2) Line 23, overstatement. Consider: "we demonstrated that HSPA2 levels correlate with ICMassociated genes and that it interacts with the CARM1.".

(3) Line 25 confusing wording, "via the execution of commitment and differentiation phases.".

(4) Line 37, replace "that" with "of;" replace "cell-fate decisions" with "cell-fate decision".

(5) Line 40: needs space before (CARM1).

(6) Line 43: the wording is confusing, consider "can result in higher expression levels of".

(7) Line 45: wording, consider "Recent [studies have] further suggested".

(8) Line 70: plurality, consider "analyzed gene expression pattern".

(9) Line 73 typo: "prevents its".

(10) Line 76-77 wording, consider "Hspa2 expression patterns can bias cell fate in the mouse embryo".

(11) Line 276: remove "in whole embryos," since MII eggs are not embryos.

(12) Line 617 "There" should be "Three".

(13) Axis label in Fig 3b "Totle" should be "Total".

(14) Lines 417, 419 missing spaces.

(15) Line 448 missing word, "interfering [with] the cell cycle".

(16) Line 462 incorrect word, "[a]polar cells being specified as ICM".

(17) Line 469 incorrect plural, "cell differentiation".

Thanks for your advisable comment. We revised the whole manuscript carefully according to the reviewers' suggestions.

Author response:

The following is the authors’ response to the original reviews

Reviewer #1:

Summary:

The manuscript by Zhang et al describes the use of a protein language model (pLM) to analyse disordered regions in proteins, with a focus on those that may be important in biological phase separation. While the paper is relatively easy to read overall, my main comment is that the authors could perhaps make it clearer which observations are new, and which support previous work using related approaches. Further, while the link to phase separation is interesting, it is not completely clear which data supports the statements made, and this could also be made clearer.

We thank the reviewer for their thoughtful evaluation of our manuscript and for the supportive comments. As outlined in the responses below, we have made substantial revisions to clarify the novel observations presented in our study and to strengthen the connection between sequence conservation and phase separation.

Comment 1: With respect to putting the work in a better context of what has previously been done before, this is not to say that there is not new information in it, but what the authors do is somewhat closely related to work by others. I think it would be useful to make those links more directly.

We have addressed the specific comments as outlined below.

Comment 1a: Alderson et al (reference 71) analysed in detail the conservation of IDRs (via pLDDT, which is itself related to conservation) to show, for example, that conserved residues fold upon binding. This analysis is very similar to the analysis used in the current study (using ESM2 as a different measure of conservation). Thus, the result that "Given that low ESM2 scores generally reflect mutational constraint in folded proteins, the presence of region a among disordered residues suggests that certain disordered amino acids are evolutionarily conserved and likely functionally significant" is in some ways very similar to the results of that (Alderson et al) paper .

We thank the reviewer for the comment. However, we would like to clarify that our findings show subtle but important differences from those reported by Alderson et al. Specifically, Alderson et al. used AlphaFold2 predictions to identify IDRs that undergo disorder-to-order transitions, which the authors termed as conditionally folded IDRs. These regions could potentially be functionally important, assuming that function of IDRs necessitate folding.

We argue, however, that, the validity of this structure-function relationship for IDRs remains to be tested. In our opinion, The most direct way to evaluate the functional significance is via evaluating the evolutionary conservation.

As shown in Author response image 1, the correlation between pLDDT scores and the conservation score, while noticable, is significantly weaker than that between the ESM2 score and the conservation score.

Author response image 1.

Comparison of the correlation between AlphaFold2 pLDDT scores and conservation scores with the correlation between ESM2 scores and conservation scores. Calculations were performed using proteins in the MLO-hProt dataset. (A) Correlation between the mean AlphaFold2 pLDDT scores and conservation scores for various amino acids. Pearson correlation coefficients (r) are indicated in the figure legends. The four panels on the right present analogous correlation plots for amino acids grouped by structural order, as defined by their pLDDT scores. (B) Similar as in part A but for ESM2 scores.

Therefore, we believe that ESM2 score is a better indicator than AlphaFold2 pLDDT score for functional relevance.

Furthermore, for the human IDRs, we explicitly selected amino acids with pLDDT scores ≤ 70.

These would be classified as structureless, disordered amino acids, according to the study by Alderson et al. Nevertheless, as shown in Figures 2 and 3 of the main text, our analyses still identifies conserved regions. Therefore, these regions may function via distinct mechanisms than the disorder to order transition.

We now discuss the novelty of our work in the context of existing studies in the newly added Conclusions and Discussion: Related Work, as quoted below.

“Numerous studies have sought to identify functionally relevant amino acid groups within IDRs [cite]. For instance, using multiple sequence alignment, several groups have identified evolutionarily conserved residues that contribute to phase separation [cite]. Alderson et al. employed AlphaFold2 to detect disordered regions with a propensity to adopt structured conformations, suggesting potential functional relevance [alderson et al].

In contrast, our approach based on ESM2 is more direct: it identifies conserved residues without relying on alignment or presupposing that functional significance requires folding into stable 3D structures. Notably, many of the conserved residues identified in our analysis exhibit low pLDDT scores (Figure 2), implying potential functional roles independent of stable conformations.”

Comment 1b: Dasmeh et al, Lu et al and Ho & Huang analysed conservation in IDRs, including aromatic residues and their role in phase separation.

We thank the reviewer for bringing these works to our attention! We now explicitly discuss these studies in both the Discussion section as mentioned above and in the Introduction as quoted below.

“Evolutionary analysis of IDRs is challenging due to difficulties in sequence alignment [cite], though several studies have attempted alignment of disordered proteins with promising results [Dasmeh et al, Lu et al and Ho & Huang].”

Comment 1c: A number of groups have performed proteomewide saturation scans using pLMs, including variants of the ESM family, including Meier (reference 89, but cited about something else) and Cagiada et al (https://doi.org/10.1101/2024.05.21.595203) that analysed variant effects in IDRs using a pLM. Thus, I think statements such as "their applicability to studying the fitness and evolutionary pressures on IDRs has yet to be established" should possibly be qualified.

We added a new paragraph in the Introduction to discuss the application of protein language models to IDRs and cited the suggested references.

“While protein language models have been widely applied to structured proteins [cite], it is important to emphasize that these models themselves are not inherently biased toward folded domains. For example, the Evolutionary Scale Model (ESM2) [cite] is trained as a probabilistic language model on raw protein sequences, without incorporating any structural or functional annotations. Its unsupervised learning paradigm enables ESM2 to capture statistical patterns of residue usage and evolutionary constraints without relying on explicit structural information. Thus, the success of ESM2 in modeling the mutational landscapes of folded proteins [cite] reflects the model’s ability to learn sequence-level constraints imposed by natural selection — a property that is equally applicable to IDRs if those regions are also under functional selection. Indeed, protein language models are increasingly been used to analyze variant effects in IDRs [cite].”

Comment 2: On page 4, the authors write, "The conserved residues are primarily located in regions associated with phase separation." These results are presented as a central part of the work, but it is not completely clear what the evidence is.

We thank the reviewer this insightful comment. We realized that our wording is not as precise as we should have been. We meant to state that the regions associated with phase separation are significantly enriched in these conserved residues. This is a significant finding and indicates that phase separation could be a source of evolutionary pressure in dictating IDP sequence conservation. However, we do not intend to suggest that phase separation is the only evolutionary pressure.

The sentence has been revised to

“Notably, regions associated with phase separation are significantly enriched in these conserved residues.”

We further replaced the section title "Conserved, Disordered Residues Localize in Regions Driving Phase Separation" with "Regions Driving Phase Separation Are Enriched with Conserved, Disordered Residues" to further clarify our findings and avoid overinterpretation.

Finally, we revised the following sentence in the discussion

“Notably, these conserved, disordered residues are predominantly located in regions actively involved in phase separation, contributing to the formation of membraneless organelles.”

to

“Notably, regions actively involved in phase separation are enriched with these conserved, disordered residues, supporting their potential role in the formation of membraneless organelles.”

The submitted manuscript provides clear evidence supporting the enrichment of conserved residues in MLO-driving IDRs. Specifically, Figures 4A and 4C demonstrate that these IDRs exhibit a substantially higher fraction of conserved residues compared to other IDRs involved in phase separation.

In this analysis, the nMLO-hIDR group serves as a baseline, representing the distribution of conservation in disordered regions lacking MLO-related functions. In contrast, IDRs from MLOassociated groups show a pronounced lower shift in their median and interquartile ranges, indicating stronger evolutionary constraints. Within the dMLO cohort, the degree of conservation follows a distinct gradient: driving residues exhibit the highest levels of conservation, followed by participant residues, with non-participant residues showing values closer to the nMLO baseline. This pattern reflects the relative functional importance of each group in phase separation, with conservation levels corresponding to their roles in MLO scaffolding.

To further support this, we computed, for each IDR, the fraction of conserved amino acids. As shown in Figure S11B, for IDRs that actively contribute to phase separation, the fraction is indeed higher than those not involved in phase separation. This analysis is now included in SI.

During the revision, we explicitly evaluated whether conserved residues are preferentially located in regions associated with phase separation. To this end, for each protein in the MLO-hProt dataset, we calculated the probability p of finding conserved residues within regions contributing to phase separation. These regions include both "driving" and "participating" segments as defined in Figure 4 of the main text.

Figure S11A presents the distribution of p across all proteins. For comparison, we also include the distribution of 1− p, representing the probability of finding conserved residues in regions not associated with phase separation. On average, p exceeds 0.5, suggesting a tendency for conserved residues to be more frequently located in phase-separating regions. However, the difference between the two distributions is not statistically significant. This result may be due to the generally low density of conserved residues in IDRs, which makes the estimation of p challenging for individual proteins. Additionally, some conserved sites may be involved in functions unrelated to phase separation.

We added the following text to the Discussion section of the main text.

“We emphasize that the results presented in Figure 4 do not directly demonstrate that conserved residues are preferentially located in regions associated with phase separation. Although these regions are more enriched in conserved amino acids, their total sequence length can be smaller than that of non-phase-separating regions. As a result, the absolute number of conserved residues may still be higher outside phase-separating regions. To quantitatively assess this, we calculated, for each protein in the MLO-hProt dataset, the probability p of finding conserved residues within regions contributing to phase separation. These regions include both "driving" and "participating" segments, as defined in Figure 4 of the main text. Figure S11 shows the distribution of p across all proteins. For comparison, we also present the distribution of 1− p, which reflects the probability of finding conserved residues in non-phase-separating regions. While the average value of p exceeds 0.5, indicating a trend toward conserved residues being more frequently located in phase-separating regions, the difference between the two distributions is not statistically significant. Future studies with expanded datasets may be necessary to clarify this trend.”

Comment 3: It would be useful with an assessment of what controls the authors used to assess whether there are folded domains within their set of IDRs.

We acknowledge that our previous labeling may have caused some confusion. Protein sequences used in Figures 2 and 3 include both folded and disordered domains. Results presented in these figures were constructed using full-length protein sequences to highlight the similarities and differences in ESM2 scores between folded and disordered domains.

In contrast, the analyses presented in Figures 4 and 5 focus exclusively on IDRs to examine their role in phase separation.

To prevent further confusion, we have renamed the dataset used in Figures 2 and 3 as MLO-hProt, emphasizing that the analysis pertains to entire protein sequences. The term MLO-hIDR is now reserved for a new dataset that includes only disordered residues, as used in Figures 4 and 5, and corresponding SI Figures.

For the dMLO-IDR dataset, all except one amino acid (P40967, residue G592) are annotated as disordered in the MobiDB database (https://mobidb.org/). This database characterizes disordered regions based on a combination of predictive algorithms and experimental data. As illustrated in Figure S5A, 25.5% of the proteins in the dataset have direct experimental evidence supporting their disorderedness. These experimental annotations are derived from a diverse range of techniques (Figure S5B). For the remaining proteins, disorder was predicted by one or more computational tools. Although not all tools were applied to every protein, each protein in the dataset was identified as disordered by at least one method.

For human proteins, IDRs were identified based on AlphaFold2 pLDDT scores, using a threshold of 70. As established in prior studies [1, 2], the pLDDT score provides a quantitative measure of local structural confidence, with lower values indicating greater structural disorder. IDRs associated with conditional folding or disorder-to-order transitions generally exhibit high pLDDT values (e.g., >70).

Author response image 2 shows a violin plot of AlphaFold2 pLDDT scores for the various MLO-hIDR groups. The consistently low scores support the conclusion that these regions are structurally disordered.

We also cross-checked the MLO-hIDR regions against the MobiDB database. As shown in Figure S6, approximately 76% of the proteins in the dataset are predicted to contain disordered regions. Among the non-labeled segments with pLDDT scores ≤ 70, the majority are relatively short, with segments of 1–5 residues accounting for approximately 80%.

Author response image 2.

AlphaFold pLDDT scores of hIDRs in different MLO-related groups.

In addition to renaming the dataset, we also revised the manuscript to highlight the validation of disorderedness in section of Results: Regions Driving Phase Separation Are Enriched with Conserved, Disordered Residues.

“The presence of evolutionarily conserved disordered residues raises the question of their functional significance. To explore this, we identified disordered regions of MLO-hProt using a pLDDT score less than 70 and partitioned these regions into two categories: drivers (dMLO-hIDR), which actively drive phase separation, and clients (cMLO-hIDR), which are present in MLOs under certain conditions but do not promote phase separation themselves [cite]. Additionally, IDRs from human proteins not associated with MLOs, termed nMLO-hIDR, were included as a control. To enhance statistical robustness, we extended our dataset by incorporating driver proteins from additional species [cite], resulting in the expanded dMLO-IDR dataset. Beyond the pLDDT-based classification, the majority of residues in these datasets are also predicted to be disordered by various computational tools and supported by experimental evidence (Figures S5 and S6).”

Recommendation 1: The authors use the terms "evolutionary fitness of IDRs" (abstract and p. 5, for example), "fitness of amino acids" (p. 4), and "quantify the fitness of particular residues at specific sites" (p. 5). It is not clear what is meant by fitness in this context.

We thank the reviewer for pointing out the ambiguity in the term fitness. To enhance clarity, we have replaced “fitness" with “mutational tolerance" to more directly emphasize the evolutionary conservation of specific residues.

Recommendation 2: The authors write (p. 6) "Previous studies have demonstrated a strong correlation between ESM2 scores and changes in free energy related to protein structure stability". While that may be true, it might be worth noting that ESM2 scores report on the effects of mutations and function more broadly than stability because these models have previously been shown to capture conservation effects beyond stability.

We fully agree with the reviewer’s comment and have revised the main text accordingly. Specifically, the referenced sentence has been revised and relocated, as shown below.

“Our analysis demonstrated that HP1_α_’s structured domains consistently yield low ESM2 scores, reflecting strong mutational constraints characteristic of folded regions. These constraints are further evident in the local LLR predictions, as shown in Figure 2B, where we illustrate the folded region G120-T130. Given the functional importance of preserving the 3D of structured domains, mutations with greater detrimental effects are likely to disrupt protein folding substantially. This interpretation is consistent with previous studies reporting a significant correlation between ESM2 LLRs and changes in free energy associated with protein structural stability [cite].”

Recommendation 3: p. 10: The authors write "To exclude sequences that no longer qualify as homologs, we filtered for sequences with at least 20% identity to the reference". How did they decide on 20% and why? And over which residues are these 20% calculated.

We apologize for the earlier lack of clarity. Sequence alignment was performed using the full-length protein sequences, encompassing both folded and disordered regions. For each sequence, we calculated the percent identity by counting the number of positions, denoted as n, at which the amino acid matches the reference. The percent identity was then computed as n/N, where N represents the total length of the aligned reference sequence. This total includes residues in folded and disordered regions, as well as gap positions introduced during alignment.

We updated the Methods section of the main text to clarify.

“We performed multi-sequence alignment (MSA) analysis using HHblits from the HH-suite3 software suite [citations], a widely used open-source toolkit known for its sensitivity in detecting sequence similarities and identifying protein folds. HHblits builds MSAs through iterative database searches, sequentially incorporating matched sequences into the query MSA with each iteration. Sequence alignment was performed using the full-length protein sequences, encompassing both folded and disordered regions.

...

To refine alignment quality by focusing on closely related homologs, we filtered out sequences with ≤ 20% identity to the query, excluding weakly related sequences where only short segments show similarity to the reference. For each sequence, we calculated the percent identity by counting the number of positions, denoted as n, at which the amino acid matches the reference. The percent identity was then computed as n/N, where N represents the total length of the aligned reference sequence. This total includes residues in folded and disordered regions, as well as gap positions introduced during alignment.”

We selected a 20% sequence identity threshold to balance inclusion of true homologs with exclusion of distant matches that may not share functional relevance. To determine this cutoff, we compared identity thresholds of 0%, 10%, 20%, and 40% and examined the resulting distributions of conservation and ESM2 scores across aligned residues for MLO-hProt dataset (Author response image 3). Thresholds of 10%, 20%, and 40% produced qualitatively similar results, with a consistent correspondence between low ESM2 scores and high conservation. Lower thresholds introduced highly divergent sequences that added noise to the alignment, resulting in reduced overall conservation scores. In contrast, higher thresholds excluded homologs with potentially meaningful conservation, particularly in disordered regions where conservation scores tend to be relatively low.

Author response image 3.

Histograms of the ESM2 score and the conservation score, presented in a format consistent with Figure 3B of the main text. The conservation scores were computed using aligned sequences with identity thresholds of ≥0, ≥10%, ≥20%, and ≥40% (left to right). Contour lines represent different levels of −log_P_(CS,ESM2), where P is the joint probability density of conservation score (CS) and ESM2 score. Contours are spaced at 0.5-unit intervals, highlighting regions of distinct density.

Recommendation 4: In their description of "motif" searching algorithm (p. 20) I think that the search algorithm would give a different result whether the search is performed N->C or C->N (because the first residue (i) needs to have a score <0.5 but the last (j) could have a score >0.5 as long as the average is below 0.5. Is that correct? And if so, why did they choose an asymmetric algorithm? .

We thank the reviewer for highlighting the asymmetry in our motif-search algorithm.

To investigate this issue, we repeated the algorithm starting from the C-terminus and compared the resulting motifs with those obtained from the N-terminal scan. We found that the two sets of motifs overlap entirely: each motif identified from the C-terminal direction has a corresponding counterpart from the N-terminal scan. However, the motifs are not identical. The directionality of the search introduces additional amino acids—referred to here as peripheral residues—at the motif boundaries, which differ between the two sets.

As shown in Author response image 4, the number of peripheral residues is small relative to the total motif length.

To eliminate asymmetry and ambiguity, we have revised our method to perform bidirectional scans—from both the N- and C-termini—and define each motif as the overlapping region identified by both directions. This approach emphasizes the conserved core and avoids the inclusion of spurious terminal residues. The updated procedure is described in Methods: Motif Identification.

“To identify motifs within a given IDR, we implemented the following iterative procedure. Starting from either the N– or C–terminus of the sequence, we first locate the initial residue i whose ESM2 score falls within 0.5. From i, residues are sequentially appended…”

Author response image 4.

Number of peripheral residues and their relative length to the full-motif length identified from both sides. (A). The unique motifs identified from N-to-C terminal direction. (B) The unique motifs identified from C-to-N terminal direction.

“…in the direction toward the opposite terminus until the segment’s average ESM2 score exceeds 0.5; the first residue to breach this threshold is denoted j. The segment (i,i+1,..., j−1) is then recorded as a candidate motif. This process repeats starting from j until the end of the IDR is reached.

We perform this full procedure independently from both termini and designate the final motif as the intersection of the two candidate-motif sets. This bidirectional overlap strategy excludes terminal residues that might transiently satisfy the average-score criterion only due to adjacent low-scoring regions, thereby isolating the conserved core of each motif. All other residues—those not included in either directional pass—are classified as non-motif regions, minimizing peripheral artifacts.”

Accordingly, we have updated the Supplementary material: ESM2_motif_with_exp_ref.csv for the new identified motifs commonly exited from both N-terminal and C-terminal searches. Minor changes were observed in the set of motifs as being discussed, but these do not affect the main conclusions. Figures 5C, 5D, and S6 have been revised accordingly.

Reviewer #2:

Summary:

Unfortunately, I do not believe that the results can be trusted. ESM2 has not been validated for IDRs through experiments. The authors themselves point out its little use in that context. In this study, they do not provide any further rationale for why this situation might have changed. Furthermore, they mention that experimental perturbations of the predicted motifs in in vivo studies may further elucidate their functional importance, but none of that is done here. That some of the motifs have been previously validated does not give any credibility to the use of ESM2 here, given that such systems were probably seen during the training of the model.

We thank the reviewer for their detailed and thoughtful critique of our manuscript. We recognize the importance of careful model validation, especially in the context of IDRs, and appreciate the opportunity to clarify the scope and rationale of our study. Below, we respond point-by-point to the main concerns.

(1) The use of ESM2 is not validated for IDRs, and the authors provide no rationale for its applicability in this context.

We thank the reviewer for raising this important point.

First, we emphasize that ESM2 is a probabilistic language model trained entirely on amino acid sequences, without any structural supervision. The model does not receive any input about protein structure — folded or disordered — during training. Instead, it learns to estimate the likelihood of each amino acid at a given position, conditioned on the surrounding sequence context. This makes ESM2 agnostic to whether a sequence is folded or disordered; the model’s capacity to identify patterns of residue usage arises solely from the statistics of natural sequences.

As such, ESM2 is not inherently biased toward folded proteins, even though previous studies have demonstrated its usefulness in identifying conserved and functionally constrained residues in structured domains [3–9]. These findings support the broader utility of language models for uncovering evolutionary constraints — and by extension, suggest that similar signatures could exist in IDRs, particularly if they are under functional selection.

Indeed, if certain residues or motifs in IDRs are conserved due to their importance in biological processes (e.g., phase separation), we would expect such selection to be reflected in sequence-based features, which ESM2 is designed to detect. The model’s applicability to IDRs, then, is a natural extension of its core probabilistic architecture.

To further evaluate this, we carried out an independent in silico validation using multiple sequence alignments (MSAs). This analysis allowed us to compute the evolutionary conservation of individual amino acids without any reliance on ESM2. We then compared these conservation scores to ESM2 scores and found a strong correlation between the two. This provides direct, quantitative support for the idea that ESM2 is capturing biologically meaningful sequence constraints — even in disordered regions.

While we agree that experimental testing would ultimately provide the most compelling validation, we believe that our MSA-based comparison constitutes a strong and arguably ideal computational validation of the model’s predictions. It offers an orthogonal measure of evolutionary pressure that confirms the biological plausibility of ESM2 scores.

We added the following text in the introduction to highlight the applicability of ESM2 to IDRs.

“While protein language models have been widely applied to structured proteins, it is important to emphasize that these models themselves are not inherently biased toward folded domains. For example, the Evolutionary Scale Model (ESM2) [cite] is trained as a probabilistic language model on raw protein sequences, without incorporating any structural or functional annotations. It operates by estimating the likelihood of observing a given amino acid at a particular position, conditioned on the entire surrounding sequence context. This unsupervised learning paradigm enables ESM2 to capture statistical patterns of residue usage and evolutionary constraints without relying on explicit structural information. Thus, the success of ESM2 in modeling fitness landscapes of folded proteins reflects the model’s ability to learn sequence-level constraints imposed by natural selection — a property that is equally applicable to IDRs if those regions are also under functional selection. Indeed, protein language models are increasingly been used to analyze variant effects in IDRs [cite].”

(2) There is no experimental validation of the ESM2-based predictions in this study.

We agree that experimental validation would provide definitive support for the utility of ESM2 in IDRs, and we explicitly state this as a limitation in the revised manuscript as quoted below.

“Limitations: Despite the promising findings, our study has several limitations. Most notably, our analysis is purely computational, relying on ESM2-derived predictions and sequence-based conservation without accompanying experimental validation. While the strong correlation between ESM2 scores and evolutionary conservation provides compelling evidence that the identified motifs are functionally constrained, the precise biological roles of these motifs remain uncharacterized. ESM2 is well-suited for highlighting regions under selective pressure, but it does not provide mechanistic insights into how conserved motifs contribute to specific molecular functions such as phase separation, molecular recognition, or dynamic regulation. Determining these roles will require targeted experimental investigations, including mutagenesis and biophysical characterization.”

In addition, we revised the manuscript title from “Protein Language Model Identifies Disordered, Conserved Motifs Driving Phase Separation" to “Protein Language Model Identifies Disordered, Conserved Motifs Implicated in Phase Separation". This revision softens the original claim to better reflect the absence of direct experimental evidence for the motifs’ role in phase separation.

However, we also emphasize that the goal of our study is not to claim definitive predictive power, but rather to explore whether ESM2-derived mutational profiles align with known biological features of IDRs — and in doing so, to generate new, testable hypotheses.

In addition, while no in vivo experiments were performed, our study does include an in silico validation step, as detailed in the response to the previous comment. The strong correlation between ESM2 scores and conservation scores provides direct support for the utility of ESM2 in identifying residues under evolutionary constraint in disordered regions.

(3) The overlap between predicted motifs and known ones may be due totraining data leakage.

We respectfully clarify that training data leakage is not possible in this case, as ESM2 is trained using unsupervised learning on raw protein sequences alone. The model has no access to experimental annotations, functional labels, or knowledge of which motifs are involved in phase separation. It only models statistical sequence patterns derived from evolutionarily observed proteins.

Therefore, any agreement between ESM2-derived predictions and previously validated motifs arises not from memorization of experimental data, but from the model’s ability to learn meaningful sequence constraints from the natural distribution of proteins.

(4) The authors should revamp the study with a testable predictive framework.

We respectfully suggest that a full revamp is not necessary or appropriate in this context.

As outlined in our previous responses, we believe that certain misunderstandings about the nature and capabilities of ESM2 may have influenced the reviewer’s assessment.

Importantly, both Reviewer 1 and Reviewer 3 express strong support for the significance and novelty of this work, and recommend publication following minor revisions.

In this context, we believe the manuscript provides a useful contribution as a first step toward understanding disordered regions using language models, and that it has value even in the absence of direct experimental testing. We have now better positioned the manuscript in this light, clarified limitations, and suggested concrete next steps for follow-up research.

We hope these clarifications and revisions address the reviewer’s concerns, and we thank them again for helping us strengthen the framing, rigor, and clarity of our study.

Reviewer #3:

Summary:

This is a very nice and interesting paper to read about motif conservation in protein sequences and mainly in IDRs regions using the ESM2 language model. The topic of the paper is timely, with strong biological significance. The paper can be of great interest to the scientific community in the field of protein phase transitions and future applications using the ESM models. The ability of ESM2 to identify conserved motifs is crucial for disease prediction, as these regions may serve as potential drug targets. Therefore, I find these findings highly significant, and the authors strongly support them throughout the paper. The work motivates the scientific community towards further motif exploration related to diseases.

Strengths:

(1) Revealing conserved regions in IDRs by the ESM-2 language model.

(2) Identification of functionally significant residues within protein sequences, especially in IDRs.

(3) Findings supported by useful analyses.

We appreciate the reviewer’s thoughtful words and support for our work.

Weaknesses:

(1) Lack of examples demonstrating the potential biological functions of these conserved regions.

As detailed in the Response to Recommendation 6, we conducted additional analyses to connect the identified conserved regions with their biological functions.

(2) Very limited discussion of potential future work and of limitations.

We have substantially revised the Conclusions and Discussion section to provide a detailed analysis of the study’s limitations and to propose several directions for future research, as elaborated in our Response to Recommendation 5 below.

Recommendation 1: The authors describe the ESM2 score such that lower scores are associated with conserved residues, stating that "lower scores indicate higher mutational constraint and reduced flexibility, implying that these residues are more likely essential for protein function, as they exhibit fewer permissible mutational states." However, when examining intrinsically disordered regions (IDRs), which are known to drive phase separation, I observe that the ESM2 score is relatively high (Figure 3C, pLDDT < 50, and Supplementary Figure S2). Could the authors clarify how this relatively high score aligns with the conservation of motifs that drive phase separation?

We thank the reviewer for this insightful comment. We would like to clarify that most amino acids in the IDRs are not conserved, even for IDRs that contribute to phase separation. Only a small set of amino acids in these IDRs, which we term as motifs, are evolutionarily conserved with low ESM2 scores. Therefore, the ESM2 scores exhibit bimodal distribution at high and low values, as shown in Figures 4A and 4C of the manuscript. When averaged over all the amino acids, the mean ESM2 scores, plotted in Figure 3C, are relatively high due to dominant population of non-conserved amino acids.

Recommendation 2: The authors mention: "We first analyzed the relationship between ESM2 and pLDDT scores for human Heterochromatin Protein 1 (HP1, residues 1-191)". I appreciate this example as a demonstration of amino acid conservation in IDRs. However, it is questionable whether the authors could provide some more examples to support amino acid conservation particularly within the IDRs along with lower ESM2 score (e.g, Could the authors provide some additional examples of "conserved disordered" regions in various proteins which are associated with relatively low ESM2 score as appear in Figure 2A).

We thank the reviewer for this valuable suggestion. We want to kindly noted that the conserved residues on IDRs are prevalent as indicated in Figures 2D and 3B. To further illustrate the prevalence of “conserved disordered” regions, we generated ESM2 versus pLDDT score plots for the full dMLO–hProt dataset (82 proteins) in Figure S2. In these plots, residues with pLDDT ≤ 70 are highlighted in blue to denote structural disorder (dMLO-hIDR), and these disordered residues with ESM2 score ≤ 1.5 are shown in purple to indicate conserved disordered segments.

Recommendation 3: Could the authors plot a Violin conservation score plot for Figure 4A to emphasise the relationship between ESM2 scores and conservation scores of disordered residues?

We thank the reviewer for this suggestion. We included a violin plot illustrating the distribution of conservation scores for disordered residues across all four IDR groups, shown in Author response image 5. Consistent with the findings in Figure 4A, the phase separation drivers (dMLO-hIDR and dMLOIDR) exhibit a higher proportion of conserved amino acids compared to the client group (cMLOhIDR).

We also note that the nMLO-hIDR group may contain conserved residues due to functions unrelated to MLO formation, which could contribute to the higher observed levels of conservation in this group.

Author response image 5.

Violin plots illustrating the distribution of conservation scores for disordered residues across the nMLO–hIDR, cMLO–hIDR, dMLO–hIDR, and dMLO–IDR datasets. Pairwise statistical comparisons were conducted using two-sided Mann–Whitney U tests on the conservation score distributions (null hypothesis: the two groups have equal medians). P-values indicate the probability of observing the observed rank differences under the null hypothesis. Statistical significance is denoted as follows: ***: p < 0.001; **: p < 0.01; *:p < 0.05;

Recommendation 4: It will be appreciated if the authors could add to Figure 4 Violin plots, a statistical comparison between the different groups.

We thank the reviewer for this valuable suggestion. We included the p-values for Figures 4A and 4C to quantify the statistical significance of differences in the distributions.

Most comparisons are highly significant (p < 0.001), while the largest p-value (p = 0.089) between the conservation score of driving and non-participating groups (Figure 4C) still suggests a marginally significant trend.

Recommendation 5: Could the authors expand more on potential future research directions using ESM2, given its usefulness in identifying conserved motifs? Specifically, how do the authors envision conserved motifs will contribute to future discoveries/applications/models using ESM (e.g, discuss the importance of conserved motifs, especially in IDRs motifs, in protein phase transition prediction in relation to diseases).

We thank the reviewer for this insightful comment. To further assess the functional relevance of the conserved motifs, we incorporated pathogenic variant data from ClinVar [10, 11] to evaluate mutational impacts. As shown in Figure S12A and B, a substantial number of pathogenic variants in MLO-hProt proteins are associated with low ESM2 LLR values. This pattern holds for both folded and disordered residues.

Moreover, we observed that variants located within motifs are more frequently pathogenic compared to those outside motifs (Figure S12C). In the main text, motifs were defined only for driver proteins; however, the available variant data for this subset are limited (6 data points). To improve statistical power, we extended motif identification to include both client and driver human proteins, following the same methodology described in the main text. Consistent with previous findings, variants within motifs in this expanded set are also more likely to be pathogenic. These results further support the functional importance of both low ESM2-scoring residues and the conserved motifs in which they reside.

The following text was added in the Discussion section of the manuscript to discuss these results and outline future research directions.

“Several promising directions could extend this work, both to refine our mechanistic understanding and to explore clinical relevance. One avenue is testing the hypothesis that conserved motifs in scaffold proteins act as functional stickers, mediating strong intermolecular interactions. This could be evaluated computationally via free energy calculations or experimentally via interaction assays. Deletion of such motifs in client proteins may also reduce their partitioning into condensates, illuminating their roles in molecular recruitment.

To explore potential clinical implications, we analyzed pathogenicity data from Clin-Var [10, 11]. As shown in Figure S12A, single-point mutations with low LLR values—indicative of constrained residues—are enriched among clinically reported pathogenic variants, while benign variants typically exhibit higher LLR values. Moreover, mutations within conserved motifs are significantly more likely to be pathogenic than those in non-motif regions (Figure S12B). These findings highlight the potential of ESM2 as a first-pass screening tool for identifying clinically relevant residues and suggest that the conserved motifs described here may serve as priorities for future studies, both mechanistic and therapeutic.”

Moreover, the functional significance of conserved motifs, particularly their implications in disease and pathology, warrants further investigation. As an initial analysis, we incorporated ClinVar pathogenic variant data [citation] to assess mutational effects within our datasets. As illustrated in Figure R12A, single-point mutations with low LLR values are enriched among clinically reported pathogenic variants, whereas benign variants are more commonly associated with higher LLR values. Notably, mutations within conserved motifs are substantially more likely to be pathogenic compared to those in non-motif regions. These findings highlight the potential of ESM2 as a firstpass tool for identifying residues of clinical relevance. The conserved motifs identified here may be prioritized in future studies aimed at elucidating their biological roles and evaluating their viability as therapeutic targets.

Recommendation 6: The authors mention: "Our findings provide strong evidence for evolutionary pressures acting on specific IDRs to preserve their roles in scaffolding phase separation mechanisms, emphasizing the functional importance of entire motifs rather than individual residues in MLO formation." They also present a word cloud of functional motifs in Figure 5D. Although it makes sense that evolutionarily conserved motifs, especially within the IDRs regions, act as functional units, I think there is no direct evidence for such functionality (e.g., examples of biological pathways associated with IDRs and phase separation). Hence, there is no justification to write in the figure caption: "ESM2 Identifies Functional Motifs in driving IDRs" unless the authors provide some examples of such functionality. This will even make the paper stronger by establishing a clear connection to biological pathways, and hence these motifs can serve as potential drug targets.

We thank the reviewer for this insightful suggestion. We have replaced “functional motifs" with “conserved motifs" in the figure caption.

Identifying the precise biological pathways associated with the conserved motifs is a complex task and a comprehensive investigation lies beyond the scope of this study. Nonetheless, as an initial effort, we explored the potential functions of these motifs using annotations available in DisProt (https://disprot.org/).

DisProt is the leading manually curated database dedicated to IDPs, providing both structural and functional annotations. Expert curators compile experimentally validated data, including definitions of disordered regions, associated functional terms, and supporting literature references. Author response image 6 presents a representative DisProt entry for DNA topoisomerase 1 (UniProt ID: P11387), illustrating its structural and biological annotation.

For each motif, we located the corresponding DisProt entry and assigned a functional annotation based on the annotated IDR from which the motif originates. We emphasize that this functional assignment should be regarded as an approximation. Because experimental annotations often pertain to the entire IDR, regions outside the motif may also contribute to the reported function.

Nevertheless, the annotations provide valuable insights.

Author response image 6.

Screenshot of information provided by the DisProt database. Detailed annotations of biological functions and structural features, along with experimental references, are accessible via mouse click.

Approximately 50% of ESM2-predicted IDR motifs lack functional annotations. Among those that are annotated, motifs from the dMLO-IDR dataset are predominantly associated with “molecular condensate scaffold activity,” followed by various biomolecular binding functions (Author response image 7A). These findings support the role of these motifs in MLO formation.

For comparison, we applied the same identification procedure (described in Methods: Motif Identification) to motifs from the nMLO-hIDR dataset. In contrast to the dMLO-IDR motifs, these exhibit a broader range of annotated functions related to diverse cellular processes. Collectively, these results suggest that motifs identified by ESM2 are aligned with biologically relevant functions captured in current databases.

Finally, as illustrated in Figure S12 and discussed in the Response to Recommendation 5, variants occurring within identified motifs are more likely to be pathogenic than those in non-motif regions, further underscoring their functional importance.

Author response image 7.

Biological functions of ESM2-predicted motifs. (A) Distribution of biological functions associated with all identified motifs from dMLO-IDR driving groups. (B) Distribution of biological functions associated with all identified motifs from nMLO-hIDR groups.

Recommendation 7: In Figure 2C the authors present FE (I assume this is free energy), some discussion about the difference in the free energy referring to the "a" region is missing (i.e. both "Folded" and "Disordered" regions are associated with low ESM score but with low and high free energy (FE), respectively.

We thank the reviewer for the comments. FE indeed abbreviates free energy. To improve clarify and avoid confusion, we have updated all figure captions by replacing “FE” with “−logP” to explicitly denote the logarithm of probability in the contour density plots.

We used “a" in Figures 2C and 2D to refer to regions with low ESM2 scores, which appears a local minimum in both plots. Since most residues in folded regions are conserved, region a has lower free energy than region b in Figure 2C. On the other hand, as most residues in disordered regions are not conserved, as we elaborated in Response to Recommendation 1, region a has lower population and higher free energy than region b.

To avoid confusion, we have replaced “a" and “b" in Figure 2D with “I" and “II".

Recommendation 8: Figure S2: It would be useful to plot the same figure for structured and disordered regions as well.

We are not certain we fully understood this comment, as we believe the requested analysis has already been addressed. In Figure S2, we used the AlphaFold2 pLDDT score to represent the structural continuum of different protein regions, where residues with pLDDT > 70 (red and lightred bars) are classified as structured, while those with pLDDT ≤ 70 (blue and light-blue bars) are classified as disordered.

Minor suggestion 1: Could the authors clarify the meaning of the abbreviation "FE" in the colorbar of the contour line? I assume this is free energy.

We have updated all contour density plot figure captions by replacing “FE” with “−logP” to explicitly denote the logarithm of probability.

Minor suggestion 2: In Figure 2A - do the authors mean "Conserved folded" instead of just "Folded"? If so, could the authors indicate this?

We thank the reviewer for this comment. The ESM2 scores indeed suggest that, within folded regions, there may be multiple distinct groups exhibiting varying degrees of evolutionary conservation. However, as our primary focus is on IDRs, we chose not to investigate these distinctions further.

Figure 2A illustrates a randomly selected folded region based on AlphaFold2 pLDDT scores.

References

(1) Ruff, K. M.; Pappu, R. V. AlphaFold and Implications for Intrinsically Disordered Proteins. Journal of Molecular Biology 2021, 433, 167208.

(2) Alderson, T. R.; Pritišanac, I.; Kolaric, Ð.; Moses, A. M.; Forman-Kay, J. D. Systematic´ Identification of Conditionally Folded Intrinsically Disordered Regions by AlphaFold2. Proceedings of the National Academy of Sciences of the United States of America, 120, e2304302120.

(3) Brandes, N.; Goldman, G.; Wang, C. H.; Ye, C. J.; Ntranos, V. Genome-Wide Prediction of Disease Variant Effects with a Deep Protein Language Model. Nature Genetics 2023, 55, 1512–1522.

(4) Lin, Z. et al. Evolutionary-Scale Prediction of Atomic-Level Protein Structure with a Language Model. 2023.

(5) Zeng, W.; Dou, Y.; Pan, L.; Xu, L.; Peng, S. Improving Prediction Performance of General Protein Language Model by Domain-Adaptive Pretraining on DNA-binding Protein. Nature Communications 2024, 15, 7838.

(6) Gong, J. et al. THPLM: A Sequence-Based Deep Learning Framework for Protein Stability Changes Prediction upon Point Variations Using Pretrained Protein Language Model. Bioinformatics 2023, 39, btad646.

(7) Lin, W.; Wells, J.; Wang, Z.; Orengo, C.; Martin, A. C. R. Enhancing Missense Variant Pathogenicity Prediction with Protein Language Models Using VariPred. Scientific Reports 2024, 14, 8136.

(8) Saadat, A.; Fellay, J. Fine-Tuning the ESM2 Protein Language Model to Understand the Functional Impact of Missense Variants. Computational and Structural Biotechnology Journal 2025, 27, 2199–2207.

(9) Chu, S. K. S.; Narang, K.; Siegel, J. B. Protein Stability Prediction by Fine-Tuning a Protein Language Model on a Mega-Scale Dataset. PLOS Computational Biology 2024, 20, e1012248.

(10) Landrum, M. J.; Lee, J. M.; Riley, G. R.; Jang, W.; Rubinstein, W. S.; Church, D. M.; Maglott, D. R. ClinVar: Public Archive of Relationships among Sequence Variation and Human Phenotype. Nucleic Acids Research 2014, 42, D980–D985.

(11) Landrum, M. J. et al. ClinVar: Improving Access to Variant Interpretations and Supporting Evidence. Nucleic Acids Research 2018, 46, D1062–D1067.

Author response:

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

Reviewer 1:

(1) A major issue throughout the paper is that Hox expression analysis is done exclusively through quantitative PCR, with values ranging from 2-fold to several thousand-fold upregulation, with no antibody validation for any Hox protein (presumably they are all upregulated).

Thank you for your comment.

We tried to verify the stimulated Hox expression pattern by in situ hybridization. Although in early embryos (E9.5) we could detect clearly hox (i.e. Hox8 and Hox9 in Author response image 1) expression patterns in the neural tube by whole mount in situ hybridization, we failed to detect a clear pattern in the brain stem at E18.5 either in whole mount tissue or on sections. That’s one reason that we turned to single nuclear RNA-seq instead.

This is likely due to their low expression levels at late developmental stages and need to be detected by more sensitive method. However, we estimated that the stimulated expression levels of the representative Hox genes are at least comparable to the physiological levels at posterior spinal cord to evoke a functional effect.

Author response image 1.

Some Hox8 and Hox9 expression pattern in E9.5 embryos.

(2) In Figure 1, massive upregulation of most Hox genes in the brainstem is shown after e16.5 but the paper quickly focuses on analysis of PN nuclei. What are the other consequences of this broad upregulation of Hox genes in the brainstem? There is no discussion of the overall phenotype of the mice, the structure of the brainstem, the migration of neurons, etc. The very narrow focus on motor cortex projections to PN nuclei seems bizarre without broad characterization of the mice, and the brainstem in particular. There is only a mention of "severe motor deficits" from previous studies, but given the broad expression of Rnf220, the fact that is a global knockout, and the effects on spinal cord populations shown previously the justification for focusing on PN nuclei does not seem strong.

Thank you for your comment.

Although RNF220 is important for the dorsal-ventral patterning of the spinal cord as well as the hindbrain during embryonic development, the earlier neural patterning and differentiation are normal in the Rnf220+/- mice (Wang et al., 2022). However, these mice showed reduced survival and motility to various degree postnatally (Ma et al., 2019; Ma et al., 2021), likely suggesting a dosage dependent role of RNF220 in maintaining late neural development. As our microarray assay showed the deregulation of the Hox genes in the brain, we followed this direction in this study and narrowed down the affected region to the pons. Our single nuclear RNA-Seq (snRNA-seq) data further shows that the Hox de-regulation mainly occurred in 3 clusters of neurons. However, the pons is complex and contains tens of nuclei. And the current resolution of our data does not support to assign a clear identity to each of them. Although it is clear that more nuclei are likely affected, the PN (cluster7) is the only cluster we can identify to follow in the current study. 

As to general effect of RNF220 haploinsufficiency on the brainstem, we carried out Nissl staining assays and found no clear difference in neuronal cell organization between WT and Rnf220+/- pons (revised Figure 2-figure supplement 2).

(3) It is stated that cluster 7 in scRNA-seq corresponds to the PN nuclei. The modest effect shown on Hox3-5 expression in that data in Figure 1 is inconsistent with the larger effect shown in Figure 2.

Thank you for your comment.

Due to the low efficiency of snRNA-seq and the depth of the sequencing, the quantification of the Hox expression based on the snRNA-seq data is likely less accurate as the qRT-PCR. In addition, only mRNAs in the nuclear could be captured by snRNA-seq, while mRNAs in both the nuclear and cytoplasm were reversed-transcribed and examined for qRT-PCR assays in Figure 2A.

(4) Presumably, Hox genes are not the only targets of Rnf220 as shown in the microarray/RNA-sequencing data. There is no definitive evidence that any phenotypes observed (which are also not clear) are specifically due to Hox upregulation. The only assay the authors use to look at a Hox-dependent phenotype in the brainstem is the targeting of PN nuclei by motor cortex axons. This is only done in 2 animals and there are no details as to how the data was analyzed and quantified. The only 2 images shown are not convincing of a strong phenotype, they could be taken at slightly different levels or angles. At the very least, serial sections should be shown and the experiment repeated in more animals. There is also no discussion of how these phenotypes, if real, would relate to previous work by the Rijli group which showed very precise mechanisms of synaptic specificity in this system.

Thank you for your comments and suggestions.

The deregulation of Hox is the most obvious phenomena observed from the RNA-seq data, and we tried to assign its specific phenotypic effect in this study. As the roles of Hox in PN patterning and circuit formation is well established, we focused on the PN in the following study. Based on literature, we carried out the circuit analysis to examine the targeting of PN neurons by the motor cortex axons. A cohort of additional animals with different genotypes (n=10 for WT and n=9 for Rnf220+/-) were used to repeat the experiment and we got the same conclusion. More detailed information on data analysis and serial images were included in the revised manuscript and figure legends.

(5) The temporal aspect of this regulation in vivo is not clear. The authors show some expression changes begin at e16.5 but are also present at 2 months. Is the presumed effect on neural circuits a result of developmental upregulation at late embryonic stages or does the continuous overexpression in adult mice have additional influence? Are any of the Hox genes upregulated normally expressed in the brainstem, or PN specifically, at 2 months? Why perform single-cell sequencing experiments at 2 months if this is thought to be mostly a developmental effect? Similarly, the significance of the upregulated WRD5 in the pons and pontine nuclei at 2 months in Figure 3 is not clear.

Thank you for your comment.

The spatial and temporal expression pattern of Hox genes is established at early embryonic stages and then maintained throughout developmental stage in mammals. As we have shown, the de-repression of Hox genes is a long-lasting defect in Rnf220+/- mice beginning at late embryonic stages. Since the neuronal circuit is established after birth in mice, we speculated that the neuronal circuit defects from motor cortex to PN neurons were due to the long-lasting up-regulation of Hox genes in PN neurons. We could not distinguish the effect on neural circuit a result of Hox genes developmental upregulation or continuous overexpression in adult mice. An inducible knockout mouse model may help to answer this question in the future. The discussion on this point was included in the revised manuscript.

We carried out snRNA-seq analysis using pons tissues from adult mice aiming to identify the specific cell population with Hox up-regulation, which we failed to specify by in situ hybridization.

We repeated the related experiments in the original Figure 3 and some of the blot images were replaced and quantified.

(6) In Figure 3C, the levels of RNF220 in wt and het don't seem to be that different.

We repeated the experiments and changed the related image in the revised Figure 3C.

(7) Based on the single-cell experiments, and the PN nuclei focus, the rescue experiments are confusing. If the Rnf220 deletion has a sustained effect for up to 2 months, why do the injections in utero? If the focus is the PN nuclei why look at Hox9 expression and not Hox3-5 which are the only Hox genes upregulated in PN based on sc-sequencing? No rescue of behavior or any phenotype other than Hox expression by qPCR is shown and it is unclear whether upregulation of Hox9 paralogs leads to any defects in the first place. The switch to the Nes-cre driver is not explained. Also, it seems that wdr5 mRNA levels are not so relevant and protein levels should be shown instead (same for rescue experiments in P19 cells).

Thank you for your comments.

Since our data suggest that the upregulation of Hox genes expression is a long-lasting effect beginning at the late embryonic stage of E16.5, we conducted the rescue experiments by in utero injection of WDR5 inhibitor at E15.5 and examined the expression of Hox genes at E18.5. Although it is also necessary to examine whether the rescue effect by WDR5 inhibitor injection is also a long-lasting effect at adult stages, it is difficult to distinguish the embryos or pups when they were given birth. As a supplement, rescue assays with genetic ablation of Wdr5 gene were conducted and the results showed that genetic ablation of a single copy of Wdr5 allele could revere the upregulation of Hox genes by RNF220 haploinsufficiency in the hindbrains at P15.

Most of the upregulated Hox genes including both Hox9 and Hox3-5 were examined in our rescue experiments. Since this study focuses on the PN nuclei, the results of Hox3-5 genes were shown in the revised main Figure 6.

We conducted rescue experiments by deleting Wdr5 in neural tissue using Nestin-Cr_e mice because _Wdr5+/- mice is embryonic lethal. And the up-regulation of Hox genes could be also observed in the hindbrains of Rnf220fl/wt; Nestin-Cre mice. Although Rnf220fl/wt; Wdr5fl/wt; Nestin-Cre mice are viable and could survive to adult stages, developmental defects in the forebrains, including cerebral cortex and hippocampus, were observed in Rnf220fl/wt;Wdr5fl/wt;Nestin-Cre mice. Therefore, no rescue of behavior tests was conducted in this study. We believe that it is out of the scope of this study to discuss the role of WDR5 in the development of forebrains.

The potential defects due to the up-regulation of Hox9 paralogs awaits further investigations.

Wdr5 mRNA levels were firstly examined to confirm the genetic deletion or siRNA mediated knockdown of Wdr5 genes. We have carried out western blot to examine the WDR5 protein levels and the results were included in the revised Figure 3.

(8) What is the relationship between Retinoic acid and WRD5? In Figure 3E there is no change in WRD5 levels without RA treatment in Rnf KO but an increase in expression with RA treatment and Rnf KO. However, the levels of WRD5 do not seem to change with RA treatment alone. Does Rnf220 only mediate WDR5 degradation in the presence of RA? This does not seem to be the case in experiments in 293 cells in Figure 4.

Thank you for your comment.

We believe that the regulation of WDR5 and Hox expression by RNF220 is context dependent and precisely controlled in vivo, depending on the molecular and epigenetic status of the cell, which is fulfilled by RA treatment in P19 cells. In Figure 4, the experiment is based on exogenous overexpression assays, which might not fully reflect the situation in vivo.

(9) Why are the levels of Hox upregulation after RA treatment so different in Figure 5 and Figure Supplement 5?

In Figure.5C, the Hox expression levels were normalized against the control group in the presence of RA; while in Figure Supplement 5 they were normalized to the control group without RA treatment.

(10) In Figures 4B+C which lanes are input and which are IP? There is no quantitation of Figure 4D, from the blot it does look that there is a reduction in the last 2 columns as well. The band in the WT flag lane seems to have a bubble. Need to quantitate band intensities. Same for E, the effect does not seem to be completely reversed with MG132.

Thanks for pointing this out. The labels were included in the revised Figure 4B and 4C.

We repeated the experiments for Figure 4D and 4E. Some of bot images were replaced and quantified in the revised Figure 4D and 4E.

Reviewer 2:

(1) Figure 1E shows that Rnf220 knockdown alone could not induce an increase in Hox expression without RA, which indicates that Rnf220 might endogenously upregulate Retinoic acid signaling. The authors should test if RA signaling is downstream of Rnf220 by looking at differences in the expression of Retinaldehyde dehydrogenase genes (as a proxy for RA synthesis) upon Rnf220 knockdown.

Thank you for your comment and suggestion.

Two sequential reactions are required for RA synthesis from retinol, which catalyzed by alcohol dehydrogenases (ADHs)/ retinol dehydrogenase (RDH) and retinaldehyde dehydrogenase (RALDHs also known as ALDHs) respectively. When RA is no longer needed, it is catabolized by cytochrome enzymes (CYP26 enzymes) (Niederreither, et al.,2008; Kedishvili et al., 2016). Here, we test ADHs、ALDHs and CYP26 enzymes in E16.5 WT and Rnf220-/- embryos.

The results are as follows. ADH7 and ADH10 are slightly upregulated. ALDH1 and ALDH3 are upregulated and downregulated in Rnf220-/- embryos, respectively, but there is no significant change in the expression of ALDH2, which plays a key role in RA synthesis during embryonic development (Niederreither, et al.,2008). Furthermore, Cyp26a1 which responsible for RA catabolism was upregulated in Rnf220-/- embryos. Collectively, these data do not support a clear effect on RA signaling by RNF220.  

Author response image 2.

The effect of Rnf220 on RA synthesis and degradation pathways

(2) In Figure 2C-D further explanation is required to describe what criteria were used to segment the tissue into Rostral, middle, and caudal regions. Additionally, it is unclear whether the observed change in axonal projection pattern is caused due to physical deformation and rearrangement of the entire Pons tissue or due to disruption of Hox3-5 expression levels. Labeling of the tissue with DAPI or brightfield image to show the structural differences and similarities between the brain regions of WT and Rnf220 +/- will be helpful.

Thank you for your comment and suggestion.

More information on the quantification of the results shown in Figure 2C-D was included in our revised manuscript. We carried out Nissl staining assays using coronal sections of the brainstem and found that there is no significant difference in neuronal cell organization between WT and Rnf220+/- (revised Figure 2-figure supplement 2).

(3) Line 192-195. These roles of PcG and trxG complexes are inconsistent with their initial descriptions in the text - lines 73-74.

We are sorry for the mistake. We carefully revised the related descriptions to avoid such mistake. Thank you.

(4) In Figure 4D, the band in the gel seems unclear and erased. Please provide a different one. These data show that neither Rnf220 nor wdr5 directly regulates Hox gene expressions. The effect of double knockdown in the presence of RA suggests that they work together to suppress Hox gene expression via a different downstream target. This point should be addressed in the text and discussion section of the paper. example for the same data which shows a full band with lower intensity.

Thank you for your suggestion.

We repeated the experiment of Figure 4D and some of the blot images were replaced in the revised Figure 4D.

Indeed, in the presence of RA, knockdown of Rnf220 alone can upregulate the expression Hox genes (Figure 5C). Knockdown of Wdr5 could reverse the upregulation of Hox genes in RNF220 knockdown cells, suggesting that Rnf220 regulated Hox gene expression in a Wdr5 dependent manner. However, in the absence of RA, none of Rnf220 knockdown, Wdr5 knockdown or Rnf220 and Wdr5 double knockdown had a significant effect on the expression of Hox genes in P19 cells. It seems that RA signaling plays a crucial role for the regulation of RNF220 to WDR5 in P19 cells and discussion on this point was included in the revised manuscript.

(5) In Figure 4G the authors could provide some form of quantitation for changes in ubiquitination levels to make it easier for the reader. They should also describe the experimental procedures and conditions used for each of the pull-down and ubiquitination assays in greater detail in the methods section.

Thank you for your suggestion.

The quantitation and statistics for the original Figure 4G were included in the revised Figure 4. More information on the biochemical assays was included in the “Methods and Materials” section of our revised manuscript.

(6) Figure 5 shows that neither Rnf220 nor wdr5 directly regulate Hox gene expressions. The effect of double knockdown in the presence of RA suggests that they work together to suppress Hox gene expression via a different downstream target.

Thank you for your comment.

In fact, knockdown of Rnf220 alone can upregulate the expression Hox genes in the presence of RA (Figure 5C). Furthermore, knockdown of Wdr5 could reverse the upregulation of Hox genes in Rnf220 knockdown cells, which suggest that Rnf220 regulated Hox gene expression in a Wdr5 dependent manner. However, in the absence of RA, none of Rnf220 knockdown, Wdr5 knockdown or Rnf220 and Wdr5 double knockdown had a significant effect on the expression of Hox genes in P19 cells. It seems that RA signaling plays a crucial role for the regulation of RNF220 to WDR5 in P19 cells and discussion on this point was included in the revised manuscript.

(7) In Figure 6, while the reversal of changes in Hox gene expression upon concurrent Rnf220; Wdr5 inhibition highlights the importance of Wdr5 in this regulatory process, the mechanistic role of wdr5 and its functional consequences are unclear. To answer these questions, the authors need to: (i) Assay for activated and repressive epigenetic modifications upon double knockdown of Rnf220 and Wdr5 similar to that shown in Figure 3- supplement 1. This will reveal if wdr5 functions according to its intended role as part of the TrxG complex. (ii) The authors need to assay for changes in axon projection patterns in the double knockdown condition to see if Wdr5 inhibition rescues the neural circuit defects in Rnf220 +/- mice.<br />

Thank you for your suggestion.

Although it is also necessary to examine whether the rescue effect by WDR5 inhibitor injection in uetro is also a long-lasting effect for neuronal cirtuit at adult stages, it is difficult to distinguish the embryos or pups when they were given birth. Although Rnf220fl/wt;Wdr5fl/wt;Nestin-Cre mice are viable and could survive to adult stages, developmental defects in the forebrains, including cerebral cortex and hippocampus, were observed in Rnf220fl/wt;Wdr5fl/wt;Nestin-Cre mice. Therefore, no rescue effect on defects of behavior and neuronal circuit were examined in this study. Maybe, a PN nuclei specific inducible Cre mouse line could help toward this direction in the future.

We carried out ChIP-qPCR and tested activated and repressive epigenetic modifications upon double knockdown of Rnf220 and Wdr5 in P19 cell line and found Rnf220 and Wdr5 double knockdown recured Hox epigenetic modification to a certain degree (Figure 6-figure supplement 1).

References

Kedishvili, N.Y. 2016. Retinoic acid synthesis and degradation. Subcell Biochem, 81:127-161. DOI: 10.1007/978-94-024-0945-1_5, PMID: 2783050

Ma, P., Li, Y., Wang, H., Mao, B., Luo, Z.-G. 2021. Haploinsufficiency of the TDP43 ubiquitin E3 ligase RNF220 leads to ALS-like motor neuron defects in the mouse. Journal of Molecular Cell Biology, 13: 374-382. DOI: 10.1093/jmcb/mjaa072, PMID: 33386850

Ma, P., Song, N.-N., Li, Y., Zhang, Q., Zhang, L., Zhang, L., Kong, Q., Ma, L., Yang, X., Ren, B., Li, C., Zhao, X., Li, Y., Xu, Y., Gao, X., Ding, Y.-Q., Mao, B. 2019. Fine-Tuning of Shh/Gli Signaling Gradient by Non-proteolytic Ubiquitination during Neural Patterning. Cell Rep, 28: 541-553.e544. DOI: 10.1016/j.celrep.2019.06.017, PMID: 31291587

Niederreither, K., Dollé, P. 2008. Retinoic acid in development: towards an integrated view. Nat Rev Genet, 9: 541-53. DOI: 10.1038/nrg2340, PMID: 18542081

Wang, Y.-B., Song, N.-N., Zhang, L., Ma, P., Chen, J.-Y., Huang, Y., Hu, L., Mao, B., Ding, Y.-Q. 2022. Rnf220 is Implicated in the Dorsoventral Patterning of the Hindbrain Neural Tube in Mice. Front Cell Dev Biol, 10. DOI: 10.3389/fcell.2022.831365, PMID: 35399523

Author Response

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

Thank you for the thoughtful consideration of our work, including both reviewers’ constructive comments. Our apologies for taking some extra time for this revision, but we wanted to adress comments thoroughly with new analyses, not to mention a PhD defense, parental leave and my teaching ultimately being the bottleneck for the team’s work!

Reviewer #1 (Public Review):

The authors use a combination of structural and MD simulation approaches to characterize phospholipid interactions with the pentameric ligand-gated ion channel, GLIC. By analyzing the MD simulation data using clusters of closed and open states derived previously, the authors also seek to compare lipid interactions between putative functional states. The ultimate goal of this work is to understand how lipids shape the structure and function of this channel.

The strengths of this article include the following:

1) The MD simulation data provide extensive sampling of lipid interactions in GLIC, and these interactions were characterized in putative closed and open states of the channel. The extensive sampling permits confident delineation of 5-6 phospholipid interaction sites per subunit. The agreement in phospholipid binding poses between structures and the all-atom MD simulations supports the utility of MD simulations to examine lipid interactions.

2) The study presents phospholipid binding sites/poses that agree with functionally-important lipid binding sites in other pLGICs, supporting the notion that these sites are conserved. For example, the authors identify interactions of POPC at an outer leaflet intersubunit site that is specific for the open state. This result is quite interesting as phospholipids or drugs that positively modulate other pLGICs are known to occupy this site. Also, the effect of mutating W217 in the inner leaflet intersubunit site suggests that this residue, which is highly conserved in pLGICs, is an important determinant of the strength of phospholipid interactions at this site. This residue has been shown to interact with phospholipids in other pLGICs and forms the binding site of potentiating neurosteroids in the GABA(A) receptor.

Weaknesses of this article include the following:

1) The authors describe in detail state-dependent lipid interactions from the MD simulations; however, the functional significance of these findings is unclear. GLIC function appears to be insensitive to lipids, although this understanding is based on experiments where GLIC proteoliposomes were fused to oocyte membranes, which may not be optimal to control the lipid environment. Without functional studies of GLIC in model membranes, the lipid dependence of GLIC function is not definitively known. Therefore, it is difficult to interpret the meaning of these state-dependent lipid interactions in GLIC.

2) It is unlikely that the bound phospholipids in the GLIC structures, which are co-purified from e. coli membranes, are POPC. Rather, these are most like PE or PG lipids. While it is difficult to accommodate mixed phospholipid membranes in all-atom MD simulations, the choice of POPC for this model, while practically convenient, seems suboptimal, especially since it is not known if PE or PG lipids modulate GLIC function. Nevertheless, it is striking that the overall binding poses of POPC from the simulations agree with those identified in the structures. It is possible that the identity of the phospholipid headgroup will have more of an impact on the strength of interactions with GLIC rather than the interaction poses (see next point).

3) The all-atom MD simulations provide limited insight into the strength of the POPC interactions at each site, which is important to interpret the significance of these interactions. It is unlikely that the system has equilibrated within the 1.7 microseconds of simulation for each replicate preventing a meaningful assessment of the lipid interaction times. Although the authors report exchange of up to 4 POPC interacting at certain residues in M4, this may not represent binding/unbinding events (depending on how binding/interaction is defined), since the 4 Å cutoff distance for lipid interactions is relatively small. This may instead be a result of small movements of POPC in and out of this cutoff. The ability to assess interaction times may have been strengthened if the authors performed a single extended replicate up to, for example, 10-20 microseconds instead of extending multiple replicates to 1.7 microseconds.

Reviewer #2 (Public Review):

The authors convincingly show multiple inner and outer leaflet non-protein (lipid) densities in a cryo-EM closed state structure of GLIC, a prokaryotic homologue of canonical pentameric ligand-gated ion channels, and observe lipids in similar sites during extensive simulations at both resting and activating pH. The simulations not only corroborate structural observations, but also suggest the existence of a state-dependent lipid intersubunit site only occupied in the open state. These important findings will be of considerable interest to the ion channel community and provide new hypotheses about lipid interactions in conjunction with channel gating.

Recommendations for the authors: please note that you control which, if any, revisions, to undertake

In particular, a discussion of whether the timescale of the simulations permit measurements of residence or interaction times of the lipids should be addressed.

Reviewer #1 (Recommendations for the authors):

Comment 1.1: The authors may consider expanding the discussion about the significance of state-dependent lipid interactions. On the one hand, they emphasize state-dependent interactions of POPC with closed and open states in the outer leaflet in the results. On the other hand, they state that GLIC is insensitive to its lipid environment. What is the significance of the state-dependent interactions of POPC in GLIC, if any? It is possible that GLIC agonist responses are sensitive to phospholipids (such as PE or PG found in e. coli)? The state-dependent differences in lipid interaction identified in this study support this possibility and suggest the need to better understand the effects of phospholipids on GLIC function.

Response 1.1: We agree with the reviewer that this is an interesting question and we have therefore extended the discussion with additional references on the functional effects on GLIC of various lipid membranes:

p. 11 (Discussion)

“Sampling was further simplified by performing simulations in a uniform POPC membrane. Prior experiments have been conducted to assess the sensitivity of GLIC in varying lipid environments (Labriola et al., 2013; Carswell et al., 2015; Menny et al., 2017), indicating that GLIC remains fully functional in pure POPC bilayers. In our cryo-EM experiments, the protein was recombinantly expressed from E. coli, which means that the experimental density would likely represent phosphatidylglycerol or phosphatidylethanolamine lipids. However, as the molecular identities of bound lipids could not be precisely determined, POPC lipids were built for straightforward comparison with simulation poses. While it appears that GLIC is capable of gating in a pure POPC bilayer, it remains plausible that its function could be influenced by different lipid species, especially due to the presence of multiple charged residues around the TMD/ECD interface which might interact differently with different lipid head groups. Further experiments would be needed to confirm whether the state dependence observed in simulations is also lipid-dependent. It is possible that certain types of lipids bind in one but not the other state, or that certain states are stabilized by a particular lipid type.”

Comment 1.2: It would be helpful to state in the discussion that the co-purified lipids from GLIC structures are likely PE or PG from e. coli membranes. Nevertheless, it is interesting that the phospholipid poses from the structures generally agree with those identified from the MD simulations using PC.

Response 1.2: Good point. We have clarified in the discussion that the native lipids in the cryo-EM structure are likely PG or PE lipids, as quoted in the preceding Response.

Comment 1.3: The authors describe a more deeply penetrating interaction of POPC in the outer intrasubunit cleft in the open state, but this is difficult to appreciate from the images in Fig. 4B, 4E or S3B. The same is true of the deep POPC interaction at the outer intersubunit site. It may be helpful to show these densities from a different perspective to appreciate the depth of these binding poses.

Response 1.3: We have added Figure 4 – figure supplement 1 to better show the depth of lipid binding poses, especially the ones in the outer leaflet intrasubunit cleft and at the inner intersubunit site, and cited the figure on p. 7 (Results).

Comment 1.4: The representation of the lipid densities in Fig. 4B is not easy to interpret. First, the meaning of resting versus activating conditions and closed versus open states can be easily missed for readers who are not familiar with the author's previous study. It may be helpful to describe this (i.e. how open and closed state clusters were generated from structures determined in resting and activating conditions) in greater detail in either the figure legend, results or methods. Second, the authors state that there are differences in lipid poses between the closed and open states but not resting and activating conditions. With the exception of the intersubunit density, this is difficult to appreciate from Fig. 4B. As stated in point #3, the difference, for example, in the complementary intrasubunit site may be better appreciated with an image from a different perspective.

Response 1.4: Acknowledged - the distinction between resting and activating conditions v.s. open and closed states can be confusing. We have tried to clarify these differences at the beginning of the results section, the methods section, and in the caption of Figure 4. Regarding differences in lipid poses between open and closed states, we agree it is difficult to appreciate from Figure 4, but here we refer the reader to Figure 4 – figure supplement 2 for an overlay between open and closed densities. Additionally, we now added Figure 1 – figure supplement 1 which provides lipid densities for all five subunits and overlays with the build cryo-EM lipids, possibly making differences easier to appreciate. Regarding images from different perspectives, we trust the new figure supplement described in Response 1.3 provides a better perspective.

p. 3 (Results)

“For computational quantification of lipid interactions and binding sites, we used molecular simulations of GLIC conducted under either resting or activating conditions (Bergh et al., 2021a). As described in Methods, resting conditions corresponded to neutral pH with most acidic residues deprotonated; activating conditions corresponded to acidic pH with several acidic residues protonated. Both open and closed conformations were present in both conditions, albeit with different probabilities.”

p. 8 (Figure 4)

“Overlaid densities for each state represent simulations conducted under resting (dark shades) or activating (light shades) conditions, which were largely superimposable within each state.”

p. 24 (Methods)

“We analyzed previously published MSMs of GLIC gating under both resting and activating conditions (Bergh et al., 2021a). Resting conditions corresponded to pH 7, at which GLIC is nonconductive in functional experiments, with all acidic residues modeled as deprotonated. Activating conditions corresponded to pH 4.6, at which GLIC is conductive and has been crystallized in an open state (Bocquet et al., 2009). These conditions were modeled by protonating a group of acidic residues (E26, E35, E67, E75, E82, D86, D88, E177, E243; H277 doubly protonated) as previously described (Nury et al., 2011).”

Comment 1.5: The new closed GLIC structure was obtained by merging multiple datasets. What were the conditions of the datasets used? Was it taken from samples in resting or also activating conditions?

Response 1.5: We have updated the Results, Discussion, and Methods to clarify this important point, in particular by merging datasets and rerunning the classification:

p. 3 (Results)

“In our cryo-EM work, a new GLIC reconstruction was generated by merging previously reported datasets collected at pH 7, 5, and 3 (Rovšnik et al., 2021). The predominant class from the merged data corresponded to an apparently closed channel at an overall resolution of 2.9 Å, the highest resolution yet reported for GLIC in this state (Figure 1 – figure supplement 2, Table 1).”

p. 11 (Discussion)

“Interestingly, the occupational densities varied remarkably little between resting and activating conditions (Figure 1 – figure supplement 1), indicating state- rather than pH- dependence in lipid interactions, also further justifying the approach of merging closed- state GLIC cryo-EM datasets collected at different pH conditions to resolve lipids.”

p. 14 (Methods)

“After overnight thrombin digestion, GLIC was isolated from its fusion partner by size exclusion in buffer B at pH 7, or in buffer B with citrate at pH 5 or 3 substituted for Tris. The purified protein was concentrated to 3–5 mg/mL by centrifugation. [...] Data from three different grids, at pH 7, 5, and 3, were merged and processed together.”

Comment 1.6: In Fig. 3D, do the spheres represent the double bond? If so, please state in the legend

Response 1.6: We have clarified in the legend of Figure 3D that the yellow spheres on the lipid tails represent a double bond.

Comment 1.7: In Fig. 3E, what is the scale of the color representation?

Response 1.7: We have clarified in the legend of Figure 3E that colors span 0 (white) to 137015 contacts (dark red).

Reviewer #2 (Recommendations For The Authors):

Comment 2.1: I'm not sure I fully understand how the final lipids were modeled (built). Fig. 1 caption suggests they may have been manually built? I understand that the idea was to place them in the overlap of simulation densities and structure densities, but can the authors please clarify if there were any quantifiable conditions that were employed during this process or if this was entirely manual placement in a pose that looked good? Regardless, it would be helpful to see an overlay of the built lipids with both the cryo and simulation densities (e.g., overly of Fig. 1F/H and G/H) to better visualize how the final built lipids compare.

Response 2.1: We thank the reviewer for pointing out unclarities regarding our methods. We have extended the methods section to clarify how the lipids were manually built in the cryo-EM structure. We have also added Figure 1 – figure supplement 1 showing overlays of the computational densities and built cryo-EM lipids.

p. 15 (Methods)

“Lipids were manually built in COOT by importing a canonical SMILES format of POPC (Kim et al., 2021) and adjusting it individually into the cryo-EM density in each of the sites associated with a single subunit, based in part on visual inspection of lipid densities from simulations, as described above. After building, 5-fold symmetry was applied to generate lipids at the same sites in the remaining four subunits.”

Comment 2.2: Regarding the state-dependent lipid entry to the outer leaflet intersubunit site associated with channel opening, if the authors could include a movie depicting this process that would be great. The current short explanation does not do this justice. Also, what were the dynamics of this process? Beyond the correlation between site occupancy and the pore being open, how did the timing of lipid entry/exit and pore opening/closing correlate?

Response 2.2: The point regarding the timing of state-dependent lipid binding at the subunit interface and pore opening is indeed an interesting one. We have added Figure 4 – figure supplement 3D showing that the state-dependent P250 lipid interaction precedes pore opening, as quantified by pore hydration levels, indicating a potential role in gating. The interaction between lipid binding and conformational change of the protein is also depicted in the newly added Figure 4 - video supplement 1, which we hope will be able to better communicate the conclusions regarding state-dependent interactions. We have also expanded the results and discussion to better explain these results:

p. 9 (Results)

“The lipid head made particularly close contacts with residue P250 on the M2-M3 loop, which undergoes substantial conformational change away from the pore upon channel opening, along with outer-leaflet regions of M1–M3 (Figure 4E, Figure 4—figure Supplement 3A,B,C, Figure 4—video 1). These conformational changes were accompanied by a flip of M1 residue F195, which blocked the site in the closed state but rotated inward to allow closer lipid interactions in the open state (Figure 4—figure Supplement 3C, Figure 4—video 1). Indeed, P250 was predominantly located within 3 Å of the nearest lipid atom in open- but not closed-state frames (Figure 4F). Despite being restricted to the open state, interactions with P250 were among the longest duration in all simulations (Figure 2C) and as these binding events preceded pore opening, it is plausible to infer a role for this state-dependent lipid interaction in the gating process (Figure 4 – figure supplement 3D).”

p. 12 (Discussion)

“The state-dependent binding event at this site preceded pore opening in MSMs, where lipid binding coincided with crossing a smaller energy barrier between closed and intermediate states, followed by pore opening at the main energy barrier between intermediate and open states (Figure 4 – figure supplement 3D). Further, since the P250- lipid interaction was characterized by relatively long residence times (Figure 2), it is possible this lipid interaction has a role to play in GLIC gating.”

Comment 2.3: Although the interaction times are helpful, I didn't get a great sense of how mobile the lipids are during the simulations. Can the authors discuss this a bit more. For example, are interaction times dominated by lipids that jiggle a bit away from a residue and then back again, vs how often are lipids exchanging with other lipids initially further away from the protein?

Response 2.3: We have now added various measures of lipid diffusion, both for initially interacting lipids and for bulk lipids, which are summarized in the new Figure 2 – figure supplement 1. We have further addressed the question of simulation timescales in Results, Discussion, and Methods. These numbers highlight that it is possible for lipids several nanometers away from the protein surface to exchange with lipids of the first lipid shell.

p. 3,6 (Results)

“Lateral lipid diffusion coefficients were estimated to 1.47 nm2/µs for bulk lipids and 0.68 nm2/µs for lipids of the first lipid shell (Figure 2 – figure supplement 1A), which is relatively slow compared to the timescales of each trajectory (1.7 µs). However, multiple residues throughout the M1, M3, and M4 helices exchanged contacts with 2-4 different lipid molecules in individual simulations (Figure 2C). Furthermore, 1.7-µs root mean square displacement of lipids originally in the first lipid shell was 2.15 nm, and 3.16 nm in the bulk bilayer, indicating such exchanges are not limited to nearby lipids (Figure 2 – figure supplement 1B). Thus, exchange events and diffusion estimates indicate that the duration of lipid contacts observed in this work can be at least partly attributed to interaction stabilities and not solely to sampling limitations.”

p. 11 (Discussion)

“Indeed, the unrestrained atomistic MD simulations studied here were not expected to capture the maximal duration of stable contacts, as indicated by some interaction times approaching the full 1.7-µs trajectory (Figure 2}). Nevertheless, simulations were of sufficient length to sample exchange of up to four lipids, particularly around the M4 helix. Calculation of lipid lateral diffusion coefficients resulted in average displacements at the end of simulations of 2.15 nm for lipids initially interacting with the protein surface, roughly corresponding to lipids diffusing out to the 4th lipid shell. Diffusion of bulk lipids was faster, allowing lipids originally 3.16 nm away from the protein surface to ingress the first lipid shell. This observation underscores the potential for lipid exchange events even among lipids initially distant from the protein surface. Of course, duration of exceptionally stable interactions, such as those involving T274 (Figure 2C), inevitably remain bounded by the length of our simulations. Still, diffusion metrics, supported by robust statistical analysis encompassing diverse starting conditions (500 trajectories), enable confident estimation of relative interaction times.“

p. 13 (Methods)

“Time-based measures of protein-lipid interactions, such as mean duration times and exchange of interactions, were calculated for the 100 x 1.7 µs-long simulations using prolintpy (Sejdiu and Tieleman, 2021) with a 4 Å interaction cutoff. Analysis of lateral lipid diffusion in individual simulations was carried out for two disjoint sets of lipids: the first lipid shell defined as lipids with any part within 4 Å of the protein surface (~90 lipids), and bulk lipids consisting of all other lipids (~280 lipids). Mean square displacements of each lipid set were calculated using GROMACS 2021.5 (Abraham et al., 2015b) with contributions from the protein center of mass removed. Diffusion coefficients for each set, DA, were calculated using the Einstein relation (Equation 1) by estimating the slope of the linear curve fit to the data.

where ri(t) is the coordinate of the center of mass of lipid i of set A at time t and DA is the self-diffusion coefficient.”

Comment 2.4: How symmetric or asymmetric are the cryo and simulation densities across subunits and was there subunit asymmetry in the final build lipids? I could not tell from any of the figures beyond the casual observation that they maybe look somewhat similar in Fig. 1?

Response 2.4: We thank the reviewer for this useful remark. We have clarified in the methods that the cryo-EM lipids were built in C5-symmetry, and thus the positions are symmetric. The computational densities were calculated independently for each subunit and are thus not necessarily symmetric. We have added Figure 1 – figure supplement 1 showing densities for all five subunits, also serving as an indication of convergence of the results.

p. 3 (Results) “Although the stochastic nature of simulations resulted in nonidentical lipid densities associated with the five GLIC subunits, patterns of lipid association were notably symmetric (Figure 1 – figure supplement 1).”

p. 14-15 (Methods)

“A smaller subset of particles was used to generate an initial model. All subsequent processing steps were done using 5-fold symmetry. […] A monomer of that model was fit to the reconstructed density and 5-fold symmetry was applied with PHENIX 1.19.2-4158 through NCS restraints detected from the reconstructed cryo-EM map, to generate a complete channel. […] After building, 5-fold symmetry was applied to generate lipids at the same sites in the remaining four subunits.”

Minor comments:

Comment 2.5: Fig. 1 is probably not easy to follow for the general reader and the caption is very brief. I suggest adding an additional explanation to the caption and/or additional annotations to the figure to help a general reader step through this.

Response 2.5: We have expanded the caption of Figure 1 and clarified the meanings of colors, labels, and annotations.

Comment 2.6: Fig. 1B - Caption is confusing. I would not call the state separation lines outlines as they are not closed loops. Also, I see red/orange and two shades of blue whereas the caption mentions orange and blue only. The caption should also explicitly say what the black lines are (other cluster separations).

Response 2.6: We have edited the caption to better describe colors, annotations, and the meaning of the data:

p. 4 (Figure 1)

“(B) Markov state models were used to cluster simulations conducted under resting (R) or activating (A) conditions into five states, including closed (left of the light or dark orange lines) and open (right of the light or dark blue lines). Black lines mark edges of other state clusters derived from MSM eigenvectors. Experimental structures are highlighted as white circles.”

Comment 2.7: Fig. 3F caption appears to conflict with data where interaction with W217A appears longer than W217. I think the authors want to suggest here that W217A reduces contact time with T274 as stated in the main text.

Response 2.7: We have clarified in this legend that “Mutation of residue W217, lining this pocket, reveals shortened interactions at the T274 binding site” (p. 6, Figure 3).

Comment 2.8: Ref 25 and 26 are the same.

Response 2.8: Apologies; this mistake has been corrected.

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

Summary: 

This study investigated the role of CD47 and TSP1 in extramedullary erythropoiesis by utilization of both global CD47-/- mice and TSP1-/- mice. 

Strengths:  

Flow cytometry combined with spleen bulk and single-cell transcriptomics were employed. The authors found that stress-induced erythropoiesis markers were increased in CD47-/- spleen cells, particularly genes that are required for terminal erythroid differentiation. Moreover, CD47 dependent erythroid precursors population was identified by spleen scRNA sequencing. In contrast, the same cells were not detected in TSP1-/- spleen. These findings provide strong evidence to support the conclusion that the differential role of CD47 and TSP1 in extramedullary erythropoiesis in mouse spleen. 

Weaknesses: 

Methods and data analysis are appropriate. However, some clarifications are required. The discussion section needs to be expanded.  

(1) The sex of mice that were used in the study is unknown.  

(2) In the method of Single-cell RNA sequencing (page 10), it mentioned that single cell suspensions from mouse spleens were depleted of all mature hematopoietic cell lineages by passing through CD8a microbeads and CD8a+ T cell isolation Kit. As described, it is confusing what cell types are obtained for performing scRNAseq. More information is required for clarity.  

(3) The constitutive CD47 knockout mouse model is utilized in this study. The observed accumulation of erythroid precursors in the spleens of CD47-/- mice suggests a chronic effect of CD47 on spleen function. Can the current findings be extrapolated to acute scenarios involving CD47 knockdown or loss, as this may have more direct relevance to the potential side effects associated with an-CD47-mediated cancer therapy? Please expand on this topic in the discussion section.  

(1) The missing mouse gender information is incorporated into the revised manuscript. For flow cytometry, two male and two female mice of each genotype were used. For single cell RNA sequencing, two female and one male mouse of each genotype were used. For the bulk RNA sequencing four male cd47−/− mice and four male wildtype mice were used.

(2) We apologize for the confusing presentation, which has been corrected. The bulk RNA sequencing analysis identified elevated expression of erythropoietic genes in CD8+ spleen cells from cd47−/− versus wildtype mice that were obtained using magnetic bead depletion of all other lineages. Therefore, we used the same Miltenyi negative selection kit as the first step to prepare the cells for single cell RNA sequencing. These untouched cells were then depleted of most mature CD8 T cells using a Miltenyi CD8a(Ly2) antibody positive selection kit. An important consideration underlying this approach was recognizing that the commercial magnetic bead depletion kits used for preparing specific immune cell types are optimized to give relatively pure populations of the intended immune cells using wildtype mice. Our previous experience studying NK cell development in the cd47−/− mice taught us that NK precursors, which are rare in wildtype mouse spleens, accumulate in cd47−/− spleens and were not removed by the antibody cocktail optimized for wildtype spleen cells (Nath et al Front Immunol 2018). The present data indicate that erythroid precursors behave similarly.

(3) The Discussion was edited as recommended. Anemia is a prevalent side effect of several CD47 therapeutic antibodies being developed for cancer therapy. This anemia would be expected to induce erythropoiesis in bone marrow and possibly at extramedullary sites. Human spleen cells are not accessible to directly evaluate extramedullary erythropoiesis in cancer patients, but analysis of circulating erythroid precursors or liquid biopsy methods could be useful to detect induction of extramedullary erythropoiesis by these therapeutics. We are currently investigating the ability of CD47 antibodies to directly induce erythropoiesis using a human in vitro model.

Reviewer #2 (Public Review):

Summary: 

The authors used existing mouse models to compare the effects of ablating the CD47 receptor and its signaling ligand Thrombospondin. The CD47-KO model used in this study was generated by Kim et al, 2018, where hemolytic anemia and splenomegaly was reported. This study analyzes the cell composition of the spleens from CD47-KO and Thsp-KO, focusing on early hematopoietic and erythroid populations. The data broadly shows that splenomegaly in the CD47-KO is largely due to an increase in committed erythroid progenitors as seen by Flow Cytometry and single-cell sequencing, whereas the Thsp-KO shows a slight depletion of committed erythroid progenitors but is otherwise similar to WT in splenic cell composition.  

Strengths:

The techniques used are appropriate for the study and the data support the main conclusions of the study. This study provides novel insights into a putative role of Thsp-CD47 signaling in triggering definitive erythropoiesis in the mouse spleen in response to anemic stress and constitutes a good resource for researchers seeking to understand extramedullary erythropoiesis.  

Weaknesses:

The Flow cytometry data alone supports the authors' main conclusion and single-cell sequencing confirms them but does not add further information, other than those already observed in the Flow data. The single-cell sequencing analysis and presentation could be improved by using alternate clustering methods as well as separating the data by genotype and displaying them in order for readers to fully grasp the nuanced differences in marker expression between the genotypes. Further, it is not clear from the authors' description of their results whether the increased splenic erythropoiesis is a direct consequence of CD47-KO or a response to the anemic stress in this mouse model. The enrichment of cKit+ Ter119+ Sca1- cells in CD47-KO indicates that these are likely stress erythroid progenitors. Another CD47-KO mouse model (Lindberg et al 1996) has no reported erythroid defects and was also not examined in this study.  

(1) The reviewer asked, “whether the increased splenic erythropoiesis is a direct consequence of CD47-KO or a response to the anemic stress in this mouse model.” Our data supports both a direct role for CD47 and an indirect role resulting from the response to anemic stress. We cited our previous publications describing increased Sox2+ stem cells in spleens of Cd47 and Thbs1 knockout mice, but we neglected to emphasize another study where we found that bone marrow from cd47−/− mice subjected to the stress of ionizing radiation exhibited more colony forming units for erythroid (CFU-E) and burst-forming unit-erythroid (BFU-E) progenitors compared to bone marrow from irradiated wildtype mice (Maxhimer Sci Transl Med 2009). Taken together, our published data demonstrates that loss of CD47 results in an intrinsic protection of hematopoietic stem cells from genotoxic stress. This function of CD47 is thrombospondin-1-dependent and is consistent with the up-regulation of early erythroid precursors in the spleens of both knockout mice but cannot explain why the Thbs1−/−  mice have fewer committed erythroid precursors than wildtype. We cited studies that documented increased red cell turnover in cd47−/− mice but less red cell turnover in Thbs1−/−  mice compared to wildtype mice. Increased red cell clearance in cd47−/− mice is mediated by loss of the “don’t eat me” function of CD47 on red cells. In wildtype mice, clearance is augmented by thrombospondin-1 binding to the clustered CD47 on aging red cells (Wang, Aging Cell 2020). Thus, anemic stress in the mouse strains studied here decreases in the order cd47−/− > WT > Thbs−/−. This is consistent with the increased committed erythroid progenitors reported here in cd47−/− spleens and decreased committed progenitors in the Thbs1−/− spleens. 

(2) Based on the reviewer’s question regarding alternative mechanisms and the publication of Yang et al 2022 identifying a role for CD47 in stress erythropoiesis though transfer of mitochondria to erythroblasts, we asked whether cd47-/- erythroid precursors  would show decreased mRNA expression for mitochondrial chromosome genes (new Figure 4−figure supplement 3C). Some of these mRNAs were more abundant in cd47-/- and thbs1-/- erythroid cells, which is the opposite of what we expected based on Yang 2022 but consistent with our previous publications identifying thrombospondin-1 and CD47 as negative regulators of mitochondrial homeostasis in muscle cells and T cells.

(3) The cd47−/− mice used for the current study are the same strain as those reported by Lindberg et al in 1996, with additional backcrossing onto a C57BL/6 background.

Recommendations For The Authors:

Reviewer #2 (Recommendations For The Authors):

Suggestions for improved or additional experiments, data, or analyses.  

Significant efforts went into analyzing the type of erythroid progenitors by marker expression, but typical Flow cytometry strategies using Ter119 and CD44 combined with forward scatter can be used to stage the committed erythroid progenitors precisely.  

We appreciate this suggestion to extend the flow data. However, the upcoming retirement of the PI required closing our breeding colony, and the mice are no longer available.  

How can the difference between the erythroid phenotypes of the Lindberg et al 1996 CD47-KO (exon2 Neo knock-in) and Kim et al 2018 CD47-ko (exon1 26bp indel) be explained?  

We are not convinced that the erythroid phenotypes of the Lindberg and Kim CD47-KO mice differ at the age used in our studies. Kim et al. focused on progressive hemolytic anemia and changes in T cells in spleen that emerge at 26 weeks age, whereas the mice used here were younger. The Lindberg and Kim mice have similar spleen enlargement at the age we used.

Another manuscript under review from our lab suggests that cis-regulation of an adjacent colinear gene could contribute to some phenotypes observed when perturbing the Cd47 gene. The Lindberg mouse exhibits minimal perturbation of that adjacent gene, but we have no data regarding the Kim et al mouse. The reviewer’s question brought to our attention that we neglected to state in the Methods that the mice used here are the Lindberg mice, not the Kim mice. This omission is now corrected.

The authors used Lindberg mouse for 2018 study on NK cells and observed splenomegaly. Did they check for extramedullary erythropoiesis there?  

Retrospective examination of the RNAseq data for the spleen cells enriched in NK precursors used in our 2018 publication (Nath, 2018) reveals significantly elevated expression for a majority of the extramedullary erythroid markers listed in Table 1, but they were generally less abundant than observed for the lineage-depleted spleen cells used in the present manuscript.   

Author response table 1.

To clarify the stress erythropoiesis issue, it might be helpful to examine the sc-seq data for the expression of specific stress erythropoiesis markers in CD47-KO. Targets of BMP4 and Hedgehog signaling can also be examined. Further colony assays can help determine if stress BFU-Es are prevalent in the CD47-KO spleens and depleted in Thsp-KO  

As noted in Table 1, twelve of the genes we studied are established markers of stress-induced extramedullary erythropoiesis, and most of these were included in the scRNA seq data presented. Our previous publication demonstrated that bone marrow from cd47−/− mice subjected to the stress of ionizing radiation exhibited more colony forming units for erythroid (CFU-E) and burst-forming unit-erythroid (BFU-E) progenitors compared to bone marrow from irradiated wildtype mice (Maxhimer Sci Transl Med 2009). We have not performed colony formation assays using spleen.

To address the reviewer’s question regarding BMP4 and hedgehog signaling we performed gene set enrichment analysis for known BMP4 and hedgehog signaling signatures. Using GSE26351_UNSTIM_VS_BMP_PATHWAY_STIM_HEMATOPOIETIC_PROGENITORS, cd47-/- cells in cluster 12 or their CD34+ orCD34- subsets did not show significant enrichment for BMP4 targets compared to WT. Thbs1-/- cells in clusters 12 and 14 showed marginally significant depletion of the BMP4 signature (p=0.04 and p=0.023, respectively). Using the KEGG_HEDGEHOG_SIGNALING_PATHWAY, we did not find any significant enrichment. However, only a few genes in this pathway were detectable in the scRNAseq data. These data suggest that the BMP4 signaling may be regulated by thrombospondin-1, but properly testing this hypothesis would require achieving greater sequencing depth combined with a cell isolation method that better enriches the early hematopoietic progenitors that are known to utilize the BMP4 pathway.

In the reclustering of erythroid progenitors in Figure 5, inclusion of Gata1 as a selection marker may help capture more of the early erythroid progenitors from the dataset and provide a more complete picture of the erythroid populations. 

We thank the reviewer for suggesting inclusion of Gata1. We repeated the reclustering including Gata1 and found the selected cell count increased from 876 cells to 1007 cells. However, most of the increase was not in the erythroid cluster, which increased from 413 cells to 419 cells. Most of the increase represented Gata1+ T cells (548 cells including Gata1 versus 463 cells without). The revised manuscript presents genotype-dependent differential gene expression based on including Gata1 selection, but none of the specific conclusions were changed from the initial submission. The new Table 4 and Figure 7−figure supplement 1 enabled us to compare differential expression of erythropoietic genes obtained using supervised and unsupervised clustering and show that both methods yield comparable results.

Just out of curiosity, was there an attempt to make a CD47 Thsp double KO? . Is it viable?  

Cd47 KO mice are somewhat difficult breeders, and several previous attempts to cross with other transgenics have produced viable homozygous offspring that could not be propagated.

Recommendations for improving the wring and presentation.  

Perhaps readers would find it more intriguing if the paper led with the single-cell sequencing showing enrichment of erythroid populations in CD47-KO, and later confirmed with Flow Cytometry (even if this was not necessarily the order in which the experiments were done). 

We considered this suggestion but believe that some of the flow cytometry data is needed to understand why we focused on CD34+ and CD34- subsets and proliferation markers when analyzing the scRNAseq data

The single-cell sequencing data in Figure 3 might benefit from UMAP clustering as well. In addition, it would greatly help readers if the data points were separated by genotype and displayed after clustering. A similar analysis has been done in this paper: doi:10.1038/s41556-022-00898-9 by clustering different conditions together but displaying them separately by condition. 

We initially explored tSNE and UMAP clustering and obtained similar results. We have added violin plots separated by genotype in Figure 4-figure supplement 2. We also included improved clusters separated by genotype in the revised Figure 3 panels C and D and for the reclustering in Figure 6D. UMAP plots provided better presentation for the reclustering (revised Figure 7). All data have been updated to the latest pipeline as noted in the Methods.

Minor corrections to the text and figures.  

Figure 4: Labels and plot legends are illegible in general, please relabel manually and if possible, redo plots with bigger font size and legends (relatively easy using ggplot2) 

All figure panels were relabeled using larger fonts

Figure 5D: Individual plots are stacked randomly atop each other and in many cases, gene names are not visible. Please restack the layers and ensure that the gene names are visible 

Panel D was made a separate figure with enlarged labels (now Figure 7).

Supp Fig 2: Layout can be organized a little better. Consider splitting into two figures for better organization  

The figure was split as recommended. Now Figure 1-figure supplement 2 and Figure 2-figure supplement

1.

Abstract Line 10: "...mRNA expression of Kit, Ermap, and Tfrc, Induction of committed erythroid precursors is...". Replace comma after "Tfrc" with period   

Done.

Discussion Page 9 Line 8: "...WT spleens, s. mRNAs for some markers of committed erythroid cells including Nr3c1 mRNA...". Remove ", s" after spleens.   

Done.

Author Response

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

Reviewer #1 (Public Review):

The paper from Hsu and co-workers describes a new automated method for analyzing the cell wall peptidoglycan composition of bacteria using liquid chromatography and mass spectrometry (LC/MS) combined with newly developed analysis software. The work has great potential for determining the composition of bacterial cell walls from diverse bacteria in high-throughput, allowing new connections between cell wall structure and other important biological functions like cell morphology or host-microbe interactions to be discovered. In general, I find the paper to be well written and the methodology described to be useful for the field. However, there are areas where the details of the workflow could be clarified. I also think the claims connecting cell wall structure and stiffness of the cell surface are relatively weak. The text for this topic would benefit from a more thorough discussion of the weak points of the argument and a toning down of the conclusions drawn to make them more realistic.

Thank you for your thorough and insightful review of our manuscript. We greatly appreciate your positive and constructive feedbacks on our methodology. We have carefully reviewed your comments and have responded to each point as follows:

Specific points:

1) It was unclear to me from reading the paper whether or not prior knowledge of the peptidoglycan structure of an organism is required to build the "DBuilder" database for muropeptides. Based on the text as written, I was left wondering whether bacterial samples of unknown cell wall composition could be analyzed with the methods described, or whether some preliminary characterization of the composition is needed before the high-throughput analysis can be performed. The paper would be significantly improved if this point were explicitly addressed in the main text. We apologize for not making it clearer. The prior knowledge of the peptidoglycan structure of an organism is indeed required to build the “DBuilder” database to accurately identify muropeptides; otherwise, the false discovery rate might increase. While peptidoglycan structures of certain organisms might not have been extensively studied, users still remain the flexibility to adapt the muropeptide compositions based on their study, referencing closely related species for database construction. We have addressed this aspect in the main text to ensure a clearer understanding.

“(Section HAMA platform: a High-throughput Automated Muropeptide Analysis for Identification of PGN Fragments) …(i) DBuilder... Based on their known (or putative) PGN structures, all possible combinations of GlcNAc, MurNAc and peptide were input into DBuilder to generate a comprehensive database that contains monomeric, dimeric, and trimeric muropeptides (Figure 1b)."

2) The potential connection between the structure of different cell walls from bifidobacteria and cell stiffness is pretty weak. The cells analyzed are from different strains such that there are many possible reasons for the change in physical measurements made by AFM. I think this point needs to be explicitly addressed in the main text. Given the many possible explanations for the observed measurement differences (lines 445-448, for example), the authors could remove this portion of the paper entirely. Conclusions relating cell wall composition to stiffness would be best drawn from a single strain of bacteria genetically modified to have an altered content of 3-3 crosslinks.

We understand your concern regarding the weak connection between cell wall structure and cell stiffness. We will make a clear and explicit statement in the main text to acknowledge that the cells analyzed are derived from different strains, introducing the possibility of various factors influencing the observed changes in physical measurements as determined by AFM. Furthermore, we greatly appreciate your suggestion to consider genetically modified strains to investigate the role of cross-bridge length in determining cell envelope stiffness. In this regard, we are in the process of developing a CRISPR/Cas genome editing toolbox for Bifidobacterium longum, and we plan on this avenue of investigation for future work.

Reviewer #2 (Public Review):

The authors introduce "HAMA", a new automated pipeline for architectural analysis of the bacterial cell wall. Using MS/MS fragmentation and a computational pipeline, they validate the approach using well-characterized model organisms and then apply the platform to elucidate the PG architecture of several members of the human gut microbiota. They discover differences in the length of peptide crossbridges between two species of the genus Bifidobacterium and then show that these species also differ in cell envelope stiffness, resulting in the conclusion that crossbridge length determines stiffness.

We appreciate your thoughtful review of our manuscript and your recognition of the potential significance of our work in elucidating the poorly characterized peptidoglycan (PGN) architecture of the human gut microbiota.

The pipeline is solid and revealing the poorly characterized PG architecture of the human gut microbiota is worthwhile and significant. However, it is unclear if or how their pipeline is superior to other existing techniques - PG architecture analysis is routinely done by many other labs; the only difference here seems to be that the authors chose gut microbes to interrogate.

We apologize if this could have been clearer. The HAMA platform stands apart from other pipelines by utilizing automatic analysis of LC-MS/MS data to identify muropeptides. In contrast, most of the routine PGN architecture analyses often use LC-UV/Vis or LC-MS platform, where only the automatic analyzing PGFinder software is supported. To our best knowledge, a comparable pipeline on automatically analyzing LC-MS/MS data was reported by Bern et al., which they used commercial Byonic software with an in-house FASTA database and specific glycan modifications. They achieved accurate and sensitive identification on monomer muropeptides, but struggled with cross-linked muropeptides due to the limitations of the Byonic software. We believe that our pipeline introducing the automatic and comprehensive analysis on muropeptide identification (particularly for Gram-positive bacterial peptidoglycans) would be a valuable addition to the field. To enhance clarity, we have adjusted the context as follows:

(Introduction) … Although they both demonstrated great success in identifying muropeptide monomers, the accurate identification of muropeptide multimers and other various bacterial PGN structures still remains unresolved. This is because deciphering the compositions requires MS/MS fragmentation, but it is still challenging to automatically annotate MS/MS spectra from these complex muropeptide structures."

I do not agree with their conclusions about the correlation between crossbridge length and cell envelope stiffness. These experiments are done on two different species of bacteria and their experimental setup therefore does not allow them to isolate crossbridge length as the only differential property that can influence stiffness. These two species likely also differ in other ways that could modulate stiffness, e.g. turgor pressure, overall PG architecture (not just crossbridge length), membrane properties, teichoic acid composition etc.

Regarding the conclusions drawn about the correlation between cross-bridge length and cell envelope stiffness, we understand your point and appreciate your feedback. We revisit this section of our manuscript and tone down the conclusions drawn from this aspect of the study. We also recognize the importance of considering other potential factors that could influence stiffness, as you mentioned above. In light of this, we mentioned the need for further investigations, potentially involving genetically modified strains, in the main text to isolate and accurately determine the impact of bridge length on cell envelope stiffness.

Reviewer #1 (Recommendations For The Authors):

Minor points:

1) One thing to consider would be testing the robustness of the analysis pipeline with one the well-characterized bacteria studied, but genetically modifying them to change the cell wall composition in predictable ways. Does the analysis pipeline detect the expected changes?

We appreciate the reviewer's suggestion and would like to provide a clear response. Regarding to testing the pipeline with genetically modified strains, our lab previously worked on genetically modified S. maltophilia (KJΔmrdA).1 Inactivation of mrdA turned out the increasing level of N-acetylglucosaminyl-1,6-anhydro-N-acetylmuramyl-L-alanyl-D-glutamyl-meso-diamnopimelic acid-D-alanine (GlcNAc-anhMurNAc tetrapeptide) in muropeptide profiles, which is the critical activator ligands for mutant strain ΔmrdA-mediated β-lactamase expression. In this case, our platform could provide rapid PGN analysis for verifying the expected change of muropeptide profiles (see Author response image 1). Besides, if the predictable changes involve genetically modifications on interpeptide bridges within the PGN structure, for example, the femA/B genes of S. aureus, which are encoded for the synthesis of interpeptide bridges,2 our current HAMA pipeline is capable of detecting these anticipated changes. However, if the genetically modifications involve the introduce of novel components to PGN structures, then it would need to create a dedicated database specific to the genetically modified strain.

Author response image 1.

2) Line 368: products catalyzed > products formed

The sentence has been revised.

“(Section Inferring PGN Cross-linking Types Based on Identified PGN Fragments) …Based on the muropeptide compositional analysis mentioned above, we found high abundances of M3/M3b monomer and D34 dimer in the PGNs of E. faecalis, E. faecium, L. acidophilus, B. breve, B. longum, and A. muciniphila, which may be the PGN products formed by Ldts.”

3) Lines 400-402: Is it possible the effect is related to porosity, not "hardness".

Thank you for the suggestion. The possibility of the slower hydrolysis rate of purified PGN in B. breve being related to porosity is indeed noteworthy. While this could be a potential factor, we would like to acknowledge the limited existing literature that directly addresses the relation between PGN architecture and porosity. It is plausible that current methods available for assessing cell wall porosity may have certain limitations, contributing to the scarcity of relevant studies. In light of this, we would like to propose a speculative explanation for the observed effect. It is plausible that the tighter PGN architecture resulting from shorter interpeptide bridges in B. breve could contribute to its harder texture. This speculation is grounded in the concept that a more compact PGN structure might lead to increased stiffness, aligning with our observations of higher cell stiffness in B. breve.

4) Lines 403-408: See point #2 above.

Thank you for the suggestion. We have explicitly addressed this point in the main text:

“(Section Exploring the Bridge Length-dependent Cell Envelope Stiffness in B. longum and B. breve) … Taken all together, we speculate that a tight peptidoglycan network woven by shorter interpeptide bridges or 3-3 cross-linkages could give bacteria stiffer cell walls. However, it is important to note that cell stiffness is a mechanical property that also depends on PGN thickness, overall architecture, and turgor pressure. These parameters may vary among different bacterial strains. Hence, carefully controlled, genetically engineered strains with similar characteristics will be needed to dissect the role of cross-bridge length in cell envelope stiffness.”

5) Lines 428-429: It is not clear to me how mapping the cell wall architecture provides structural information about the synthetic system. It is also not clear how antibiotic resistance can be inferred. More detail is needed here to flesh out these points.

Thank you for the suggestion. To provide further clarity on these important aspects, the context in the manuscript has been revised.

“(Discussion) …Importantly, our HAMA platform provides a powerful tool for mapping peptidoglycan architecture, giving structural information on the PGN biosynthesis system. This involves the ability to infer possible PGN cross-linkages based on the type of PGN fragments obtained from hydrolysis. For instance, the identification of 3-3 cross-linkage formed by L,D-transpeptidases (Ldts) is of particular significance. Unlike 4-3 cross-linkages, the 3-3 cross-linkage is resistant to inhibition by β-Lactam antibiotics, a class of antibiotics that commonly targets bacterial cell wall synthesis through interference with 4-3 cross-linkages. Therefore, by elucidating the specific cross-linkage types within the peptidoglycan architecture, our approach offers insights into antibiotic resistance mechanisms.”

6) Line 478: "maneuvers are proposed for" > "work is needed to generate". Also, delete "innovative". Also "in silico" > "in silico-based".

The sentence has been revised.

“(Discussion) …To achieve a more comprehensive identification of muropeptides, future work is needed to generate an expanded database, in silico-based fragmentation patterns, and improved MS/MS spectra acquisition.”

7) Line 485: "Its" > "It has potential"

The sentence has been revised.

“(Discussion) …It has potential applications in identifying activation ligands for antimicrobial resistance studies, characterizing key motifs recognized by pattern recognition receptors for host-microbiota immuno-interaction research, and mapping peptidoglycan in cell wall architecture studies.”

8) Figure 1 legend: Define Gb and Pb.

Gb and Pb are the abbreviations of glycosidic bonds and peptide bonds. We have revised the Figure legend 1 as follow:

“(Figure legend 1) …(b) DBuilder constructs a muropeptide database containing monomers, dimers, and trimers with two types of linkage: glycosidic bonds (Gb) and peptide bonds (Pb).”

9) Figure 2: It is hard to see what is going on in panel a and c with all the labels. Consider removing them and showing a zoomed inset with labels in addition to ab unlabeled full chromatogram.

We apologize for not making this clearer. The panel a and c in Figure 2 were directly generated by the Analyzer as a software screenshot of the peak annotations on chromatogram. Our intention was to present a comprehensive PGN mapping (approximately 70% of the peak area was assigned to muropeptide signals) using this platform. We understand the label density might affect clarity, so we have added the output tables of the whole muropeptide identifications as source data (Table 1–Source Data 1&2). Additionally, we have uploaded the Analyzer output files (see Additional Files), which can be better visualized in the Viewer program, and it also allows users zoom in for detailed labeling information.

10) Figure 3: It is worth pointing out what features of the MS/MS fingerprints are helping to discriminate between species.

Thank you for the suggestion. We have revised Figure 3 and the legend as follow:

“(Figure legend 3) …The sequence of each isomer was determined using in silico MS/MS fragmentation matching, with the identified sequence having the highest matching score. The key MS/MS fragments that discriminate between two isomers are labeled in bold brown.”

Author response image 2.

11) Figure 4 and 5 legend: Can you condense the long descriptions of the abbreviations - or at least only refer to them once?

Certainly, to enhance clarity and conciseness in the figure legends, we have revised Figure legend 5 as follow:

“(Figure legend 5) …(b) Heatmap displaying …. Symbols: M, monomer; D, dimer; T, trimer (numbers indicate amino acids in stem peptides). Description of symbol abbreviations as in Figure legend 4, with the addition of "Glycan-T" representing trimers linked by glycosidic bonds.”

Reviewer #2 (Recommendations For The Authors):

1. Please read the manuscript carefully for spelling errors.

We appreciate your careful review of our manuscript. We have thoroughly rechecked the entire manuscript for spelling errors and have made the necessary corrections to ensure the accuracy and quality of the text.

1. Line 46 - "multilayered" is likely only true for Gram-positive bacteria.

We thank reviewer #2 for bringing up this concern. Indeed, Gram-negative bacteria mostly possess single layer of peptidoglycan, but could be up to three layers in some part of the cell surface.3, 4 In order to reduce the confusion, we have rewritten the context as follow: “(Introduction) …PGN is a net-like polymeric structure composed of various muropeptide molecules, with their glycans linearly conjugated and short peptide chains cross-linked through transpeptidation.”

1. Methods section: It seems like pellets from a 10 mL bacterial culture were ultimately suspended in 1.5 L (750 mL water + 750 mL tris) - why such a large volume? And how were PG fragments subsequently washed (centrifugation? There is no information on this in the Methods).

We apologize for the mislabeling on the units. The accurate volume should be “1.5 mL (750 µL water + 750 µL tris)”. We have updated the correct volume in the Methods section (lines 99-100). For the washing process of purified PGN, we added 1 mL water, centrifuged at 10,000 rpm for 5 minutes, and removed supernatant. This information has added to the Methods section (lines 95-98).

1. Line 183 - why were 6 modifications chose as the cutoff? Please make rationale more clear.

We thank reviewer #2 for the comments. We set the maximum modification number of 6 in the assumption of one modification on each sugar of a trimeric muropeptide. A lower cutoff could effectively limit the identification of muropeptides with unlikely numbers of modifications, whereas a higher cutoff could allow for having multiple modifications on a muropeptide. In our hand, muropeptide modifications of E. coli are mostly N-deacetyl-MurNAc and anhydro-MurNAc, and modifications of gut microbes used here are mostly N-deacetyl-GlcNAc, anhydro-MurNAc, O-acetyl-MurNAc, loss of GlcNAc, and amidated iso-Glu. While we recommend starting data analysis with the cutoff of 6 modifications, users are free to adjust this based on their studies.

1. Line 339 - define donor vs. acceptor here (can be added in parentheses after explaining the relevant chemical reactions further above in the text)

Thank you for the suggestion. To provide greater clarity regarding the roles of the donor and acceptor substrates in the transpeptidation process, we have revised the content in the manuscript as follows:

“(Section Inferring PGN Cross-linking Types Based on Identified PGN Fragments) …In general, there are two types of PGN cross-linkage…. Transpeptidation involves two stem peptides which function as acyl donor and acceptor substrates, respectively. As the enzyme names imply, the donor substrates that Ddts and Ldts bind to are terminated as D,D-stereocenters and L,D-stereocenters, which structurally means pentapeptides and tetrapeptides. During D,D-transpeptidation, Ddts recognize D-Ala4-D-Ala5 of the donor stem (pentapeptide) and remove the terminal D-Ala5 residue, forming an intermediate. The intermediate then cross-links the NH2 group in the third position of the neighboring acceptor stem, forming a 4-3 cross-link.”

1. Line 366 following - can you calculate % crosslinks based on these numbers? What does "high abundance" of 3,3 crosslinks mean in this context? Is this the majority of PG?

Thank you for your questions. Calculating the percentage of crosslinks based on the muropeptide compositional numbers is a valid consideration. However, it's important to note that the muropeptides we analyzed were hydrolyzed by mutanolysin, and as such, deriving an accurate % crosslink value from these data might not provide a true representation of the crosslinking percentage within the PGN network. For a more precise determination of % crosslinks, methods such as solid-phase NMR on purified peptidoglycan would be required. Our research provides insights into the characterization of PGN fragments and allows us to infer potential PGN cross-linkage types and the enzymes involved based on the dominant muropeptide fragments. Regarding the phrase "high abundance" in the context, it indicates that the M3b/M4b monomer and D34 dimer muropeptides represent a significant portion of the hydrolysis products. These muropeptides are major constituents within the PGN fragments obtained from the enzymatic hydrolysis.

1. Line 375 - I am not sure PG is a meaningful diffusion barrier for drugs and signaling molecules, give that even larger proteins can apparently diffuse through the pores.

Thank you for raising this point. Peptidoglycan indeed possesses relatively wide pores that allow for the diffusion of larger molecules, including proteins.5 Research has provided a rough estimate of the porosity of the PGN meshwork, suggesting that it allows for the diffusion of proteins with a maximum molecular mass of around 50 kDa.6 Considering this, we acknowledge that PGN may not serve as a significant diffusion barrier for drugs and signaling molecules. The porosity of the PGN scaffold, which is defined by the degree of cross-linking, plays a role in influencing the transport of molecules to the cell membrane. Thus, while PGN may not serve as a strict diffusion barrier, its structural characteristics still impact bacterial cell mechanics and interactions. We have revised the manuscript to reflect this understanding:

“(Section Exploring the Bridge Length-dependent Cell Envelope Stiffness in B. longum and B. breve) …The porosity of the PGN scaffold, defined by the degree of cross-linking, influences the transport of larger molecules such as proteins. Therefore, modifications to PGN structure are anticipated to significantly affect bacterial cell mechanics and interactions.”

1. Line 400 - what does "slower hydrolysis rate" refer to, is this chemical hydrolysis or enzymatic (autolysins?). also, I am not sure hydrolysis rate of either modality allows for solid conclusions about how hard (line 402) the PG is.

Thank you for your comments. The hydrolysis rate here refers to the enzymatic hydrolysis, specifically the mutanolysin cleaving the β-N-acetylmuramyl-(1,4)-N-acetylglucosamine linkage. Indeed, there is no direct correlation between the hydrolysis rate and the hardness of PGN architecture, although the structure rigidity is a key determinant in protein digestion.7 Considering the enzymatic hydrolysis rate depending on the accessibility of the substrate to the enzyme, we proposed that the tighter PGN architecture could also lead to a slower hydrolysis rate. This speculation aligns with our observations of higher cell stiffness or more compact PGN structure of B. breve and its slower hydrolysis rate. We understand this is indirect proof, so the revised sentence now reads:

“(Section Exploring the Bridge Length-dependent Cell Envelope Stiffness in B. longum and B. breve) …Furthermore, B. breve also showed a slower enzymatic hydrolysis rate in purified PGNs, implying that the cell wall structure of B. breve is characterized by a compact PGN architecture.”

1. Line 424 - I am not convinced this pipeline can detect PG architectures that other pipelines cannot; likely, the difference between previous analyses and theirs is due to different growth conditions (3,3 crosslink formation is often modulated by environmental factors/growth stage). In the next sentence, it sounds like mutanolysin treatment is a novelty in PG analysis (which it is not).

We apologize if this could have been clearer and we have revised the paragraph to describe our study more accurately. We agree that different growth conditions could influence PGN architecture and other pipelines could manually identify the PGN architectures or automatically identify them if they are not too complex. Our original intention was to highlight the ability of the HAMA program to automatically identify unreported PGN structure. Here are the revised sentences:

“(Discussion) …We speculate that this finding may be influenced by the comprehensive mass spectrometric approaches we employed or by variations in growth conditions. Moreover, we utilized the well-established enzymatic method involving mutanolysin to cleave the β-N-acetylmuramyl-(1,4)-N-acetylglucosamine linkage, which preserves the original peptide linkage in intact PGN subunits.”

1. Line 440- 442: As outlined in more detail above: I don't think you can conclude something about the relationship between bridge length and envelope stiffness based on these data. Thank you for your valuable feedback. We agree that our data may not definitively support the direct conclusion about the relationship between bridge length and envelope stiffness in Bifidobacterium species. Instead, we will rephrase this section to accurately present the observed correlations without overgeneralizing:

“(Discussion) … Notably, our study suggested a potential correlation between the cell stiffness and the compactness of bacterial cell walls in Bifidobacterium species (Figure 5). B. longum, which predominantly harbors tetrapeptide bridges (Ser-Ala-Thr-Ala), exhibits a trend towards lower stiffness, whereas B. breve, characterized by PGN cross-linked with monopeptide bridges (Gly), demonstrates a trend towards higher stiffness. These findings suggested that it may be correlated between the increased rigidity and the more compact PGN architecture built by shorter cross-linked bridges.”

References: 1. Huang, Y.-W.; Wang, Y.; Lin, Y.; Lin, C.; Lin, Y.-T.; Hsu, C.-C.; Yang, T.-C., Impacts of Penicillin Binding Protein 2 Inactivation on β-Lactamase Expression and Muropeptide Profile in Stenotrophomonas maltophilia. mSystems 2017, 2 (4), 00077-00017.

1. Jarick, M.; Bertsche, U.; Stahl, M.; Schultz, D.; Methling, K.; Lalk, M.; Stigloher, C.; Steger, M.; Schlosser, A.; Ohlsen, K., The serine/threonine kinase Stk and the phosphatase Stp regulate cell wall synthesis in Staphylococcus aureus. Sci. Rep. 2018, 8 (1), 13693.

2. Labischinski, H.; Goodell, E. W.; Goodell, A.; Hochberg, M. L., Direct proof of a "more-than-single-layered" peptidoglycan architecture of Escherichia coli W7: a neutron small-angle scattering study. J. Bacteriol. 1991, 173 (2), 751-756.

3. Rohde, M., The Gram-Positive Bacterial Cell Wall. Microbiol. Spectr. 2019, 7 (3), gpp3-0044-2018.

4. Vollmer, W.; Höltje, J. V., The architecture of the murein (peptidoglycan) in gram-negative bacteria: vertical scaffold or horizontal layer(s)? J. Bacteriol. 2004, 186 (18), 5978-5987.

5. Vollmer, W.; Blanot, D.; De Pedro, M. A., Peptidoglycan structure and architecture. FEMS Microbiol. Rev. 2008, 32 (2), 149-167.

6. Li, Q.; Zhao, D.; Liu, H.; Zhang, M.; Jiang, S.; Xu, X.; Zhou, G.; Li, C., "Rigid" structure is a key determinant for the low digestibility of myoglobin. Food Chem.: X 2020, 7, 100094.

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

Summary

Type 1 diabetes mellitus (T1DM) progression is accelerated by oxidative stress and apoptosis. Eugenol (EUG) is a natural compound previously documented as anti-inflammatory, anti-oxidative, and anti-apoptotic. In this manuscript by Jiang et al., the authors study the effects of EUG on T1DM in MIN6 insulinoma cells and a mouse model of chemically induced T1DM. The authors show that EUG increases nuclear factor E2-related factor 2 (Nrf2) levels. This results in a reduction of pancreatic beta-cell damage, apoptosis, oxidative stress markers, and a recovery of insulin secretion. The authors highlight these effects as indicative of the therapeutic potential of EUG in managing T1DM.

Strengths

Relevant, timely, and addresses an interesting question in the field. The authors consistently observe enhanced beta cell functionality following EUG treatment, which makes the compound a promising candidate for T1DM therapy.

Weaknesses

(1) The in vivo experiments have too few biological replicates. With an n=3 (as all figure legends indicate) in complex mouse studies such as these, drawing robust conclusions becomes challenging. It is important to reproduce these results in a larger cohort, to validate the conclusions of the authors.

Thanks for your comments. In the figure legends of the first draft manuscript, n=3 means at least 3 biological replicates, and in the section of material and methods, n=30 means sample size. The number of mice in each group is 30 and there were 150 mice used in this study, and mice are assigned as follows for the whole in vivo experiments. The relative information has been added in the revised manuscript.

Author response image 1.

(2) Another big concern is the lack of quantifications and statistical analysis throughout the manuscript. Although the authors claim statistical significance in various experiments, the limited information provided makes it difficult to verify. The authors use vague and minimal descriptions of their experiments, which further reduces the reader's comprehension and the reproducibility of the experiments.

Thanks for your constructive suggestion. We conducted quantitative and statistical analysis of the entire manuscript through GraphPad Prism software again. Additionally, we have improved the experimental description in the revised manuscript.

(3) Finally, the use of Min6 cells as a model for pancreatic beta cells is a strong limitation of this study. Future studies should seek to reproduce these findings in a more translational model and use more relevant in vitro cell systems (eg. Islets).

Thanks for your professional comments. Mouse insulinoma cells (MIN6 cell line) are permanent cell lines isolated from mouse islet β cell tumors, which can reflect the functional changes of islet β cells. As mature islet cells, MIN6 cells have been widely used in the study of type 1 diabetes mellitus[1-4], so in this study, MIN6 cells were used as the cell model in vitro. In our future studies, we will try to conduct our findings using more relevant in vitro cell systems (eg. Islets).

References:

(1) WU M, CHEN W, ZHANG S, et al. Rotenone protects against β-cell apoptosis and attenuates type 1 diabetes mellitus [J]. Apoptosis, 2019, 24(11-12): 879-91.

(2) LUO C, HOU C, YANG D, et al. Urolithin C alleviates pancreatic β-cell dysfunction in type 1 diabetes by activating Nrf2 signaling [J]. Nutr Diabetes, 2023, 13(1): 24.

(3) LAKHTER A J, PRATT R E, MOORE R E, et al. Beta cell extracellular vesicle miR-21-5p cargo is increased in response to inflammatory cytokines and serves as a biomarker of type 1 diabetes [J]. Diabetologia, 2018, 61(5): 1124-34.

(4) LIN Y, SUN Z. Antiaging Gene Klotho Attenuates Pancreatic β-Cell Apoptosis in Type 1 Diabetes [J]. Diabetes, 2015, 64(12): 4298-311.

Reviewer #3 (Public Review):

Summary:

This study by Jiang et al. aims to establish the streptozotocin (STZ)-induced type 1 diabetes mellitus (T1DM) mouse model in vivo and the STZ-induced pancreatic β cell MIN6 cell model in vitro to explore the protective effects of Eugenol (EUG) on T1DM. The authors tried to elucidate the potential mechanism by which EUG inhibits the NRF2-mediated anti-oxidative stress pathway. Overall, this study is well executed with solid data, offering an intriguing report from animal studies for a potential new treatment strategy for T1DM.

Strengths:

The in vivo efficacy study is comprehensive and solid. Given that STZ-induced T1DM is a devastating and harsh model, the in vivo efficacy of this compound is really impressive.

Weaknesses:

(1) The Mechanism is linked with the anti-oxidant property of the compound, which is common for many natural compounds, such as flavonoids and polyphenol. However, rarely, this kind of compound has been successfully developed into therapeutics in clinical usage. Indeed, if that is the case, Vitamin C or Vitamin E could be used here as the positive control.

Thanks for your comments. In fact, many anti-oxidant drugs are used for the treatment of type 1 diabetes mellitus in the clinical. For example, lipoic acid was used to treat diabetic peripheral neuropathy[5]. Vitamin E could effectively eliminate free radicals, protect cell membranes, and significantly reduce the risk of cardiovascular disease in patients with SPACE or ICARE diabetes[6]. Glutathione played crucial roles in the detoxification and anti-oxidant systems of cells and has been used to treat acute poisoning and chronic liver diseases by intravenous injection[7]. Therefore, eugenol enhances the management of type 1 diabetes mellitus by modulating oxidative stress pathways and holds potential as a future therapeutic choice for clinical application. In the future relevant studies, we will try to use Vitamin C or Vitamin E as the positive control.

References:

(5) ZIEGLER D, PAPANAS N, SCHNELL O, et al. Current concepts in the management of diabetic polyneuropathy [J]. J Diabetes Investig, 2021, 12(4): 464-75.

(6) VARDI M, LEVY N S, LEVY A P. Vitamin E in the prevention of cardiovascular disease: the importance of proper patient selection [J]. J Lipid Res, 2013, 54(9): 2307-14.

(7) HONDA Y, KESSOKU T, SUMIDA Y, et al. Efficacy of glutathione for the treatment of nonalcoholic fatty liver disease: an open-label, single-arm, multicenter, pilot study [J]. BMC Gastroenterol, 2017, 17(1): 96.

Reviewer #1 (Recommendations For The Authors):

• For each of the figure panels the authors should indicate the exact number of biological replicates (how many mice or how many independent in vitro experiments). For IF panels, the number of mice, the number of histology slides per mouse, number of fields analyzed should be indicated.

Thanks for your constructive suggestion. These details had been added in the revised manuscript.

• The methods state n=30 and Figure 1 states n=3. N=3 is too little for such a complex in vivo study and would severely reduce the reliability of the in vivo experiments.

Thanks for your suggestion. In the figure legends of the first draft manuscript, n=3 means at least 3 biological replicates, and in the section of material and methods, n=30 means sample size. The number of mice in each group is 30 and there were 150 mice used in this study, and mice are assigned as follows for the whole in vivo experiments. The in vivo experimental data of Figure 1 were supplemented in the revised manuscript.

• Individual data points should be included in each of the graphs from this manuscript.

Thanks for your reminder. The revised manuscript have shown the individual data points in each of the graphs.

• The quantifications and statistics in the manuscript need improvement. Several experiments are missing quantifications and/or statistical tests (e.g. Figure 1J). Other experiments show a quantification but without any explanation of replicates (e.g. Figures 2B and 2G). None of the experiments show individual data points, and as in the previous comment, these should be included.

Thanks for your comments. In the revised manuscript, statistics and repetitions of experimental data have been supplemented, and individual data points were shown in each graph.

• What is the reason for intragastric administration? The previous studies on which the dosages were based used oral administration (gavage). (Discussed in methods 4.2).

Thanks for your professional comments. The intervention treatment of T1DM mice is conducted through two methods: oral administration[8] and oral gavage[9-11]. Due to limited experimental conditions, it is not feasible to feed a single mouse in a single cage, which makes it challenging to precisely control the actual daily intervention dose for each mouse when using oral administration. To ensure that each mouse receives an intervention dose according to its weight and expected dosage, we employ a method of gavage. In addition, oral gavage is more convenient and easier to operate than oral administration. Therefore, in vivo experiment of this study used eugenol gavage intervention as a treatment method. These details had been added in the revised manuscript.

References:

(8) ZHAO H, WU H, DUAN M, et al. Cinnamaldehyde Improves Metabolic Functions in Streptozotocin-Induced Diabetic Mice by Regulating Gut Microbiota [J]. Drug Des Devel Ther, 2021, 15: 2339-55.

(9) XING D, ZHOU Q, WANG Y, et al. Effects of Tauroursodeoxycholic Acid and 4-Phenylbutyric Acid on Selenium Distribution in Mice Model with Type 1 Diabetes [J]. Biol Trace Elem Res, 2023, 201(3): 1205-13.

(10) SUDIRMAN S, LAI C S, YAN Y L, et al. Histological evidence of chitosan-encapsulated curcumin suppresses heart and kidney damages on streptozotocin-induced type-1 diabetes in mice model [J]. Sci Rep, 2019, 9(1): 15233.

(11) YAO H, SHI H, JIANG C, et al. L-Fucose promotes enteric nervous system regeneration in type 1 diabetic mice by inhibiting SMAD2 signaling pathway in enteric neural precursor cells [J]. Cell Commun Signal, 2023, 21(1): 273.

• Urine volume cannot be specified per mouse (methods 4.4) unless the mice were single-housed or if the different groups were not mixed, both are not ideal study set-ups. Please clarify in the methods section.

Thanks for your constructive suggestion. After successful modeling of T1DM mice, the successful modeling mice were grouped based on method 4.2 as follows Control, T1DM, T1DM + EUG (5 mg/kg/day), T1DM + EUG (10 mg/kg/day), and T1DM + EUG (20 mg/kg/day). To ensure consistency among groups, each group consisted of 5 mice and had equal amounts of diet (100 g), drinking water (250 mL), and environmental conditions for feeding. The urine-soaked area of mice in each group was recorded to quantify the urine volume. The conditions are the same for each group. The description of Method 4.4 has been improved in the revised manuscript.

• OGTT (Figure 1H) of week 2 is missing. This is an important control time point, as it would show the effect of STZ before EUG treatment.

Thanks for your careful review. OGTT (Figure 1H) of week 2 has been added in the revised manuscript.

• In Figure 1J, the control group does not follow the expected ITT trajectory. If possible, add the 120-minute time point to see if the blood glucose levels return to baseline in the control group. The graph shows increased basal glucose levels in the experimental groups, but no differences in insulin tolerance. It also misses the AUC calculations. It is probably not significantly different, which should be noted in the text.

Thanks for your suggestion. T1DM primarily manifests as pancreatic β cell damage and the absolute reduction of insulin secretion, resulting in the disorder of glucose metabolism in vivo. The oral glucose tolerance test (OGTT) is a series of plasma glucose concentrations measured within 2 h after oral gavage of a certain amount of glucose. It is a standard method to evaluate an individual's blood glucose regulation ability and to understand the function of islet β cells. Insulin resistance means reducing the efficiency of insulin to promote glucose uptake and utilization for various reasons, and the body's compensatory secretion of excessive insulin leads to hyperinsulinemia to maintain the stability of blood glucose. The insulin resistance test (ITT) is commonly employed to detect insulin resistance in T2DM. However, it was found that the ITT experiment had little correlation with T1DM. Therefore, the ITT experiment of Figure 1J and related description have been removed from the revised manuscript.

• The staining and FACS data on the effects of STZ+EUG+/- ML385 are not convincing (Figure 6 and Figure 7) and do not seem to align with the bar graphs and the conclusions in the text. It would be good to include immunofluorescent staining for insulin to further validate the effects of STZ+EUG+/- ML385 on insulin expression.

Thanks for your comments.

(1) In the revised manuscript, between the statistical results and the pictures, so we re-conducted the statistics of the immunofluorescence results of NRF2 and HO-1, as follows：

(1) NRF2 immunofluorescence staining:

Author response image 2.

Group 1

Author response image 3.

Group 2

Author response image 4.

Group 3

Author response image 5.

Group 4

Author response image 6.

Group 5

Author response image 7.

NRF2 immunofluorescence staining statistics:

(2) HO-1 immunofluorescence staining:

Author response image 8.

Group 1

Author response image 9.

Group 2

Author response image 10.

Group 3

Author response image 11.

Group 4

Author response image 12.

Group 5

Author response image 13.

HO-1 immunofluorescence staining statistics:

(2) The meanings represented by each quadrant of cell flow analysis are as follows: Q1 represents a group of necrotic cells, characterized by positive PI staining and negative Anenexin V staining; Q2 represents late apoptotic cells, with both PI and Anenexin V staining negative; Q3 represents early apoptotic cells, with both PI and Anenexin V staining positive; Q4 represents living cells, characterized by positive Anenexin V staining and negative PI staining. In the experiment, the number of apoptotic cells were calculated as the sum of late apoptotic cells in Q2 and early apoptotic cells in Q3. As shown in Figure 9F-G, these results were consistent with those observed in Figure 6G, 6J and Figure 7D-F.

(3) MIN6 cells, as mouse islet β cell line, has the function of secreting insulin. The intervention of STZ was an absolute decrease in the number of islet β cells, so the result of insulin immunofluorescence staining was only a decrease in the number of MIN6 cells in each cell group. In addition, the detection of insulin protein expression level is always through ELISA method to assess the secretion of insulin protein in the cell supernatant. Figure 6E is the ELISA results of insulin protein secretion in the cell supernatant.

• The experimental design for the in vitro experiments was unclear from the text. Consider including a schematic to show when cells were treated with STZ, EUG, and ML385.

Thanks for your suggestion. The experimental design for the in vitro experiments of this study has been added in Figure 6A of the revised manuscript.

• As stated in the Discussion, the use of the insulinoma line Min6 as a model instead of primary pancreatic beta cells is a clear limitation of the study. The mechanistic data would be stronger if validated on a more relevant system (eg. untransformed Islets).

Thanks for your comments. Mouse insulinoma cells (MIN6 cell line) are permanent cell lines isolated from mouse islet β cell tumors, which can reflect the functional changes of islet β cells. As mature islet cells, MIN6 cells have been widely utilized as an in vitro cellular model for diabetes research to investigate the functionality of β cells within pancreatic islets[1, 2, 12]. So in this study, MIN6 cells were used as the cell model in vitro. In our future studies, we will try to conduct our findings using more relevant in vitro cell systems (eg. Islets).

References:

(1) WU M, CHEN W, ZHANG S, et al. Rotenone protects against β-cell apoptosis and attenuates type 1 diabetes mellitus [J]. Apoptosis, 2019, 24(11-12): 879-91.

(2) LUO C, HOU C, YANG D, et al. Urolithin C alleviates pancreatic β-cell dysfunction in type 1 diabetes by activating Nrf2 signaling [J]. Nutr Diabetes, 2023, 13(1): 24.

(12) CHEN H, LOU Y, LIN S, et al. Formononetin, a bioactive isoflavonoid constituent from Astragalus membranaceus (Fisch.) Bunge, ameliorates type 1 diabetes mellitus via activation of Keap1/Nrf2 signaling pathway: An integrated study supported by network pharmacology and experimental validation [J]. J Ethnopharmacol, 2024, 322: 117576.

• The use of small molecule inhibitors such as ML385 can have unspecific effects. Genetic manipulation or the use of siRNAs to inhibit the NRF2 pathway would have been preferable for the in vitro experiments.

Thanks for your constructive suggestion. ML385 is a commonly used and stable inhibitor of the NRF2 and has been used in a variety of disease studies[13-15]. The MIN6 cells utilized in this study were cultured under challenging conditions and exhibited a sluggish growth rate. Owing to the cytotoxicity associated with siRNAs transfection reagents, a significant proportion of MIN6 cells succumbed following transfection. Consequently, small molecule inhibitors ML385 were employed in this investigation. In our future studies, we will try to conduct our findings using siRNAs.

References:

(13) DANG R, WANG M, LI X, et al. Edaravone ameliorates depressive and anxiety-like behaviors via Sirt1/Nrf2/HO-1/Gpx4 pathway [J]. J Neuroinflammation, 2022, 19(1): 41.

(14) WANG Z, YAO M, JIANG L, et al. Dexmedetomidine attenuates myocardial ischemia/reperfusion-induced ferroptosis via AMPK/GSK-3β/Nrf2 axis [J]. Biomed Pharmacother, 2022, 154: 113572.

(15) LI J, DENG S H, LI J, et al. Obacunone alleviates ferroptosis during lipopolysaccharide-induced acute lung injury by upregulating Nrf2-dependent antioxidant responses [J]. Cell Mol Biol Lett, 2022, 27(1): 29.

• The study proposes a mechanism in which EUG-induced disruption of KEAP1 and NRF2 interaction leads to NRF2 translocation to the nucleus and upregulation of proteins required to prevent oxidative stress. In Figure 6H it is unclear whether the nuclear NRF2 increases. Please add quantifications of the immunostainings.

Thanks for your reminder. Figure 6J shows the quantifications of the immunostainings of NRF2 in the revised manuscript.

• Some of the figure legends lack important information. In Figure 5A, 6E for instance, what is the protein expression normalized to?

Thanks for your constructive suggestion. Protein normalization refers to the standardization of proteins from different sources and with different properties, so as to facilitate the comparison of protein content and expression in different samples. In WB experiment, protein expression normalization is one of the essential steps. Western blot of nuclear protein generally cannot be performed using β-Actin as an internal reference. Lamin B was chosen because β-Actin is an intrinsic parameter not found in the nucleus. N-NRF2, as a nuclear protein, requires Lamin B as a reference for protein normalization. The lack important information of WB in Figure have been supplemented in figure legends of the revised manuscript.

• Please acknowledge previous literature on the effects of EUG/clove oil in diabetes models. The meta-analytical review by Carvalho et al. (DOI: 10.1016/j.phrs.2020.105315) should be cited and discussed.

Thanks for your suggestion. It has been cited and discussed in the revised manuscripts.

• Consider revising the text for grammar, language mistakes, and readability. The text is not always precise (e.g. in the explanation of gamma-H2AX in the results), does not explain terminology (e.g. the oxidative stress markers - line 204+205), or simplifies conclusions (e.g. "improved islet function" based on glucose tolerance test", line 129).

Thanks for your comments. The above problem has been solved in the revised manuscripts. In addition, we had send our manuscript to the professional English language editing company to improve our paper, and the editorial certificate had been submitted as a supplement document.

• In the current format, some figures are out of focus. Please make sure to upload a high-quality version for publication.

Thanks for your suggestion. A high quality version figures has been uploaded. Perhaps due to the excessive content of the file after upload, the file is compressed, and the figures is not focused. So, all figures in this study have been uploaded separately for download in the review system.

Reviewer #2 (Recommendations For The Authors):

Below are specific points of criticism on the experiments presented.

(1a) There is no comparison among eugenol treatments with regards to fasting weight, blood glucose, water intake, food intake, and, crucially, OGTT. All three treatments appear to show very similar effects but has this been statistically assessed? Shown statistical significance of ketonuria between no and high eugenol treatments seems exaggerated.

Thanks for your comments. EUG intervention has a dose-dependent effect on T1DM. According to Figure 1B-I, 20 mg/kg EUG has the best effect. Fasting body weight, blood glucose, water intake, food intake, and OGTT were statistically assessed in Figure 1 of the revised manuscript. In addition, we performed statistical analyse of ketonuria between no and high eugenol treatments again in the revised manuscript. In the revised manuscript, we have also made objective revisions to the expression of eugenol's efficacy.

(b) ITT is not used to detect T1DM (line 126).

Thanks for your suggestion. T1DM primarily manifests as pancreatic β cell damage and the absolute reduction of insulin secretion, resulting in the disorder of glucose metabolism in vivo. The oral glucose tolerance test (OGTT) is a series of plasma glucose concentrations measured within 2 h after oral gavage of a certain amount of glucose. It is a standard method to evaluate an individual's blood glucose regulation ability and to understand the function of islet β cells. Insulin resistance means reducing the efficiency of insulin to promote glucose uptake and utilization for various reasons, and the body's compensatory secretion of excessive insulin leads to hyperinsulinemia to maintain the stability of blood glucose. The insulin resistance test (ITT) is commonly employed to detect insulin resistance in T2DM. However, it was found that the ITT experiment had little correlation with T1DM. Therefore, the ITT experiment and related description have been removed in the revised manuscript.

(2) Here it is hard to reconcile the gradual increase of Ins protein levels in (STZ) and (STZ + increasing eugenol) samples with(a) results in 1 suggesting that the dose of eugenol does not significantly affect the outcome and(b) Ins expression, which is essentially undetectable in both STZ and STZ+EUG mice. A likely explanation is that EUG just postpones beta cell death. I assume that these analyses were done in week 10 but it is not stated.

Thanks for your professional suggestion. Perhaps because the file is compressed, the gray value of WB strip is not obvious, so the expression of INS is not seen clearly. In fact, the intervention of STZ resulted in a significant decrease in INS expression compared with the Control group, which could be alleviated by the treatment of EUG. However, due to the large difference in INS between the STZ group, EUG treatment, and the Control group, the gray values of INS in the STZ group and the STZ + EUG group were not clear. As mentioned in the method 4.12-4.13, our WB and PCR samples were from 10 week mice.

(3) The γH2Ax stainings provided are weak and do not fully correspond to the quantitation - the 5 mg/Kg EUG treatment appears less severe than the 10 mg/Kg. In contrast, changes in the PCD pathway are convincingly demonstrated.

Thanks for your reminder. γH2AX immunohistochemical staining is required to be located in the islets. It measured the number of β cells stained with brown, not the brown area. The ZOOM image of γH2AX staining showed that the EUG improvement effect of 10 mg/kg was better than that of 5 mg/kg. γH2AX, as a marker of DNA damage, exhibits nuclear localization and is absent in the cytoplasmic compartment. Therefore, in Figure 4C-D, we quantified the proportion of cells exhibiting brown staining. In Figure 4C, black arrows were employed to highlight the presence of brown-stained islet β cells.

(4) Is there a reason for looking at mRNA levels of Ho-1 but not KEAP1 or NQO-1 ? What is the expression of Nrf2 itself at the RNA level? Please give in the text what the abbreviations MDA, SOD, CAT GSH-Px stand for. Are these protein levels or activity assays? Units in the y-axis of graphs?

Thanks for your constructive suggestion.The required KEAP1 and NQO-1 primers have been synthesized, and the relevant data have been supplemented in the revised manuscript. The expression of Nrf2 itself at the RNA level is T-NRF2 (Total NRF2). The MDA, SOD, CAT and GSH-Px abbreviations stand for Malondialdehyde, Superoxide dismutase, Catalase, Glutathione peroxidase, and the relevant information, which have been supplemented in the revised manuscript. These are activity assays of serum, and units in the y-axis of graphs have been added in the revised manuscripts.

(5) The Ins levels in the culture medium of STZ + ML treated cells are much lower than the levels in STZ treated cells (6D). This is not consistent with the results of Ins cell content or Ins expression as stated (6B and D).

Thanks for your careful review. The experimental samples in Figure 6C in the revised manuscript represent the proteins extracted from cells of each group, while the experimental samples in Figure 6E represent the supernatant of cells from each group. ML385 is an inhibitor of NRF2, which effectively suppresses the NRF2 signaling pathway and aggravates MIN6 cell damage, resulting in lower INS expression observed in both the STZ+ML385 group depicted in Figures 6C and 6E compared to that in the STZ group. Although the sample sources of the two groups differ and there are slight variations in the trend, it can be observed that the overall trend of the STZ+ML385 group is comparatively lower than that of the STZ group.

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

Chen et al. identified the role of endocardial id2b expression in cardiac contraction and valve formation through pharmaceutical, genetic, electrophysiology, calcium imaging, and echocardiography analyses. CRISPR/Cas9 generated id2b mutants demonstrated defective AV valve formation, excitation-contraction coupling, reduced endocardial cell proliferation in AV valve, retrograde blood flow, and lethal effects.

Strengths:

Their methods, data and analyses broadly support their claims.

Weaknesses:

The molecular mechanism is somewhat preliminary.

We thank the reviewer for the positive assessment of our work. A detailed point-by-point response has been incorporated in the response to “Recommendations for the authors” section.

Reviewer #2 (Public review):

Summary:

Biomechanical forces, such as blood flow, are crucial for organ formation, including heart development. This study by Shuo Chen et al. aims to understand how cardiac cells respond to these forces. They used zebrafish as a model organism due to its unique strengths, such as the ability to survive without heartbeats, and conducted transcriptomic analysis on hearts with impaired contractility. They thereby identified id2b as a gene regulated by blood flow and is crucial for proper heart development, in particular, for the regulation of myocardial contractility and valve formation. Using both in situ hybridization and transgenic fish they showed that id2b is specifically expressed in the endocardium, and its expression is affected by both pharmacological and genetic perturbations of contraction. They further generated a null mutant of id2b to show that loss of id2b results in heart malformation and early lethality in zebrafish. Atrioventricular (AV) and excitation-contraction coupling were also impaired in id2b mutants. Mechanistically, they demonstrate that Id2b interacts with the transcription factor Tcf3b to restrict its activity. When id2b is deleted, the repressor activity of Tcf3b is enhanced, leading to suppression of the expression of nrg1 (neuregulin 1), a key factor for heart development. Importantly, injecting tcf3b morpholino into id2b-/- embryos partially restores the reduced heart rate. Moreover, treatment of zebrafish embryos with the Erbb2 inhibitor AG1478 results in decreased heart rate, in line with a model in which Id2b modulates heart development via the Nrg1/Erbb2 axis. The research identifies id2b as a biomechanical signaling-sensitive gene in endocardial cells that mediates communication between the endocardium and myocardium, which is essential for heart morphogenesis and function.

Strengths:

The study provides novel insights into the molecular mechanisms by which biomechanical forces influence heart development and highlights the importance of id2b in this process.

Weaknesses:

The claims are in general well supported by experimental evidence, but the following aspects may benefit from further investigation:

(1) In Figure 1C, the heatmap demonstrates the up-regulated and down-regulated genes upon tricane-induced cardiac arrest. Aside from the down-regulation of id2b expression, it was also evident that id2a expression was up-regulated. As a predicted paralog of id2b, it would be interesting to see whether the up-regulation of id2a in response to tricane treatment was a compensatory response to the down-regulation of id2b expression.

We thank the reviewer for the comment. As suggested, we performed qRT-PCR analysis to assess id2a expression in tricaine-treated heart. Our results demonstrate a significant upregulation of id2a following the inhibition of cardiac contraction, suggesting a potential compensatory response to the decreased id2b. These new results have been incorporated into the revised manuscript (Figure 1D).

(2) The study mentioned that id2b is tightly regulated by the flow-sensitive primary cilia-klf2 signaling axis; however aside from showing the reduced expression of id2b in klf2a and klf2b mutants, there was no further evidence to solidify the functional link between id2b and klf2. It would therefore be ideal, in the present study, to demonstrate how Klf2, which is a transcriptional regulator, transduces biomechanical stimuli to Id2b.

We have examined the expression levels of id2b in both klf2a and klf2b mutants. The whole mount in situ results clearly demonstrate a decrease in id2b signal in both mutants (Figure 3E). As noted by the reviewer, klf2 is a transcriptional regulator, suggesting that the regulation of id2b may occur at the transcriptional level. However, dissecting the molecular mechanisms underlying the crosstalk between klf2 and id2b is beyond the scope of the present study.

(3) The authors showed the physical interaction between ectopically expressed FLAG-Id2b and HA-Tcf3b in HEK293T cells. Although the constructs being expressed are of zebrafish origin, it would be nice to show in vivo that the two proteins interact.

We thank the reviewer for this insightful comment. As suggested, we synthesized Flag-id2b and HA-tcf3b mRNA and co-injected them into 1-cell stage zebrafish embryos. We collected 100-300 embryos at 12, 24, and 48 hpf and performed western blot analysis using the same anti-HA and anti-Flag antibodies validated in HEK293 cell experiments. Despite multiple independent attempts, we were unable to detect clear bands of the tagged proteins in zebrafish embryos. We speculate that this could be due to mRNA instability, translational efficiency, or the low abundance of Id2b and Tcf3b proteins. We have acknowledged these technical limitations in the revised manuscript and clarified that the HEK293 cell data support a potential interaction between Id2b and Tcf3b, while confirming their endogenous interaction will require further investigations (Lines 295-296).

Reviewer #3 (Public review):

Summary:

How mechanical forces transmitted by blood flow contribute to normal cardiac development remains incompletely understood. Using the unique advantages of the zebrafish model system, Chen et al make the fundamental discovery that endocardial expression of id2b is induced by blood flow and required for normal atrioventricular canal (AVC) valve development and myocardial contractility by regulating calcium dynamics. Mechanistically, the authors suggest that Id2b binds to Tcf3b in endocardial cells, which relieves Tcf3b-mediated transcriptional repression of Neuregulin 1 (NRG1). Nrg1 then induces expression of the L-type calcium channel component LRRC1. This study significantly advances our understanding of flow-mediated valve formation and myocardial function.

Strengths:

Strengths of the study are the significance of the question being addressed, use of the zebrafish model, and data quality (mostly very nice imaging). The text is also well-written and easy to understand.

Weaknesses:

Weaknesses include a lack of rigor for key experimental approaches, which led to skepticism surrounding the main findings. Specific issues were the use of morpholinos instead of genetic mutants for the bmp ligands, cilia gene ift88, and tcf3b, lack of an explicit model surrounding BMP versus blood flow induced endocardial id2b expression, use of bar graphs without dots, the artificial nature of assessing the physical interaction of Tcf3b and Id2b in transfected HEK293 cells, and artificial nature of examining the function of the tcf3b binding sites upstream of nrg1.

We thank the reviewer for the positive assessment and the constructive suggestions. We have performed additional experiments and data analysis to address these issues. A detailed point-by-point response has been incorporated in the response to “Recommendations for the authors” section.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Questions/Concerns:

(1) In the introduction, it would be beneficial to include background information on the id2b gene, what is currently known about its function in heart development/regeneration and in other animal models than just the zebrafish.

We thank the reviewer for the constructive suggestion. In the revised manuscript, we have added a paragraph in the Introduction to provide background on id2b and its role in heart development. Specifically, we discuss its function as a member of the ID (inhibitor of DNA binding) family of helix-loop-helix (HLH) transcriptional regulators and highlight its involvement in cardiogenesis in both zebrafish and mouse models. These additions help place our findings in a broader developmental and evolutionary context (Lines 91-100).

(2) Of the 6 differentially expressed genes identified in Figure 1C, why did the authors choose to focus on id2b and not the other 5 downregulated genes?

We thank the reviewer for the comments. As suggested, we have added a sentence in the revised manuscript to clarify the rationale for selecting id2b as the focus of the present study (Lines 117-121).

(3) As the authors showed representative in situ images for id2b expression with blebbistatin treatment in Figure 1E, and tnn2a MO in Figure 1F, it would also be beneficial to show relative mRNA expression levels for id2b in conditions of blebbistatin treatment and tnn2a MO knockdown. In Fig. 1C: id2b is downregulated with tricaine, but id2a is upregulated with tricaine. Do these genes perform similar or different functions, results of gene duplication events?

We thank the reviewer for the thoughtful suggestion. Our in situ hybridization results demonstrate reduced id2b expression following tricaine, blebbistatin, and tnn2 morpholino treatment. To further validate these observations and enhance cellular resolution, we generated an id2b:eGFP knockin line. Analysis of this reporter line confirmed a significant reduction in id2b expression in the endocardium upon inhibition of cardiac contraction and blood flow (Figure 3A-D), supporting our in situ results. The divergent expression patterns of id2a and id2b in response to tricaine treatment likely reflect functional specification following gene duplication in zebrafish. While our current study focuses on characterizing the role of id2b in zebrafish heart development, the specific function of id2a remains to be determined. 

(4) In Fig. 2b, could the authors compare the id2b fluorescence with RNAscope ISH at 24, 48, and 72 hpf? RNAscope ISH allows for the visualization of single RNA molecules in individual cells. The authors should at least compare these in the heart to demonstrate that id2b accurately reflects the endogenous id2b expression. In Fig. 2E: Suggest showing the individual fluorescent images for id2b:eGFP and kdrl:mCherry in the same colors as top panel images instead of in black and white. In Fig. 2F: The GFP fluorescence from id2b:eGFP signals looks overexposed.

We thank the reviewer for the valuable comment. In response, we attempted RNAscope in situ hybridization on embryos carrying the id2b:eGFP reporter to directly compare fluorescent reporter expression with endogenous id2b transcripts. However, we encountered a significant reduction in id2b:eGFP fluorescence following the RNAscope procedure, and even subsequent immunostaining with anti-GFP antibodies yielded only weak signals. Despite this technical limitation, the RNAscope results independently confirmed id2b expression in endocardial cells (Figure 2E), supporting the specificity and cell-type localization observed with the reporter line. As suggested by the reviewer, we have updated Figure 2G to display id2b:eGFP and kdrl:mCherry images in the same color scheme as the top panel to improve consistency and clarity. Additionally, we have replaced the images in Figure 2F to avoid overexposure and better represent the spatial distribution of id2b:eGFP in adult heart.

(5) In Fig. 3A: are all the images in panel A taken with the same magnification? In Fig. 3e, could the authors show the localization of klf2 and id2b and confirm their expression in the same endocardial cells? In Fig. 3, the authors conclude that klf2-mediated biomechanical signaling is essential for activating id2b expression. This statement is somewhat overstated because they only demonstrated that knockout of klf2 reduced id2b expression.

We thank the reviewer for these constructive comments. All images presented in Figure 3A were captured using the same magnification, as now clarified in the revised figure legend. We appreciate the reviewer’s question regarding the localization of klf2 and id2b. While we were unable to directly visualize both markers in the same embryos due to the current unavailability of klf2 reporter lines, prior studies using klf2a:H2B-eGFP transgenic zebrafish have demonstrated that klf2a is broadly expressed in endocardial cells, with enhanced expression in the atrioventricular canal region (Heckel et al., Curr Bio 2015, PMID: 25959969; Gálvez-Santisteban et al., Elife 2019, PMID: 31237233). Our id2b:eGFP reporter analysis revealed a similarly broad endocardial expression pattern. These independent observations support the likelihood that klf2a and id2b are co-expressed in the same endocardial cell population.   

We also appreciate the reviewer’s comments regarding the connection between biomechanical signaling and id2b expression. Previous studies have already established that biomechanical cues directly regulate klf2 expression in zebrafish endocardial cells (Vermot et al., Plos Biol 2009, PMID: 19924233; Heckel et al., Curr Bio 2015, PMID: 25959969). In the present study, we observed a significant reduction in id2b expression in both klf2a and klf2b mutants, suggesting that id2b acts downstream of klf2. These observations together establish the role of biomechanical cues-klf2-id2b signaling axis in endocardial cells. Nevertheless, we agree with the reviewer that further investigation is required to elucidate the precise mechanism by which klf2 regulates id2b expression.

(6) In Fig. 4: What's the mRNA expression for id2b in WT and id2b mutant fish hearts?

We performed qRT-PCR analysis on purified zebrafish hearts and observed a significant reduction in id2b mRNA levels in id2b mutants compared to wild-type controls. These new results have been incorporated into the revised manuscript (Figure 4A).

(7) In Fig. 5E, the heart rate shows no difference between id2b+/+ and id2b-/- fish according to echocardiography analysis. However, Fig. 5B indicates a difference in heart rate. Could the authors explain this discrepancy?

We thank the reviewer for this insightful observation. In our study, we observed a reduction in heart rate in id2b mutants during embryonic stages (120 hpf), as shown in Figure 5B. However, this difference was not evident in adult fish based on echocardiography analysis (Figure 5E). While the exact reason for these changes during development remains unclear, it is possible that the reduction in cardiac output observed in id2b mutants during early development triggers compensatory mechanisms over time, ultimately restoring heart rate in adulthood. Given that heart rate is primarily regulated by pacemaker activity, further investigation will be required to determine whether such compensatory adaptations occur and to elucidate the underlying mechanisms.

(8) In Fig. 6A: it's a little hard to read the gene names in the left most image in the panel. In Fig. 6B, the authors conducted qRT-PCR analysis of 72 hpf embryonic hearts and validated decreased nrg1 levels in id2b-/- compared to control. Since nrg1 is not specifically expressed in endocardial cells in the developing heart, the authors should isolate endocardial cells and compare nrg1 expression in id2b-/- to control. This would ensure that the loss of id2b affects nrg1 expression derived from endocardial cells rather than other cell types. In Supp Figure S6: Suggest adding an image of the UMAP projection to show tcf3b expression in endocardial cells from sequencing analysis.

We thank the reviewer for these helpful suggestions. In response, we have increased the font size of gene names in the leftmost panel of Figure 6A to improve readability. Regarding nrg1 expression, we acknowledge the importance of assessing its cell-type specificity. Unfortunately, due to the lack of reliable transgenic or knock-in tools for nrg1, its precise expression pattern in embryonic hearts remains unclear. We attempted to isolate endocardial cells from embryonic hearts using FACS, but the limited number of cells obtained at this stage precluded reliable qRT-PCR analysis. Nonetheless, our data show that id2b is specifically expressed in endocardial cells, and publicly available single-cell RNA-seq datasets also support that nrg1 is predominantly expressed in endocardial, but not myocardial or epicardial cells during embryonic heart development (Figure 6-figure supplement 1). These findings suggest that id2b may regulate nrg1 expression in a cell-autonomous manner within the endocardium. As suggested, we have also added a UMAP image to Figure 7-figure supplement 1 to show tcf3b expression in endocardial cells, further supporting the cell identity in single-cell dataset.

(9) In Fig. 6, Nrg1 knockout shows no gross morphological defects and normal trabeculation in larvae. Could the authors explain why they propose that endocardial id2b promotes nrg1 synthesis, thereby enhancing cardiomyocyte contractile function? Did Nrg1 knockdown with Mo lead to compromised calcium signaling and cardiac contractile function? Nrg2a has been reported to be expressed in endocardial cells in larvae, and its loss leads to heart function defects. Perhaps Nrg2a plays a more important role than Nrg1.

We thank the reviewer for raising this important point. Although we did not directly test nrg1 knockout in our study, previous reports have shown that genetic deletion of nrg1 in zebrafish does not impair cardiac trabeculation during embryonic stages (Rasouli et al., Nat Commun 2017, PMID: 28485381; Brown et al., J Cell Mol Med 2018, PMID: 29265764). However, reduced trabecular area and signs of arrhythmia were observed in juvenile and adult fish (Brown et al., J Cell Mol Med 2018, PMID: 29265764), suggesting a potential role for nrg1 in maintaining cardiac structure and function later in development. Whether calcium signaling and cardiac contractility are affected at these stages remains to be determined. Given that morpholino-induced knockdown is limited to early embryonic stages, it is not suitable for assessing nrg1 function in juvenile or adult hearts.

As noted by the reviewer, nrg2a is expressed in endocardial cells, and its deletion has been associated with cardiac defects (Rasouli et al., Nat Commun 2017, PMID: 28485381). To assess its potential involvement in our model, we performed qRT-PCR analysis and observed increased nrg2a expression in id2b mutant hearts (Author response image 1). This upregulation may reflect a compensatory response to the loss of id2b. Therefore, nrg2a is unlikely to play an essential role in mediating the depressed cardiac function in this context.

Author response image 1.

Expression levels of nrg2a. qRT-PCR analysis of nrg2a mRNA in id2b^+/+ and id2b^-/- adult hearts. Data were normalized to the expression of actb1. N=5 biological replicates, with each sample containing two adult hearts.

(10) In Fig. 7A of the IP experiment, it is recommended that the authors establish a negative control using control IgG corresponding to the primary antibody source. This control helps to differentiate non-specific background signal from specific antibody signal.

As suggested, we have included an IgG control corresponding to the primary antibody species in the immunoprecipitation (IP) experiment to distinguish specific from non-specific binding. The updated data are presented in Figure 7A of the revised manuscript.

(11) In Pg. 5, line 115: there is no reference included for previous literature on blebbistatin.

We have added the corresponding reference (Line 126, Reference #5).

In Pg. 5, lines 118-119; pg. 6 line 144: It would be beneficial to include a short sentence describing why choosing a tnnt2a morpholino knockdown to help provide mechanistic context to readers.

We thank the reviewer for the constructive suggestion. In cardiomyocytes, tnnt2a encodes a sarcomeric protein essential for cardiac contraction, and its knockdown is a well-established method for abolishing heartbeat and blood flow in zebrafish embryos, thereby allowing investigation of flow-dependent gene regulation. In the revised manuscript, we have added a sentence and corresponding reference to clarify the rationale for using tnnt2a morpholino in our study (Lines 128-129, Reference #35).

In Pg. 6, line 140: Results title of "Cardiac contraction promotes endocardial id2b expression through primary cilia but not BMP" is misleading and contradicts the results presented in this section and corresponding figure. For example, the bmp Mo knockdown experiments led to decreased id2b fluorescence and the last statement of this results section contradicts the title that BMP does not promote endocardial id2b in lines 179-180: "Collectively, these results suggest that BMP signaling and blood flow modulate id2b expression in a developmental-stage-dependent manner." It would be helpful to clarify whether BMP signaling is involved in id2b expression or not.

We apologize for any confusion caused by the section title. Our results demonstrate that id2b expression is regulated by both BMP signaling and biomechanical forces in a developmental-stage-specific manner. Specifically, morpholino-mediated knockdown of bmp2b, bmp4, and bmp7a at the 1-cell stage significantly reduced id2b:eGFP fluorescence at 24 hpf (Figure 3-figure supplement 1A, B), suggesting that id2b is responsive to BMP signaling during early embryonic development. However, treatment with the BMP inhibitor Dorsomorphin during later stages (24-48 or 36-60 hpf) did not significantly alter id2b:eGFP fluorescence intensity in individual endocardial cells, although a modest reduction in total endocardial cell number was noted (Figure 3-figure supplement 1C, D). These results suggest that BMP signaling is required for id2b expression during early development but becomes dispensable at later stages, when biomechanical cues may play a more prominent role. To address this concern and better reflect the data, we have revised the Results section title to: "BMP signaling and cardiac contraction regulate id2b expression". This revised title more accurately reflects the dual regulation of id2b expression (Line 153).

In line 205: Any speculation on why the hemodynamics was preserved between id2b mutant and WT siblings at 96 hpf?

As suggested, we have included a sentence to address this observation. “Surprisingly, the pattern of hemodynamics was largely preserved in id2b^-/- embryos compared to id2b^+/+ siblings at 96 hpf (Figure 4-figure supplement 1E, Video 1, 2), suggesting that the reduced number of endocardial cells in the AVC region was not sufficient to induce functional defects.” (Lines 223-225)

In line 246: Fig. 6k and 6j are referenced, but should be figure 5k and 5j.

We have corrected this in the revised manuscript.

Reviewer #2 (Recommendations for the authors):

he manuscript was overall well explained, aside from a few minor points that would help facilitate reader comprehension:

(1) The last paragraph of the introduction could be a brief summary of the study.

We thank the reviewer for this constructive suggestion. As recommended, we have included a paragraph in the Introduction section summarizing our key findings to provide clearer context for the study (Lines 96-100).

(2) Lines 127-128: 'revealed a substantial recapitulation of the... of endogenous id2b expression' may need to be rephrased.

We thank the reviewer for the valuable suggestion. In the revised manuscript, we have changed the sentence to: “Comparison of id2b:eGFP fluorescence with in situ hybridization at 24, 48, and 72 hpf revealed that the reporter signal closely recapitulates the endogenous id2b expression pattern.” (Lines 137-139)

(3) Line 182: '... in a developmental-stage-dependent manner' sounds a bit ambiguous, may need to slightly elaborate/ clarify what this means.

We thank the reviewer for the helpful comment. To improve clarity, we have revised the statement to: “Collectively, these results suggest that id2b expression is regulated by both BMP and biomechanical signaling, with the relative contribution of each pathway varying across developmental stages.” (Lines 195-197)

Reviewer #3 (Recommendations for the authors):

(1) The conclusion that BMP signaling prior to 24 hpf is necessary for id2b expression is not fully supported by the data. How do the authors envision pre-linear heart tube BMP signaling impacting endocardial id2b expression during later chamber stages? Id2b reporter fluorescence can be clearly visualized in the linear heart tube in panel B from Figure 1. Does id2b expression initiate prior to contraction? Can the model be refined by showing when id2b endocardial reporter fluorescence is first observed, and whether this early/pre-contractile expression is dependent on BMP signaling?

We thank the reviewer for the important comment. As suggested, we performed morpholino-mediated knockdown of bmp2b, bmp4, and bmp7a at the 1-cell stage. Live imaging at 24 hpf showed significantly reduced id2b:eGFP fluorescence compared to controls (Figure 3-figure supplement 1A, B), suggesting that id2b is responsive to BMP signaling during early embryonic development. However, treatment with the BMP inhibitor Dorsomorphin during 24-48 or 36-60 hpf did not significantly impact id2b:eGFP fluorescence intensity in individual endocardial cells, although a reduction in endocardial cell number was observed (Figure 3-figure supplement 1C, D). These results suggest that BMP signaling is essential for id2b expression during early embryonic development, while it becomes dispensable at later stages, when biomechanical cues exert a more significant role.

(2) Overexpressing tagged versions of TCF3b and Id2b in HEK293 cells is a very artificial way to make the major claim that these two proteins interact in endogenous endocardial cells. Can this be done in zebrafish embryonic or adult hearts?

We thank the reviewer for this insightful comment. As suggested, we synthesized Flag-id2b and HA-tcf3b mRNA and co-injected them into 1-cell stage zebrafish embryos. We collected 100-300 embryos at 12, 24, and 48 hpf and performed western blot analysis using the same anti-HA and anti-Flag antibodies validated in HEK293 cell experiments. Despite multiple independent attempts, we were unable to detect clear bands of the tagged proteins in zebrafish embryos. We speculate that this could be due to mRNA instability, translational efficiency, or the low abundance of Id2b and Tcf3b proteins. We have acknowledged these technical limitations in the revised manuscript and clarified that the HEK293 cell data support a potential interaction between Id2b and Tcf3b, while confirming their endogenous interaction will require further investigations (Lines 295-296).

(3) The data presented are consistent with the claim that the tcf3b binding sites are functional upstream of nrg1 to repress its transcription. To fully support this idea, those two sites should be disrupted with gRNAs if possible.

We thank the reviewer for the valuable suggestion. In response, we attempted to disrupt the tcf3b binding sites using sgRNAs. However, we encountered technical difficulties in identifying sgRNAs that specifically and efficiently target these binding sites without affecting adjacent regions. Despite these challenges, our luciferase reporter assay, using tcf3b mRNA overexpression and morpholino knockdown, clearly demonstrated that tcf3b binds to and regulates nrg1 promoter region. Nevertheless, we acknowledge that future study using genome editing will be necessary to validate the direct binding of tcf3b to nrg1 promoter.

Minor Points:

(1) Must remove all of the "data not shown" statements and add the primary data to the Supplemental Figures.

As suggested, we have removed all of the “data not shown” statements and added the original data to the revised manuscript (Figure 4E, middle panels, and Figure 4-figure supplement 1F)

(2) Must present the order of the panels in the figure as they are presented in the text. One example is Figure 6 where 6E is discussed in the text before 6C and 6D.

We thank the reviewer for bring up this important point. In the revised manuscript, we have carefully revised the manuscript to ensure that the order of figure panels matches the sequence in which they are discussed in the text. Specifically, we have reorganized the presentation of Figure 6 panels to align with the text flow, discussing panels 6C and 6D before panel 6E. The updated figure and corresponding text have been corrected accordingly in the revised manuscript.

(3) Change the italicized gene names (e.g. tcf3b) to non-italicized names with the first letter capitalized (e.g. Tcf3b) when referencing the protein.

As suggested, we have revised the manuscript to use non-italicized names with the first letter capitalized when referring to proteins.

(4) All bar graphs should be replaced with dot bar graphs.

We have replaced all bar graphs with dot bar graphs throughout the manuscript.

(5) The new id2b mutant allele should be validated as a true null using quantitative RT-PCR to show that the message becomes destabilized through non-sense mediated decay or by immunostaining/western blot analysis if there is a zebrafish Id2b-specific antibody available.

We thank the reviewer for this important suggestion. We have performed qRT-PCR analysis and detected a significant reduction in id2b mRNA levels in id2b^-/- compared to id2b^+/+ controls. These new results are presented in Figure 4A of the revised manuscript.

(6) Was tricaine used to anesthetize embryos for capturing heart rate and percent fractional area change? This analysis should be performed with no or very limited tricaine as it affects heart rate and systolic function. These parameters were captured at 120 hpf, but the authors should also look earlier at 72 hpf at a time when valves are not present by calcium transients are necessary to support heart function.

We thank the reviewer for this important comment. When performing live imaging to assess cardiac contractile function, we used low-dose tricaine (0.16 mg/mL) to anesthetize the zebrafish embryos. We have included this important information in the Methods section (Line 503). As suggested, we have also included the heart function results at 72 hpf, which are now presented in Figure 5-figure supplement 2A-C of the revised manuscript.

(7) The alpha-actinin staining in Figure 5-supplement 2D is very pixelated and unconvincing. This should be repeated and imaged at a higher resolution.

As suggested, we have re-performed the α-actinin staining and acquired higher-resolution images. The updated results are now presented in Figure 5-figure supplement 2G of the revised manuscript.

(8) The authors claim that reductions in id2b mutant heart contractility are due to perturbed calcium transients instead of sarcomere integrity. Why do the authors think that regulation of calcium dynamics was not observed in the DEG enriched GO-terms? Was significant downregulation of cacna1 identified in the bulk RNAseq?

We thank the reviewer for raising this important point. In our bulk RNAseq dataset comparing id2b mutant and control hearts, GO term enrichment was primarily associated with pathways related to cardiac muscle contraction and heart contraction (Figure 5-figure supplement 1B). We speculate that the transcriptional changes related to calcium dynamics may be relatively subtle and thus were not captured as significantly enriched GO terms. In addition, our qRT-PCR analysis revealed a significant reduction in cacna1c expression in id2b mutant hearts compared to controls, suggesting that id2b deletion impairs calcium channel expression. However, this change was not detected by RNA-seq, likely due to limitations in sensitivity.

(9) In line 277, the authors say, "To determine whether this interaction occurs in zebrafish, Flag-id2b and HA-tcf3b were co-expressed in HEK293 cells...". This should be re-phrased to, "To determine if zebrafish Id2b and Tcf3b interact in vitro, Flag-id2b and HA-tcf3b were co-expressed in HEK293 cells for co-immunoprecipitation analysis." The sentence in line 275 should be changed to, "....heterodimer with Tcf3b to limit its function as a potent transcriptional repressor."

We thank the reviewer for these constructive comments and have revised the text accordingly (Lines 291-294).

(10) Small text corrections or ideas:

Line 63: emphasized

We have corrected this in the revised manuscript.

Line 71: studied signaling pathways

We have corrected this in the revised manuscript.

Line 106: the top 6 DEGS (I think that the authors mean top 6 GO-terms) and is Id2b in one of the enriched GO categories?

id2b is one of the top DEGs. This point has been clarified in the revised manuscript (Lines 116-117).

Line 125: a knockin id2b:eGFP reporter line

We have corrected this in the revised manuscript (Line 136).

Line 138: This paragraph could use a conclusion sentence.

We have added a conclusion sentence in the revised manuscript (Lines 150-151).

Line 190: id2b-/- zebrafish experienced early lethality

We have revised the statement as suggested (Line 206).

Line 193: The prominent enlargement of the atrium with a smaller ventricle has characterized as cardiomyopathy in zebrafish (Weeks et al. Cardiovasc Res, 2024, PMID: 38900908), which has also been associated with disruptions in calcium transients (Kamel et al J Cardiovasc Dev Dis, 2021, PMID: 33924051 and Kamel et al, Nat Commun 2021, PMID: 34887420). This information should be included in the text along with these references.

We thank the reviewer for this helpful suggestion. We have incorporated these important references into the revised manuscript and included the relevant information to acknowledge the established link between atrial enlargement, cardiomyopathy, and disrupted calcium transients in zebrafish models (Reference #41, 42, and 45; Lines 210 and 260).

Author Response

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

Reviewer #1 (Public Review):

[...] Weaknesses

Showing that A-2 and especially A-3 are outliers in the PCA analysis is useful, but it may be hiding other interesting signals in the data. The other strains are remarkably colinear on these plots, hinting that if the outliers were removed, one main component would emerge along which they are situated. It also seems possible that this additional analysis step would allow the second dimension to better differentiate them in a way that is interesting with respect to their mutator status or mutations in key metabolic or regulatory genes.

We thank the reviewer for their positive comments and their constructive feedback on the manuscript. Following reviewer’s recommendation, we performed the PCA analysis on metabolism data after removing A-2 and A-3 data. We have detailed those results below. Consistent with a similar analysis performed on RNA-seq datasets in our previous publication, we find that removing these outliers has only a modest effect on separating mutators from non-mutators. We find that, while the new PC2 separates most mutators from the non-mutators, the separation is rather weak. Moreover, we do not see a similar distinction when looking at metabolic data in the Stationary phase. In the interest of improving the readability of the manuscript, we recommend not including these analysis in the final manuscript. We have presented the data for the reviewer’s benefit in Author response image 1, 2 and 3.

Author response image 1.

Author response image 2.

Author response image 3.

There is a missed opportunity to connect some key results to what is known about LTEE mutations that reduce the activity of pykF (pyruvate kinase I). This gene is mutated in all 12 LTEE populations, and often these mutations are frameshifts or transposon insertions that should completely knock out its activity. At first glance, inactivating an enzyme for a step in glycolysis does not make sense when the nutrient source in the growth medium is glucose, even though PykF is only one of two isozymes E. coli encodes for this reaction. There has been speculation that inactivating pykF increases the concentration of phosphoenolpyruvate (PEP) in cells and that this can lead to increased rates of glucose import because PEP is used by the phosphotransferase system of E. coli to import glucose (see https://doi.org/10.1002/bies.20629). The current study has confirmed the higher PEP levels, which is consistent with this model.

We thank the reviewer for pointing out this missed opportunity. We have expanded the discussion around the role of pykF mutations and the elevated concentrations of PEP observed in our data in section 3.4.

In the introduction, the papers cited to show the importance of changes in metabolism for adaptation do not seem to fit the focus of this study very well. They stress production of toxins and secondary metabolites, which do not seem to be mechanisms that are at work in the LTEE. I can think of two areas of background that would be more relevant: (1) studies of how bacterial metabolism evolves in adaptive laboratory evolution (ALE) experiments to optimize metabolic fluxes toward biomass production (for example, https://doi.org/10.1038/nature01149), and (2) discussions of how cross-feeding, metabolic niche specialization, and metabolic interdependence evolve in microbial communities, including in other evolution experiments (for example, https://doi.org/10.1073/pnas.0708504105 and https://doi.org/10.1128/mBio.00036-12).

We thank the reviewer for pointing out missed citations in our introduction. We agree that these papers are relevant to the topic and have added their citations. Additionally, following the suggestion of another reviewer, we have reorganized the introduction so that the concept of the role of metabolism in evolution is presented first and the LTEE second.

Reviewer #2 (Public Review):

[...] Overall, this is a significant and well-executed research study. It offers new insights into the complex relationship between genetic changes and observable traits in evolving populations and utilizes metabolomics in the LTEE, a novel approach in combination with RNA-seq and mutation datasets.

However, the paper's overall clarity is lacking. It is spread too thin and covers many topics without a clear focus. I strongly recommend a substantial rewrite of the manuscript, emphasizing structure and readability. The science is well executed, but the current writing does not do it justice.

We thank the reviewer for their positive comments and their constructive feedback on the lack of clarity in writing. Following the reviewer’s suggestions, we have rewritten parts of the manuscript and reorganizd a few sections to improve readability. We hope the revised manuscript is significantly improved.

Recommendations for the authors

Reviewer #1 (Recommendations For The Authors):

1) Title and Abstract: Add the study organism to the abstract, and probably also the title. Currently, E. coli is not mentioned in either! I'm also not sure that the LTEE is a sufficiently well-known acronym to abbreviate this in the title.

We have revised the title of the manuscript and now spell out LTEE and included E. coli in the title and the abstract.

2) Abstract: I would switch the usage of metabolome to metabolism in a few more places. For example, "changes in its metabolism", "networked and convoluted nature of metabolism". The metabolome, the concentrations of all metabolites, is what is being measured, but I think of this as a phenotypic readout of how metabolism evolving.

We have changed “metabolome” to “metabolism” in cases where we refer to what is evolving and use “metabolome” when we refer to what is being measured.

3) Line 16: Technically, the 12 LTEE populations were not initially identical. The Ara- differed from the Ara+ ancestors by one intentional mutation and one unintentional mutation that was not discovered until whole genomes were sequenced. I would rephrase this to "where 12 replicate populations of E. coli are propagated" or something similar so that it can be correct without needing to describe this unnecessary detail.

The line has been rephrased as suggested.

4) General Note: The text refers to populations as Ara-3 but the figures use A-3. I'd suggest going with A-3 and similar throughout for consistency.

Instances of Ara have been changed to A+/-, and a sentence specifying as such has been added to the intro to make mention of this.

5) Lines 43-44, 97-98. My understanding is that both S and L ecotypes in A-2 can use both glucose and acetate, but that the differentiation is related to their specialization that leads to each one being better on one or the other nutrient. The descriptions make it sound like each grows at a different time. Also, by definition, cells are not growing during "stationary phase". The change from glucose utilization (and acetate secretion) to acetate utilization during one cycle of growth is better described as a diauxic shift.

We have reworded this part to remove mention of “growth” during stationary phase and changed the wording such that it no longer sounds like they grow at different times.

6) Line 54: The statement "provide the ability to test hypotheses from previous data" is vague. Either provide an example or delete.

We have removed this sentence as suggested.

7) Lines 71-72: The terms "interphase" and "intraphase" sound too much like parts of the cell cycle. I'd suggest describing the comparisons as between and within growth phases.

The use of intra and interphase have been changed as suggested.

8) Line 79: The citrate is presumably still a chelating agent, so change phrasing to "Citrate is present in the medium because it was originally added as a chelating agent" or something similar.

This sentence has been rewritten as suggested.

9) Line 83: Write out "mutation accumulations" so it is easier to understand as "the number of mutations that have accumulated".

The phrase has been changed as suggested.

10) Line 116: It's unclear whether the abundances of metabolites are "strategies of survival" in stationary phase. An equally valid explanation is that there is less selection on the metabolome to have a specific composition during stationary phase to have high fitness.

We have added a line about the possibility for alternative hypotheses.

11) Figure 1: There seems to be some information missing from the legend. What are R06 and R07 in Panels A and B? Is panel D exponential phase and panel E stationary phase?

This information was inadvertently missing from the caption and has been added.

12) Figures 2 and 3: Gene names should be in italics. To me, the gray for deleted genes is hard to tell apart from the blue/red. Perhaps you could put a little X in these boxes instead? I think that having a little triangle pointing from each gene or metabolite name its corresponding abundance panel would help the reader track which information goes with which features. In Fig. 3 the placement of L-aspartate is a bit awkward. I'd suggest moving it down so the dashed line does not have to go through the abundance panel.

These figures have been edited to include small triangles that link a gene or metabolite and its heatmap. Additionally, an X has been added where genes have suffered inactivating mutations and the placement of some elements has been moved to improve overall clarity.

13) Lines 183-185: It would be easier to see and judge the consistency of these argR related relationships if a correlation graph of some kind was shown, probably as a supplemental figure. This plot could, for example, have genes/metabolites across the x-axis and fold-change on the y-axis with lines connecting points corresponding to each of the twelve populations across these categories (like Fig S8 but with lines added). Alternatively, it could be a heat map with the populations across one axis and the genes/metabolites across the other axis (like Fig S3).

We have added a supplementary figure consisting of heatmaps showing the consistency of these changes within an evolved line. It is now figure S9.

14) Line 195: I think adding a sentence elaborating on what exactly mutation accumulation means in this context would be helpful to readers.

We have attempted to clarify the meaning of this by specifically stating that it is due to the accumulation of deleterious mutations.

15) Line 293: Is standard LTEE medium DM25? These omics experiments with the LTEE sometimes use similar media with different glucose concentrations, and this is a very important detail to precisely specify.

We reference “standard” LTEE medium in the methods section and have additionally specified the amount of sugar to make it clear that we are not supplementing the media with additional sugar.

16) Figure S8B. Is "cystine" used instead of "cysteine" on purpose here since the compound is oxidized in the metabolomics treatment?

The use of cystine is intentional, we detect the oxidized compound.

Reviewer #2 (Recommendations For The Authors):

Title:

The abbreviation "LTEE" should not be in the title. Most readers will not recognize what it means. Instead, either the full name of the experiment, "Long-Term Evolution Experiment with E. coli," should be used, or the title should be rephrased to "Linking genotypic and phenotypic changes during a long-term evolution experiment using metabolomics."

We have spelled out LTEE and included E. coli in the title.

Abstract:

Sentence 1: Consider softening the statement: "Do changes in an organism's environment, genome, or gene expression patterns often lead to changes in its metabolome?"

We have rephrased this sentence to “Changes in an organism's environment, genome, or gene expression patterns can lead to changes in its metabolism”.

Sentence 4: Use a hyphen for "Long-Term."

This addition has been made.

Sentence 4: Replace "transduce" with a more appropriate term: "...how the effects of mutations can be distributed through a cellular network to eventually affect metabolism and fitness."

We have rewritten this sentence as “to understand how mutations can eventually affect metabolism and perhaps fitness”.

Sentence 5: Clarify the use of "both" to refer to the ancestor of the LTEE and its descendant populations as two classes.

We have reworded this sentence so it’s clear that the ancestors and evolved lines are two separate classes “We used mass-spectrometry to broadly survey the metabolomes of the ancestral strains and all 12 evolved lines…”.

Sentence 6: Reverse the order for better emphasis: "Our work provides a better understanding of how mutations might affect fitness through the metabolome in the LTEE, and thus provides a major step in developing a complete genotype-phenotype map for this experimental system."

We have rearranged this sentence per the reviewers suggestion.

Introduction:

Revise the introduction for clarity, readability, and logical narrative progression. Start with the second paragraph to set up the basic scientific principles being studied and then transition to describing the LTEE as a model system to examine those principles.

The introduction has been rearranged and reworded in parts to increase clarity.

Sentence 1: Revise for clarity: "The Long-Term Evolution Experiment (LTEE) has studied 12 initially identical populations of Escherichia coli as they have evolved in a carbon-limited, minimal glucose medium under a daily serial transfer regime."

Sentence 2: Suggestion: "Begun in 1988, the LTEE populations have evolved for more than 75,000 generations, making it the longest-running experiment of its kind."

Paragraph 2, sentence 2: Italicize "Drosophila."

Paragraph 3, sentence 2: Make an important distinction: "Ara-3 is unique in that it evolved the ability to grow aerobically on citrate."

Paragraph 3, sentence 4: Introduce the IS-mediated loss of the rbs operon in the LTEE as if it has not been described elsewhere.

These suggestions have been incorporated into the manuscript.

Results:

Section 3.1: The use of samples from hours 2 and 24 to represent exponential and stationary phase may present some issues. For instance, capturing Ara-3 during its exponential growth on glucose, but not citrate, at hour 2. Furthermore, except for Ara-3, the LTEE populations reach stationary phase after approximately 4 hours, and there could be significant differences between early, mid, and late stationary phase. This possibility should be acknowledged, and future follow-up work should consider exploring these differences.

We have added sentences in the first paragraph of the results section to include these details. We have also added a short paragraph to the conclusions suggesting additional studies of stationary phase, citing work on evolution of E. coli during long term stationary phase.

Paragraph 3: While Turner et al. 2017 is an essential reference regarding resource use differences between Ara-3 and other LTEE populations, it would be more suitable to reference Blount et al. 2012 for the mutations that enabled access to citrate. Also, it is important to note that the difference lies in the ability to grow aerobically on citrate, rather than the ability to metabolize it.

This citation has been added.

Paragraph 4: As mentioned elsewhere, most LTEE populations exhibit balanced polymorphisms. Therefore, it is more appropriate to state that Ara-2 is the best-understood example of long-term diversity. It is likely that there are important metabolic differences between co-existing lineages in other LTEE populations.

We now refer to Ara-2 as being the best-understood example of long term diversity..

Paragraph 5: The first sentence of this paragraph should likely end with "levels."

The word “levels” was added to the end of this sentence.

Figure 3: It is preferable to refer to the "Superpathway of arginine and polyamine biosynthesis," citing EcoCyc as a reference, rather than a descriptor.

This has been changed to a reference.

Section 3.3, Paragraph 3: While higher intracellular amino acid abundances may facilitate higher translation rates and faster growth, the higher abundances themselves do not evaluate the hypothesis. To evaluate the hypothesis, it is necessary to demonstrate that higher abundances are associated with higher translation or growth rates. Therefore, the final sentence of this paragraph is not meaningful.

We have reworded this sentence to say that it’s not possible to tell what the additional amino acids are being used for given only this data and that additional experiments are needed to confirm this hypothesis.

Section 3.4: The first paragraph of this section misstates how evolution works. The low level of glucose in the LTEE does not drive innovation; instead, innovation occurs at random through the introduction of variation by mutation. Although the existence of the citrate resource acts as a reward that selects for variation that provides access to it, it is essential to remember that evolution is blind to such a reward. Moreover, regarding the evolution of the Cit+ trait, it is incorrect to assert that low glucose contributed to its evolution. As shown by Quandt et al. (2015), it seems probable that Cit+ evolution was potentiated by adaptation to specialization on acetate, which is produced by overflow metabolism resulting from rapid growth on glucose. This rapid growth only occurs when glucose is relatively abundant. The level of glucose seems low to us because it is low relative to traditional levels in bacteriological media, but not to the bacteria.

We agree that this is a semantical, but important distinction. We have reworded this part as to not suggest that evolution has any forward thinking properties and is indeed blind to any rewards that might occur as the result of adaptation.

In general, all instances of "utilize" and its cognates should be replaced with "use" and its cognates.

Instances of “utilize” have been changed to use and its cognates.

There is some uncertainty about the expectation of ramping up the TCA cycle in the LTEE. Overflow metabolism and acetate production appear to be prevalent in the LTEE, suggesting that many lineages only partially oxidize carbon derived from glucose, thereby bypassing the TCA cycle. While it is possible that this interpretation is incorrect, it would be helpful to see it addressed in the manuscript.

We agree that this is a plausible hypothesis, we have added a paragraph at the end of this section that discusses the implications of overflow metabolism as an alternative hypothesis.

Author Response

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

Reviewer #1 (Public Review):

Several concerns are raised from the current study.

1) Previous studies showed that iTregs generated in vitro from culturing naïve T cells with TGF-b are intrinsically unstable and prone to losing Foxp3 expression due to lack of DNA demethylation in the enhancer region of the Foxp3 locus (Polansky JK et al, Eur J Immunol., 2008, PMID: 18493985). It is known that removing TGF-b from the culture media leads to rapid loss of Foxp3 expression. In the current study, TGF-b was not added to the media during iTreg restimulation, therefore, the primary cause for iTreg instability should be the lack of the positive signal provided by TGF-b. NFAT signal is secondary at best in this culturing condition.

In restimulation, void of TGFb is necessary to cause iTreg instability. Otherwise, the setup is similar to the iTreg-inducing environment (Author response image 1). On the other hand, the ultimate goal of this study is to provide a scenario that bears some resemblance of clinical treatment, where TGFb may not be available. The reviewer is correct in stating that TGFb is essential for iTreg stability, we are studying the role played by NFAT in iTreg instability in vitro, and possibly in potential clinical use of iTreg .

Author response image 1.

Restimulation with TGFb will persist iTreg inducing environment, resulting in less pronounced instability. Sorted Foxp3-GFP+ iTregs were rested for 1d, and then rested or restimulated in the presence of TGF-β for 2 d. Percentages of Foxp3+ cells were analyzed by intracellular staining of Foxp3 after 2 d.

2) It is not clear whether the NFAT pathway is unique in accelerating the loss of Foxp3 expression upon iTreg restimulation. It is also possible that enhancing T cell activation in general could promote iTreg instability. The authors could explore blocking T cell activation by inhibiting other critical pathways, such as NF-kb and c-Jun/c-Fos, to see if a similar effect could be achieved compared to CsA treatment.

We thank the reviewer for this suggestion. We performed this experiment according to see extent of the role that NFAT plays, or whether other major pathways are involved. As Author response image 2 shows, solely inhibiting NFAT effectively rescued the instability of iTreg. The inhibition of NFkB (BAY 11-7082), c-Jun (SP600125), or a c-Jun/c-Fos complex (T5224) had no discernable effect, or in one case, possibly further reduction in stability. These results may indicate that NFAT plays a crucial and special role in TCR activation, which leads to iTreg instability. Other pathways, as far as how this experiment is designed, do not appear to be significantly involved.

Author response image 2.

Comparing effects of NFAT, NF-kB and c-Jun/c-Fos inhibitors on iTreg instability. Sorted Foxp3-GFP+ iTregs were rested for 1d, then restimulated by anti-CD3 and CD28 in the presence of listed inhibitors. Percentages of Foxp3+ cells were analyzed by intracellular staining after 2d restimulation.

3) The authors linked chromatin accessibility and increased expression of T helper cell genes to the loss of Foxp3 expression and iTreg instability. However, it is not clear how the former can lead to the latter. It is also not clear whether NFAT binds directly to the Foxp3 locus in the restimulated iTregs and inhibits Foxp3 expression.

T helper gene activation is likely to cause instability in iTregs by secreting more inflammatory cytokines, as shown in Figure Q9, for example, IL-21 secretion. Further investigation is needed to understand how these genes contribute to Foxp3 gene instability exactly. With our limited insight, there may be two possibilities. 1. IL-21 directly affects Foxp3 through its impact on certain inflammation-related transcription factors (TFs). 2. There could be an indirect relationship where NFAT has a greater tendency to bind to those inflammatory TFs when iTreg instability appears, promoting the upregulation of these Th genes like in activated T cells, while being less likely to bind to SMAD and Foxp3, representing a competitive behavior. We at the moment cannot comprehend the intricacies that lead to the differential effects on T helper genes and Treg related genes.

With that said, we have previously attempted to explore the direct effect of NFAT on Foxp3 gene locus. Foxp3 transcription in iTregs primarily relies on histone modifications such as H3K4me3 (Tone et al., 2008; Lu et al., 2011) rather than DNA demethylation (Ohkura et al., 2012; Hilbrands et al., 2016). Previous studies have reported that NFAT and SMAD3 can together promote the histone acetylation of Foxp3 genes (Tone et al., 2008). In our previous set of experiments, we simultaneously obtained information of NFAT binding sites and H3K4me3. In Foxp3 locus, we observed a decreasing trend in NFAT binding to the CNS3 region of Foxp3 in restimulated iTregs compared to resting iTregs (Author response image 3). Additionally, the H3K4me3 modification in the CNS3 region of Foxp3 decreased upon iTreg restimulation, but inhibiting NFAT nuclear translocation with CsA could maintain this modification at its original level (Author response image 3).

Author response image 3.

The NFAT binding and histone modification on Foxp3 gene locus. Genome track visualization of NFAT binding profiles and H3K4me3 profiles in Foxp3 CNS3 locus in two batches of dataset.

Based on these preliminary explorations, it is concluded that NFAT can directly bind to the Foxp3 locus, and it appears that NFAT decreases upon restimulation, resulting in a decrease in H3K4me3, ultimately leading to the close association of NFAT and Foxp3 instability. However, due to limited sample replicates, these data need to be verified for more solid conclusions. We speculate that during the induction of iTregs, NFAT may recruit histone-modifying enzymes to open the Foxp3 CNS3 region, and this effect is synergistic with SMAD. When instability occurs upon restimulation, NFAT binding to Foxp3 weakens due to the absence of SMAD's assistance, subsequently reducing the recruitment of histone modifications enzyme and ultimately inhibiting Foxp3 transcription.

Reviewer #2 (Public Review):

(1) Some concerns about data processing and statistic analysis.

The authors did not provide sufficient information on statistical data analysis; e.g. lack of detailed descriptions about

-the precise numbers of technical/biological replicates of each experiment

-the method of how the authors analyze data of multiple comparisons... Student t-test alone is generally insufficient to compare multiple groups; e.g. figure 1.

These inappropriate data handlings are ruining the evidence level of the precious findings.

We thank the reviewer for pointing out this important aspect. In the figure legend, numbers of independently-performed experiment repeats are shown as N, biological replicates of each experiment as n. Student’s t test was used for comparing statistical significance between two groups. In this manuscript, all calculations of significant differences were based on comparisons between two groups. There were no multiple conditions compared simultaneously within a single group, and thus, no other calculation methods were used.

(2) Untransparent data production; e.g. the method of Motif enrichment analysis was not provided. Thus, we should wait for the author's correction to fully evaluate the significance and reliability of the present study.

Per this reviewer’s request, we have provided detailed descriptions of the data analysis for Fig 5, including both the method section and the Figure legend, as presented below:

“The peaks annotations were performed with the “annotatePeak” function in the R package ChIPseeker (Yu et al, 2015).

The plot of Cut&Tag signals over a set of genomic regions were calculated by using “computeMatrix” function in deepTools and plotted by using “plotHeatmap” and “plotProfile” functions in deepTools. The motif enrichment analysis was performed by using the "findMotifsGenome.pl" command in HOMER with default parameters.

The motif occurrences in each peak were identified by using FIMO (MEME suite v5.0.4) with the following settings: a first-order Markov background model, a P value cutoff of 10-4, and PWMs from the mouse HOCOMOCO motif database (v11).”

Additionally, we have also supplemented the method section with further details on the analysis of RNA-seq and ATAC-seq data.

(3) Lack of evidence in human cells. I wonder whether human PBMC-derived iTreg cells are similarly regulated.

This is a rather complicated issue, human T cells express FoxP3 upon TCR stimulation (PNAS, 103(17): 6659–6664), whose function is likely to protect T cells from activation induced cell death, and does not offer Treg like properties. In contrast in mice, FoxP3 can be used as an indicator of Treg. Currently, this is not a definitive marker for Treg in human, our FoxP3 based readouts do not apply. Nevertheless, we have now investigated whether inhibiting calcium signaling or NFAT could enhance the stability of human iTreg. As shown in Author response image 4, we found that the proportion of Foxp3-expressing cells did not show significant changes across the different conditions, while the MFI analysis revealed that CsA-treated iTreg exhibited higher Foxp3 expression levels compared to both restimulated iTreg and rest iTreg. However, CM4620 had no significant effect on Foxp3 stability, consistent with the observation of its limited efficacy in suppressing human iTreg long term activation. In summary, our results suggest that inhibiting NFAT signaling through CsA treatment can help maintain higher levels of Foxp3 expression in human iTreg.

Author response image 4.

Effect of inhibiting NFAT and calcium on human iTreg stability. Human naïve CD4 cells from PBMC were subjected to a two-week induction process to generate human iTreg. Subsequently, human iTreg were restimulated for 2 days with dynabeads followed by 2 days of rest in the prescence of CsA and CM-4620. Four days later, percentages of Foxp3+ cells and Foxp3 mean fluorescence intensity (MFI) were analyzed by intracellular staining.

(4) NFAT regulation did not explain all of the differences between iTregs and nTregs, as the authors mentioned as a limitation. Also, it is still an open question whether NFAT can directly modulate the chromatin configuration on the effector-type gene loci, or whether NFAT exploits pre-existing open chromatin due to the incomplete conversion of Treg-type chromatin landscape in iTreg cells. The authors did not fully demonstrate that the distinct pattern of chromatin regional accessibility found in iTreg cells is the direct cause of an effector-type gene expression.

To our surprise, the inhibition of NFkB (BAY 11-7082), c-Jun (SP600125), and the c-Jun/c-Fos complex (T5224) resulted in minimal alterations, as shown in Fig Q1. This seems to argue that NFAT may play a more special role in events leading iTreg instability.

We hypothesize that NFAT takes advantage of pre-existing open chromatin state due to the incomplete conversion of chromatin landscape in iTreg cells. Because iTreg cells, after induction, already exhibit inherent chromatin instability, with highly-open inflammatory genes. Furthermore, when iTreg cells were restimulated, the subsequent change in chromatin accessibility was relatively limited and not rescued by NFAT inhibitor treatment (Author response image 5). Therefore, in the case of iTreg cells, we propose that NFAT exploits the easy access of those inflammatory genes, leading to rapid destabilization of iTreg cells in the short term.

In contrast, tTreg cells possess a relatively stable chromatin structure in the beginning, it would be interesting to investigate whether NFAT or calcium signaling could disrupt chromatin accessibility during the activation or expansion of tTreg cells. It is possible that NFAT might cause the loss of the originally established demethylation map and open up inflammatory loci, thereby inducing a shift in gene transcriptional profiles, equally leading to instability.

Author response image 5.

Chromatin accessibility of Rest, Retimulated, CsA/ORAIinh treated restimulated iTreg. PCA visualization of chromatin accessibility profiles of different cell types. Color indicates cell type.

To establish a direct relationship between gene locus accessibility and its overexpression, a controlled experimental approach can be employed. One such method involves precise manipulation of the accessibility of a specific genomic locus using CRISPR-mediated epigenetic modifications at targeted loci. Subsequently, the impact of this manipulation on the expression level of the target gene can be precisely examined. By conducting these experiments, it will be possible to determine whether the augmented gene accessibility directly causes the observed gene overexpression.

Reviewer #1 (Recommendations For The Authors):

1) It might be helpful to add TGF-b to the iTreg restimulation culture to remove the influence of the lack of TGF-b from the equation, and measure the influence of SOCE/NFAT on iTreg instability.

Please refer to Author response image 1.

2) Alternatively, authors can also culture iTreg cells with TGF-b for 2 weeks when they undergo epigenetic changes and become more stabilized (Polansky JK et al, Eur J Immunol., 2008, PMID: 18493985). At this point, the stabilized iTregs can be used to measure the influence of SOCE/NFAT on iTreg instability.

In the study conducted by Polansky, it was observed in Figure 1 that prolonged exposure to TGF-β fails to induce stable Foxp3 expression and demethylation of the Treg-specific demethylated region (TSDR). Based on this finding, we could consider exploring alternative approaches to obtain a more stabilized iTreg population. One such approach could be isolating Foxp3+helios-Nrp1- iTreg cells directly from the peripheral in vivo, which are also known as pTregs. Generally, pTreg cells generated in vivo tend to be more stable compared to iTreg cells induced in vitro, and they already exhibit partial demethylation of the Treg signature, as shown in Fig 6C (Polansky JK et al, Eur J Immunol., 2008, PMID: 18493985). Investigating the role of NFAT and calcium signaling in pTreg cells would provide further insights into the additional roles of NFAT in Treg phenotypical transitions, particularly its role in chromatin accessibility.

3) In Figure 3, NFAT binding to the inflammatory genes in iTreg cells was even stronger than in activated T conventional cells. This is possibly due to Tconv cells being stimulated only once while iTregs were restimulated. A fair comparison should be conducted with restimulated activated conventional T cells.

Figure 3 demonstrates the accessibility of inflammatory gene loci, rather than NFAT binding. Comparing restimulated Tconvs with restimulated iTreg cells is indeed a valuable suggestion, as their activation state and polarization in iTreg directions could lead to distinct chromatin accessibility. Although one is activated long term regularly and the other is activated long term under iTreg polarization, it is highly likely that the chromatin state of both activated Tconvs and iTreg cells is highly open, especially in terms of the accessibility of inflammatory genes. This may provide us with a new perspective to understand iTreg cells, but will unlikely affect our central conclusion.

4) In the in vivo experiment in Figure 6, a control condition without OVA immunization should be included as a baseline.

We have performed this experiment in the absence of OVA, as depicted in Author response image 6. In the absence of OVA immunization, both WT-ORAI and DN-ORAI iTreg exhibited substantial stability, although DN-ORAI demonstrated a slightly less stable trend. Upon activation with 40ug and 100ug of OVA, DN-ORAI iTreg demonstrated enhanced stability than WT-ORAI iTreg, maintaining a higher proportion of Foxp3 expression.

Author response image 6.

Stability of DN-ORAI iTreg in vivo with or without OVA immunization. WT-ORAI/DN-ORAI-GFP+-transfected CD45.2+ Foxp3-RFP+ OT-II iTregs were transferred i.v. into CD45.1 mice. Recipients were left or immunized with OVA323-339 in Alum adjuvant. On day 5, mLN were harvested and analyzed for Foxp3 expression by intracellular staining.

Reviewer #2 (Recommendations For The Authors):

Major

Some concerns about the data processing and statistic analysis, as mentioned in the public review. In the figure legend, what does it mean e.g. n=3, N=3? Technical triplicate experiments? Three mice? Independently-performed three experiments? The authors should define it at least in the "Statistical analysis" in the method section otherwise the readers cannot determine the reason why they mainly use SEM for the data description.

Moreover, in some cases, the number of experiments was not sure; e.g., Fig.1B, Fig. 5.

How did the authors analyze data including multiple comparisons? Student t-test alone is generally insufficient to compare multiple groups; e.g. figure 1.

We thank the reviewer for pointing out this omission. Now, in the figure legend, numbers of independently-performed experiment repeats are shown as N, biological replicates of each experiment as n. For Fig. 1B, N=2, and for Fig 5, we have acquired NFAT Cut&Tag data for 2 times, N=2. Student’s t test was used for comparing statistical significance between two groups. In this manuscript, all calculations of significant differences were based on comparisons between two groups. There were no multiple conditions compared simultaneously within a single group, and thus, no other calculation methods were involved apart from the Student's t-test.

In Figure 1A, the difference in suppressiveness seemed subtle. Data collection of multiple doses of Tconv:Treg ratio will enhance the reliability of such kind of analysis.

We have now attempted the suppression assay with varying Treg:Tconv ratios and observed that the suppressive effect of iTreg was more obvious than that of tTreg when co-cultured at a 1:1 ratio with Tconv cells. However, as the cell number of tTreg and iTreg decreased, the inhibitory effects converged.

Author response image 7.

Compare multiple dose of Tconv:Treg ratio in suppression function CFSE-labelled OT-II T cells were stimulated with OVA-pulsed DC, then different number of Foxp3-GFP+ iTregs and tTregs were added to the culture to suppress the OT-II proliferation. After 4 days, CFSE dilution were analyzed. Left, Representative histograms of CFSE in divided Tconvs. Right, graph for the percentage of divided Tconvs.

In Figure 3F, to which group did the shaded peaks belong? In this context, the authors should focus on "Activation Region" peaks (open chromatin signature in both TcAct & iTreg defined in Fig. 4D) but I did not find the peak in the focusing DNA regions in TcAct (e.g. the shaded regions in IL-4 loci). The clear attribution of the peaks to the heatmap will enhance the visibility and understanding of readers.

We have selected some typical peaks that belong to Fig 3D. These genes encompass some T-cell activation-associated transcription factors, such as Irf4, Atf3, as well as multiple members of the Tnf family including Lta, Tnfsf4, Tnfsf8, and Tnfsf14. Additionally, genes related to inflammation such as Il12rb2, Il9, and Gzmc are included. These genes show elevated accessibility upon T-cell activation, partially open in activated nTreg cells, referred to as the "Activation Region." They collectively exhibit high accessibility in iTreg cells, which may contribute to their instability.

Author response image 8.

Chromatin accessibility of some “Activation Region”. Genomic track showing chromatin accessibility of Irf4, Atf3, Lta, Tnfsf8, Tnfsf4, Tnsfsf14, Il12rb2, Il9, Gzmc in activated Tconv and iTreg.

In Figure 4A/S4A, the information on cell death will help the understanding of readers because the sustained SOCE is associated with cell survival as shown in Fig. S2. The authors can discuss the relationships between cell death and Foxp3 retention, which potentially leads to a further interesting question; e.g. the selective/resistance to activation-induced cell death as the identity of Treg cells.

As shown in Author response image 9, activated iTreg cells indeed exhibit a certain degree of cell death compared to resting iTreg cells. The inhibition of NFAT by CsA enhances the survival rate of iTreg cells, but the inhibition of ORAI by CM-4620 leads to more severe cell death. The cell death induced by CsA and CM-4620 is not consistent, indicating that there may not be a direct proportional relationship between cell death and the expression of Foxp3 and Treg identity.

Author response image 9.

Relationship of cell death and Foxp3 stability in restimulated iTregs. Sorted Foxp3-GFP+ iTregs were rested for 1d, then restimulated by anti-CD3 and CD28 in the presence of CsA or CM-4620. After 2d restimulation, live cell percentage were analyzed by staining of Live/Dead fixable Aqua, and percentages of Foxp3+ cells were analyzed by intracellular staining of Foxp3. Upper, live cell percentage of iTregs. Lower, percentages of Foxp3 in iTregs.

In Figure 5, the information for the data interpretation was insufficient.

We have provided detailed descriptions of the data analysis for Fig 5, including both the method section and the Figure legend, as presented below:

“The peaks annotations were performed with the “annotatePeak” function in the R package ChIPseeker (Yu et al, 2015). The plot of Cut&Tag signals over a set of genomic regions were calculated by using “computeMatrix” function in deepTools and plotted by using “plotHeatmap” and “plotProfile” functions in deepTools. The motif enrichment analysis was performed by using the "findMotifsGenome.pl" command in HOMER with default parameters. The motif occurrences in each peak were identified by using FIMO (MEME suite v5.0.4) with the following settings: a first-order Markov background model, a P value cutoff of 10-4, and PWMs from the mouse HOCOMOCO motif database (v11).”

Additionally, we have also supplemented the method section with further details on the analysis of RNA-seq and ATAC-seq data.

The correlation between the open chromatin status of the gene loci described in Fig.5E and the expression at mRNA level? e.g.; Do iTreg-Act cells produce a higher level of IL-21 than nTreg-act? The analysis in Fig.5F-G should be performed in parallel with nTreg cells to emphasize the distinct NFAT-chromatin regulation in iTreg cells.

We have now compared the secretion levels of IL-21 in tTreg and iTreg upon activation and treated with CsA by ELISA. As shown in Author response image 10, tTreg did not secrete IL-21 regardless of activation status (undetectable), while iTreg did not secrete IL-21 at resting state but exhibited IL-21 secretion after 48 h of activation. Moreover, the secretion of IL-21 was inhibited by CsA and CM-4620 treatment. This observation aligns with our earlier findings where we observed nuclear binding of NFAT to gene loci of these cytokines, enhancing their expression and pushing iTreg unstable under inflammatory conditions. These findings further underscore the likelihood that the inhibition of calcium and NFAT signaling might contribute to the stabilization of iTreg by suppressing the secretion of inflammatory cytokines.

Author response image 10.

IL-21 secretion in tTreg and iTreg upon activation. iTregs and tTregs were sorted and restimulated with anti-CD3 and anti-CD28 antibodies, in the presence of CsA and CM-4620. Cell culture supernatant were harvested after 2 d restimulation and IL-21 secretion was analyzed by ELISA.

Performing a parallel comparison of NFAT activity between tTreg and iTreg cells was initially part of our experimental plan. However, it proved challenging in practice, as we encountered difficulties in efficiently infecting tTreg cells with NFAT-flag. Consequently, we could not obtain a sufficient number of tTreg cells for conducting Cut&Tag experiments.

Based on our observations, we speculate that there might be substantial differences in the accessibility of genes in tTreg cells, leading to considerable variations in the repertoire of genes available for NFAT to regulate. As a result, we expect significant differences in the nuclear localization and activity of NFAT between iTreg and tTreg cells.

In Figure 6C, what does the FCM plot between Foxp3-CFSE look like?

The authors can discuss the mechanism of ORAI-DN-mediated through such analysis; e.g. the possibility that selective proliferation defect by ORAI-DN in Foxp3- cells led to an increased percentage of Foxp3, not only just unstable transcription of Foxp3.

This is an in vitro experiment to assess the suppressive effect of iTreg on Tconv proliferation. Therefore, CFSE is used to stain Tconv cells, but not iTreg cells, so we did not detect proliferation feature of iTreg.

Minor

Confusing terminology of "tTreg" at line 47, etc. "natural Treg" contains both thymic-derived Treg and periphery-derived Treg cells. (A Abbas et al. Nat Immunol. 2013)

We have now changed the designation to tTreg at line 47. tTreg refers to thymus-derived regulatory T cells, while nTreg includes both tTreg and pTreg. However, it is important to note that the Treg cells used in our study were isolated from the spleen of 2-4-month-old Foxp3-GFP or Foxp3-RFP mice. The CD4+ T cells were first enriched using the CD4 Isolation kit, and the FACSAriaII was utilized to collect CD4+ Foxp3-GFP/RFP+ Treg cells. Subsequently, Helios and Nrp-1 staining revealed that the majority of these cells were nTreg, with only approximately 6% being pTreg. Overall, we consider the cells we used as tTreg.

In all FCM analyses, the authors should clarify how to detect Foxp3 expression; Foxp3-GFP/Foxp3-RFP/Intracellular staining like Figure S5A (but not specified in the other FCM plots)

All Foxp3 expressions in the article were assessed using intracellular staining, as described in the methods section, and we have added specific descriptions to each figure legend. The reason for employing intracellular staining is that we used Foxp3-IRES-GFP mice, where GFP and Foxp3 are not fused into a single protein, existing as separate proteins after expression. Therefore, during induction, the appearance of GFP protein might potentially represent the presence of Foxp3. However, in cases of Foxp3 instability, the degradation of GFP protein may not be entirely synchronized with that of Foxp3 protein, making GFP an unreliable indicator of Foxp3 expression levels. As a result, for the purification of pure iTreg cells, we used Foxp3-GFP/RFP fluorescence, while for observing instability, we employed intranuclear staining of Foxp3.

In Figure 6B, the captions were lacking in the two graphs on the right side

The two restimulation conditions, 0.125+0.25 and 0.25+0.5, have been added into Fig 6B right side.

In Figure S2, the annotation of the x-y axis was missing.

Added.

Lack of reference at line 292.

Reference 42-46 were added.

In the method section, the authors should note the further product information of antibodies and reagents to enhance reproducibility and transparency. Making a list that clarifies the suppliers, Ab clone, product IDs, etc. is encouraged. The authors did not specify the supplier of recombinant proteins and which type of TGF-beta (TGF-beta 1, 2, or 3?).

A detailed description of the mice, antibodies, Peptide recombinant protein, commercial kit, and software has been provided and incorporated into the methods section.

In the method section, the authors should clarify which Foxp3-reporter strain. There are many strains of Foxp3-reporter mice in the world. In line 373, is the "FoxP3-IRES-GFP transgenic mice" true? Knock-in strain or BAC-transgene?

This mouse is a gift from Hai Qi Lab in Tsinghua University. They acquired this mouse strain from Jackson Laboratory, and the strain name is B6.Cg-Foxp3tm2Tch/J, Strain #:006772. An IRES-EGFP-SV40 poly A sequence was inserted immediately downstream of the endogenous Foxp3 translational stop codon, but upstream of the endogenous polyA signal, generating a bicistronic locus encoding both Foxp3 and EGFP.

The age of mice used in the experiments should be specified, and confusing words such as "young" should not be used in any method descriptions; e.g. line 405.

The detailed mouse age has been added in the methods section. “To prepare Tconv, tTreg and iTreg for experiments, spleen was isolated from 2-4-month-old Foxp3-GFP mice for Tconv and tTreg sorting, and 6-week-old mice for iTreg induction.”

The method of how the original ATAC-seq/Cut & Tag data were generated was not described in the method section.

Added in method section.

The reference section was incomplete, and the style was not unified. e.g.; ref 7, 24, 25, 26 ... I gave up checking all.

The style of ref 7, 22, 24, 26, 28, 31, 33, 35 were modified.

Changes in manuscript:

Author Name: “Huiyun Lv” to “Huiyun Lyu”.

Fig 1A was updated according to Reviwer 2’s suggestion.

Fig S3E and associated description was added according to Reviwer 2’s suggestion.

Fig S4C and associated description was added according to Reviwer 1’s suggestion.

Fig 5H and associated description was added according to Reviwer 2’s suggestion.

Fig 6D were updated according to Reviwer 1’s suggestion.

Fig 2D was corrected, the labels for gapdh and actin in the iTreg panel were inadvertently switched. The mistake has been rectified, and the original gel image will be provided.

Fig 2A and Fig 4A was updated.

The style of Fig 6B and Fig S2A was modified.

Method:

Mice: FoxP3-IRES-GFP with more description.

Flow Cytometry sorting and FACS: the detailed mouse age has been added. RNA-seq analysis, ATAC-sequencing, ATAC-seq analysis, Cut&Tag assay, Cut&Tag data analysis: more description was added.

Statistical analysis: “Numbers of independently-performed experiment repeats are shown as N, biological replicates of each experiment as n.” were added.

Reference: Ref 42-46 and 49-52 were added. The style of ref 7, 22, 24, 26, 28, 31, 33, 35 were corrected.

A detailed description of the mice, antibodies, Peptide recombinant protein, commercial kit, and software has been provided.

Author Response

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

eLife assessment

This study provides potentially important, new information about the combination of information from the two eyes in humans. The data included frequency tagging of each eye's inputs and measures reflecting both cortical (EEG) and sub-cortical processes (pupillometry). Binocular combination is of potentially general interest because it provides -in essence- a case study of how the brain combines information from different sources and through different circuits. The strength of supporting evidence appears to be solid, showing that temporal modulations are combined differently than spatial modulations, with additional differences between subcortical and cortical pathways. However, the manuscript's clarity could be improved, including by adding more convincing motivations for the approaches used.

We thank the editor and reviewers for their detailed comments and suggestions regarding our paper. We have implemented most of the suggested changes. In doing so we noticed a minor error in our analysis code that affected the functions shown in Figure 2e (previously Figure 1e), and have fixed this and rerun the modelling. Our main results and conclusions are unaffected by this change. We have also added a replication data set to the Appendix, as this bears on one of the points raised by a reviewer, and included a co-author who helped run this experiment.

Reviewer #1 (Public Review):

In this paper, the interocular/binocular combination of temporal luminance modulations is studied. Binocular combination is of broad interest because it provides a remarkable case study of how the brain combines information from different sources. In addition, the mechanisms of binocular combination are of interest to vision scientists because they provide insight into when/where/how information from two eyes is combined.

This study focuses on how luminance flicker is combined across two eyes, extending previous work that focused mainly on spatial modulations. The results appear to show that temporal modulations are combined in different ways, with additional differences between subcortical and cortical pathways.

1. Main concern: subcortical and cortical pathways are assessed in quite different ways. On the one hand, this is a strength of the study (as it relies on unique ways of interrogating each pathway). However, this is also a problem when the results from two approaches are combined - leading to a sort of attribution problem: Are the differences due to actual differences between the cortical and subcortical binocular combinations, or are they perhaps differences due to different methods. For example, the results suggest that the subcortical binocular combination is nonlinear, but it is not clear where this nonlinearity occurs. If this occurs in the final phase that controls pupillary responses, it has quite different implications.

At the very least, this work should clearly discuss the limitations of using different methods to assess subcortical and cortical pathways.

The modelling asserts that the nonlinearity is primarily interocular suppression, and that this is stronger in the subcortical pathway. Moreover the suppression impacts before binocular combination. So this is quite a specific location. We now say more about this in the Discussion, and also suggest that fMRI might avoid the limits on the conclusions we can draw from different methods.

1. Adding to the previous point, the paper needs to be a better job of justifying not only the specific methods but also other details of the study (e.g., why certain parameters were chosen). To illustrate, a semi-positive example: Only page 7 explains why 2Hz modulation was used, while the methods for 2Hz modulation are described in detail on page 3. No justifications are provided for most of the other experimental choices. The paper should be expanded to better explain this area of research to non-experts. A notable strength of this paper is that it should be of interest to those not working in this particular field, but this goal is not achieved if the paper is written for a specialist audience. In particular, the introduction should be expanded to better explain this area of research, the methods should include justifications for important empirical decisions, and the discussion should make the work more accessible again (in addition to addressing the issues raised in point 1 above). The results also need more context. For example, why EEG data have overtones but pupillometry does not?

We now explain the choice of frequency in the final paragraph of the introduction as follows:

‘We chose a primary flicker frequency of 2Hz as a compromise between the low-pass pupil response (see Barrionuevo et al., 2014; Spitschan et al., 2014), and the relatively higher-pass EEG response (Regan, 1966).’

We also mention why the pupil response is low-pass:

‘The pupil response can be modulated by periodic changes in luminance, and is temporally low-pass (Barrionuevo et al., 2014; Spitschan et al. 2014), most likely due to the mechanical limitations of the iris sphincter and dilator muscles’.

Reviewer #2 (Public Review):

Previous studies have extensively explored the rules by which patterned inputs from the two eyes are combined in the visual cortex. Here the authors explore these rules for un-patterned inputs (luminance flicker) at both the level of the cortex, using Steady-State Visual Evoked Potentials (SSVEPs) and at the sub-cortical level using pupillary responses. They find that the pattern of binocular combination differs between cortical and sub-cortical levels with the cortex showing less dichoptic masking and somewhat more binocular facilitation.

Importantly, the present results with flicker differ markedly from those with gratings (Hou et al., 2020, J Neurosci, Baker and Wade 2017 cerebral cortex, Norcia et al, 2000 Nuroreport, Brown et al., 1999, IOVS). When SSVEP responses are measured under dichoptic conditions where each eye is driven with a unique temporal frequency, in the case of grating stimuli, the magnitude of the response in the fixed contrast eye decreases as a function of contrast in the variable contrast eye. Here the response increases by varying (small) magnitudes. The authors favor a view that cortex and perception pool binocular flicker inputs approximately linearly using cells that are largely monocular. The lack of a decrease below the monocular level when modulation strength increase is taken to indicate that previously observed normalization mechanism in pattern vision does not play a substantial role in the processing of flicker. The authors present a computational model of binocular combination that captures features of the data when fit separately to each data set. Because the model has no frequency dependence and is based on scalar quantities, it cannot make joint predictions for the multiple experimental conditions which is one of its limitations.

A strength of the current work is the use of frequency-tagging of both pupil and EEG responses to measure responses for flicker stimuli at two anatomical levels of processing. Flicker responses are interesting but have been relatively neglected. The tagging approach allows one to access responses driven by each eye, even when the other eye is stimulated which is a great strength. The tagging approach can be applied at both levels of processing at the same time when stimulus frequencies are low, which is an advantage as they can be directly compared. The authors demonstrate the versatility of frequency tagging in a novel experimental design which may inspire other uses, both within the present context and others. A disadvantage of the tagging approach for studying sub-cortical dynamics via pupil responses is that it is restricted to low temporal frequencies given the temporal bandwidth of the pupil. The inclusion of a behavioral measure and a model is also a strength, but there are some limitations in the modeling (see below).

The authors suggest in the discussion that luminance flicker may preferentially drive cortical mechanisms that are largely monocular and in the results that they are approximately linear in the dichoptic cross condition (no effect of the fixed contrast stimulus in the other eye). By contrast, prior research using dichoptic dual frequency flickering stimuli has found robust intermodulation (IM) components in the VEP response spectrum (Baitch and Levi, 1988, Vision Res; Stevens et al., 1994 J Ped Ophthal Strab; France and Ver Hoeve, 1994, J Ped Ophthal Strab; Suter et al., 1996 Vis Neurosci). The presence of IM is a direct signature of binocular interaction and suggests that at least under some measurement conditions, binocular luminance combination is "essentially" non-linear, where essential implies a point-like non-linearity such as squaring of excitatory inputs. The two views are in striking contrast. It would thus be useful for the authors could show spectra for the dichoptic, two-frequency conditions to see if non-linear binocular IM components are present.

This is an excellent point, and one that we had not previously appreciated the importance of. We have generated a figure (Fig 8) showing the IM response in the cross frequency conditions. There is a clear response at 0.4Hz in the pupillometry data (2-1.6Hz), and at 3.6Hz in the EEG data (2+1.6Hz). We therefore agree that this shows the system is essentially nonlinear, despite the binocular combination appearing approximately linear. We now say in the Discussion:

‘In the steady-state literature, one hallmark of a nonlinear system is the presence of intermodulation responses at the sums and differences of fundamental flicker frequencies (Baitch & Levi, 1988; Tsai et al., 2012). In Figure 8 we plot the amplitude spectra of conditions from Experiment 1 in which the two eyes were stimulated at different frequencies (2Hz and 1.6Hz) but at the same contrast (48%; these correspond to the binocular cross and dichoptic cross conditions in Figures 2d,e and 3d,e). Consistent with the temporal properties of pupil responses and EEG, Figure 8a reveals a strong intermodulation difference response at 0.4Hz (red dashed line), and Figure 8b reveals an intermodulation sum response at 3.6Hz (red dashed line). The presence of these intermodulation terms is predicted by nonlinear gain control models of the type considered here (Baker and Wade, 2017; Tsai et al., 2012), and indicates that the processing of monocular flicker signals is not fully linear prior to the point at which they are combined across the eyes.’

If the IM components are indeed absent, then there is a question of the generality of the conclusions, given that several previous studies have found them with dichoptic flicker. The previous studies differ from the authors' in terms of larger stimuli and in their use of higher temporal frequencies (e.g. 18/20 Hz, 17/21 Hz, 6/8 Hz). Either retinal area stimulated (periphery vs central field) or stimulus frequency (high vs low) could affect the results and thus the conclusions about the nature of dichoptic flicker processing in cortex. It would be interesting to sort this out as it may point the research in new directions.

This is a great suggestion about retinal area. As chance would have it, we had already collected a replication data set where we stimulated the periphery, and we now include a summary of this data set as an Appendix. In general the results are similar, though we obtain a measurable (though still small) second harmonic response in the pupillometry data with this configuration, which is a further indication of nonlinear processing.

Whether these components are present or absent is of interest in terms of the authors' computational model of binocular combination. It appears that the present model is based on scalar magnitudes, rather than vectors as in Baker and Wade (2017), so it would be silent on this point. The final summation of the separate eye inputs is linear in the model. In the first stage of the model, each eye's input is divided by a weighted input from the other eye. If we take this input as inhibitory, then IM would not emerge from this stage either.

We have performed the modelling using scalar values here for simplicity and transparency, and to make the fitting process computationally feasible (it took several days even done this way). This type of model is quite capable of processing sine waves as inputs, and producing a complex output waveform which is Fourier transformed and then analysed in the same way as the experimental data (see e.g. Tsai, Wade & Norcia, 2012, J Neurosci; Baker & Wade, 2017, Cereb Cortex). However our primary aim here was to fit the model, and make inferences about the parameter values, rather than to use a specific set of parameter values to make predictions. We now say more about this family of models and how they can be applied in the methods section:

“Models from this family can handle both scalar contrast values and continuous waveforms (Tsai et al., 2012) or images (Meese and Summers, 2007) as inputs. For time-varying inputs, the calculations are performed at each time point, and the output waveform can then be analysed using Fourier analysis in the same way as for empirical data.This means that the model can make predictions for the entire Fourier spectrum, including harmonic and intermodulation responses that arise as a consequence of nonlinearities in the model (Baker and Wade, 2017). However for computational tractability, we performed fitting here using scalar contrast values.”

As a side point, there are quite a lot of ways to produce intermodulation terms, meaning they are not as diagnostic as one might suppose. We demonstrate this in Author response image 1, which shows the Fourier spectra produced by a toy model that multiplies its two inputs together (for an interactive python notebook that allows various nonlinearities to be explored, see here). Intermodulation terms also arise when two inputs of different frequencies are summed, followed by exponentiation. So it would be possible to have an entirely linear binocular summation process, followed by squaring, and have this generate IM terms (not that we think this is necessarily what is happening in our experiments).

Author response image 1

Related to the model: One of the more striking results is the substantial difference between the dichoptic and dichoptic-cross conditions. They differ in that the latter has two different frequencies in the two eyes while the former has the same frequency in each eye. As it stands, if fit jointly on the two conditions, the model would make the same prediction for the dichoptic and dichoptic-cross conditions. It would also make the same prediction whether the two eyes were in-phase temporally or in anti-phase temporally. There is no frequency/phase-dependence in the model to explain differences in these cases or to potentially explain different patterns at the different VEP response harmonics. The model also fits independently to each data set which weakens its generality. An interpretation outside of the model framework would thus be helpful for the specific case of differences between the dichoptic and dichoptic-cross conditions.

As mentioned above, the limitations the reviewer highlights are features of the specific implementation, rather than the model architecture in general. Furthermore, although this particular implementation of the model does not have separate channels for different phases, these can be added (see e.g. Georgeson et al., 2016, Vis Res, for an example in the spatial domain). In future work we intend to explore the phase relationship of flicker, but do not have space to do this here.

Prior work has defined several regimes of binocular summation in the VEP (Apkarian et al.,1981 EEG Journal). It would be useful for the authors to relate the use of their terms "facilitation" and "suppression" to these regimes and to justify/clarify differences in usage, when present. Experiment 1, Fig. 3 shows cases where the binocular response is more than twice the monocular response. Here the interpretation is clear: the responses are super-additive and would be classed as involving facilitation in the Apkarian et al framework. In the Apkarian et al framework, a ratio of 2 indicates independence/linearity. Ratios between 1 and 2 indicate sub-additivity and are diagnostic of the presence of binocular interaction but are noted by them to be difficult to interpret mechanistically. This should be discussed. A ratio of <1 indicates frank suppression which is not observed here with flicker.

Operationally, we use facilitation to mean an increase in response relative to a monocular baseline, and suppression to mean a decrease in response. We now state this explicitly in the Introduction. Facilitation greater than a factor of 2 indicates some form of super-additive summation. In the context of the model, we also use the term suppression to indicate divisive suppression between channels, however this feature does not always result in empirical suppression (it depends on the condition, and the inhibitory weight). We think that interpretation of results such as these is greatly aided by the use of a computational modelling framework, which is why we take this approach here. The broad applicability of the model we use in the domain of spatial contrast lends it credibility for our stimuli here.

Can the model explore the full range of binocular/monocular ratios in the Apkarian et al framework? I believe much of the data lies in the "partial summation" regime of Apkarian et al and that the model is mainly exploring this regime and is a way of quantifying varying degrees of partial summation.

Yes, in principle the model can produce the full range of behaviours. When the weight of suppression is 1, binocular and monocular responses are equal. When the weight is zero, the model produces linear summation. When the weight is greater than 1, suppression occurs. It is also possible to produce super-additive summation effects, most straightforwardly by changing the model exponents. However this was not required for our data here, and so we kept these parameters fixed. We agree that the model is a good way to unify the results across disparate experimental paradigms, and that is our main intention with Figure 7i.

Reviewer #3 (Public Review):

This manuscript describes interesting experiments on how information from the two eyes is combined in cortical areas, sub-cortical areas, and perception. The experimental techniques are strong and the results are potentially quite interesting. But the manuscript is poorly written and tries to do too much in too little space. I had a lot of difficulty understanding the various experimental conditions, the complicated results, and the interpretations of those results. I think this is an interesting and useful project so I hope the authors will put in the time to revise the manuscript so that regular readers like myself can better understand what it all means.

Now for my concerns and suggestions:

The experimental conditions are novel and complicated, so readers will not readily grasp what the various conditions are and why they were chosen. For example, in one condition different flicker frequencies were presented to the two eyes (2Hz to one and 1.6Hz to the other) with the flicker amplitude fixed in the eye presented to the lower frequency and the flicker amplitude varied in the eye presented to the higher frequency. This is just one of several conditions that the reader has to understand in order to follow the experimental design. I have a few suggestions to make it easier to follow. First, create a figure showing graphically the various conditions. Second, come up with better names for the various conditions and use those names in clear labels in the data figures and in the appropriate captions. Third, combine the specific methods and results sections for each experiment so that one will have just gone through the relevant methods before moving forward into the results. The authors can keep a general methods section separate, but only for the methods that are general to the whole set of experiments.

We have created a new figure (now Fig 1) that illustrates the conditions from Experiment 1, and is referenced throughout the paper. We have kept the names constant, as they are rooted in a substantial existing literature, and it will be confusing to readers familiar with that work if we diverge from these conventions. We did consider separating out the methods section, but feel it helps the flow of the results section to keep it as a single section.

I wondered why the authors chose the temporal frequencies they did. Barrionuevo et al (2014) showed that the human pupil response is greatest at 1Hz and is nearly a log unit lower at 2Hz (i.e., the change in diameter is nearly a log unit lower; the change in area is nearly 2 log units lower). So why did the authors choose 2Hz for their primary frequency? And why did the authors choose 1.6Hz which is quite close to 2Hz for their off frequency? The rationale behind these important decisions should be made explicit.

We now explain this in the Introduction as follows:

‘We chose a primary flicker frequency of 2Hz as a compromise between the low-pass pupil response (see Barrionuevo et al., 2014; Spitschan et al., 2014), and the relatively higher-pass EEG response (Regan, 1966).’

It is a compromise frequency that is not optimal for either modality, but generates a measurable signal for both. The choice of 1.6 Hz was for similar reasons - for a 10-second trial it is four frequency bins away from the primary frequency, so can be unambiguously isolated in the spectrum.

By the way, I wondered if we know what happens when you present the same flicker frequencies to the two eyes but in counter-phase. The average luminance seen binocularly would always be the same, so if the pupil system is linear, there should be no pupil response to this stimulus. An experiment like this has been done by Flitcroft et al (1992) on accommodation where the two eyes are presented stimuli moving oppositely in optical distance and indeed there was no accommodative response, which strongly suggests linearity.

We have not tried this yet, but it’s on our to-do list for future work. The accommodation work is very interesting, and we now cite it in the manuscript as follows:

‘Work on the accommodative response indicates that binocular combination there is approximately linear (Flitcroft et al. 1992), and can even cancel when signals are in antiphase (we did not try this configuration here).’

Figures 1 and 2 are important figures because they show the pupil and EEG results, respectively. But it's really hard to get your head around what's being shown in the lower row of each figure. The labeling for the conditions is one problem. You have to remember how "binocular" in panel c differs from "binocular cross" in panel d. And how "monocular" in panel d is different than "monocular 1.6Hz" in panel e. Additionally, the colors of the data symbols are not very distinct so it makes it hard to determine which one is which condition. These results are interesting. But they are difficult to digest.

We hope that the new Figure 1 outlining the conditions has helped with interpretation here.

The authors make a strong claim that they have found substantial differences in binocular interaction between cortical and sub-cortical circuits. But when I look at Figures 1 and 2, which are meant to convey this conclusion, I'm struck by how similar the results are. If the authors want to continue to make their claim, they need to spend more time making the case.

Indeed, it is hard to make direct comparisons across figures - this is why Figure 4 plots the ratio of binocular to monocular conditions, and shows a clear divergence between the EEG and pupillometry results at high contrasts.

Figure 5 is thankfully easy to understand and shows a very clear result. These perceptual results deviate dramatically from the essentially winner-take-all results for spatial sinewaves shown by Legge & Rubin (1981); whom they should cite by the way. Thus, very interestingly the binocular combination of temporal variation is quite different than the binocular combination of spatial variation. Can the pupil and EEG results also be plotted in the fashion of Figure 5? You'd pick a criterion pupil (or EEG) change and use it to make such plots.

We now cite Legge & Rubin. We see what you mean about plotting the EEG and pupillometry results in the same coordinates as the matching data, but we don’t think this is especially informative as we would end up only with data points along the axes and diagonal of the plot, without the points at other angles. This is a consequence of how the experiments were conducted.

My main suggestion is that the authors need to devote more space to explaining what they've done, what they've found, and how they interpret the data. I suggest therefore that they drop the computational model altogether so that they can concentrate on the experiments. The model could be presented in a future paper.

We feel that the model is central to the understanding and interpretation of our results, and have retained it in the revised version of the paper.

Reviewer #2 (Recommendations For The Authors):

I found the terms for the stimulus conditions confusing. I think a simple schematic diagram of the conditions would help the reader.

Now added (the new Fig 1).

In reporting the binocular to monocular ratio, please clarify whether the monocular data was from one eye alone (and how that eye was chosen) or from both eyes and then averaged, or something else. It would be useful to plot the results from the dichoptic condition in this form, as well.

These were averaged across both eyes. We now say in the Methods section:

‘We confirmed in additional analyses that the monocular consensual pupil response was complete, justifying our pooling of data across the eyes.’

Also, clarify whether the term facilitation is used as above throughout (facilitation being > 2 times monocular response under binocular condition) or if a different criterion is being used. If we take facilitation to mean a ratio > 2, then facilitation depends on temporal frequency in Figure 4.

We now explain our use of these terms in the final paragraph of the Introduction:

‘Relative to the response to a monocular signal, adding a signal in the other eye can either increase the response (facilitation) or reduce it (suppression).’

The magnitude of explicit facilitation attained is interesting, but not without precedent. Ratios of binocular to mean monocular > 2, have been reported previously and values of summation depend strongly on the stimulus used (see for example Apkarian et al., EEG Journal, 1981, Nicol et al., Doc Ophthal, 2011).

We now mention this in the Discussion as follows:

‘(however we note that facilitation as substantial as ours has been reported in previous EEG work by Apkarian et al. (1981))’

In Experiment 3, the authors say that the psychophysical matching results are consistent with the approximately linear summation effects observed in the EEG data of Experiment 1. In describing Fig. 3, the claim is that the EEG is non-linear, e.g. super-additive - at least at high contrasts. Please reconcile these statements.

We think that the ‘superadditive’ effects are close enough to linear that we don’t want to make too much of a big deal about them - this could be measurement error, for example. So we use terms such as near-linear, or approximately linear, when referring to them throughout.

Reviewer #3 (Recommendations For The Authors):

Let me make some more specific comments using a page/paragraph/line format to indicate where in the text they're relevant.

1/2 (middle)/3 from end. "In addition" seems out of place here.

Removed.

1/3/4. By "intensities" do you mean "contrasts"?

Fixed.

1/3/last. "... eyes'...".

Fixed.

2/5/3. By "one binocular disc", you mean into "one perceptually fused disc".

Rewritten as: ‘to help with their perceptual fusion, giving the appearance of a single binocular disc’

3/1/1. "calibrated" seems like the wrong word here. I think you're just changing the vergence angle to enable fusion, right?

Now rewritten as: ‘Before each experiment, participants adjusted the angle of the stereoscope mirrors to achieve binocular fusion’

3/1/1. "adjusting the angles...". And didn't changing the mirror angles affect the shapes of the discs in the retinal images?

Perhaps very slightly, but this is well within the tolerance of the visual system to compensate for in the fused image, especially for such high contrast edges.

3/3/5. "fixed contrast" is confusing here because it's still a flickering stimulus if I follow the text here. Reword.

Now ‘fixed temporal contrast’

3/4/1. It would be clearer to say "pupil tracker" rather than "eye tracker" because you're not really doing eye tracking.

True, but the device is a commercial eye tracker, so this is the appropriate term regardless of what we are using it for.

3/5/6. I'm getting lost here. "varying contrast levels" applies to the dichoptic stimulus, right?

Yes, now reworded as ‘In the other interval, a target disc was displayed, flickering at different contrast levels on each trial, but with a fixed interocular contrast ratio across the block.’

3/5/7. Understanding the "ratio of flicker amplitudes" is key to understanding what's going on here. More explanation would be helpful.

Addressed in the above point.

4/3/near end. Provide some explanation about why the Fourier approach is more robust to noise.

Added ‘(which can make the phase and amplitude of a fitted sine wave unstable)’

Figure 1. In panel a, explain what the numbers on the ordinate mean. What's zero, for example? Which direction is dilation? Same question for panel b. It's interesting in panel c that the response in one eye to 2Hz increases when the other eye sees 1.6Hz. Would be good to point that out in the text.

Good idea about panel (a) - we have changed the y-axis to ‘Relative amplitude’ for clarity, and now note in the figure caption that ‘Negative values indicate constriction relative to baseline, and positive values indicate dilation.’ Panel (b) is absolute amplitude, so is unsigned. Panel (c) only contains 2Hz conditions, but there is some dichoptic suppression across the two frequencies in panels (d,e) - we now cover this in the text and include statistics.

6/2/1. Make clear in the text that Figure 1c shows contrast response functions for the pupil.

Now noted in the caption.

Figure 3. I'm lost here. I feel like I should be able to construct this figure from Figures 1 and 2, but don't know how. More explanation is needed at least in the caption.

Done. The caption now reads:

‘Ratio of binocular to monocular response for three data types. These were calculated by dividing the binocular response by the monocular response at each contrast level, using the data underlying Figures 2c, 3c and 3f. Each value is the average ratio across N=30 participants, and error bars indicate bootstrapped standard errors.’

9/1/1-2. I didn't find the evidence supporting this statement compelling.

We now point the reader to Figure 4 as a reminder of the evidence for this difference.

9/1/6-9. You said this. But this kind of problem can be fixed by moving the methods sections as I suggested above.

As mentioned, we feel that the results section flows better with the current structure.

Figure 4. Make clear that this is EEG data.

Now added to caption.

Figure 5 caption. Infinite exponent in what equation?

Now clarified as: ‘models involving linear combination (dotted) or a winner-take-all rule (dashed)’

Figure 6. I hope this gets dropped. No one will understand how the model predictions were derived. And those who look at the data and model predictions will surely note (as the authors do) that they are rather different from one another.

As noted above, we feel that the model is central to the paper and have retained this figure. We have also worked out how to correct the noise parameter in the model for the number of participants included in the coherent averaging, which fixes the discrepancy at low contrasts. The correspondence between the data and model in is now very good, and we have plotted the data points and curves in the same panels, which makes the figure less busy.

12/1. Make clear in this paragraph that "visual cortex" is referring to EEG and perception results and that "subcortical" is referring to pupil. Explain clearly what "linear" would be and what the evidence for "non-linear" is.

Good suggestion, we have added qualifiers linking to both methods. Also tidied up the language to make it clearer that we are talking about binocular combination specifically in terms of linearity, and spelled out the evidence for each point.

12/2/6-9. Explain the Quaia et al results enough for the reader to know what reflexive eye movements were studied and how.

We now specify that these eye movements are also known as the ‘ocular following response’ and were measured using scleral search coils.

12/2/9-10. Same for Spitchan and Cajochen: more explanation.

Added:

“(melatonin is a hormone released by the pineal gland that regulates sleep; its production is suppressed by light exposure and can be measured from saliva assays)”

12/3/2-3. Intriguing statements about optimally combining noisy signals, but explain this more. It won't be obvious to most readers.

We have added some more explanation to this section.

13/1. This is an interesting paragraph where the authors have a chance to discuss what would be most advantageous to the organism. They make the standard argument for perception, but basically punt on having an argument for the pupil.

Indeed, we agree that this point is necessarily speculative, however we think it is interesting for the reader to consider.

13/2/1. "Pupil size affects the ..." is more accurate.

Fixed.

13/2/2 from end. Which "two pathways"? Be clear.

Changed to ‘the pupil and perceptual pathways’

Author response:

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

Reviewer #1 (Public review):

(1) The mechanism by which STAMBPL1 mediates GRHL3 transcription through its interaction with FOXO1 is not sufficiently discussed, especially in relation to how STAMBPL1 regulates FOXO1. Some reported effects are modest.

We appreciate the reviewer’s comments. In response, we have added a discussion on the potential mechanisms by which STAMPBL1 regulates FOXO1 transcriptional activity in Discussion, highlighted in red on page 18, lines 342 to 352. The specific reply content is as follows: “The transcriptional activity of FOXO1 is primarily regulated by its nucleocytoplasmic shuttling process (Van Der Heide, Hoekman et al. 2004). The PI3K/AKT pathway promotes the phosphorylation of FOXO1, resulting in the formation of a complex with members of the 14-3-3 family (including 14-3-3σ, 14-3-3ε, and 14-3-3ζ), which facilitates its export from the nucleus and inhibits its transcriptional activity (Huang and Tindall 2007, Tzivion, Dobson et al. 2011). It’s reported that TDAG51 prevents the binding of 14-3-3ζ to FOXO1 in the nucleus by interacting with FOXO1, thereby enhancing its transcriptional activity through increased accumulation within the nucleus (Park, Jeon et al. 2023). Our results indicate that the overexpression of STAMBPL1 and STAMBPL1-E292A did not affect the protein levels of FOXO1 (Fig.7E and Fig.S5E), but STAMBPL1 co-localizes with FOXO1 in the nucleus (Fig.7M) and interacts with it (Fig.7N and Fig.S5I-J). This suggests that STAMBPL1 enhances the transcriptional activity of FOXO1 on GRHL3 by interacting with nuclear FOXO1.” The result was added to Supplementary Figure 5 as Fig.S5E.

Reviewer #2 (Public review):

(1) A potential limitation of the study is the reliance on specific cellular and animal models, which may constrain the extrapolation of these findings to the broader spectrum of human TNBC biology. Furthermore, while the study provides evidence for a novel regulatory axis involving STAMBPL1, FOXO1, and GRHL3, the multifaceted nature of angiogenesis may implicate additional regulatory factors not exhaustively addressed in this research.

We appreciate the valuable suggestions provided by the reviewer. In Discussion, we have added an in-depth discussion of the limitations of the study, as well as an analysis of the regulatory factors related to tumor angiogenesis, which highlighted in red on pages 20 to 21, lines 396 to 412. The relevant content added is as follows: “In this study, we utilized two triple-negative breast cancer cell lines, HCC1806 and HCC1937, along with human primary umbilical vein endothelial cells (HUVECs) and a nude mouse breast orthotopic transplantation tumor model to investigate the regulatory mechanism by which STAMBPL1 activates the GRHL3/HIF1α/VEGFA signaling pathway through its interaction with FOXO1, thereby promoting angiogenesis in TNBC. The results of this study have certain limitations regarding their applicability to human TNBC biology. Furthermore, in addition to the HIF1α/VEGFA signaling pathway emphasized in this study, tumor cells can continuously release or upregulate various pro-angiogenic factors, such as Angiopoietin and FGF, which activate endothelial cells, pericytes (PCs), cancer-associated fibroblasts (CAFs), endothelial progenitor cells (EPCs), and immune cells (ICs). This leads to capillary dilation, basement membrane disruption, extracellular matrix remodeling, pericyte detachment, and endothelial cell differentiation, thereby sustaining a highly active state of angiogenesis (Liu, Chen et al. 2023). It is important to collect clinical TNBC tissue samples in the future to analyze the expression of the STAMBPL1/FOXO1/GRHL3/HIF1α/VEGFA signaling axis. Furthermore, patient-derived organoid and xenograft models are useful to elucidate the regulatory relationship of this axis in TNBC angiogenesis”

Reviewer #3 (Public review):

The main weaknesses of this work are that the relevance of this molecular axis to the pathogenesis of TNBC is not clear, and it is not clearly established whether this is a regulatory pathway that occurs in hypoxic conditions or independently of oxygen levels.

(1) With respect to the first point, both FOXO1 and GRHL3 have been previously described as tumor suppressors, with reports of FOXO1 inhibiting tumor angiogenesis. Therefore, this works describes an apparently contradictory function of these proteins in TNBC. While it is not surprising that the same genes perform divergent functions in different tumor contexts, a stronger evidence in support of the oncogenic function of these two genes should be provided to make the data more convincing. As an example, the data in support of high STAMBPL1, FOXO and GRHL3 gene expression in TNBC TCGA specimens provided in Figure 8 is not very strong and it is not clear what the non-TNBC specimens are (whether other breast cancers or other tumors, perhaps those tumors whether these genes perform tumor suppressive functions). To strengthen the notion that STAMBPL1, FOXO and GRHL3 are overexpressed in TNCB, the authors could provide a comparison with normal tissue, as well as the analysis of other publicly available datasets (like the NCI Clinical Proteomic Tumor Analysis Consortium as an example). Finally, is it not clear what are the basal protein expression levels of STAMBPL1 in the cell lines used in this study, as based on the data presented in Figures 2D and F it appears that the protein is not expressed if not exogenously overexpressed. It would be helpful if the authors addressed this issue and provided further evidence of STAMBPL1 expression in TNBC cell lines.

We appreciate the suggestions. In this study, we utilized the BCIP online tool to analyze the Metabric database, incorporating adjacent normal tissues as controls. Although the expression levels of STAMBPL1, FOXO1, and GRHL3 in breast cancer tissues are not uniformly higher than those in adjacent tissues, their expression levels in triple-negative breast cancer (TNBC) are significantly elevated compared to non-TNBC. The results of this re-analysis have been added in Supplementary Figure 6 as Fig.S6A-C.

About the question of the basal protein expression levels of STAMBPL1 in the cell lines used in this study, our response is that Fig. 2A showed the endogenous level of STAMBPL1 in HCC1806 and HCC1937. For Fig. 2D and 2F, the overexpressed STAMBPL1 was fused with a 3xFlag tag, resulting in a higher molecular weight compared to the endogenous STAMBPL1. In the revised Figure 2, we have indicated the positions of the endogenous (Endo.) and exogenous (OE.) STAMBPL1 bands with arrows.

(2) Linked to these considerations is the second major criticism, namely that it is not made clear if this new regulatory axis is proposed to act in normoxic or hypoxic conditions. The experiments presented in this paper are performed in both conditions but a clear explanation as to why cells are exposed to hypoxia is not given and would be necessary being that HIF-1a transcription and not protein stability is being analyzed. Also, different hypoxic conditions are sometimes used, resulting in different mRNA levels of HIF-1a and its downstream targets and quite significant fluctuations within the same cell line from one experimental setting to the next. The authors should provide an explanation as to why experimental conditions are changed and, more importantly, the experiments presented in Figure 2 should be performed also in normoxia.

Thanks for the comments. Under normoxic conditions, HIF1α is recognized by pVHL due to hydroxylation and is rapidly degraded via the proteasomal pathway. In contrast, under hypoxic conditions, HIF1α protein is accumulated. To investigate the effect of STAMBPL1 knockdown on HIF1A gene transcription levels, we conducted experiments under hypoxic conditions to avoid interference from the rapid degradation of HIF1α at the protein level, as shown in Figures 2B-C. Furthermore, under normoxic conditions, the overexpression of STAMBPL1 had been demonstrated to significantly enhance the protein levels of HIF1α and upregulate the transcription of VEGFA through HIF1α. To avoid the potential impact of excessive accumulation of HIF1α protein under hypoxic conditions on its protein level detection and the transcription of downstream VEGFA, the related experiments shown in Figure 2D-G were performed under normoxic conditions. We have explained the corresponding experimental conditions in the “Result” and “Figure legends” according to the reviewer's comments, highlighted in red.

(3) Another critical point is that necessary experimental controls are sometimes missing, and this is reducing the strength of some of the conclusions enunciated by the authors. As examples, experiments where overexpression of STAMBPL1 is coupled to silencing of FOXO1 to demonstrate dependency lack FOXO1 silencing the absence of STAMBPL1 overexpression. Because diminishing FOXO1 expression affects HIF-1a/VEGF transcription even in the absence of STAMBPL1 (shown in Figure 7C, D), it is not clear if the data presented in Figure 7G are significant. The difference between HIF-1a expression upon FOXO1 silencing should be compared in the presence or absence of STAMBPL1 overexpression to understand if FOXO1 impacts HIF-1a transcription dependently or independently of STAMBPL1.

Thank you for this comment. For Fig.7G-H, our experimental objective was to determine whether the activation of HIF1A/VEGFA transcription by STAMBPL1 via FOXO1. Therefore, under STAMBPL1 overexpression, we knocked down FOXO1 to investigate whether FOXO1 silencing could reverse the upregulation of HIF1A/VEGFA transcription induced by STAMBPL1 overexpression.

(4) In addition, some minor comments to improve the quality of this manuscript are provided.

(4.1) As a general statement, the manuscript is extremely synthetic. While this is not necessarily a negative feature, sometimes results are discussed in the figure legends and not in the main text (as an example, western blots showing HIF-1a expression) and this makes it hard to read thought the data in an easy and enjoyable manner.

Thank you for this suggestion. We have revised the figure legends to make them clearer and more concise, highlighted in red.

(4.2) The effect of STAMBPL1 overexpression on HIF-1a transcription is minor (Figure 2) The authors should explain why they think this is the case and whether hypoxia may provide a molecular environment that is more permissive to this type of regulation.

Thank you for the comment. Under normoxic conditions, we conducted WB to examine the protein expression of HIF1α after the overexpression of STAMBPL1 and the knockdown of HIF1α. To visually illustrate the impact of STAMBPL1 overexpression on HIF1A protein levels, as well as the effectiveness of HIF1α knockdown, we annotated the grayscale analysis results of the bands in Figures 2D and 2F. As the reviewer pointed out, under normoxic conditions, HIF1α is rapidly degraded, which may explain why the upregulation of HIF1α protein levels by STAMBPL1 overexpression is not very pronounced.

(4.3) HIF-1a does not appear upregulated at the protein level protein by STAMBPL1 or GRLH3 overexpression, even though this is stated in the legends of Figures 2 and 6. The authors should show unsaturated western blots images and provide quantitative data of independent experiments to make this point.

Thank you for this comment. We have added the unsaturated image of HIF1α into Fig.2D, and performed a grayscale analysis of the HIF1α bands in Fig.2D and Fig.6A to indicate the relative protein level of HIF1α.

Reviewer #1 (Recommendations for the authors):

(1) The authors previously reported that STAMBPL1 stabilizes MKP1 in TNBC. However, in this study, they focus on HIF1a. Given that STAMBPL1 affects HIF1a expression, it would be valuable to examine the levels of ROS in TNBC cells with or without STAMBPL1, as ROS is known to influence HIF1a stability.

Thank you for your comments. It’s known that STAMBPL1 functions as a deubiquitinating enzyme. However, our study reveals that the upregulation of HIF1α by STAMBPL1 is independent of its deubiquitinating activity. This conclusion is supported by the observation that overexpression of the deubiquitinase active site mutant, STAMBPL1-E292A, also upregulated HIF1α expression (Figure 1F). Moreover, STAMBPL1 overexpression enhanced HIF1α transcription (Figures 4E and S3E), while STAMBPL1 knockdown was able to inhibit the transcription of HIF1α (Figures 2B-C). These results indicate that STAMBPL1 mediates the transcription of HIF1α but does not affect the stability of HIF1α. For these reasons, we think that it is unnecessary to examine the ROS levels.

(2) Figure 1A: The regulation of HIF1a mRNA by STAMBPL1, but not its protein levels, could be better addressed by using MG132 to rule out the impact of protein degradation.

Thanks for this comment. Under normoxic conditions, the oxygen-sensitive prolyl hydroxylases PHD1-3 act on HIF1α, specifically inducing hydroxylation at the proline 402 and 564 residues. These hydroxylated residues are recognized by the pVHL/E3 ubiquitin ligase complex, leading to ubiquitination and subsequent degradation via the proteasome pathway. Conversely, under hypoxic conditions, PHD1-3 are inactivated, and non-hydroxylated HIF1α is not recognized by the pVHL/E3 ubiquitin ligase complex, thereby avoiding ubiquitination and proteasomal degradation (DOI: 10.1073/pnas.95.14.7987, DOI: 10.1515/BC.2004.016, and DOI: 10.1042/BJ20040620). The mechanism of HIF1α accumulation under hypoxia is analogous to the action of the proteasome inhibitor MG132. When we treated cells with hypoxia, the ubiquitination and proteasomal degradation pathway of HIF1α was blocked. At this time, STAMBPL1 knockdown could downregulate the expression of HIF1α (Fig.1A). Meanwhile, since the knockdown of STAMBPL1 significantly downregulated the mRNA level of HIF1α under hypoxia (Fig.2B-C), we concluded that STAMBPL1 affects the expression of HIF1α by mediating its transcription. In addition, MG132 will block all proteasomal substrate degradation and may affect HIF1α mRNA levels indirectly.

(3) Figure 2D and 2F: The effect of STAMBPL1 in promoting HIF1a expression is quite mild, and the effect of HIF1a knockdown is also modest. Given the high levels of STAMBPL1 in TNBC cell lines (Figure 2A), it would be better to repeat these experiments in a STAMBPL1-knockdown setting for clearer insights.

We appreciate this insightful suggestion. Considering that the regulation of HIF1α expression by STAMBPL1 occurs at the transcriptional level, and to prevent excessive accumulation of HIF1a during hypoxia that could confound the effect of STAMBPL1 overexpression on HIF1α regulation, we opted to overexpress STAMBPL1 under normoxic conditions and subsequently knock down HIF1α, as shown in Fig.2D and Fig.2F. This approach allowed us to observe that STAMBPL1 overexpression can upregulate HIF1a expression to some extent. Additionally, in response to the reviewer's suggestion to knock down STAMBPL1, we have conducted the corresponding experiments, with results presented in Fig.1A-E and Fig.2B-C.

(4) Figure 4A: Why does the RNA-seq pattern differ significantly between the two siRNAs? Additionally, the authors should clarify why they focus primarily on transcription factors, as other mechanisms, such as mRNA stability and RNA modification, could also influence gene transcription.

Thank you for this comment. Two siRNAs for STAMBPL1 were designed and synthesized by a biotechnology company. Although both siRNAs target STAMBPL1, they target different sequences. While both siRNAs effectively knocked down STAMBPL1 (Fig. 1A and Fig. 2A), the possibility of off-target effects cannot be completely ruled out. Therefore, we needed to use two siRNAs simultaneously for RNA-seq, ensuring that the gene expression changes observed are due to the knockdown of STAMBPL1 by focusing on genes downregulated by both two siRNAs. Additionally, among the 27 genes downregulated by both two siRNAs, only 18 genes were annotated. Of these 18 genes, except for GRHL3, which is a transcription factor reported to be involved in gene transcription regulation, the remaining 17 genes have no documented association with RNA transcription, stability, or modification. Therefore, we focused on the GRHL3 gene.

(5) Figure 5G: To investigate whether STAMBPL1 and GRHL3 function epistatically in the pathway, a double knockdown of STAMBPL1 and GRHL3 should be examined. Additionally, a double knockdown of STAMBPL1 and FOXO1 should be assessed.

Thank you for your comment. In Figure 5G, we aimed to assess the knockdown efficiency of GRHL3 using siRNAs. To determine whether STAMBPL1 upregulates the HIF1a/VEGFA axis via GRHL3, we overexpressed STAMBPL1 and subsequently knocked down GRHL3. Our findings indicated that STAMBPL1 overexpression indeed enhanced the HIF1a/VEGFA axis, which was rescued by the knockdown of GRHL3, as shown in Figures 4E-F and S3E-F. Similarly, upon overexpressing STAMBPL1 and knocking down FOXO1, we observed that STAMBPL1 overexpression increased the GRHL3/HIF1a/VEGFA axis, which could also be rescued by knocking down FOXO1, as shown in Figures 7F-H. These results suggest that STAMBPL1 upregulates the GRHL3/HIF1a/VEGFA axis through FOXO1. We do not think it is a right way to double knock down STAMBPL1 and FOXO1 or GRHL3.

(6) Figure 7: It remains unclear how STAMBPL1 regulates FOXO1. The authors show that STAMBPL1 increases the transcriptional activation of FOXO1 at the GRHL3 promoter, but it is not clear if STAMBPL1 is required for FOXO1 binding to the GRHL3 promoter. To address this, STAMBPL1-knockdown should be included to examine its effect on FOXO1 binding to the GRHL3 promoter. Furthermore, it would be important to determine whether the STAMBPL1-FOXO1 interaction is essential for GRHL3 transcription. Since the interaction sites of STAMBPL1-FOXO1 have been mapped, a mutant disrupting the interaction would provide better insight into how STAMBPL1 promotes GRHL3 transcription by interacting with FOXO1.

Thank you for this comment. It has been reported that FOXO1 promotes the transcription of the GRHL3 gene by interacting with its promoter (DOI: 10.1093/nar/gkw1276). We also verified through ChIP assay that FOXO1 can bind to the promoter of GRHL3 gene (Fig.7I) and mediate its transcription. Specifically, knocking down FOXO1 significantly down-regulated the mRNA level of GRHL3 (Fig.7B), and the GRHL3 promoter lacking FOXO1 binding site almost completely lost transcriptional activity (Fig.7J), indicating that FOXO1 is crucial for the transcriptional activity of the GRHL3 promoter. Overexpression of STAMBPL1 enhances the activating effect of FOXO1 on the transcriptional activity of the GRHL3 promoter (Fig.7K). However, the up-regulation of GRHL3 transcription by overexpression of STAMBPL1 is completely blocked by FOXO1 knockdown (Fig.7F), and the knockdown of FOXO1 essentially blocks the binding of STAMBPL1 to the GRHL3 promoter (Fig.7L), suggesting that STAMBPL1 affects the transcriptional expression of GRHL3 based on FOXO1. As we added in Discussion, the transcription factor activity of FOXO1 is mainly regulated by its nucleoplasm shuttling process, and the accumulation of FOXO1 in nucleus can enhance its transcription factor activity (DOI: 10.1042/BJ20040167; DOI: 10.15252/embj.2022111867). In our research, neither STAMBPL1 nor its mutant of deubiquitinating enzyme site affected the expression of FOXO1 (Fig.S5E), but STAMBPL1 and FOXO1 co-located in the nucleus (Fig.7M), and they interacted with each other (Fig.7N, Fig.S5I-J). Therefore, we speculate that STAMBPL1 interacts with FOXO1 in the nucleus, obstructs the binding of FOXO1 with the members of 14-3-3 family, inhibits the export of FOXO1, thereby enhancing its transcriptional activity. This interaction between STAMBPL1 and FOXO1 does not necessarily affect the binding of FOXO1 with DNA, including the GRHL3 promoter.

(7) Figure 8 A-C: What is the correlation among the expressions of STAMBPL1, FOXO1, and GRHL3 in TNBC tumors compared to non-TNBC tumors?

Thank you for your comment. In Figure 8A-C, we analyzed the expression levels of STAMBPL1, FOXO1, and GRHL3 in both TNBC and non-TNBC samples using the BCIP. The results indicate that the expression levels of these three genes are significantly higher in TNBC compared to non-TNBC samples. To investigate the correlation among the expressions of STAMBPL1, FOXO1, and GRHL3 in TNBC versus non-TNBC, we further utilized the Metabric data. Besides the positive correlation trend between STAMBPL1 and GRHL3 expression in TNBC clinical samples (Pearson R = 0.27), no significant correlation was observed in the expression levels of STAMBPL1, FOXO1, and GRHL3 in TNBC and non-TNBC clinical samples (as shown in Author response image 1 below). Since STAMBPL1 and FOXO1 are involved as protein molecules in the transcriptional regulation of GRHL3 gene, and the data obtained from the Metabric database are the transcriptional levels of these three genes, this might be the reason why the correlation between their expressions was not observed.

Author response image 1.

Reviewer #2 (Recommendations for the authors):

The authors have thoroughly elucidated the role of STAMBPL1 in TNBC. However, it would be beneficial to discuss the potential clinical implications of these findings, such as how targeting STAMBPL1 or FOXO1 might impact current treatment strategies for TNBC. However, several issues need to be addressed.

Major:

(1) While the study provides an exhaustive analysis of the molecular mechanisms, a comparison with other subtypes of breast cancer could enhance our understanding of the specificity of the STAMBPL1/FOXO1/GRHL3/HIF1α/VEGFA axis in TNBC.

Thank you for your comment. According to report, STAMBPL1 is significantly associated with the mesenchymal characteristics of breast cancer (DOI: 10.1038/s41416-020-0972-x). We utilized cBioPortal (http://www.cbioportal.org/) to analyze the expression of STAMBPL1 across various clinical subtypes of breast cancer. The results indicated that STAMBPL1 is highly expressed in invasive breast cancer, which has been added to Supplementary Figure 6 as Fig.S6D. Given that TNBC is an aggressive type of invasive breast cancer, we further examined the expression of STAMBPL1 in TNBC compared to non-TNBC using BCIP (http://omicsnet.org/bcancer/database). Our findings revealed that the expression level of STAMBPL1 in TNBC was elevated relative to its levels in non-TNBC (Fig.8A). Additionally, since tumor angiogenesis is a critical factor influencing the metastasis of cancer cells, our study focused specifically on the pro-angiogenic effects of STAMBPL1 in TNBC.

(2) The authors might consider discussing any potential off-target effects of the siRNA and shRNA used in the study to bolster the conclusions drawn from the knockdown experiments.

We appreciate the reviewer's suggestion. It is well-known that siRNA or shRNA have off-target effects. To address this concern, we employed two siRNAs for each gene knockdown in our study. Specifically, we knocked down genes such as STAMBPL1, FOXO1, GRHL3, and HIF1A in two TNBC cell lines, HCC1806 and HCC1937, using two siRNAs. Except for siRNA#1 targeting HIF1A, which did not show a significant knockdown effect in HCC1806 cells (Fig.2D and Fig.6A), the knockdown effects of other siRNAs on their respective genes were effective, and the resulting phenotypes were consistent. As shown in Fig.2F and Fig.S4H, siRNA#1 targeting HIF1A had a significant knockdown effect in HCC1937 cells. The lower knockdown efficiency of this siRNA in HCC1806 cell line might be attributed to cell-specific factors.

(3) It would be advantageous if the authors could provide further details on the patient demographics and tumor characteristics in the TCGA database analysis to better comprehend the clinical relevance of their findings.

Thanks for the reviewer's suggestions. We have now indicated the number of clinical samples in each group in the legend of Fig.8A-C. Since we utilized the BCIP online database to analyze and compare the expression levels of the three genes STAMBPL1, FOXO1, and GRHL3 in TNBC and non-TNBC, we are unable to obtain more specific information regarding the tumor characteristics of each sample. However, our analysis clearly shows that the expression levels of these three genes are significantly higher in TNBC compared to non-TNBC.

(4) The authors should consider discussing any limitations regarding the generalizability of their findings, such as potential variations among different TNBC subtypes or the specificity of their observations to certain stages of the disease.

We appreciate the reviewer's comment. Accordingly, we have added a discussion on the limitation of this study in Discussion, highlighted in red font on pages 20 to 21, lines 396 to 412. In addition, we utilized the bc-GenExMiner online database to conduct a comparative analysis of STAMBPL1 expression in different subtypes of non-TNBC and TNBC. The result indicates that STAMBPL1 is highly expressed in mesenchymal-like and basal-like TNBC, which has been added into Supplementary Figure 6 as Fig.S6E. Since these two subtypes of TNBC are highly invasive and metastatic, it suggests that targeting the signaling pathway of STAMBPL1/FOXO1/GRHL3/HIF1α/VEGFA may offer clinical benefits for patients with invasive TNBC.

Minor:

The paper is generally well-written, but it's crucial to maintain vigilance for subject-verb agreement, proper use of tense, and consistent terminology.

Thank you for this suggestion. We have thoroughly revised the article for issues such as grammar, including tense, subject-verb agreement, and terminology.

Author response:

The following is the authors’ response to the original reviews

Reviewer #1:

Summary:

In this manuscript (eLife-RP-RA-2024-103904), the authors identified that NOLC1 was upregulated in gastric cancer samples, which promoted cancer progression and cisplatin resistance. They further found that NOLC1 could bind to p53 and decrease its nuclear transcriptional activity, then inhibit p53-mediated ferroptosis. There are several major concerns regarding the conclusions.

Strengths:

This study identified that NOLC1 could bind to p53 and decrease its nuclear transcriptional activity, then inhibit p53-mediated ferroptosis in gastric cancer.

Weaknesses:

The major conclusions were not sufficiently supported by the results. The experiments were not conducted in a comprehensive manner.

Major concerns

(1) The authors investigated NOLC1 expression in gastric cancer (GC) using clinical samples, which is valuable; however, the sample array includes only 3 patients. This sample size is insufficient to support conclusions for human samples. Please increase the sample size and apply a more robust statistical analysis. Additionally, specify the statistical methods used in the figure legend.

Thanks very much for the kind comments and great suggestions. As suggested, we have increased the sample size of GC patients, and the new data (six pair samples) was shown in Fig. S1A, further reflecting that NOLC1 was upregulate in gastric cancer (GC). Moreover, the statistical methods have been added in each figure legend.

(2) These data are not sufficient to support the key conclusion of this study "NOLC1 is significantly upregulated in GC tissues and Cis-resistant GC cells". There is no convincing data showing that NOLC1 upregulation is specific to cancer cells or any other cell types. Based on the following results that NOLC1 expressed in cancer cells can support cancer cell survival and drug resistance, the authors switched to investigating the role of NOLC1 in cancer cells without demonstrating cancer cells indeed highly upregulate NOLC1.

Thanks for raising this good question. As shown in Fig. 1E-F, the TCGA database have shown that NOLC1 was upregulated in GC. Moreover, we further analyzed the NOLC1 expression level in other cancer type, according to the Human Protein Atlas (https://www.proteinatlas.org/). The results indicated that NOLC1 mRNA level was much higher in almost all cancers except acute myeloid leukemia (LAML). In addition, according to the gene expression profiling interactive analysis (GEPIA, http://gepia.cancer-pku.cn/index.html), NOLC1 mRNA level was above 100 nTPM in most gastric cancer cell lines, however in most non-cancerous cell lines was below 100 nTPM, indicating that NOLC1 was up-regulated in gastric cancer.

Author response image 1.

The mRNA level of NOLC1 in different GC cells and non-cancerous cells.

(3) The authors primarily use MGC-803 cells for experiments; however, MGC-803 is known to be a HeLa-contaminated cell line. Could the authors explain this choice of using this cell line only? Did they validate key findings with additional cell lines? This is particularly important for assays such as cisplatin resistance validation, in vivo experiments, TEM imaging, and MitoPeDPP fluorescence imaging.

Thanks for raising this good question. We are not only use MGC-803 cells, the key findings in vitro was also validated in MKN-45 cells (Fig. 2), and in vivo experiment also validated in Mouse Forestomach Carcinoma cells (MFC)-tumor bearing 615 mice model (Fig 7). Furthermore, we further added some experiments in MKN-45 cells. The TEM imaging showed that NOLC1 could significantly inhibit cisplatin (Cis) induced lipid membrane damage in MKN-45 cells (Fig. S6A). Moreover, MitoPeDPP fluorescence assay analyzed by FCAs also indicating that rapid ROS was enriched in mitochondria in MKN-45 cells (Fig. 4E, Fig. S6J).

(4) In Figure 2, did the authors perform assays with NOLC1 overexpression? If so, please include these results to strengthen the conclusions.

Thanks very much for the kind comments and great suggestions. As suggested, we added new data about NOLC1 overexpression assay Cell counting kit-8 assay shows that NOLC1-overexpression group is more resistance to Cis compared to vector group (Fig. S4E, S5A).

(5) The authors show in Figures 2A-B that shNOLC1 without cisplatin treatment does not affect cell viability. However, Figures 2D-E suggest increased apoptosis in shNOLC1 cells without cisplatin treatment. Additionally, in vivo studies in Figure 3 show no significant difference between the shNC+PBS and shNOLC1+PBS groups, which appears contradictory to the apoptosis assays. Similarly, Ki67 staining shows decreased scores in the shNOLC1 group compared to shNC. Could the authors clarify this inconsistency?

Thanks for raising this good question. In Fig 2D-E, the difference in proportion of death cells between shNOLC1 and shNC treated with PBS groups were only 3% (MGC-803) and 7% (MKN-45) which is much lower than that treated with cisplatin in vitro. Moreover, in vivo analysis indicated that the average tumor volume in NOLC1+PBS group was smaller than that in NC group, but there was no statistical significance (p value = 0.3962). Moreover, tumor proliferation is a complex process regulated by many factors [1,2], thus the level of Ki67 is by no means the same as the rate of tumor proliferation, might be positively correlated.

(6) In Figure 4, NOLC1 knockdown appears to enhance cisplatin-induced ferroptosis rather than apoptosis. Given p53's role in apoptosis, did the authors compare the effects of NOLC1 on cisplatin-induced apoptosis vs. ferroptosis? If so, please clarify whether NOLC1 predominantly regulates apoptosis or ferroptosis.

Thanks for raising this good question. We do have compared the effects of NOLC1 on cisplatin-induced apoptosis vs. ferroptosis. As shown in Fig. 5A, NOLC1 knockdown obviously increased the BCL-2 protein level which is an anti-apoptotic protein and mediated by p53 via protein interaction in cytoplasm[3,4], this phenomenon may cause by the increasing level of p53 in cytoplasm (Fig. 6I). Also, the TEM imaging showed the classic ferroptotic morphological changes rather than apoptosis (Fig. 5A, S6A). Taken together, NOLC1 mainly regulates p53 mediated ferroptosis rather than apoptosis.

(7) Did the authors perform co-IP assays with p53 or HA antibodies to immunocapture NOLC1? If not, please add this experiment to support protein interactions. The mechanistic correlation between p53 and NOLC1 can be supported by adding experiments using multiple GC cell lines with various p53 alterations (such as loss-of- function or gain-of-function mutations/deletions). This is critical because the authors specifically claimed that NOLC1 can inhibit p53-mediated ferroptosis, but not other tumor suppressors.

Thanks very much for the kind comments and great suggestions. As suggested, we had performed Co-IP assay with anti-HA antibodies to immunocapture NOLC1-FLAG. As shown in Fig. 5K, p53 DNA binding domain (DBD)-HA could immunocapture with NOLC1, further indicated that NOLC1 could binding to p53 DBD. Moreover, we concur with the reviewer that adding experiments using multiple p53 alterations, however considering that different p53 mutants have completely different functional changes. Therefore, we using siRNA to knockdown p53 level in MGC-803 cells, the results shown that NOLC1 mediated resistance was disappear and the GPX4 level was increased (Fig. S10). These data have shown that NOLC1 promotes GC resistance via mediated p53 functions.

(8) In Figure S5B, the LDH release can be blocked by Fer-1?

Thanks for raising this good question. As suggested, Fer-1 (20 μmol/mL) significantly blocked the LDH release in NOLC1 knockdown group (Fig S6E). This data further confirmed that NOLC1 suppressed Cis-induced ferroptosis.

(9) How about the ubiquitination assay in MGC-803 cells?

Thanks for raising this good question. As suggested, we also analyzed the ubiquitination assay in MGC-803 cells. As the result showed that NOLC1 also could increasing level of ubiquitination of p53 (Fig. 6H).

(10) In Figure 6H, the DBD domain of NOLC1 is required for inhibiting P53 ubiquitination.

Thanks for your opinion. However, in our paper, we only mentioned that p53 DBD domain, rather than NOLC1 DBD domain. Also, we did not find any DNA binding function of NOLC1 in the Pubmed database. Therefore, we would like to ask whether the revised opinion is correct.

(11) In Figure 8B, the CD3 antibody is not specific, please change it to a new one.

Thanks very much for the kind comments and great suggestions. As suggested, we have used new CD3 antibody and the new data was added in Fig. 8B.

(12) The authors report that NOLC1 influences peripheral blood lymphocytes with cisplatin treatment, with or without PD-1. Could the authors explain why NOLC1 would affect peripheral blood lymphocytes? Additionally, did they assess immune cell infiltration in the tumor microenvironment (TME) by flow cytometry?

Thanks for raising good question. The tumor size of the knockdown group treated with Cis + PD-1 was too small (less than 100 mg) to extract enough infiltrated immune cells (less than 10000 CD45⁺ cells), thus we chose to detect immune cells in the blood of the mice. Considering that the infiltrating immune cells including CTLs were originate from peripheral blood by circulation. Under the normal conditions, serval tumor biology behavior impact the TME to limit immune responses and present barriers to cancer therapy. For example, tumor could express or secret lots of negative regulator like PD-L1. Causing immune cells cannot recognize tumor cells and infiltrate into tumor tissue. Ferroptosis, as a new from of ICD, could damage tumor cell plasm and release amount of tumor associated antigen and tumor-specific antigens causing immune cells priming and activation. Eventually, the activated immune cells in peripheral blood travel towards the tumor site, infiltrating the tumor tissue under favorable co-stimulatory conditions and guided by chemokine gradients. Once within the tumor microenvironment, these activated T cells can control tumor growth through direct tumor cell destruction and cytokine-mediated processes [5–8]

To assess immune cell infiltration in the TME, we analyzed the tumor infiltrated CD3⁺ and CD8⁺ immune cells in tumor tissue by immunofluorescence (Fig. 8B). Thus, the peripheral blood lymphocytes could reflect the infiltration of immune cells in the tumor.

Minor concerns:

(1) Please clarify the statistical methods in each figure legend.

Thanks for your opinion. We have added statistical methods in each figure legend.

(2) In Figure 2D, please provide statistical data of cleaved-caspase3 expression.

Thanks for your opinion. As is shown in Fig. S5B-C, the relative cleaved-caspase3 were provided.

(3) Please ensure that the canonical expressions used in the research paper are adhered to.

Thanks for your opinion. We have carefully modified our expressions in our paper.

(4) Please pay more attention to the grammar and formatting of texts.

Thanks for your opinion. We revised our manuscript through the American Journal Experts (AJE) service.

Reviewer #2:

Summary:

Shengsheng Zhao et al. investigated the role of nucleolar and coiled-body phosphoprotein 1 (NOLC1) in relegating gastric cancer (GC) development and cisplatin-induced drug resistance in GC. They found a significant correlation between high NOLC1 expression and the poor prognosis of GC. Meanwhile, upregulation of NOLC1 was associated with cis-resistant GC. Experimentally, the authors demonstrate that knocking down NOLC1 increased GC sensitivity to Cis possibly by regulating ferroptosis. Mechanistically, they found NOLC1 suppressed ferroptosis by blocking the translocation of p53 from the cytoplasm to the nucleus and promoting its degradation. In addition, The authors also evaluated the effect of combinational treatment of anti- PD-1 and cisplatin in NOLC1-knockdown tumor cells, revealing a potential role of NOLC1 in the targeted therapy for GC.

Strengths:

Chemoresistance is considered a major reason causing failure of tumor treatment and death of cancer patients. This paper explored the role of NOLC1 in the regulation of Cis-mediated resistance, which involves a regulated cell death named ferroptosis. These findings provide more evidence highlighting the study of regulated cell death to overcome drug resistance in cancer treatment, which could give us more potential strategies or targets for combating cancer.

Weaknesses:

More evidence supporting the regulation of ferroptosis induced by Cisplatin by NOLC1 should be added. Particularly, the role of ferroptosis in the cisplatin-resistance should be verified and whether NOLC1 regulates ferroptosis induced by additional FINs should be explored. Besides, the experiments to verify the regulation of ferroptosis sensitivity by NOLC1 are sort of superficial. The role of MDM2/p53 in ferroptosis or cisplatin resistance mediated by NOLC1 should be further studied by genetic manipulation of p53, which is the key evidence to confirm its contribution to NOLC1 regulation of GC and relative cell death.

Major points:

(1) More evidence supporting the regulation of ferroptosis induced by Cisplatin by NOLC1 should be added. Particularly, the role of ferroptosis in the cisplatin-resistance should be verified and whether NOLC1 regulates ferroptosis induced by additional FINs should be explored.

Thanks very much for the kind comments and great suggestions. As suggested, we have further analyzed the ferroptosis inhibit ability of NOLC1 in MGC-45 cells treated with Erastin, a common used ferroptosis activator. As shown in Fig. S6B, the ferroptosis activated by Erastin was also blocked by NOLC1.

(2) In Figure 1J, the CR cell line should obviously have less apoptosis-maker c-PARP expression, which means these cells are resistant to apoptosis induced by CR. Thus, it would be more rational to study the role of apoptosis regulation by NOLC1. Why did the later data shift to the study of ferroptosis?

Thanks for raising this good question. In the CR cells, the expression levels of many genes were changed, so it is uncertain whether the decreased expression level of cleaved-PARP in the resistant cells is caused by NOLC1 up-regulated. To explore the specific mechanism of NOLC1 mediated resistant, we performed the TEM imaging (Fig. 4A, S6A) and the results showed that cells exhibited classic ferroptosis morphological changes. Moreover, the BCL-2 (an anti-apoptotic protein, and regulated by p53 via protein interaction in cytoplasm) was increased after NOLC1 knockdown (Fig S5A). This phenomenon may cause by the increasing p53 levels in the cytoplasm[3,4] (Fig 5I). Taken together we shift to study of cisplatin induced ferroptosis.

(3) Besides, how about the regulation of apoptosis during cis-resistance by NOLC1 in GC?

Thanks for raising this good question. As mentioned above the Cis induced apoptosis was not as significant as ferroptosis, caused by BCL-2 (a key anti-apoptosis protein) increasing which is mediated by p53 via protein interaction in cytoplasm. NOLC1 increased plasm p53 level subsequently increased BCL-2 level.

(4) The experiments to verify the regulation of ferroptosis sensitivity by NOLC1 are sort of superficial. The role of MDM2/p53 in ferroptosis or cisplatin resistance mediated by NOLC1 should be further studied by genetic manipulation of p53, which is the key evidence to confirm its contribution to NOLC1 regulation of GC and relative cell death.

Thanks for raising this good question. As is shown in Fig S10, after knockdown p53 protein level by using siRNA, NOLC1 could not promote Cis-resistance and the GPX4 level was increased reflecting that NOLC1 promotes Cis resistance via mediate p53 function.

(5) In Figure 2, the data indicated that the knockdown of NOLC1 increased rH2Ax in the presence of Cisplatin, which indicated that NOLC1 might regulate DNA damage-related cellular function. These functions should be more relevant to cisplatin resistance, considering the fundamental effect of this chemo drug.

Thanks very much for the kind comments and great suggestions. Indeed, we found that DNA damage was more obvious in knockdown groups, but the ferroptotic changes like ROS and mitochondrial membrane damage were also significantly different in knockdown groups. Considering that as a chemo drug, cisplatin not only induces damage DNA but also acts as a stress which could activates various signal pathways including apoptosis, ferroptosis, pyroptosis, necroptosis, etc., under different drug concentrate or time [9–11]. Therefore, it is important to find out the NOLC1 predominantly blocked pathway in GC.

(6) In Figure.4, ferroptosis inhibitors like Ferr-1 or DFO should be used to verify the regulation of ferroptosis by Cisplatin and NOLC1.

Thanks very much for the kind comments and great suggestions. As suggested, we performed additional LDH release assay. The results showed that Fer-1 also could block cisplatin induced LDH release in NOLC1 knockdown groups (Fig. S6E).

(7) In Figure 4H, Cisplatin decreased FSP1 and GPX4, which could be enhanced in the NOLC1-konckdown cell line. Meanwhile, the knockdown of NOLC1 increased the ACSL4 level. These findings could be the key reason for the regulation of ferroptosis by NOLC1 rather than p53 since they all are direct regulators of ferroptosis.

Thanks very much for the kind comments and great suggestions. We rewrote the text as you suggested. Recently, it also has been reported that ACSL4-regulated ferroptosis is related to p53, but the exact mechanism is still unclear [12]. Moreover, further studies of specific relation between NOLC1 and FSP1/ACSL4 will be conducted in the further

(8) Whether p53 mediates the regulation of ferroptosis and cisplatin resistance by NOLC1 should be thoroughly studied using p53-KO cell lines.

Thanks very much for the kind comments and great suggestions. As previously mentioned, by using si-RNA to knockdown p53, the NOLC1 mediate Cis-resistance were blocked (Fig. S10). Meanwhile, the GPX4 level was also increased in p53/NOLC1 double-knockdown groups compared to the NOLC1 knockdown group. These data indicating that NOLC1 suppresses ferroptosis via mediating p53 functions.

Reviewer #3:

The authors have put forth a compelling argument that NOLC1 is indispensable for gastric cancer resistance in both in vivo and in vitro models. They have further elucidated that NOLC1 silencing augments cisplatin-induced ferroptosis in gastric cancer cells. The mechanistic underpinning of their findings suggests that NOLC1 modulates the p53 nuclear/plasma ratio by engaging with the p53 DNA Binding Domain, which in turn impedes p53-mediated transcriptional regulation of ferroptosis. Additionally, the authors have shown that NOLC1 knockdown triggers the release of ferroptosis-induced damage-associated molecular patterns (DAMPs), which activate the tumor microenvironment (TME) and enhance the efficacy of the anti-PD-1 and cisplatin combination therapy.

Strengths:

The manuscript presents a robust dataset that substantiates the authors' conclusion. They have identified NOLC1 as a potential oncogene that confers resistance to immuno-chemotherapy in gastric cancer through the mediation of ferroptosis and subsequent TME reprogramming. This discovery positions NOLC1 as a promising therapeutic target for gastric cancer treatment. The authors have delineated a novel mechanistic pathway whereby NOLC1 suppresses p53 transcriptional functions by reducing its nuclear/plasma ratio, underscoring the significance of p53 nuclear levels in tumor suppression over total protein levels.

Weaknesses:

While the overall findings are commendable, there are specific areas that could benefit from further refinement. The authors have posited that NOLC1 suppresses p53- mediated ferroptosis; however, the mRNA levels of ferroptosis genes regulated by p53 have not been quantified, which is a critical gap in the current study. In Figure 4A, transmission electron microscopy (TEM) results are reported solely for the MGC-803 cell line. It would be beneficial to include TEM data for the MKN-45 cell line to strengthen the findings. The authors have proposed a link between NOLC1-mediated reduction in the p53 nuclear/plasma ratio and gastric cancer resistance, yet the correlation between this ratio and patient prognosis remains unexplored, which is a significant limitation in the context of clinical relevance.

Thanks very much for the kind comments and great suggestions. As suggested, recently studies have reported that CDKN1A (also called p21, a p53 transcriptional mediated protein) could promotes ferroptosis[13], the mRNA levels of ferroptosis genes regulated by p53 have were quantified in Fig. S8G-H. Moreover, we further proceed TEM imaging in MKN-45 cells, the result was consistent to MGC-803 cells, reflecting that NOLC1 has a broad spectrum of promoting drug resistance in gastric cancer. Also, recently studies have reported that p53 transcriptional active and p53 transcriptional inactive types include patients with intermediate prognosis and recurrence rates, with the p53-acvtie group showing better prognosis[14]. Considering p53 transcriptional activity depends on p53 nuclear accumulation, we assume that the low level of p53 nuclear/plasma may cause poor prognosis in gastric cancer. Meanwhile we will further collect enough samples and their prognostic information to analysis NOLC1-mediated reduction in the p53 nuclear/plasma ratio and gastric cancer resistance.

References

(1) Z. Seferbekova, A. Lomakin, L.R. Yates, M. Gerstung, Spatial biology of cancer evolution, Nat Rev Genet 24 (2023) 295–313. https://doi.org/10.1038/s41576-022-00553-x.

(2) T. Matsuoka, M. Yashiro, Molecular Mechanism for Malignant Progression of Gastric Cancer Within the Tumor Microenvironment, IJMS 25 (2024) 11735. https://doi.org/10.3390/ijms252111735.

(3) Y. Liu, Z. Su, O. Tavana, W. Gu, Understanding the complexity of p53 in a new era of tumor suppression, Cancer Cell (2024) S1535610824001338. https://doi.org/10.1016/j.ccell.2024.04.009.

(4) R. Pan, V. Ruvolo, H. Mu, J.D. Leverson, G. Nichols, J.C. Reed, M. Konopleva, M. Andreeff, Synthetic Lethality of Combined Bcl-2 Inhibition and p53 Activation in AML: Mechanisms and Superior Antileukemic Efficacy, Cancer Cell 32 (2017) 748-760.e6. https://doi.org/10.1016/j.ccell.2017.11.003.

(5) E. Catanzaro, M. Beltrán-Visiedo, L. Galluzzi, D.V. Krysko, Immunogenicity of cell death and cancer immunotherapy with immune checkpoint inhibitors, Cell Mol Immunol 22 (2024) 24–39. https://doi.org/10.1038/s41423-024-01245-8.

(6) G. Lei, L. Zhuang, B. Gan, The roles of ferroptosis in cancer: Tumor suppression, tumor microenvironment, and therapeutic interventions, Cancer Cell 42 (2024) 513–534. https://doi.org/10.1016/j.ccell.2024.03.011.

(7) E. Catanzaro, R. Demuynck, F. Naessens, L. Galluzzi, D.V. Krysko, Immunogenicity of ferroptosis in cancer: a matter of context?, Trends in Cancer 10 (2024) 407–416. https://doi.org/10.1016/j.trecan.2024.01.013.

(8) X. Jiang, B.R. Stockwell, M. Conrad, Ferroptosis: mechanisms, biology and role in disease, Nat Rev Mol Cell Biol 22 (2021) 266–282. https://doi.org/10.1038/s41580-020-00324-8.

(9) J.-L. Roh, E.H. Kim, H. Jang, D. Shin, Nrf2 inhibition reverses the resistance of cisplatin-resistant head and neck cancer cells to artesunate-induced ferroptosis, Redox Biology 11 (2017) 254–262. https://doi.org/10.1016/j.redox.2016.12.010.

(10) X. Wang, Y. Zhou, D. Wang, Y. Wang, Z. Zhou, X. Ma, X. Liu, Y. Dong, Cisplatin-induced ototoxicity: From signaling network to therapeutic targets, Biomedicine & Pharmacotherapy 157 (2023) 114045. https://doi.org/10.1016/j.biopha.2022.114045.

(11) J. Liang, G. Bi, Y. Huang, G. Zhao, Q. Sui, H. Zhang, Y. Bian, J. Yin, Q. Wang, Z. Chen, C. Zhan, MAFF confers vulnerability to cisplatin-based and ionizing radiation treatments by modulating ferroptosis and cell cycle progression in lung adenocarcinoma, Drug Resistance Updates 73 (2024) 101057. https://doi.org/10.1016/j.drup.2024.101057.

(12) M.Y. Kosim, T. Fukazawa, M. Miyauchi, N. Hirohashi, K. Tanimoto, p53 status modifies cytotoxic activity of lactoferrin under hypoxic conditions, Front. Pharmacol. 13 (2022) 988335. https://doi.org/10.3389/fphar.2022.988335.

(13) Q. Gao, J. Chen, C. Li, J. Zhan, X. Yin, B. Li, H. Dong, L. Luo, Z. Li, CDKN1A promotes Cis-induced AKI by inducing cytoplasmic ROS production and ferroptosis, Food and Chemical Toxicology 193 (2024) 115003. https://doi.org/10.1016/j.fct.2024.115003.

(14) R. Cristescu, Molecular analysis of gastric cancer identifies subtypes associated with distinct clinical outcomes, Nature Medicine (2015).

Author response:

The following is the authors’ response to the original reviews

Reviewer #1 (Public review):

Summary:

This study aimed at replicating two previous findings that showed (1) a link between prediction tendencies and neural speech tracking, and (2) that eye movements track speech. The main findings were replicated which supports the robustness of these results. The authors also investigated interactions between prediction tendencies and ocular speech tracking, but the data did not reveal clear relationships. The authors propose a framework that integrates the findings of the study and proposes how eye movements and prediction tendencies shape perception.

Strengths:

This is a well-written paper that addresses interesting research questions, bringing together two subfields that are usually studied in separation: auditory speech and eye movements. The authors aimed at replicating findings from two of their previous studies, which was overall successful and speaks for the robustness of the findings. The overall approach is convincing, methods and analyses appear to be thorough, and results are compelling.

Weaknesses:

Linking the new to the previous studies could have been done in more detail, and the extent to which results were replicated could have been discussed more thoroughly.

Eye movement behavior could have been presented in more detail and the authors could have attempted to understand whether there is a particular component in eye movement behavior (e.g., microsaccades) that drives the observed effects.

We would like to thank you for your time and effort in reviewing our work and we appreciate the positive comments!

We extended our manuscript, now providing intermediate results on individual prediction tendency, which can be compared to our results from Schubert et al., (2023).

Furthermore, we expanded our discussion now detailing the extent to which our results (do not) replicate the previous findings (e.g. differences in horizontal vs. vertical ocular speech tracking, lack of distractor tracking, link between ocular speech tracking and behavioral outcomes).

While we agree with the reviewer that it is an important and most interesting question, to what extent individual features of gaze behavior (such as microsaccades, blinks etc.) contribute to the ocular speech tracking effect, it is beyond the scope of the current manuscript. It will be methodologically and conceptually challenging to distinguish these features from one another and to relate them to diverse cognitive processes. We believe that a separate manuscript is needed to give these difficult questions sufficient space for new methodological approaches and control analyses. The primary goal of this manuscript was to replicate the findings of Gehmacher et al. (2024) using similar methods and to relate them to prediction tendencies, attention, and neural speech tracking. 

Reviewer #2 (Public review):

Summary

Schubert et al. recorded MEG and eye-tracking activity while participants were listening to stories in single-speaker or multi-speaker speech. In a separate task, MEG was recorded while the same participants were listening to four types of pure tones in either structured (75% predictable) or random (25%) sequences. The MEG data from this task was used to quantify individual 'prediction tendency': the amount by which the neural signal is modulated by whether or not a repeated tone was (un)predictable, given the context. In a replication of earlier work, this prediction tendency was found to correlate with 'neural speech tracking' during the main task. Neural speech tracking is quantified as the multivariate relationship between MEG activity and speech amplitude envelope. Prediction tendency did not correlate with 'ocular speech tracking' during the main task. Neural speech tracking was further modulated by local semantic violations in the speech material, and by whether or not a distracting speaker was present. The authors suggest that part of the neural speech tracking is mediated by ocular speech tracking. Story comprehension was negatively related to ocular speech tracking.

Strengths

This is an ambitious study, and the authors' attempt to integrate the many reported findings related to prediction and attention in one framework is laudable. The data acquisition and analyses appear to be done with great attention to methodological detail (perhaps even with too much focus on detail-see below). Furthermore, the experimental paradigm used is more naturalistic than was previously done in similar setups (i.e. stories instead of sentences).

Weaknesses

For many of the key variables and analysis choices (e.g. neural/ocular speech tracking, prediction tendency, mediation) it is not directly clear how these relate to the theoretical entities under study, and why they were quantified in this particular way. Relatedly, while the analysis pipeline is outlined in much detail, an overarching rationale and important intermediate results are often missing, which makes it difficult to judge the strength of the evidence presented. Furthermore, some analysis choices appear rather ad-hoc and should be made uniform and/or better motivated.

We would like to thank you very much for supporting our paper and your thoughtful feedback!

To address your concerns, that our theoretical entities as well as some of our analytical choices lack transparency, we expanded our manuscript in several ways:

(1) We now provide the intermediate results of our prediction tendency analysis (see new Figure 2 of our manuscript). These results are comparable to our findings from Schubert et al. (2023), demonstrating that on a group level there is a tendency to pre-activate auditory stimuli of high probability and illustrating the distribution of this tendency value in our subject population.

(2) We expanded our methods section in order to explain our analytical choices (e.g. why this particular entropy modulation paradigm was used to measure individual prediction tendency).

(3) We now provide an operationalisation of the terms “neural speech tracking” and “ocular speech tracking” at their first mention, to make these metrics more transparent to the reader.

(4) We are summarizing important methodological information ahead of each results section, in order to provide the reader with a comprehensible background, without the necessity to read through the detailed methods section. 

(5) We expanded our discussion section, with a special emphasis on relating the key variables of the current investigation to theoretical entities.

Reviewer #3 (Public review):

Summary:

In this paper, the authors measured neural activity (using MEG) and eye gaze while individuals listened to speech from either one or two speakers, which sometimes contained semantic incongruencies.

The stated aim is to replicate two previous findings by this group: (1) that there is "ocular speech tracking" (that eye-movements track the audio of the speech), (2) that individual differences in neural response to tones that are predictable vs. not-predictable in their pitch is linked to neural response to speech. In addition, here they try to link the above two effects to each other, and to link "attention, prediction, and active sensing".

Strengths:

This is an ambitious project, that tackles an important issue and combines different sources of data (neural data, eye-movements, individual differences in another task) in order to obtain a comprehensive "model" of the involvement of eye-movements in sensory processing.

The authors use many adequate methods and sophisticated data-analysis tools (including MEG source analysis and multivariate statistical models) in order to achieve this.

Weaknesses:

Although I sympathize with the goal of the paper and agree that this is an interesting and important theoretical avenue to pursue, I am unfortunately not convinced by the results and find that many of the claims are very weakly substantiated in the actual data.

Since most of the analyses presented here are derivations of statistical models and very little actual data is presented, I found it very difficult to assess the reliability and validity of the results, as they currently stand. I would be happy to see a thoroughly revised version, where much more of the data is presented, as well as control analyses and rigorous and well-documented statistical testing (including addressing multiple comparisons).

We thank you for your thoughtful feedback. We appreciate your concerns and will address them below in greater detail.

These are the main points of concern that I have regarding the paper, in its current format.

(1) Prediction tendencies - assessed by listening to sequences of rhythmic tones, where the pitch was either "predictable" (i.e., followed a fixed pattern, with 25% repetition) or "unpredictable" (no particular order to the sounds). This is a very specific type of prediction, which is a general term that can operate along many different dimensions. Why was this specific design selected? Is there theoretical reason to believe that this type of prediction is also relevant to "semantic" predictions or other predictive aspects of speech processing?

Theoretical assumptions and limitations of our quantification of individual prediction tendency are now shortly summarized in the first paragraph of our discussion section. With this paradigm we focus on anticipatory “top-down” predictions, whilst controlling for possibly confounding “bottom-up” processes. Since this study aimed to replicated our previous work we chose the same entropy-modulation paradigm as in other studies from our group (e.g. Demarchi et al. 2019, Schubert et al. 2023;2024, Reisinger et al. 2024), which has proven to give reproducible findings of feature-specific preactivations of sounds in a context of low entropy. One advantage of this design is that it gives us the opportunity to directly compare the processing of “predictable” and “unpredictable” sounds of the same frequency in a time-resolved manner (this argument is now also included in the Methods section).

Regarding the question to what extent this type of prediction might also be relevant to “semantic” predictions we would like to refer to our previous study (Schubert et al., 2023), where we explicitly looked at the interaction between individual prediction tendency and encoding of semantic violations in the cortex. (In short, there we found a spatially dissociable interaction effect, indicating an increased encoding of semantic violations that scales with prediction tendency in the left hemisphere, as well as a disrupted encoding of semantic violations for individuals with stronger prediction tendency in the right hemisphere.) We did not aim to replicate all our findings in the current study, but instead we focused on merging the most important results from two independent phenomena in the domain of speech processing and bringing them into a common framework. However, as now stated in our discussion, we believe that “predictions are directly linked to the interpretation of sensory information. This interpretation is likely to occur at different levels along the cognitive (and anatomical) hierarchy…” and that “this type of prediction is relevant for acoustic processing such as speech and music, whose predictability unfolds over time.”

(2) On the same point - I was disappointed that the results of "prediction tendencies" were not reported in full, but only used later on to assess correlations with other metrics. Even though this is a "replication" of previous work, one would like to fully understand the results from this independent study. On that note, I would also appreciate a more detailed explanation of the method used to derive the "prediction tendency" metric (e.g, what portion of the MEG signal is used? Why use a pre-stimulus and not a post-stimulus time window? How is the response affected by the 3Hz steady-state response that it is riding on? How are signals integrated across channels? Can we get a sense of what this "tendency" looks like in the actual neural signal, rather than just a single number derived per participant (an illustration is provided in Figure 1, but it would be nice to see the actual data)? How is this measure verified statistically? What is its distribution across the sample? Ideally, we would want enough information for others to be able to replicate this finding).

We now included a new figure (similar to Schubert et al. 2023) showing the interim results of the “prediction tendency” effect as well as individual prediction tendency values of all subjects.

Furthermore we expanded the description of the “prediction tendency” metric in the Methods section, where we explain our analytical choices in more detail. In particular we used a pre-stimulus time window in order to capture “anticipatory predictions”. The temporally predictably design gives us the opportunity to capture this type of predictions. The integration across channels is handled by the multivariate pattern analysis (MVPA), which inherently integrates multidimensional data (as mentioned in the methods section we used data from 102 magnetometers) and links it to (in this case) categorical information.

(3) Semantic violations - half the nouns ending sentences were replaced to create incongruent endings. Can you provide more detail about this - e.g., how were the words selected? How were the recordings matched (e.g., could they be detected due to audio editing?)? What are the "lexically identical controls that are mentioned"? Also, is there any behavioral data to know how this affected listeners? Having so many incongruent sentences might be annoying/change the nature of listening. Were they told in advance about these?

We expanded the Methods section and included the missing information: 

“We randomly selected half of the nouns that ended a sentence (N = 79) and replaced them with the other half to induce unexpected semantic violations. The swap of nouns happened in the written script before the audio material was recorded in order to avoid any effects of audio clipping. Narrators were aware of the semantic violations and had been instructed to read out the words as normal. Consequently all target words occurred twice in the text, once in a natural context (serving as lexical controls) and once in a mismatched context (serving as semantic violations) within each trial, resulting in two sets of lexically identical words that differed greatly in their contextual probabilities (see Figure 1F for an example). Participants were unaware of these semantic violations.” Since we only replaced 79 words with semantic violations in a total of ~ 24 minutes of audio material we believe that natural listening was not impaired. In fact none of the participants mentioned to have noticed the semantic violations during debriefing (even though they had an effect on speech tracking in the brain). 

(4) TRF in multi-speaker condition: was a univariate or multivariate model used? Since the single-speaker condition only contains one speech stimulus - can we know if univariate and multivariate models are directly comparable (in terms of variance explained)? Was any comparison to permutations done for this analysis to assess noise/chance levels?

For mTRF models it depends on the direction (“encoding” vs. “decoding”) whether or not the model is comparable to a univariate model. In our case of an encoding model the TRFs are fitted to each MEG channel independently. This gives us the possibility to explore the effect over different areas (whereas a multivariate “decoding” model would result in only one speech reconstruction value).

In both conditions (single and multi speaker) a single input feature (the envelope of the attended speech stream) was used. Of course it would be possible to fit the model to use a multivariate encoding model, predicting the brain’s response to the total input of sounds. This would, however, target a slightly different question than ours as we aimed to investigate how much of the attended speech is tracked.

Regarding your suggestion of a comparison to permutations to assess noise levels we would like to point out that we chose the same methodological approach as in our previous studies, that we aimed to replicate here. Indeed in these original studies no permuted versions (with exception of the mediation analysis where comparing a model with an additional input predictor to a single predictor model would not result in a fair comparison) have been used. We conducted the mTRF approach considering the guidelines of Crosse et al. (2016) to the best of our knowledge and in accordance with similar studies in this field.

Crosse, M. J., Di Liberto, G. M., Bednar, A., & Lalor, E. C. (2016). The multivariate temporal response function (mTRF) toolbox: a MATLAB toolbox for relating neural signals to continuous stimuli. Frontiers in human neuroscience, 10, 604.

(5) TRF analysis at the word level: from my experience, 2-second segments are insufficient for deriving meaningful TRFs (see for example the recent work by Mesik & Wojtczak). Can you please give further details about how the analysis of the response to semantic violations was conducted? What was the model trained on (the full speech or just the 2-second long segments?) Is there a particular advantage to TRFs here, relative - say - to ERPs (one would expect a relatively nice N400 response, not)? In general, it would be nice to see the TRF results on their own (and not just the modulation effects).

We fully agree with the reviewers statement that 2-second segments would have been too short to derive meaningful TRFs. To investigate the effect of semantic violations, we used the same TRFs trained on the whole dataset (with 4-fold cross validation). The resulting true as well as the predicted data was segmented into single word epochs of 2 seconds. We selected semantic violations as well as their lexically identical controls and correlated true with predicted responses for every word. Thus, we conducted the same analysis as for the overall encoding effect, focusing on only part of the data. We have reformulated the Methods section accordingly to clear up this misunderstanding. Since the TRFs are identical to the standard TRFs from the overall neural speech tracking, they are not informative to the semantic violation effect. However, since the mTRF approach is the key method throughout the manuscript (and our main focus is not on the investigations of brain responses to semantic violations) we have favoured this approach over the classical ERF analysis. 

(6) Another related point that I did not quite understand - is the dependent measure used for the regression model "neural speech envelope tracking" the r-value derived just from the 2sec-long epochs? Or from the entire speech stimulus? The text mentions the "effect of neural speech tracking" - but it's not clear if this refers to the single-speaker vs. twospeaker conditions or to the prediction manipulation. Or is it different in the different analyses? Please spell out exactly what metric was used in each analysis.

As suggested we now provide a clear definition of each dependent metric for each analysis.

“Neural speech tracking” refers to the correlation coefficients between predicted and true brain responses from the aforementioned encoding model, trained and tested on the whole audio material within condition (single vs. multi-speaker).

Recommendations for the authors:

Reviewing Editor Comments:

The reviewers have provided a number of recommendations to improve the manuscript, particularly requesting that more data be reported, with an emphasis on the measurements themselves (eye movements and TRFs) rather than just the numerical outputs of mathematical models.

We appreciate all the reviewers' and editor’s comments and effort to improve our manuscript. In the revised version we provide interim findings and missing data, updated figures that include an intuitive illustration of the metrics (such as TRFs), and a thoroughly revised discussion section where we focus on the relationship between our observed quantities and theoretical entities. We now offer operationalized definitions of the relevant concepts (“prediction tendency”, “active ocular sensing” and “selective attention”) and suggest how these entities might be related in the context of speech processing, based on the current findings. We are confident that this revision has improved the quality of our paper a lot and we are grateful for all the feedback and suggestions. 

Reviewer #1 (Recommendations for the authors):

(1) Participants had to fixate throughout the tasks. How did the authors deal with large eye movements that violated the instructed fixation?

As described in the Methods section: “Participants were instructed to look at a black fixation cross at the center of a grey screen.” This instruction was not intended to enforce strict fixation but rather to provide a general reference point, encouraging participants to keep their gaze on the grey screen and avoid freely scanning the room or closing their eyes. Unlike trial-based designs, where strict fixation is feasible due to shorter trial durations, this approach did not impose rigid fixation requirements. Consequently, the threshold for "instruction violation" was inherently more flexible, and no additional preprocessing was applied to the gaze vectors.

Fixating for such an extended period of time (1.5 hours?) is hard. Did fixation behavior change over time? Could (fixation) fatigue affect the correlations between eye movements and speech tracking? For example, fatigued participants had to correct their fixation more often and this drives, in part, the negative correlation with comprehension?

Yes, participants spent approximately 2 hours in the MEG, including preparation time (~30 minutes). However, participants were given opportunities to rest their eyes between different parts and blocks of the experiment (e.g., resting state, passive listening, and audiobook blocks), which should help mitigate fatigue to some extent.

That said, we agree that it is an intriguing idea that fatigue could drive the ocular speech tracking effect, with participants potentially needing to correct their gaze more as the experiment progresses. However, our analysis suggests this is unlikely for several reasons:

(1) Cross-validation in encoding models: Ocular speech tracking effects were calculated using a 4-fold cross-validation approach (this detail has now been added to the Methods section; please see our response to public review #3). This approach reduces the influence of potential increases in gaze corrections over time, as the models are trained and validated on independent data splits.  Moreover, if there were substantial differences in underlying response magnitudes between folds - for instance, between the first and fourth fold - this would likely compromise the TRF's ability to produce valid response functions for predicting the left-out data. Such a scenario would not result in significant tracking, further supporting the robustness of the observed effects.

(2) TRF time-course stability: If fatigue were driving increased gaze corrections, we would expect this to be reflected in a general offset (capturing the mean difference between folds) in the TRF time-courses shown in Figure 4 (right panel). However, no such trend / offset is evident.

(3) Comparison of eye movement data: To directly investigate this possibility, we compared the amount of total eye movements between the first and last blocks for both the single and multi-speaker conditions. Total movement was calculated by first calculating the differences in pixel values between consecutive eye positions on both the x- and y-axes. The Euclidean distance was then computed for each difference, providing a measure of movement between successive time points. Summing these distances yielded the total movement for each block. Statistical analysis was performed separately for the single speaker (ASS) and multi-speaker (AMS) conditions. For each condition, paired comparisons were made between the first and last blocks (we resorted to non-parametric tests, if assumptions of normality were violated):

For the single speaker condition (ASS), the normality assumption was not satisfied (p≤0.05p, Kolmogorov-Smirnov test). Consequently, a Wilcoxon signedrank test was conducted, which revealed no significant difference in total movements between the first and last blocks (z=−1.330, p=0.184). For the multi-speaker condition (AMS), the data met the normality assumption (p>0.05), allowing the use of a paired t-test. The results showed no significant difference in total movements between the first and last blocks (t=−0.184, p=0.855).

The results are visualized in a bar plot (see below), where individual data points are displayed alongside the mean and standard error for each block. Statistical annotations indicate that neither condition demonstrated significant differences between the blocks. These findings suggest that total eye movements remained stable across the experimental conditions, regardless of whether participants were exposed to a single or multiple speakers.

Author response image 1.

(4) Behavioral responses: Participants’ behavioral responses did not indicate any decrease in comprehensibility for later blocks compared to earlier ones. Specifically, a comparison of comprehension scores between the first and last blocks revealed no significant difference in either the single-speaker condition (ASS; Wilcoxon signed-rank test Z=−0.5911, p=0.5545) or the multi-speaker condition (AMS; Wilcoxon signed-rank test: Z=0.5018, p=0.6158). These findings suggest that participants maintained consistent levels of comprehension throughout the experiment, regardless of the condition or block order. The results are visualized in a bar plot (see below), where individual data points are displayed alongside the mean and standard error for each block. Statistical annotations indicate that neither condition demonstrated significant differences between the blocks.

Author response image 2.

Together, these factors suggest that fatigue is unlikely to be a significant driver of the ocular speech tracking effects observed in this study.

(2) The authors should provide descriptive statistics of fixation behavior /fixational eye movements. What was the frequency and mean direction of microsaccades, do they follow the main sequence, etc., quantify drift and tremor?

Thank you for their suggestion regarding descriptive statistics. To address this, we computed the rates of microsaccades (which were extracted using the microsaccade detection algorithm as proposed in Liu, B., Nobre, A. C. & van Ede, F. Functional but not obligatory link between microsaccades and neural modulation by covert spatial attention. Nat. Commun. 13, 3503 (2022)) and fixations as these metrics are directly relevant to our study and the requests above.

Microsaccade Rates:

- Single speaker Condition: Mean = 2.306 Hz, SD = 0.363 Hz. ○ Multi speaker: Mean = 2.268 Hz, SD = 0.355 Hz.

Fixation Rates:

- Single speaker Condition: Mean = 2.858 Hz, SD = 1.617 Hz. ○ Multi speaker Condition: Mean = 2.897 Hz, SD = 1.542 Hz.

These values fall within the expected ranges reported in the literature (fixation rates: 2– 4 Hz, microsaccade rates: ~0.5–2.5 Hz) and serve as a sanity check, confirming the plausibility of our eye-tracking data. Regarding the reviewer’s request for additional metrics (e.g., microsaccade direction, main sequence analysis, drift, and tremor), extracting these features would require advanced algorithms and analyses not supported by our current preprocessing pipeline or dataset. We hope that the provided metrics, which were the main focus of this study, serve as a sufficient sanity check and highlight the robustness of our data.

Related to this, I am wondering whether microsaccades are the feature that drives speech tracking.

This is an important and pressing question that we aim to address in future publications. Currently, our understanding - and the reason microsaccades and blinks are not analysed in this manuscript - is limited by methodological constraints. Specifically, microsaccades are binary response vectors, which are not compatible with TRF analyses. Addressing this would require adapting future models to handle timecontinuous binary response data or exploring alternative approaches, such as regression-based ERFs (for example as in Heilbron et al. 2022). As the primary goal of this manuscript was to replicate the findings of Gehmacher et al. (2024) using similar methods and to integrate these findings into an initial unified framework, we did not investigate additional eye movement features here. However, we agree that microsaccades (and also blinks, see below) likely contribute, at least in part, to the observed ocular speech tracking effects, and we now suggest this in the Discussion:  

“Relatedly, it remains an open question whether microsaccades are a key feature driving ocular speech tracking. However, our current study does not analyze microsaccades due to methodological constraints: microsaccades are binary response vectors, which are incompatible with TRF analyses used here. Addressing this would require adapting models to handle time-continuous binary response data or potentially exploring alternative approaches, such as regression-based ERFs (e.g., as in Heilbron et al., 2022). While these limitations preclude microsaccade analysis in the current study, we hypothesize that they could enhance temporal precision and selectively amplify relevant sensory input, supporting auditory perception. Future studies should explore this possibility to uncover the specific contributions of microsaccades to speech tracking.”

(3) Can the authors make sure that interpolated blinks did not drive any of the effects? Can interpolated blink trials be excluded?

Using continuous audiobooks as stimuli meant that we could not exclude blink periods from the analysis without introducing substantial continuation artifacts in the TRF analysis. Importantly, the concept of covert motor routines and active sensing suggests that participants engage more strongly in motor routines - including ocular behaviors such as microsaccades and blinks - during tasks like speech tracking. These motor routines are inherently tied to individual gaze patterns, making microsaccades and blinks correlated with other ocular behaviors. This complicates efforts to disentangle their individual contributions to the observed ocular speech tracking effects.

Engagement in these motor routines, as posited by active sensing, would naturally load onto various viewing behaviors, further intertwining their roles.

Even if we were to examine correlations, such as the amount of blinks with the ocular speech tracking effect, it is unlikely to provide a clearer understanding due to these inherent overlaps. The methodological and conceptual challenge lies in distinguishing these features from one another and understanding their respective roles in driving the observed effects.

However, the aim of this manuscript was not to dissect the ocular speech tracking effect in greater detail, but rather to relate it - based on similar analytical choices as in Gehmacher et al - to prediction tendencies, attention, and neural speech tracking. While it will be crucial in future work to differentiate these patterns and their connections to diverse cognitive processes, it is beyond the scope of this study to address all these questions comprehensively.

We acknowledge that eye movements, including microsaccades and blinks (however, see challenges for this in response 2), remain underexplored in many experimental paradigms. Their interplay with cognitive processes - such as attention, prediction, and sensory integration - will undoubtedly be an important focus for future studies. 

(4) Could the authors provide more details on how time shuffling was done for the eyemovement predictor, and include a circularly shifted version (or a version that does not destroy temporal contiguity) in their model comparisons? Some types of shuffling can result in unrealistic time series, which would end up in an unfair comparison with the model that has the real eye movement traces as predictors.

We thank the reviewer for their insightful question regarding the time-shuffling procedure for the eye-movement predictor and for suggesting the inclusion of a circularly shifted version in our model comparisons. Below, we provide further details about our approach and the rationale behind it:

(1) Random Shuffling: In our analysis, the eye-movement predictor was randomly shuffled over time, meaning that individual samples were randomly replaced. This method completely disrupts the temporal structure of the signal, providing a null model that directly tests whether the temporal mediation observed is due to the specific temporal relationship between ocular movements and envelope tracking.

(2) Circular Shifting: While circular shifting maintains temporal contiguity, it introduces certain challenges in the context of TRF analysis. Specifically:

- Adaptation to Shifts: The TRF model could adapt to the introduced shift, potentially reducing the validity of the null comparison.

- Similarity due to Repetition: The broadband envelope exhibits strong repetitive patterns over time, such as rhythms inherent to speech. Circular shifting can therefore produce predictors that are very similar to the original signal. As a result, this similarity may lead to null distributions that do not adequately disrupt the temporal mediation we aim to test, making it less robust as a control.

(3) Rationale for Random Shuffling: The primary goal of our mediation analysis is to determine whether there is a temporal mediation of envelope tracking by ocular movements. By deliberately destroying the temporal structure through random shuffling, we ensure that the null model tests for the specific temporal relationship that is central to our hypothesis. Circularly shifted predictors, on the other hand, may partially preserve temporal dependencies, making them less suitable for this purpose.

In summary, while circular shifting is a valuable approach in other contexts, it is less appropriate for the specific goals of this study. We hope this explanation clarifies our methodological choices and demonstrates their alignment with the aims of our analysis.

(5) Replication: I want to point out that it is great that the previous findings were in principle replicated. However, I would like to suggest a more nuanced evaluation of the replication:

a) Instead of a (direct) replication, the present study should be called a 'conceptual replication', since modifications in design and procedure were made.

Thank you very much for this suggestion! We now use the term ‘conceptual replication’ throughout the manuscript.

b) Not all the findings from the Gehmacher et al., 2024 study were replicated to a full extent:

Did the authors find indications of a vertical vs. horizontal tracking difference in the Gehmacher 2024 data? Could they check this in the Gehmacher 2024 data?

The findings for horizontal and vertical gaze tracking in Gehmacher et al. (2024) are detailed in the supplementary material of that publication. Both single-speaker and multi-speaker target conditions showed significant speech tracking effects in both horizontal and vertical directions. However, there was a slightly stronger tracking effect for the single-speaker condition in the vertical direction. Due to the highly predictable structure of words in Gehmacher et al. effects here were probably overall boosted as compared to continuous audiobook listening, likely leading to the differentiation of horizontal and vertical gaze. See figures in Gehmacher et al. supplementary file for reference.

c) Another difference between their previous and this study is the non-existent tracking of the multi-speaker distractor in this study. The authors should point this out clearly in the discussion and potentially provide an explanation.

Thank you for highlighting this point! We now address this in the discussion:

“Importantly, in contrast to Gehmacher et al. (2024), we did not observe ocular tracking of the multi-speaker distractor in this study. This difference is likely attributable to the simplistic single-trial, 5-word task structure in Gehmacher et al., which resulted in high temporal overlap between the target and distractor speech streams and likely drove the significant distractor-tracking effects observed in that study. The absence of such an effect during continuous listening in our study suggests that ocular tracking is indeed more specific to selective attention.”

Minor:

(1) I was a little surprised to not see an indication of eyes/eye movements in Figure 6. The intention of the authors might have been to create a general schematic illustration, but I find this a bit misleading. This paper provides nice evidence for a specific ocular effect in speech tracking. There is, to my knowledge, no indication that speech would be influenced by different kinds of active sensing (if there are, please include them in the discussion). Given that the visuomotor system is quite dominant in humans, it might actually be the case that the speech tracking the authors describe is specifically ocular.

Taking into account all the reviewers' remarks on the findings and interpretations, we have updated this figure (now Fig. 7) in the manuscript to make it more specific and aligned with the revised discussion section. Throughout the manuscript, we now explicitly refer to active ocular sensing in relation to speech processing and have avoided the broader term 'active sensing' in this context. We hope these revisions address the concerns raised.

(2) I find the part in the discussion (page 2, last paragraph) on cognitive processes hard to follow. I don't agree that 'cognitive processes' are easily separable from any of the measured responses (eye and brain). Referring to the example they provide, there is evidence that eye movements are correlated with brain activity that is correlated with memory performance. How, and more importantly, why would one separate those?

Thank you for raising this important point. We have carefully considered your comments, particularly regarding the interplay between cognitive processes and measured responses (eye and brain), as well as the challenge of conceptually separating them. Additionally, we have incorporated Reviewer #2's query (13) into a unified and complementary reasoning. In response, we have rewritten the relevant paragraph in the discussion to provide a clearer and more detailed explanation of how ocular and neural responses contribute to speech processing in an interdependent manner. We hope this revision addresses your concerns and offers a more precise and coherent discussion on this topic:

“Despite the finding that eye movements mediate neural speech tracking, the behavioural relevance for semantic comprehension appears to differ between ocular and neural speech tracking. Specifically, we found a negative association between ocular speech tracking and comprehension, indicating that participants with lower comprehension performance exhibited increased ocular speech tracking. Interestingly, no significant relationship was observed between neural tracking and comprehension.

In this context, the negative association between ocular tracking and comprehension might reflect individual differences in how participants allocate cognitive resources. Participants with lower comprehension may rely more heavily on attentional mechanisms to process acoustic features, as evidenced by increased ocular tracking. This reliance could represent a compensatory strategy when higher-order processes, such as semantic integration or memory retrieval, are less effective. Importantly, our comprehension questions (see Experimental Procedure) targeted a broad range of processes, including intelligibility and memory, suggesting that this relationship reflects a trade-off in resource allocation between low-level acoustic focus and integrative cognitive tasks.

Rather than separating eye and brain responses conceptually, our analysis highlights their complementary contributions. Eye movements may enhance neural processing by increasing sensitivity to acoustic properties of speech, while neural activity builds on this foundation to integrate information and support comprehension. Together, these systems form an interdependent mechanism, with eye and brain responses working in tandem to facilitate different aspects of speech processing.

This interpretation is consistent with the absence of a difference in ocular tracking for semantic violations (e.g., words with high surprisal versus lexically matched controls), reinforcing the view that ocular tracking primarily reflects attentional engagement with acoustic features rather than direct involvement in semantic processing. This aligns with previous findings that attention modulates auditory responses to acoustic features (e.g., Forte et al., 2017), further supporting the idea that ocular tracking reflects mechanisms of selective attention rather than representations of linguistic content.

Future research should investigate how these systems interact and explore how ocular tracking mediates neural responses to linguistic features, such as lexical or semantic processing, to better understand their joint contributions to comprehension.”.  

(3) Attention vs. predictive coding. I think the authors end up with an elegant description of the observed effects, "as an "active sensing" mechanism that implements the attentional optimization of sensory precision." However, I feel the paragraph starts with the ill-posed question "whether ocular speech tracking is modulated not by predictive, but other (for example attentional) processes". If ocular tracking is the implementation of a process (optimization of sensory precision, aka attention), how could it be at the same time modulated by that process? In my opinion, adding the notion that there is a modulation by a vague cognitive concept like attention on top of what the paper shows does not improve our understanding of how speech tracking in humans works.

Thank you for raising this point. We agree that it is critical to clarify the relationship between ocular speech tracking, attention, and predictive processes, and we appreciate the opportunity to refine this discussion.  

To avoid the potential confusion that active ocular sensing represents on the one hand an implementation of selective attention on the other it seems to be modulated by it, we now use  the formulation “ocular speech tracking reflects attentional mechanisms rather than predictive processes.”

To address your concern that the conceptualization of attention seems rather vague, we have revised the whole paragraph in order to redefine the theoretical entities in question (including selective attention) and to provide a clearer and more precise picture (see also our revised version of Fig. 6, now Fig. 7). We now focus on highlighting the distinct yet interdependent roles of selective attention and individual prediction tendencies for speech tracking.:

“With this speculative framework we attempt to describe and relate three important phenomena with respect to their relevance for speech processing: 1) “Anticipatory predictions” that are created in absence of attentional demands and contain probabilistic information about stimulus features (here, inferred from frequency-specific pre-activations during passive listening to sound sequences). 2) “Selective attention” that allocates resources towards relevant (whilst suppressing distracting) information (which was manipulated by the presence or absence of a distractor speaker). And finally 3) “active ocular sensing”, which refers to gaze behavior that is temporally aligned to attended (but not unattended) acoustic speech input (inferred from the discovered phenomenon of ocular speech tracking). We propose that auditory inflow is, at a basic level, temporally modulated via active ocular sensing, which “opens the gates” in the sensory periphery at relevant timepoints. How exactly this mechanism is guided (for example where the information about crucial timepoints comes from, if not from prediction, and whether it requires habituation to a speechstream etc.) is yet unclear. Unlike predictive tendencies, active ocular sensing appears to reflect selective attention, manifesting as a mechanism that optimizes sensory precision. Individual differences with respect to anticipatory predictions on the other hand, seem to be independent from the other two entities, but nevertheless relevant for speech processing. We therefore support the notion that representational content is interpreted based on prior probabilistic assumptions. If we consider the idea that “a percept” of an (auditory) object is actually temporally and spatially distributed (across representational spacetime - see Fig. 7), the content of information depends on where and when it is probed (see for example Dennett, 1991 for similar ideas on consciousness). Having to select from multiple interpretations across space and time requires a careful balance between the weighting of internal models and the allocation of resources based on current goals. We suggest that in the case of speech processing, this challenge results in an independent adaptation of feature-based precision-weighting by predictions on the one hand and temporal precision-weighting by selective attention on the other.”

Reviewer #2 (Recommendations for the authors):

My main recommendation is outlined in the Weaknesses above: the overarching rationale for many analysis choices should be made explicit, and intermediate results should be shown where appropriate, so the reader can follow what is being quantified and what the results truly mean. Specifically, I recommend to pay attention to the following (in no particular order):

(1) Define 'neural speech tracking' early on. (e.g.: 'The amount of information in the MEG signal that can multivariately be explained by the speech amplitude envelope.' (is that correct?))

Thank you for pointing out that this important definition is missing. It is now defined at the first mention in the Introduction as follows: “Here (and in the following) “neural speech tracking” refers to a correlation coefficient between actual brain responses and responses predicted from an encoding model based solely on the speech envelope”.

(2) Same for 'ocular speech tracking'. Here even reading the Methods does not make it unambiguous how this term is used.

It is now defined at the first mention in the Introduction as follows: “Ocular speech tracking” (similarly to “neural speech tracking” refers to the correlation coefficient between actual eye movements and movements predicted from an encoding model based on the speech envelope”.

In addition also define both (neural and ocular speech tracking) metrics in the Methods Section.

(3) Related to this: for ocular speech tracking, are simply the horizontal and vertical eye traces compared to the speech envelope? If so, this appears somewhat strange: why should the eyes move more rightward/upward with a larger envelope? And the direction here depends on the (arbitrary) sign of right = positive, etc. (It would make more sense to quantify 'amount of movement' in some way, but if this is done, I missed it in Methods.)

Thank you for your insightful comments. You are correct that the horizontal and vertical traces were used for ocular speech tracking, and no additional details were included in the Methods. While we agree that the observed rightward/upward movement may seem unusual, this pattern is consistent with previous findings, including those reported in Gehmacher et al. (2024). In that study, we discussed how ocular speech tracking could reflect a broader engagement of the motor system during speech perception. For example, we observed a general right-lateralized gaze bias when participants attended to auditory speech, which we hypothesized might resemble eye movements during text reading, with a similar temporal alignment (~200 ms). We also speculated that this pattern might differ in cultures that read text from right to left.

We appreciate your suggestion to explore alternative methods for quantifying gaze patterns, such as the "amount of movement" or microsaccades. While these approaches hold promise for future studies, our primary aim here was to replicate previous findings using the same signal and analysis methods to establish a basis for further exploration.  

(4) In the Introduction, specifically blink-related ocular activity is mentioned as being related to speech tracking (for which a reference is, incidentally, missing), while here, any blink-related activity is excluded from the analysis. This should be motivated, as it appears in direct contradiction.

Thank you for pointing this out. The mention of blink-related ocular activity in the Introduction refers to findings by Jin et al. (2018), where such activity was shown to align with higher-order syntactic structures in artificial speech. We have now included the appropriate reference for clarity.

While Jin et al. focused on blink-related activity, in the present study, we focused on gaze patterns to investigate ocular speech tracking, replicating findings from

Gehmacher et al. (2024). This approach was motivated by our goal to validate previous results using the same methodology. Importantly to this point, the exclusion of blinks in our analysis was due to methodological constraints of TRF analysis, which requires a continuous response signal; blinks, being discrete and artifact-prone, are incompatible with this approach.

To address your concern, we revised the Introduction to clarify this distinction and provide explicit motivation for focusing on gaze patterns. It now reads:

“Along these lines, It has been shown that covert, mostly blink related eye activity aligns with higher-order syntactic structures of temporally predictable, artificial speech (i.e. monosyllabic words; Jin et al, 2018). In support of ideas that the motor system is actively engaged in speech perception (Galantucci et al., 2006; Liberman & Mattingly, 1985), the authors suggest a global entrainment across sensory and (oculo)motor areas which implements temporal attention. 

In another recent study from our lab (Gehmacher et al., 2024), we showed that eye movements continuously track intensity fluctuations of attended natural speech, a phenomenon we termed ocular speech tracking. In the present study, we focused on gaze patterns rather than blink-related activity, both to replicate findings from

Gehmacher et al. (2024) and because blink activity is unsuitable for TRF analysis due to its discrete and artifact-prone nature. Hence, “Ocular speech tracking” (similarly to “neural speech tracking” refers to the correlation coefficient between actual eye movements and movements predicted from an encoding model based on the speech envelope.”

Jin, P., Zou, J., Zhou, T., & Ding, N. (2018). Eye activity tracks task-relevant structures during speech and auditory sequence perception. Nature communications, 9(1), 5374.

(5) The rationale for the mediation analysis is questionable. Let speech envelope = A, brain activity = B, eye movements = C. The authors wish to claim that A -> C -> B. But it is equally possible that A -> B -> C. They reflect on this somewhat in Discussion, but throughout the rest of the paper, the mediation analysis is presented as specifically testing whether A -> B is mediated by C, which is potentially misleading.

Indeed we share your concern regarding the directionality of the relationships in the mediation analysis. Our choice of ocular movements as a mediator was motivated by the fact that the relationship between acoustic speech and neural activity is well established, as well as previous results indicating that oculomotor activity contributes to cognitive effects in auditory attention (Popov et al., 2022). 

Indeed, here we treat both interpretations (“ocular movements contribute to neural speech tracking” versus “neural activity contributes to ocular speech tracking”) as equal.  We now emphasise this point in our discussion quite thoroughly:

“It is important to note that our current findings do not allow for inference on directionality. Our choice of ocular movements as a mediator was motivated by the fact that the relationship between acoustic speech and neural activity is well established, as well as previous results indicating that oculomotor activity contributes to cognitive effects in auditory attention (Popov et al., 2022). However, an alternative model may suggest that neural activity mediates the effect of ocular speech tracking. Hence, it is possible that ocular mediation of speech tracking may reflect a) active (ocular) sensing for information driven by (top-down) selective attention or b) improved neural representations as a consequence of temporally aligned increase of sensory gain or c) (not unlikely) both. In fact, when rejecting the notion of a single bottom-up flow of information and replacing it with a model of distributed parallel and dynamic processing, it seems only reasonable to assume that the direction of communication (between our eyes and our brain) will depend on where (within the brain) as well as when we look at the effect. Thus, the regions and time-windows reported here should be taken as an illustration of oculo-neural communication during speech processing rather than an attempt to "explain" neural speech processing by ocular movements.”

(6) The mediation analysis can be improved by a proper quantification of the effect (sizes or variance explained). E.g. how much % of B is explained by A total, and how much of that can in turn be explained by C being involved? For drawing directional conclusions perhaps Granger causality could be used.

In Figure 4 (now Figure 5) of our manuscript we use standardized betas (which correspond to effect sizes) to illustrate the mediation effect. With the current mTRF approach it is however not possible (or insightful) to compare the variance explained. It is reasonable to assume that variance in neural activity will be explained better when including oculomotor behavior as a second predictor along with acoustic simulation. However this increase gives no indication to what extent this oculomotor behavior was task relevant or irrelevant (since all kinds of “arbitrary” movements will be captured with brain activity and therefore lead to an increase in variance explained). For this reason we chose to pursue the widely accepted framework of mediation (Baron & Kenny, 1986). This (correlational) approach is indeed limited in its interpretations (see prev. response), however the goal of the current study was to replicate and illustrate the triad relationship of acoustic speech input, neural activity and ocular movements with no particular hypotheses on directionality.

(7) Both prediction tendency and neural speech tracking depend on MEG data, and thus on MEG signal-to-noise ratio (SNR). It is possible some participants may have higher SNR recordings in both tasks, which may result in both higher (estimated) prediction tendency and higher (estimated) speech tracking. This would result in a positive correlation, as the authors observe. This trivial explanation should be ruled out, by quantifying the relative SNR and testing for the absence of a mediation here.

We agree that for both approaches (MVPA and mTRF models) individual MEG SNR plays an important role. This concern has been raised previously and addressed in our previous manuscript (Schubert et al., 2023). First, it should be noted that our prediction tendency value is the result of a condition contrast (rather than simple decoding accuracy) which compensates for the influence of subject specific signal-to-noise ratio (as no vacuous difference in SNR is to be expected between conditions). Second, in our previous study we also used frequency decoding accuracy as a control variable to correlate with speech tracking variables of interest and found no significant effect.

(8) Much of the analysis pipeline features temporal response functions (TRFs). These should be shown in a time-resolved manner as a key intermediate step.

We now included the Neural Speech tracking TRFs into the Figure (now Figure 3).

(9) Figure 2 shows much-condensed results from different steps in the pipeline. If I understand correctly, 2A shows raw TRF weights (averaged over some time window?), while 2B-F shows standardized mean posterior regressor weights after Bayesian stats? It would be very helpful to make much more explicit what is being shown here, in addition to showing the related TRFs.

Thank you for pointing this out! The figure description so far has been indeed not very insightful on this issue. We now adapted the caption and hope this clarifies the confusion: “ Neural speech tracking is related to prediction tendency and word surprisal, independent of selective attention. A) Envelope (x) - response (y) relationships are estimated using deconvolution (Boosting). The TRF (filter kernel, h) models how the brain processes the envelope over time. This filter is used to predict neural responses via convolution. Predicted responses are correlated with  actual neural activity to evaluate model fit and the TRF's ability to capture response dynamics. Correlation coefficients from these models are then used as dependent variables in Bayesian regression models. (Panel adapted from Gehmacher et al., 2024b). B) Temporal response functions (TRFs) depict the time-resolved neural tracking of the speech envelope for the single speaker and multi speaker target condition, shown here as absolute values averaged across channels. Solid lines represent the group average. Shaded areas represent 95% Confidence Intervals. C–H) The beta weights shown in the sensor plots are derived from Bayesian regression models in A). For Panel C, this statistical model is based on correlation coefficients computed from the TRF models (further details can be found in the Methods Section). C) In a single speaker condition, neural tracking of the speech envelope was significant for widespread areas, most pronounced over auditory processing regions. D) The condition effect indicates a decrease in neural speech tracking with increasing noise (1 distractor). E) Stronger prediction tendency was associated with increased neural speech tracking over left frontal areas. F) However, there was no interaction between prediction tendency and conditions of selective attention. G) Increased neural tracking of semantic violations was observed over left temporal areas. H) There was no interaction between word surprisal and speaker condition, suggesting a representation of surprising words independent of background noise. Marked sensors indicate ‘significant’ clusters, defined as at least two neighboring channels showing a significant result. N = 29.”

Gehmacher, Q., Schubert, J., Kaltenmaier, A., Weisz, N., & Press, C. (2024b). The "Ocular Response Function" for encoding and decoding oculomotor related neural activity. bioRxiv, 2024-11.

(10) Bayesian hypothesis testing is not done consistently. Some parts test for inclusion of 0 in 94% HDI, while some parts adopt a ROPE approach. The same approach should be taken throughout. Additionally, Bayes factors would be very helpful (I appreciate these depend on the choice of priors, but the default Bambi priors should be fine).

Our primary aim in this study was to replicate two recent findings: (1) the relationship between individual prediction tendencies and neural speech tracking, and (2) the tracking of the speech envelope by eye movements. To maintain methodological consistency with the original studies, we did not apply a ROPE approach when analyzing these replication effects. Instead, we followed the same procedures as the original work, focusing on the inclusion of 0 in the HDI for the neural effects and using the same methods for the ocular effects. Additionally, we were not specifically interested in potential null effects in these replication analyses, as our primary goal was to test whether we could reproduce the previously reported findings.

For the mediation analysis, however, we chose to extend the original approach by not only performing the analysis in a time-resolved manner but also applying a ROPE approach. This decision was motivated by our interest in gaining more comprehensive insights — beyond the replication goals — by also testing for potential null effects, which can provide valuable information about the presence or absence of mediation effects.

We appreciate your thoughtful feedback and hope this clarifies our rationale for the differing approaches in our Bayesian hypothesis testing. 

Regarding Bayes Factors, 

We understand that some researchers find Bayes Factors appealing, as they offer a seemingly simple and straightforward way to evaluate the evidence in favor of/ or against H0 in relation to H1 (e.g. BF10 > 102 =  Decisive; according to the Jeffreys Scale). However, in practice Bayes Factors are often misunderstood e.g. by interpreting Bayes Factor as posterior odds or not acknowledging the notion of relative evidence in the Bayes Factor (see Wong et al. 2022). Instead of using Bayes Factors, we prefer to rely on estimating and reporting the posterior distribution of parameters given the data, prior and model assumptions (in form of the 94% HDI). This allows for a continuous evaluation of evidence for a given hypothesis that is in our eyes easier to interpret as a Bayes Factor.

Jeffreys, Harold (1998) [1961]. The Theory of Probability (3rd ed.). Oxford, England. p. 432. ISBN 9780191589676.

Wong, T. K., Kiers, H., & Tendeiro, J. (2022). On the Potential Mismatch Between the Function of the Bayes Factor and Researchers’ Expectations. Collabra: Psychology, 8(1), 36357. https://doi.org/10.1525/collabra.36357

(11) It would be helpful if Results could be appreciated without a detailed read of Methods. I would recommend a recap of each key methodological step before introducing the relevant Result. (This may also help in making the rationale explicit.)

In addition to the short recaps of methods that were already present, and information on quantifications of neural and ocular tracking and bayes statistics (see responses 1, 2, 9), we now added the following parts below to the results sections. Please refer to them in the context of the manuscript where they should now complement a key recap of methodological steps necessary to readily understand each analysis and rational that led to the results:

Individual prediction tendency is related to neural speech tracking:

“Thus, this measure is a single value per subject, which comprises a) differences between two contextual probabilities (i.e. ordered vs. random) in b) feature-specific tone representations c) in advance of their observation (summed over a time-window of -0.3 - 0 s). Importantly, this prediction tendency was assessed in an independent entropy modulation paradigm (see Fig. 1). On a group level we found an increased tendency to pre-activate a stimulus of high probability (i.e. forward transition) in an ordered context compared to a random context (see Fig, 2A). This effect replicates results from our previous work (Schubert et al., 2023, 2024). Using the summed difference between entropy levels (ordered - random) across pre-stimulus time, one value was extracted per subject (Fig. 2B). This value was used as a proxy for “individual prediction tendency” and correlated with encoding of clear speech across different MEG sensors. [...]

Neural speech tracking, quantified as the correlation coefficients between predicted and observed MEG responses to the speech envelope, was used as the dependent variable in Bayesian regression models. These models included condition (single vs. multi-speaker) as a fixed effect, with either prediction tendency or word surprisal as an additional predictor, and random effects for participants.”

Eye movements track acoustic speech in selective attention:

“For this, we separately predicted horizontal and vertical eye movements from the acoustic speech envelope using temporal response functions (TRFs). The resulting model fit (i.e. correlation between true and predicted eye movements) is commonly referred to as “speech tracking”. Bayesian regression models were applied to evaluate tracking effects under different conditions of selective attention (single speaker, attended multi-speaker, unattended multi-speaker). Furthermore, we assessed whether individual prediction tendency or semantic word surprisal influenced ocular speech tracking.”

Neural speech tracking is mediated by eye movements:

“This model evaluates to what extent gaze behaviour functions as a mediator between acoustic speech input and brain activity.”

Neural and ocular speech tracking are differently related to comprehension: “Bayesian regression models were used to investigate relationships between neural/ocular speech tracking and comprehension or difficulty. Ocular speech tracking was analyzed separately for horizontal and vertical eye movements.”

(12) The research questions in the Introduction should be sharpened up, to make explicit when a question concerns a theoretical entity, and when it concerns something concretely measured/measurable.

We sharpened them up:

“Taking into account the aforementioned study by Schubert and colleagues (2023), the two recently uncovered predictors of neural tracking (individual prediction tendency and ocular tracking) raise several empirical questions regarding the relationship between predictive processes, selective attention, and active ocular sensing in speech processing:

(1) Are predictive processes related to active ocular sensing in the same way they are to neural speech tracking? Specifically, do individuals with a stronger tendency to anticipate predictable auditory features, as quantified through prestimulus neural representations in an independent tone paradigm, show increased or even decreased ocular speech tracking, measured as the correlation between predicted and actual eye movements? Or is there no relationship at all?

(2) To what extent does selective attention influence the relationship between prediction tendency, neural speech tracking, and ocular speech tracking? For example, does the effect of prediction tendency or ocular speech tracking on neural tracking differ between a single-speaker and multi-speaker listening condition?

(3) Are individual prediction tendency and ocular speech tracking related to behavioral outcomes, such as comprehension and perceived task difficulty? Speech comprehension is assessed through accuracy on comprehension questions, and task difficulty is measured through subjective ratings.

Although predictive processes, selective attention, and active sensing have been shown to contribute to successful listening, their potential interactions and specific roles in naturalistic speech perception remain unclear. Addressing these questions will help disentangle their contributions and establish an integrated framework for understanding how neural and ocular speech tracking support speech processing.”

(13) The negative relationship between story comprehension and ocular speech tracking appears to go against the authors' preferred interpretation, but the reflection on this in the Discussion is very brief and somewhat vague.

Thank you for pointing this out. We have taken your comments into careful consideration and also incorporated Reviewer #1's query (Minor point 2) into a unified and complementary reasoning. We have rewritten the relevant paragraph in the discussion to provide a clearer and more detailed explanation. We hope this revision offers a more precise and less vague discussion on this important point.

“Despite the finding that eye movements mediate neural speech tracking, the behavioural relevance for semantic comprehension appears to differ between ocular and neural speech tracking. Specifically, we found a negative association between ocular speech tracking and comprehension, indicating that participants with lower comprehension performance exhibited increased ocular speech tracking. Interestingly, no significant relationship was observed between neural tracking and comprehension.

In this context, the negative association between ocular tracking and comprehension might reflect individual differences in how participants allocate cognitive resources. Participants with lower comprehension may rely more heavily on attentional mechanisms to process acoustic features, as evidenced by increased ocular tracking. This reliance could represent a compensatory strategy when higher-order processes, such as semantic integration or memory retrieval, are less effective. Importantly, our comprehension questions (see Experimental Procedure) targeted a broad range of processes, including intelligibility and memory, suggesting that this relationship reflects a trade-off in resource allocation between low-level acoustic focus and integrative cognitive tasks.

Rather than separating eye and brain responses conceptually, our analysis highlights their complementary contributions. Eye movements may enhance neural processing by increasing sensitivity to acoustic properties of speech, while neural activity builds on this foundation to integrate information and support comprehension. Together, these systems form an interdependent mechanism, with eye and brain responses working in tandem to facilitate different aspects of speech processing.

This interpretation is consistent with the absence of a difference in ocular tracking for semantic violations (e.g., words with high surprisal versus lexically matched controls), reinforcing the view that ocular tracking primarily reflects attentional engagement with acoustic features rather than direct involvement in semantic processing. This aligns with previous findings that attention modulates auditory responses to acoustic features (e.g., Forte et al., 2017), further supporting the idea that ocular tracking reflects mechanisms of selective attention rather than representations of linguistic content.

Future research should investigate how these systems interact and explore how ocular tracking mediates neural responses to linguistic features, such as lexical or semantic processing, to better understand their joint contributions to comprehension.”.  

(14) Page numbers would be helpful.

We added the page numbers.

Reviewer #3 (Recommendations for the authors):

Results

(1) Figure 2 - statistical results are reported in this figure, but they are not fully explained in the text, nor are statistical values provided for any of the analyses (as far as I can tell).

Also, how were multiple comparisons dealt with (the choice of two neighboring channels seems quite arbitrary)? Perhaps for this reason, the main result - namely the effect of "prediction tendency" and "semantic violations" - is quite sparse and might not survive more a rigorous statistical criterion. I would feel more comfortable with these results if the reporting of the statistical analysis had been more thorough (ideally, including comparison to control models).

We would like to thank you again for your detailed queries, comments, and questions on our work. We first of all adapted this figure (now Figure 3 in the manuscript, please see responses 8 and 9 to Reviewer #2) to help readers understand the metrics and values within each statistical analysis. In addition, we indeed did not include the detailed statistics in the text! We now added the missing statistic reports calculated as averages over ‘clusters’:

“Replicating previous findings (Schubert et al., 2023), we found widespread encoding of clear speech (average over cluster: β = 0.035, 94%HDI = [0.024, 0.046]), predominantly over auditory processing regions (Fig. 3C), that was decreased (β = -0.018, 94%HDI = [0.029, -0.006]) in a multi-speaker condition (Fig. 3D). Furthermore, a stronger prediction tendency was associated with increased neural speech tracking (β = 0.014, 94%HDI = [0.004, 0.025]) over left frontal sensors (see Fig. 3E). We found no interaction between prediction tendency and condition (see Fig. 3F).” [...] “In a direct comparison with lexically identical controls, we found an increased neural tracking of semantic violations (β = 0.039, 94%HDI = [0.007, 0.071]) over left temporal areas (see Fig. 3G). Furthermore, we found no interaction between word surprisal and speaker condition (see Fig. 3H).”

Regarding the "prediction tendency" effect, it is important to note that this finding replicates a result from Schubert et al. (2023). The left frontal location of this effect is also consistent over studies, which convinces us of the robustness of the finding. Furthermore, testing this relationship properly requires a mixed-effects model in order to account for the variability across subjects that is not explained by fixed effects and the repeated measures design. For this reason a random Intercept had to be fitted for each subject (1|subject in the respective model formula). This statistical requirement motivated our decision to use bayesian statistics as (at least to our knowledge) there is no implementation of a cluster-based permutation mixed effects model (yet). In order to provide a more conservative criterion (as bayesian statistics don’t require a multiple comparison correction) we chose to impose in addition the requirement of a “clustered” effect.

The choice of using two neighboring channels is consistent with the default parameter settings in FieldTrip’s cluster-based permutation testing (cfg.minnbchan = 2). This parameter specifies the minimum number of neighboring channels required for a sample to be included in the clustering algorithm, ensuring spatial consistency in the identified clusters. This alignment ensures that our methodology is comparable to numerous prior studies in the field, where such thresholds are standard. While it is true that all statistical analyses involve some degree of arbitrariness in parameter selection (e.g., alpha levels or clustering thresholds), our approach reflects established conventions and ensures comparability with previous findings.

While the original study utilized source space analyses, we replicated this effect using only 102 magnetometers. This choice was made for computational simplicity, demonstrating that the effect is robust even without source-level modeling. Similarly, the "semantic violation" effect, while perceived as sparse, is based solely on magnetometer data and - in our opinion - should not be viewed as overly sparse given the methods employed. This effect aligns with the two-neighbor clustering approach, ensuring spatial consistency across magnetometers. The results reflect the robustness of the effects within the constraints of magnetometer-level analyses.

Overall, the methodological choices, including the choice of a bayesian linear mixed effects model, the use of two neighboring channels and the reliance on magnetometers, are grounded in established practices and methodological considerations. While stricter thresholds or alternative approaches might yield different results, our methods align with best practices in the field and ensure the robustness, comparability, and replicability of our findings.

(2) Figure 3 - the difference between horizontal and vertical eye-movements. This result is quite confusing and although the authors do suggest a possible interpretation for this in the discussion, I do wonder how robust this difference is or whether the ocular signal (in either direction) is simply too noisy or the effect too small to be detected consistently across conditions. Also, the ocular-TRFs themselves are not entirely convincing in suggesting reliable response/tracking of the audio - despite the small-but-significant increase in prediction accuracy.

The horizontal versus vertical comparison was conducted to explore potential differences in how these dimensions contribute to ocular tracking of auditory stimuli (please also see our response to Reviewer #1, Response 5b, that includes the vertical vs. horizontal effects of Gehmacher at al. 2024). It would indeed be interesting to develop a measure that combines the two directions into a more natural representation of 'viewing,' such as a combined vector. However, this approach would require the use of complex numbers to represent both magnitude and direction simultaneously, hence the development of novel TRF algorithms capable of modeling this multidimensional signal. While beyond the scope of the current study, this presents an exciting avenue for future research and would allow us to move closer to understanding ocular speech tracking and the robustness of these effects, above and beyond the already successful replication.

It is also important to emphasize that ocular-TRFs are derived from (viewing) behavioral data rather than neural signals, and are thus inherently subject to greater variability across participants and time. This higher variability does not necessarily indicate a small or unreliable effect but reflects the dynamic and task-dependent nature of eye movement behavior. The TRFs with shaded error margins represent this variability, highlighting how eye movements are influenced by both individual differences and moment-to-moment changes in task engagement.

Despite this inherent variability, the significant prediction accuracy improvements confirm that ocular-TRFs reliably capture meaningful relationships between eye movements and auditory stimuli. The observed differences between horizontal and vertical TRFs further support the hypothesis that these dimensions are differentially involved in the task, possibly driven by the specific roles they play in sensorimotor coupling.

(3) Figure 4 - this figure shows source distribution of 3 PCA components, derived from the results of the mediation effect of eye movements on the speech-tracking. Here too I am having difficulty in interpreting what the results actually are. For one, all three components are quite widespread and somewhat overlapping, so although they are statistically "independent" it is hard to learn much from them about the brain regions involved and whether they truly represent separable contributions. Similarly difficult to interpret are the time courses, which share some similarities with the known TRFs to speech (especially PC3). I would have expected to find a cleaner "auditory" response, and clearer separation between sensory regions and regions involved in the control of eye movements. I also wonder why the authors chose not to show the sourcelocalization of the neural and ocular speech-tracking responses alone - this could have helped us between understand what "mediation" of the neural response might look like.

We appreciate the reviewer’s interest in better understanding the source distribution and time courses of the PCA components. While we acknowledge that the widespread and overlapping nature of the components may make a more fine grained interpretation challenging, it is important to emphasize that our analysis simply reflects the data, hence we can only present and interpret what the analysis revealed.

Regarding your suggestion to show the source localization of ocular speech tracking and neural speech tracking alone, we would like to point out that ocular tracking is represented by only one channel for vertical and one channel for horizontal eye movements. Thus, in this case the estimated source of the effect are the eyes themselves. We believe that the source localization of neural speech tracking has been a thoroughly studied topic in research so far (locating it to perisylvian, auditory areas with a stronger preference for the left hemisphere) and can also be seen in Schubert et al., (2023). Nevertheless, we believe the observed PCA components still provide valuable, and most importantly novel insights into the interplay between eye movements and neural responses in speech tracking.  

Discussion/interpretation

(1) Although I appreciate the authors' attempt to propose a "unified" theoretical model linking predictions about low-level features to higher features, and the potential involvement of eye movements in 'active sensing' I honestly think that this model is overambitious, given the data presented in the current study. Moreover, there is very little discussion of past literature and existing models of active sensing and hierarchical processing of speech, that could have helped ground the discussion in a broader theoretical context. The entire discussion contains fewer than 20 citations (some of which are by these authors) and needs to be substantially enriched in order to provide context for the authors' claims.

Thank you very much for your thoughtful feedback and for appreciating our approach. We hope that the revised manuscript addresses your concerns. Specifically, we now emphasize that our proposal is a conceptual framework, with the main goal to operationale “prediction tendency”, “active ocular sensing”, and “selective attention” and to “organise these entities according to their assumed function for speech processing and to describe their relationship with each other.” We did this by thoroughly revising our discussion section with a clear emphasis on the definition of terms, for example: 

“With this speculative framework we attempt to describe and relate three important phenomena with respect to their relevance for speech processing: 1) “Anticipatory predictions” that are created in absence of attentional demands and contain probabilistic information about stimulus features (here, inferred from frequency-specific pre-activations during passive listening to sound sequences). 2) “Selective attention” that allocates resources towards relevant (whilst suppressing distracting) information (which was manipulated by the presence or absence of a distractor speaker). And finally 3) “active ocular sensing”, which refers to gaze behavior that is temporally aligned to attended (but not unattended) acoustic speech input (inferred from the discovered phenomenon of ocular speech tracking).”

Our theoretical proposals are now followed by a recap of our results that support the respective idea, for example: 

“...these predictions are formed in parallel and carry high feature-specificity but low temporal precision (as they are anticipatory in nature). This idea is supported by our finding that pure-tone anticipation is visible over a widespread prestimulus interval, instead of being locked to sound onset”

“....we suggest that active (ocular) sensing does not necessarily convey feature- or content-specific information, it is merely used to boost (and conversely filter) sensory input at specific timescales (similar to neural oscillations). This assumption is supported by our finding that semantic violations are not differentially encoded in gaze behaviour than lexical controls.”

And we put a strong focus on highlighting the boundaries of these ideas, in order to avoid theoretical confusion, misunderstandings or implicit theoretical assumption that are not grounded in data, in particular: 

“In fact, when rejecting the notion of a single bottom-up flow of information and replacing it with a model of distributed parallel and dynamic processing, it seems only reasonable to assume that the direction of communication (between our eyes and our brain) will depend on where (within the brain) as well as when we look at the effect. Thus, the regions and time-windows reported here should be taken as an illustration of oculo-neural communication during speech processing rather than an attempt to "explain" neural speech processing by ocular movements.”

“Even though the terminology [“hierarchy”] is suggestive of a fixed sequence (similar to a multi storey building) with levels that must be traversed one after each other (and even the more spurious idea of a rooftop, where the final perceptual experience is formed and stored into memory), we distance ourselves from these (possibly unwarranted) ideas. Our usage of “higher” or “lower” simply refers to the observation that the probability of a feature at a higher (as in more associative) level affects the interpretation (and thus the representation and prediction) of a feature at lower (as in more segregated) levels (Caucheteux et al., 2023).”

Additionally, we have made substantial efforts to present complementary results (see response to Reviewer #2, point 8) to further substantiate our interpretation. Importantly, we have updated the illustration of the model (see response to Reviewer #, minor point 1) and refined both our interpretations and the conceptual language in the Discussion. Furthermore, we have included additional citations where appropriate to strengthen our argument.

We would also like to briefly note that this section of the Discussion aimed to highlight existing literature that bridges the gap our model seeks to address. However, as this is a relatively underexplored area, the references available are necessarily limited.

(2) Given my many reservations about the data, as presented in the current version of the manuscript, I find much of the discussion to be an over-interpretation of the results. This might change if the authors are able to present more robust results, as per some of my earlier comments.

We sincerely hope that our comprehensive revisions have addressed your concerns and improved the manuscript to your satisfaction.

Author Response

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

Reviewer #1 (Public Review):*

The manuscript by Hariani et al. presents experiments designed to improve our understanding of the connectivity and computational role of Unipolar Brush Cells (UBCs) within the cerebellar cortex, primarily lobes IX and X. The authors develop and cross several genetic lines of mice that express distinct fluorophores in subsets of UBCs, combined with immunocytochemistry that also distinguishes subtypes of UBCs, and they use confocal microscopy and electrophysiology to characterize the electrical and synaptic properties of subsets of so-labelled cells, and their synaptic connectivity within the cerebellar cortex. The authors then generate a computer model to test the possible computational functions of such interconnected UBCs.

Using these approaches, the authors report that:

1) GRP-driven TDtomato is expressed exclusively in a subset (20%) of ON-UBCs, defined electrophysiologically (excited by mossy fiber afferent stimulation via activation of UBC AMPA and mGluR1 receptors) and immunocytochemically by their expression of mGluR1.

2) UBCs ID'd/tagged by mCitrine expression in Brainbow mouse line P079 are expressed in a similar minority subset of OFF-UBCs defined electrophysiologically (inhibited by mossy fiber afferent stimulation via activation of UBC mGluR2 receptors) and immunocytochemically by their expression of Calretinin. However, such mCitrine expression was also detected in some mGluR1 positive UBCs, which may not have shown up electrophysiologically because of the weaker fluorophore expression without antibody amplification.

This is correctly stated with the exception that the P079 mouse line itself expresses mCitrine. The Brainbow mouse line was used in the connectivity study by crossing it to the GRP-Cre or Calretinin-Cre lines.

3) Confocal analysis of crossed lines of mice (GRP X P079) stained with antibodies to mGluR1 and calretinin documented the existence of all possible permutations of interconnectivity between cells (ON-ON, ON-OFF, OFF-OFF, OFF-ON), but their overall abundance was low, and neither their absolute nor relative abundance was quantified.

They were certainly rare to observe using our approaches, but we reasoned that the densities of such connections are not possible to estimate accurately. Please see discussion below.

4) A computational model (NEURON ) indicated that the presence of an intermediary UBC (in a polysynaptic circuit from MF to UBC to UBC) could prolong bursts (MF-ON-ON), prolong pauses (MF-ON-OFF), cause a delayed burst (MF-OFF- OFF), cause a delayed pause (MF-OFF-ON) relative to solely MF to UBC synapses which would simply exhibit long bursts (MF-ON) or long pauses (MF-OFF).

The authors thus conclude that the pattern of interconnected UBCs provides an extended and more nuanced pattern of firing within the cerebellar cortex that could mediate longer-lasting sensorimotor responses.

The cerebellum's long-known role in motor skills and reflexes, and associated disorders, combined with our nascent understanding of its role in cognitive, emotional, and appetitive processing, makes understanding its circuitry and processing functions of broad interest to the neuroscience and biomedical community. The focus on UBCs, which are largely restricted to vestibular lobules of the cerebellum reduces the breadth of likely interest somewhat. The overall design of specific experiments is rigorous and the use of fluorophore expressing mouse lines is creative. The data that is presented and the writing are clear. However, the overall experimental design has issues that reduce overall interpretation (please see specific issues for details), which combined with a lack of thorough analysis of the experimental outcomes severely undermines the value of the NEURON model results and the advance in our understanding of cerebellar processing in situ (again, please see specific issues for details).

Specific issues:

1) All data gathered with inhibition blocked. All of the UBC response data (Fig. 1) was gathered in the presence of GABAAR and Glycine R blockers. While such an approach is appropriate generally for isolating glutamatergic synaptic currents, and specifically for examining and characterizing monosynaptic responses to single stimuli, it becomes problematic in the context of assaying synaptic and action potential response durations for long-lasting responses, and in particular for trains of stimuli, when feed-forward and feed-back inhibition modulates responses to afferent stimulation. That is, even for single MF stimuli, given the >500ms duration of UBC synaptic currents, there is plenty of time for feedback inhibition from Golgi cells (or feedforward, from MF to Golgi cell excitation) to interrupt AP firing driven by the direct glutamatergic synaptic excitation. This issue is compounded further for all of the experiments examining trains of MF stimuli. Beyond the impact of feedback inhibition on the AP firing of any given UBC, it would also obviously reduce/alter/interrupt that UBC's synaptic drive of downstream UBCs. This issue fundamentally undermines our ability to interpret the simulation data of Vm and AP firing of both the modeled intermediate and downstream UBC, in terms of applying it to possible cerebellar cortical processing in situ.

The goal of Figure 1 was to determine the cell types of labeled UBCs in transgenic mouse lines, which is determined entirely by their synaptic responses to glutamate (Borges-Merjane and Trussell, 2015). Thus, blocking inhibition was essential to produce clear results in the characterization of GRP and P079 UBCs. While GABAergic/glycinergic feedforward and feedback inhibition is certainly important in the intact circuit, it was not our intention, nor was it possible, to study its contribution in the present study. Leaving inhibition unblocked does not lead to a physiologically realistic stimulation pattern in acute brain slices, because electrical stimulation produces synchronous excitation and inhibition by directly exciting Golgi cells, rather than their synaptic inputs. The main inhibition that UBCs receive that are crucial to determining burst or pause durations is not via GABA/glycine, but instead through mGluR2, which lasts for 100-1000s of milliseconds. The main excitation that drives UBC firing is mGluR1 and AMPA, which both last 100-1000s of milliseconds. Thus, these large conductances are unlikely to be significantly shaped by 1-10 ms IPSCs from feedforward and feedback GABA/glycine inhibition. Recent studies that examined the duration of bursting or pausing in UBCs had inhibition blocked in their experiments, presumably for the reasons outlined above (Guo et al., 2021; Huson et al., 2023).

In Author response image 1 is an example showing the synaptic currents and firing patterns in an ON UBC before and after blocking inhibition. The GABA/glycinergic inhibition is fast, occurs soon after the stimuli and has little to no effect on the slow inward current that develops after the end of stimulation, which is what drives firing for 100s of milliseconds.

Author response image 1.

Example showing small effect of GABAergic and glycinergic inhibition on excitatory currents and burst duration. A) Excitatory postsynaptic currents in response to train of 10 presynaptic stimuli at 50 Hz before (black) and after (Grey) blocking GABA and glycine receptors. The slow inward current that occurs at the end of stimulation is little affected. B) Expanded view of the synaptic currents evoked during the train of stimuli. GABA/glycine receptors mediate the fast outward currents that occur immediately after the first couple stimuli. C) Three examples of the bursts caused by the 50 Hz stimulation in the same cell without blocking GABA and glycine receptors. D) Three examples in the same cell after blocking GABA and glycine receptors.

2) No consideration for the involvement of polysynaptic UBCs driving UBC responses to MF stimulation in electrophysiology experiments. Given the established existence (in this manuscript and Dino et al. 2000 Neurosci, Dino et al. 2000 ProgBrainRes, Nunzi and Mugnaini 2000 JCompNeurol, Nunzi et al. 2001 JCompNeurol) of polysynaptic connections from MFs to UBCs to UBCs, the MF evoked UBC responses established in this manuscript, especially responses to trains of stimuli could be mediated by direct MF inputs, or to polysynaptic UBC inputs, or possibly both (to my awareness not established either way). Thus the response durations could already include extension of duration by polysynaptic inputs, and so would overestimate the duration of monosynaptic inputs, and thus polysynaptic amplification/modulation, observed in the NEURON model.

We are confident that the synaptic responses shown are monosynaptic for several reasons. UBCs receive a single mossy fiber input on their dendritic brush, and thus if our stimulation produces a reliable, short-latency response consistent with a monosynaptic input, then there is not likely to be a disynaptic input, because the main input is accounted for by the monosynaptic response. In all cells included in our data set, the fast AMPA receptor-mediated currents always occurred with short latency (1.24 ± 0.29 ms; mean ± SD; n = 13), high reliability (no failures to produce an EPSC in any of the 13 GRP UBCs in this data set), and low jitter (SD of latency; 0.074 ± 0.046 ms; mean ± SD; n = 13). These measurements have been added to the results section. In some rare cases, we did observe disynaptic currents, which were easily distinguishable because a single electrical stimulation produced a burst of EPSCs at variable latencies. Please see example in Author response image 2. These cases of disynaptic input, which have been reported by others (Diño et al., 2000; Nunzi and Mugnaini, 2000; van Dorp and De Zeeuw, 2015) support the conclusion that UBCs receive input from other UBCs.

Author response image 2.

Example of GRP UBC with disynaptic input. Three examples of the effect of a single presynaptic stimulus (triangle) in a GRP UBC with presumed disynaptic input. Note the variable latency of the first evoked EPSC, bursts of EPSCs, and spontaneous EPSCs.

3) Lack of quantification of subtypes of UBC interconnectivity. Given that it is already established that UBCs synapse onto other UBCs (see refs above), the main potential advance of this manuscript in terms of connectivity is the establishment and quantification of ON-ON, ON-OFF, OFF-ON, and OFF-OFF subtypes of UBC interconnections. But, the authors only establish that each type exists, showing specific examples, but no quantification of the absolute or relative density was provided, and the authors' unquantified wording explicitly or implicitly states that they are not common. This lack of quantification and likely small number makes it difficult to know how important or what impact such synapses have on cerebellar processing, in the model and in situ.

As noted by the reviewer, the connections between UBCs were rare to observe. We decided against attempting to quantify the absolute or relative density of connections for several reasons. A major reason for rare observations of anatomical connections between UBCs is likely due to the sparse labeling. First, the GRP mouse line only labels 20% of ON UBCs and we are unable to test whether postsynaptic connectivity of GRP ON UBCs is the same as that of the rest of the population of ON UBCs that are not labeled in the GRP mouse line. Second, the Brainbow reporter mouse only labels a small population of Cre expressing cells for unknown reasons. Third, the Brainbow reporter expression was so low that antibody amplification was necessary, which then limited the labeled cells to those close to the surface of the brain slices, because of known antibody penetration difficulties. Therefore, we refrained from estimating the density of these connections, because each of these variables reduced the labeling to unknown degrees and we reasoned that extrapolating our rare observations to the total population would be inaccurate.

A paper that investigated UBC connectivity using organotypic slice cultures from P8 mice suggests that 2/3 of the UBC population receives UBC input, based on the observation that 2/3 of the mossy fibers did not degenerate as would be expected after 2 days in vitro if they were severed from a distant cell body (Nunzi and Mugnaini, 2000). It remains to be seen if this high proportion is due to the young age of these mice or is also the case in adult mice. Even if these connections are indeed rare, they are expected to have profound effects on the circuit, as each UBC has multiple mossy fiber terminals (Berthie and Axelrad, 1994), and mossy fiber terminals are estimated to contact 40 granule cells each (Jakab and Hamori, 1988). We have added a comment regarding this point to the discussion.

4) Lack of critical parameters in NEURON model.

A) The model uses # of molecules of glutamate released as the presumed quantal content, and this factor is constant. However, no consideration of changes in # of vesicles released from single versus trains of APs from MFs or UBCs is included. At most simple synapses, two sequential APs alters release probability, either up or down, and release probability changes dynamically with trains of APs. It is therefore reasonable to imagine UBC axon release probability is at least as complicated, and given the large surface area of contact between two UBCs, the number of vesicles released for any given AP is also likely more complex.

B) the model does not include desensitization of AMPA receptors, which in the case of UBCs can paradoxically reduce response magnitude as vesicle release and consequent glutamate concentration in the cleft increases (Linney et al. 1997 JNeurophysiol, Lu et al. 2017 Neuron, Balmer et al. 2021 eLIFE), as would occur with trains of stimuli at MF to ON-UBCs.

A) The model produces synaptic AMPA and mGluR2 currents that reproduce those we recorded in vitro. We did not find it necessary to implement changes in glutamate release during a train as the model was fit to UBC data with the assumption that the glutamate transient did not change during the train. If there is a change in neurotransmitter release during a train, it is therefore built into the model, which has the advantage of reducing its complexity. UBCs are a special case where the postsynaptic currents are mediated mostly by the total amount of transmitter released. Most of the evoked current occurs tens to hundreds of milliseconds after neurotransmitter release and is therefore much more sensitive to total release and less sensitive to how it is released during the train. Author response image 3 shows the effect of reducing the amount of glutamate released by 10% on each stimulus in the model. Despite a significant change in the pattern of neurotransmitter release, as well as a reduction in the total amount of glutamate, the slow EPSC still decays over the course of hundreds of milliseconds.

Author response image 3.

Effect of short-term depression of neurotransmitter release. A) The top trace shows the glutamate transient that drives the AMPA receptor model used in our study. No change in release is implemented, although the slow tail of each transient summates during the train. The bottom trace shows the modeled AMPA receptor mediated current. B) In this model the amount of glutamate released is reduced by 10% on each stimulus. The duration of the slow AMPA current that develops at the end of stimulation is similar, despite a profound change in the pattern of neurotransmitter exposure.

B) The detailed kinetic AMPA receptor model used here accurately reproduces desensitization, and in fact recovery from desensitization is what mediates the slow ON UBC current. This AMPA receptor is a 13-state model, including 4 open states with 1-4 glutamates bound, 4 closed states with 1-4 glutamates bound, 4 desensitized states with 1-4 glutamates bound, and 5 closed states with 0-4 glutamates bound. The forward and reverse rates between different states in the model were fit to AMPA receptor currents recorded from dissociated UBCs and they accurately reproduced the ON UBC currents evoked by synaptic stimulation in our previous work (Balmer et al., 2021).

5) Lack of quantification of various electrophysiological responses. UBCs are defined (ON or OFF) based on inward or outward synaptic response, but no information is provided about the range of the key parameter of duration across cells, which seems most critical to the current considerations. There is a similar lack of quantification across cells of AP duration in response to stimulation or current injections, or during baseline. The latter lack is particularly problematic because, in agreement with previous publications, the raw data in Fig. 1 shows ON UBCs as quiescent until MF stimulation and OFF UBCs firing spontaneously until MF stimulation, but, for example, at least one ON UBC in the NEURON model is firing spontaneously until synaptically activated by an OFF UBC (Fig. 11A), and an OFF UBC is silent until stimulated by a presynaptic OFF UBC (Fig. 11C). This may be expected/explainable theoretically, but then such cells should be observed in the raw data.

To address this reasonable concern of a general lack of quantification of electrophysiological responses we have added data characterizing the slow inward and outward currents evoked by synaptic stimulation in GRP and P079 UBCs in the results section and in new panels in Figure 1. We report the action potential pause lengths in P079 UBCs and burst lengths in ON UBCs in the results section. However, we favor the duration of the currents to the length of burst and pause, because the currents do not depend on a stable resting membrane potential, which is itself difficult to determine in intracellular recordings of these small cells. We have added peak times and decay time constants of the slow inward and outward currents in ON and OFF UBCs in the results section and have added new panels to figure 1.

In a series of recent publications that focused on UBC firing, the authors argue that cell-attached recordings are necessary to determine accurately the burst and pause lengths, as well as spontaneous firing rates (Guo et al., 2021; Huson et al., 2023). (The trade-off of these extracellular recordings is that the monosynaptic nature of the input is nearly impossible to confirm.) Spontaneous firing rates were variable within both GRP and P079 UBCs from silent to firing regularly or in bursts, as previously reported for UBCs (Kim et al., 2012; van Dorp and De Zeeuw, 2015). For clarity, we chose to model the GRP UBCs as silent unless receiving synaptic input and P079 UBCs as active unless receiving synaptic input. As the reviewer suggests, we have observed UBCs firing in the patterns similar to those shown in the model UBCs that have input from a spontaneously active presynaptic UBC. In Author response image 4 are some examples.

Author response image 4.

Examples of UBCs that receive spontaneous input. A) Three ON UBCs that had spontaneous EPSCs, suggesting the presence of an active presynaptic UBC. B) Two OFF UBCs that had spontaneous outward currents.

Reviewer #2 (Public Review):

In this paper, the authors presented a compelling rationale for investigating the role of UBCs in prolonging and diversifying signals. Based on the two types of UBCs known as ON and OFF UBC subtypes, they have highlighted the existing gaps in understanding UBCs connectivity and the need to investigate whether UBCs target UBCs of the same subtype, different subtypes, or both. The importance of this knowledge is for understanding how sensory signals are extended and diversified in the granule cell layer.

The authors designed very interesting approaches to study UBCs connectivity by utilizing transgenic mice expressing GFP and RFP in UBCs, Brainbow approach, immunohistochemical and electrophysiological analysis, and computational models to understand how the feed-forward circuits of interconnected UBCs transform their inputs.

This study provided evidence for the existence of distinct ON and OFF UBC subtypes based on their electrophysiological properties, anatomical characteristics, and expression patterns of mGluR1 and calretinin in the cerebellum. The findings support the classification of GRP UBCs as ON UBCs and P079 UBCs as OFF UBCs and suggest the presence of synaptic connections between the ON and OFF UBC subtypes. In addition, they found that GRP and P079 UBCs form parallel and convergent pathways and have different membrane capacitance and excitability. Furthermore, they showed that UBCs of the same subtype provide input to one another and modify the input to granule cells, which could provide a circuit mechanism to diversify and extend the pattern of spiking produced by mossy fiber input. Accordingly, they suggested that these transformations could provide a circuit mechanism for maintaining a sensory representation of movement for seconds.

Overall, the article is well written in a sound detailed format, very interesting with excellent discovery and suggested model, however, I have some comments/suggestions that may help to improve this manuscript:

• The discovery of UBCs innervating each other and their own subtypes, suggesting the presence of feed-forward networks in the cerebellum, is an incredibly fascinating and exciting finding followed by an intriguing model by authors. However, it is worth considering an alternative model as well. I acknowledge that visualizing such interactions using current tools and methods can be challenging ("The approaches used here were not able to determine the existence of networks of more than 2 UBCs connected one after the other. If present, 3 or more UBCs in series could extend and transform the input in even more dramatic ways. The temporal diversity that UBC circuits generate may underlie the flexibility of the cerebellum to coordinate movements over a broad range of behaviors."). Therefore, if this is the case in which more than 2 UBCs connected one after the other, then an alternative model PERHAPS resembles the basal nuclei, with its direct and indirect circuits, can be considered (maybe a type of circular model). The basal nuclei circuits are also regulated by modulators such as D1 dopamine receptors in the direct pathway, causing depolarization, and D2 dopamine receptors in the indirect pathway, resulting in hyperpolarization upon dopamine activation. This approach could involve using computational models to gain insight into potential alternatives within this pathway (may be a future direction).

Thank you for this suggestion to consider the potentially similar circuit interactions in the basal nuclei. We will certainly investigate this further as we move forward with modeling the feed-forward networks in the cerebellum.

• GRP UBCs are more densely distributed in lobes VI-IX, while P079 UBCs are more densely distributed in the dorsal leaflet of lobe X in sagittal sections. While the cerebellum is well known for its characteristic stripy pattern, are UBC distributions the same in coronal/transverse section?

UBCs of different types, based on their expression of specific proteins, have overlapping but somewhat distinct distributions in coronal sections. The densities of calretinin-expressing UBCs are higher within Zebrin II positive zones and form sagittal stripes, whereas the densities of mGluR1-expressing and PLCb4-expressing UBCs vary less but are in their highest densities at the midline (Chung et al., 2009; Sekerkova et al., 2014). The difference noted by the reviewer between the dorsal and ventral leaflets of lobe X are the most distinct that we know of in the GRP and P079 populations.

• The extension of the axons from both subtypes of UBCs show they are long enough to pass several UBCs and even projections are directed toward the white matter (e.g. Fig 9A), suggesting targeting the UBCs or granule cells in other lobules. Is it suggesting UBCs connectivity between different lobules (perhaps longitudinal connectivity)? Is there any observation or information in coronal/transverse section to visualize mediolateral connectivity?

This is certainly worth exploring in future work. UBCs have been reported to project their axons into and across the white matter (Diño et al., 2000). To our knowledge, whether UBCs project their axons out of one lobule and into another has not been examined.

• The limitation in identifying networks involving more than two sequentially connected UBCs was briefly noted. I suggest including a paragraph describing limitations and discussing the implications of the findings would enhance the overall impact of the research and broaden our understanding of cerebellar function.

• It is a pity that there is no clear conclusion to the discussion of this very interesting study. I suggest providing the key points as a conclusion.

Thank you for these suggestions. Limitations and implications are included throughout the discussion section and we feel that the summary figure and significance statement now sufficiently convey the key conclusions of the study.

• Please make the correction in Figure 2A by relabeling it as IXa, IXb, and IXc to correct the typographical error.

Fixed

• I recommend rotating Figure 7A to align its orientation with the other figures for consistency.

Fixed

Reviewer #1 (Recommendations For The Authors):

Minor comments that should be addressed for clarity:

1) In the NEURON model, why was the reversal potential for the leak conductance and Gmax for Ih different for the two types of UBCs. Relatedly, why is Erev for GABAB -95mV if Ek is -90mV?

The h-current (Ih) was estimated from a hyperpolarizing current step in both cell types and these data have been added to the result section and as a panel in Figure 1. The conductance of Ih in the model cells were adjusted accordingly, with OFF UBCs having ~3 times that of ON UBCs and approximated the measured voltage sag, as we now describe in the methods section. The reversal potential of the model mGluR2 current (which is based on a model of GABAB) has been fixed.

2) Line 69 justification for their dual genetic approach is a bit too strong: "Paired recordings not possible". It may be difficult, but it is certainly possible.

Reworded

3) Confusing wording, only one stat for two parameters? Line 93: These currents were produced by both mGluR1 and AMPA receptors, as they were blocked by their antagonists JNJ16259685 and GYKI53655, respectively (92.86% {plus minus} 3.25; paired t-test; P=0.0066; n = 9; 95 mean {plus minus} SEM) (Fig 1D-E).

Reworded

References

Balmer TS, Borges-Merjane C, Trussell LO (2021) Incomplete removal of extracellular glutamate controls synaptic transmission and integration at a cerebellar synapse. eLife 10:e63819.

Berthie B, Axelrad H (1994) Granular layer collaterals of the unipolar brush cell axon display rosette-like excrescences. A Golgi study in the rat cerebellar cortex. Neuroscience Letters 167:161–165.

Borges-Merjane C, Trussell LO (2015) ON and OFF unipolar brush cells transform multisensory inputs to the auditory system. Neuron 85:1029–1042.

Chung SH, Sillitoe RV, Croci L, Badaloni A, Consalez G, Hawkes R (2009) Purkinje cell phenotype restricts the distribution of unipolar brush cells. Neuroscience 164:1496–1508.

Diño MR, Schuerger RJ, Liu Y-B, Slater NT, Mugnaini E (2000) Unipolar brush cell: a potential feedforward excitatory interneuron of the cerebellum. Neuroscience 98:625–636.

Guo C, Huson V, Macosko EZ, Regehr WG (2021) Graded heterogeneity of metabotropic signaling underlies a continuum of cell-intrinsic temporal responses in unipolar brush cells. Nat Commun 12:5491.

Huson V, Newman LN, Regehr WG (2023) A continuum of response properties across the population of Unipolar Brush Cells in the Dorsal Cochlear Nucleus. J Neurosci Available at: https://www.jneurosci.org/content/early/2023/07/26/JNEUROSCI.0873-23.2023 [Accessed August 15, 2023].

Jakab RL, Hamori J (1988) Quantitative morphology and synaptology of cerebellar glomeruli in the rat. Anatomy and embryology 179:81–88.

Kim JA, Sekerkova G, Mugnaini E, Martina M (2012) Electrophysiological, morphological, and topological properties of two histochemically distinct subpopulations of cerebellar unipolar brush cells. Cerebellum 11:1012–1025.

Nunzi M-G, Mugnaini E (2000) Unipolar brush cell axons form a large system of intrinsic mossy fibers in the postnatal vestibulocerebellum. Journal of Comparative Neurology 422:55–65.

Sekerkova G, Watanabe M, Martina M, Mugnaini E (2014) Differential distribution of phospholipase C beta isoforms and diaglycerol kinase-beta in rodents cerebella corroborates the division of unipolar brush cells into two major subtypes. Brain structure & function 219:719–749.

van Dorp S, De Zeeuw CI (2015) Forward signaling by unipolar brush cells in the mouse cerebellum. Cerebellum 14:528– 533.

Author Response

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

First and foremost, we would like to thank all the editors and reviewers for their thoughtful and thorough evaluations of our manuscript. We greatly appreciate their assessment about the novelty and strength in this study and have revised the manuscript according to their recommendations. Below are our detailed responses and revisions based on the reviewer recommendations.

Reviewer #1 (Recommendations For The Authors):

1) It is unclear the rationale for choosing the P35-42 adolescent window for stimulating the mesofrontal dopamine system.

The dopaminergic innervation in the mesofrontal circuit exhibits a protracted maturation from P21 to P56 (Kalsbeek, Voorn et al. 1988, Niwa, Kamiya et al. 2010, Naneix, Marchand et al. 2012, Hoops and Flores 2017). P35-42 is in the center of this period and captures the mid-adolescent stage in rodents (Spear 2000). We have previously shown that increasing dopamine neuron activity by wheel running or optogenetic stimulation during this period, but not adulthood, can induce formation of mesofrontal dopaminergic boutons and enhance mesofrontal circuit activity in wild-type mice (Mastwal, Ye et al. 2014). We therefore chose the P35-P42 adolescent window to stimulate the mesofrontal dopamine circuit and test the long-term effect of this intervention on the frontal circuit and memory-guided decision-making deficits in mutant mice. We have detailed this rationale in the revised manuscript when we first introduced this intervention.

2). Please provide a justification for choosing the optical recording M2 neuronal activity instead of the prelimbic prefrontal cortex, which has been known to show the highest levels of dopamine terminals.

While the prelimbic area has the highest level of dopamine terminals among frontal cortical regions, a robust presence of dopaminergic terminals and dopamine release in the M2 frontal cortex have been well documented (Berger, Gaspar et al. 1991, Mastwal, Ye et al. 2014, Aransay, Rodriguez-Lopez et al. 2015, Patriarchi, Cho et al. 2018). The M2 cortex plays an important role in action planning, generating the earliest neural signals among frontal cortical regions that are related to upcoming choice during spatial navigation (Sul, Kim et al. 2010, Sul, Jo et al. 2011). Our chemogenetic inactivation experiments (Supplementary Fig 1) has further confirmed the involvement of M2 in the memory-guided Y-maze navigation task used in this study. Technically, M2 has the advantage of being more amendable to optical recording of neuronal activity without the tissue damage caused by implanting a lens, which would be necessary for deeper areas such as the prelimbic cortex. We have provided this justification in the revised manuscript.

3). What was the rationale for using the 3-day chemogenetic stimulation paradigm?

Our previous work in wild-type adolescent mice showed that a single optogenetic stimulation session or a 2-hr wheel running session is sufficient to induce bouton formation in mesofrontal dopaminergic axons (Mastwal, Ye et al. 2014). In this study, we sought to rescue existing structural and functional deficits in the mesofrontal dopaminergic circuits due to genetic mutations. Because previous studies suggested that an optimal level of dopamine is important for normal cognitive function (Arnsten, Cai et al. 1994, Robbins 2000, Floresco 2013), we elected to do multiple stimulation sessions to boost the potential rescue effects. We tested both a 3-day and a 3-week stimulation paradigm, and found that the 3-day, but not the 3-week paradigm led to robust functional improvement (Fig. 5). These results indicate that moderate but not excessive stimulation of dopamine neurons can provide functional improvement of a deficient mesofrontal circuit. We have revised our text to clarify the rationale for these experiments.

4). A major maturational event occurring in the prefrontal cortex is the gain of local GABAergic transmission, which is crucial for sustaining proper levels of Y-maze tasks. I am wondering if the authors have any thoughts about what is really happening at the postsynaptic level following adolescent dopamine stimulation.

The developmental increases in dopaminergic innervation to the frontal cortex and local GABAergic transmission are likely synergistic processes, which both contribute to the maturation of high-order cognitive functions supported by the frontal cortex (Caballero and Tseng 2016, Larsen and Luna 2018). Previous electrophysiological studies have suggested that dopamine can act on five different receptors expressed in both excitatory and inhibitory postsynaptic neurons (Seamans and Yang 2004, Tseng and O'Donnell 2007, O'Donnell 2010). At the network level, dopaminergic signaling can increase the signal-to-noise ratio and temporal synchrony of neural activity during cognitive tasks (Rolls, Loh et al. 2008, Vander Weele, Siciliano et al. 2018, Lohani, Martig et al. 2019). As the frontal GABAergic inhibitory network undergoes major functional remodeling during adolescence (Caballero and Tseng 2016), adolescent stimulation of dopamine neurons may interact with this maturational process to promote a network configuration conducive for synchronous and high signal-to-noise neural computation (Porter, Rizzo et al. 1999, Murty, Calabro et al. 2016, Mukherjee, Carvalho et al. 2019). The microcircuit mechanisms underlying adolescent dopamine stimulation induced changes, particularly in the GABAergic inhibitory neurons, will be an exciting direction for future research. We have extended our discussion about these points in the revised manuscript.

5). A change in the density of dopamine boutons is unlikely to be limited to the M2 region in Arc-/- mice. The authors should provide some data illustrating that similar changes are widespread across the medial prefrontal cortex, and that the optical recording in the M2 region was preferred for technical limitations and to avoid damaging areas in the frontal cortex.

As discussed above, this study focused on the M2 region of the frontal cortex because it is functionally required for memory-guided Y-maze navigation, generates behavioral choice-related neural signals during spatial navigation, and is optically most accessible. The medial prefrontal regions (anterior cingulate, prelimbic and infralimbic) ventral to M2 also receive dense dopaminergic innervation and can act in concert with M2 in decision making (Sul, Kim et al. 2010, Sul, Jo et al. 2011, Barthas and Kwan 2017). As dopaminergic innervations to the frontal cortical regions progress in a ventral-to-dorsal direction during development (Kalsbeek, Voorn et al. 1988, Hoops and Flores 2017), how the changes induced by adolescent dopamine stimulation may proceed spatial-temporally across different frontal subregions requires more extensive investigation in the future. We have added this discussion into the revised manuscript.

Reviewer #2 (Public Review):

The manuscript by Mastwal and colleagues explores how transient adolescent stimulation of ventral midbrain neurons that project to the frontal cortex may help to improve performance on certain memory tasks. The manuscript provides an interesting set of observations that DREADD-based activation over only 3 days during adolescence provides a fast-acting and long-lasting improvement in performance on Y-maze spontaneous alternation as well as aspects of neuronal function as assessed using in vivo imaging methods. While interesting, there are several weaknesses. First and foremost, it is not clear that the effects the authors are observing are mediated by dopamine. It has been clearly documented that the DAT-Cre line provides a better representation of midbrain dopamine cells in the mouse, particularly near the midline of the ventral midbrain (Lammel et al., Neuron 2015). This is precisely where the cells that project to the frontal cortex are located. Therefore, the selection of TH-Cre is problematic. It is very likely that the authors are labeling a substantial number of non-dopaminergic cells.

We agree with Review 2 that the DAT-Cre line can provide specific labeling of midbrain dopamine neurons, particularly those projecting to the striatum, as discussed in the cited study (Lammel, Steinberg et al. 2015). DAT transports the extracellularly released dopamine back into presynaptic terminals, but it is not essential for dopamine synthesis and release (Sulzer, Cragg et al. 2016). Mesocortical dopamine neurons in the ventral tegmental area (VTA) express very little DAT (Sesack, Hawrylak et al. 1998, Lammel, Hetzel et al. 2008, Li, Qi et al. 2013), which limits the use of the DAT-Cre line to target these neurons (Lammel, Steinberg et al. 2015). Because mesocortical dopamine neurons have strong expression of TH, a key enzyme involved in dopamine synthesis, TH-Cre lines have been extensively used to study the mesocortical pathway (Lammel, Lim et al. 2012, Gunaydin, Grosenick et al. 2014, Ellwood, Patel et al. 2017, Vander Weele, Siciliano et al. 2018, Lohani, Martig et al. 2019). We provide more details below about our rationales for using TH-Cre rather than DAT-Cre mice in our study and the revisions we made in response to the reviewer’s specific recommendations.

Reviewer #2 (Recommendations For The Authors):

1). The authors should rigorously demonstrate that there is a reasonable midbrain DA projection to the coordinates that they are assessing and that their effects are due to DA release from these cells. It is not clear that there is a VTA dopaminergic projection to M2 - it does not appear for example in the Allen Mouse Brain Connectivity Atlas (https://connectivity.brainmap.org/projection/experiment/siv/160540751? imageId=160541123&imageType=TWO_PHOTON,SEGMENTATION&initImage=TWO_PHOTON&x=17321&y=15284&z=3). Though there is a projection to the mPFC, at the coordinates the authors report, there does not appear to be any signal from DAT-Cre mice. However, there is much more signal when expression is not restricted to dopamine cells (https://connectivity.brain-map.org/projection/experiment/siv/165975096? imageId=165975158&imageType=TWO_PHOTON,SEGMENTATION&initImage=TWO_PHOTON&x=17950&y=11504&z=3). The argument that these cells may express less TH is not relevant for this particular issue. Therefore, it is possible that the vast majority of observed effects are not in fact mediated by dopamine but another neurotransmitter such as glutamate. While the experiment using SCH23390 does suggest DA receptors may be involved, this result in isolation doesn't alleviate this caveat - there can be, for example, DA release from NE cells (e.g., Takeuchi et al., Nature 2016). While this does not entirely invalidate the authors' results, as their effects of stimulation of ventral midbrain cells to the forebrain don't necessarily have to occur via dopamine - the mechanism by how this is occurring needs to be clear.

While the prelimbic area has the highest level of dopaminergic terminals among frontal cortical regions, a robust presence of midbrain dopaminergic projections and dopamine release in the M2 frontal cortex have been well established by immunostaining, viral labeling, single-cell axon-tracing, and in vivo imaging of recently developed dopamine biosensors (Berger, Gaspar et al. 1991, Mastwal, Ye et al. 2014, Aransay, Rodriguez-Lopez et al. 2015, Ye, Mastwal et al. 2017, Patriarchi, Cho et al. 2018). It has also been reported repeatedly that mesocortical dopamine neurons in the VTA express very little DAT, which is different from mesostriatal dopamine neurons (Sesack, Hawrylak et al. 1998, Lammel, Hetzel et al. 2008, Li, Qi et al. 2013). This limitation in the use of the DAT-Cre line to target mesocortical dopamine neurons has been acknowledged in previous studies (Lammel, Steinberg et al. 2015) and is consistent with the reviewer’s observation of DAT-Cre labeling in the Allen Brain Mouse Connectivity atlas. Additionally, and interestingly, recent extensive evaluation of the DAT-Cre line reported ectopic labeling of multiple non-dopaminergic neuronal populations (Soden, Miller et al. 2016, Stagkourakis, Spigolon et al. 2018, Papathanou, Dumas et al. 2019). Our own evaluation of the DAT-Cre line’s utility for cortical imaging also revealed sparse axonal labeling and sporadic ectopic labeling of cortical cell somas. We have included representative DAT-Cre images in Author response image 1 to highlight the limitations of this line in the study of the dopaminergic mesocortical circuit.

Author response image 1.

Example images from DAT-Cre/Ai14 mice. Left most panel shows little axonal labeling in Layer 5/6 of M2. The center panel shows sparse axonal label in Layer 1/2 of M2, but also ectopic labeling of cell soma. The right panel shows a lack of labeling in L1/2 of prelimbic cortex as well. Scale bars 50um.

We as well as others have confirmed that TH immunoreactivity in the frontal cortex can label dopaminergic axons originated from the VTA, and ablation of VTA dopaminergic neurons removes this labeling (Niwa, Jaaro-Peled et al. 2013, Ye, Mastwal et al. 2017). Because mesocortical dopamine neurons have much stronger TH expression than DAT expression (Sesack, Hawrylak et al. 1998, Lammel, Hetzel et al. 2008, Li, Qi et al. 2013, Lammel, Steinberg et al. 2015), TH-Cre lines have been frequently used to label these neurons and study the mesocortical pathway (Lammel, Lim et al. 2012, Gunaydin, Grosenick et al. 2014, Ellwood, Patel et al. 2017, Vander Weele, Siciliano et al. 2018, Lohani, Martig et al. 2019). While TH-Cre expression itself is not restricted to dopaminergic neurons, we targeted our viral injections to the VTA and optogenetic stimulation to the cortical dopaminergic projection target area in M2 (Patriarchi, Cho et al. 2018) to specifically modulate mesofrontal dopaminergic axons. In addition, we tested D1 antagonist’s effects in our manipulations. Although we targeted dopamine neurons in our adolescent stimulation, the final behavioral outcome likely includes contributions from co-released neurotransmitters such as glutamate and non-dopaminergic neurons via network effects (Morales and Margolis 2017, Lohani, Martig et al. 2019), which will be interesting directions for future research. We have revised our results and discussion sections to highlight our rationales for using the TH-Cre line and the open mechanistic questions for future studies.

2) SSFOs don't increase excitability like DREADDs, but rather, cause long-lasting hyperactivity through continuous passage of cations. What the actual firing properties are of these cells over a long period of time is not clear.

We did not measure the precise firing patterns of the dopaminergic neurons targeted by SSFOs but evaluated the effects of SSFO activation on the frontal cortex. Similar to our DREADD-Gq mediated activity changes in the mesofrontal circuit, we found increased frontal cortical activity post-light stimulation of frontal dopamine axons in our SSFO treated animals (Fig 6a-c, S6e). While quantitatively the firing patterns of DREADD-Gq and SSFO activated dopaminergic neurons likely differ, qualitatively both of these manipulations lead to increased mesofrontal circuit activity and improvements in cognitive behaviors. In our previous work with wild-type adolescent mice, both wheel running and a single 10-min session of phasic optogenetic stimulation of the VTA resulted in dopaminergic bouton outgrowth in the frontal cortex (Mastwal, Ye et al. 2014). Taken together, these results suggest that adolescent dopaminergic mesofrontal projections are highly responsive to neural activity changes and a variety of adolescent stimulation paradigms are sufficient to elicit lasting changes in this circuit. We have added this discussion of the limitations and implications of our study into the revised manuscript.

3) It is not clear what the increase in boutons means, given that DA release is thought to largely occur via non-synaptic release.

Although many of dopamine boutons are not associated with defined postsynaptic structures, these axonal boutons and the active zones they contain are the major release sites for dopamine (Goldman-Rakic, Leranth et al. 1989, Arbuthnott and Wickens 2007, Sulzer, Cragg et al. 2016, Liu, Goel et al. 2021). Past studies have established a consistent association between increased dopaminergic innervation in the frontal cortex and an increase in dopamine levels (Niwa, Kamiya et al. 2010, Naneix, Marchand et al. 2012). Our previous work also found that increasing dopaminergic boutons through adolescent VTA stimulation led to prolonged frontal local field potential responses with high-frequency oscillations (Mastwal, Ye et al. 2014), which is characteristic of increased dopaminergic signaling (Lewis and O'Donnell 2000, Gireesh and Plenz 2008, Wood, Kim et al. 2012, Lohani, Martig et al. 2019). Importantly, in our quantification of the structural changes in this study, we evaluated boutons which were labeled with synaptophysin, a molecular marker indicating the presence of synaptic vesicle release machinery (Li, Tasic et al. 2010, Oh, Harris et al. 2014). Thus, our study, taken in the context of the previous work, suggests the increased number of boutons signifying an increase in dopaminergic signaling within the mesofrontal circuit. We have added this discussion into the revised manuscript.

4) The use of Arc and DISC mutants as models of schizophrenia is perhaps a bit overstated - while deficits in prefrontal innervation certainly occur, there are many differences between these models and the human disease states. Language should be toned down accordingly, particularly in the introduction.

We strived to avoid overstating the extent to which the mouse lines are models for specific diseases, but we can appreciate that this may not have been clear in our original writing. We have adjusted our language to better distinguish between the utility of the animal models for the purposes of our study and their relationship to specific human disease states. Particularly in the introduction, we stated that: “Genetic disruptions of several genes involved in synaptic functions related to psychiatric disorders, such as Arc and DISC1, lead to hypoactive mesofrontal dopaminergic input in mice (Niwa, Kamiya et al. 2010, Niwa, Jaaro-Peled et al. 2013, Fromer, Pocklington et al. 2014, Purcell, Moran et al. 2014, Wen, Nguyen et al. 2014, Manago, Mereu et al. 2016). Although there are many differences between these mouse lines and specific human disease states, these mice offer opportunities to test whether genetic deficits in frontal cortex function can be reversed through circuit interventions.”

5) Some experiments are missing proper controls, e.g., Figure 3g-I where a WT mouse should be used as a positive control.

The goal of this experimental design (Fig 3g-i) was to evaluate the potential effects of chemogenetic VTA stimulation in the Arc-/- mice. We used Arc-/- mice with mCherry injections to control for the potential effects of CNO administration. While WT mice could be used to determine if adolescent VTA stimulation would lead to long-lasting enhancement of VTA-to-Cortical transmission, this wouldn’t necessarily be a positive control for our experiments, but rather a separate line of inquiry. As dopamine’s effects often display an inverted-U dose-response curve (Vijayraghavan, Wang et al. 2007, Floresco 2013), evaluating the effects adolescent VTA stimulation in the absence of underlying dopamine deficiency could be an interesting future research direction. We have added this discussion into the revised manuscript.

Reviewer #3 (Recommendations For The Authors):

1) Did the SSFO stimulation of the TH+ axons in PFC during adolescence lead to the same long-term change in DA bouton number the authors saw with DREADDs?

We did not examine the degree of bouton growth in the SSFO cohort, which is a limitation of this study. Accurate quantification of dopamine boutons requires the co-injection of another AAV vector encoding Synaptophysin-GFP to label the boutons. Because we used light to directly stimulate SSFO-labeled dopaminergic axons in the frontal cortex, we were concerned that co-injecting another AAV vector may dilute SSFO-labeling of axons and reduce the efficacy of optogenetic stimulation. Given the behavioral benefits we observed, we would expect an increase in bouton density after optogenetic stimulation. A systematic optimization of viral co-labeling and optogenetic stimulation protocols will facilitate examination of the impact of SSFO stimulation at the structural level in future studies. We have added a discussion of the limitation of this study in the revised manuscript.

2) The DISC1 section is far less detailed than the Arc section, and it was not completely clear to me that the mechanisms of dysfunction and rescue were the same in these mice compared with the Arc mice. For example, there was no mention of DA bouton density or the patterned firing of the PFC neurons at the time of decision making.

The initial motivation of this study was to test if adolescent dopamine stimulation can rescue the deficits in the mesofrontal dopaminergic circuit and cognitive function of Arc-/- mice, which were identified in our previous studies (Manago, Mereu et al. 2016). We first conducted multiple levels of analyses including viral tracing, in vivo calcium imaging, and behavioral tests to establish the coherent impacts of adolescent dopamine neuron stimulation on circuits and behaviors. We then examined a range of stimulation protocols to assess the efficacy requirements for cognitive improvement, which is our primary goal. Finally, we included DISC1 mice in our study to test if adolescent dopamine stimulation can also reverse the cognitive deficit in another genetic model for mesofrontal dopamine deficiency. By demonstrating a similar cognitive recuse effect of adolescent VTA stimulation in an independent mouse model, this study provides a foundation for future research to compare the detailed cellular mechanisms that underlie the functional rescue in different genetic models. We have added the discussion of the scope and limitation of this study to the revised manuscript.

References

Aransay, A., C. Rodriguez-Lopez, M. Garcia-Amado, F. Clasca and L. Prensa (2015). "Long-range projection neurons of the mouse ventral tegmental area: a single-cell axon tracing analysis." Front Neuroanat 9: 59.

Arbuthnott, G. W. and J. Wickens (2007). "Space, time and dopamine." Trends Neurosci 30(2): 62-69.

Arnsten, A. F., J. X. Cai, B. L. Murphy and P. S. Goldman-Rakic (1994). "Dopamine D1 receptor mechanisms in the cognitive performance of young adult and aged monkeys." Psychopharmacology (Berl) 116(2): 143-151.

Barthas, F. and A. C. Kwan (2017). "Secondary motor cortex: where ‘sensory’meets ‘motor’in the rodent frontal cortex." Trends in neurosciences 40(3): 181-193.

Berger, B., P. Gaspar and C. Verney (1991). "Dopaminergic innervation of the cerebral cortex: unexpected differences between rodents and primates." Trends Neurosci 14(1): 21-27.

Caballero, A. and K. Y. Tseng (2016). "GABAergic Function as a Limiting Factor for Prefrontal Maturation during Adolescence." Trends Neurosci 39(7): 441-448.

Ellwood, I. T., T. Patel, V. Wadia, A. T. Lee, A. T. Liptak, K. J. Bender and V. S. Sohal (2017). "Tonic or Phasic Stimulation of Dopaminergic Projections to Prefrontal Cortex Causes Mice to Maintain or Deviate from Previously Learned Behavioral Strategies." J Neurosci 37(35): 8315-8329.

Floresco, S. B. (2013). "Prefrontal dopamine and behavioral flexibility: shifting from an "inverted-U" toward a family of functions." Front Neurosci 7: 62.

Fromer, M., A. J. Pocklington, D. H. Kavanagh, H. J. Williams, S. Dwyer, P. Gormley, L. Georgieva, E. Rees, P. Palta, D. M. Ruderfer, N. Carrera, I. Humphreys, J. S. Johnson, P. Roussos, D. D. Barker, E. Banks, V. Milanova, S. G. Grant, E. Hannon, S. A. Rose, K. Chambert, M. Mahajan, E. M. Scolnick, J. L. Moran, G. Kirov, A. Palotie, S. A. McCarroll, P. Holmans, P. Sklar, M. J. Owen, S. M. Purcell and M. C. O'Donovan (2014). "De novo mutations in schizophrenia implicate synaptic networks." Nature 506(7487): 179-184.

Gireesh, E. D. and D. Plenz (2008). "Neuronal avalanches organize as nested theta- and beta/gamma-oscillations during development of cortical layer 2/3." Proc Natl Acad Sci U S A 105(21): 7576-7581.

Goldman-Rakic, P. S., C. Leranth, S. M. Williams, N. Mons and M. Geffard (1989). "Dopamine synaptic complex with pyramidal neurons in primate cerebral cortex." Proc Natl Acad Sci U S A 86(22): 9015-9019.

Gunaydin, L. A., L. Grosenick, J. C. Finkelstein, I. V. Kauvar, L. E. Fenno, A. Adhikari, S. Lammel, J. J. Mirzabekov, R. D. Airan, K. A. Zalocusky, K. M. Tye, P. Anikeeva, R. C. Malenka and K. Deisseroth (2014). "Natural neural projection dynamics underlying social behavior." Cell 157(7): 1535-1551.

Hoops, D. and C. Flores (2017). "Making Dopamine Connections in Adolescence." Trends Neurosci 40(12): 709-719.

Kalsbeek, A., P. Voorn, R. M. Buijs, C. W. Pool and H. B. Uylings (1988). "Development of the dopaminergic innervation in the prefrontal cortex of the rat." J Comp Neurol 269(1): 58-72.

Lammel, S., A. Hetzel, O. Hackel, I. Jones, B. Liss and J. Roeper (2008). "Unique properties of mesoprefrontal neurons within a dual mesocorticolimbic dopamine system." Neuron 57(5): 760-773.

Lammel, S., A. Hetzel, O. Haeckel, I. Jones, B. Liss and J. Roeper (2008). "Unique properties of mesoprefrontal neurons within a dual mesocorticolimbic dopamine system." Neuron 57(5): 760-773.

Lammel, S., B. K. Lim, C. Ran, K. W. Huang, M. J. Betley, K. M. Tye, K. Deisseroth and R. C. Malenka (2012). "Input-specific control of reward and aversion in the ventral tegmental area." Nature 491(7423): 212-217.

Lammel, S., E. E. Steinberg, C. Foldy, N. R. Wall, K. Beier, L. Luo and R. C. Malenka (2015). "Diversity of transgenic mouse models for selective targeting of midbrain dopamine neurons." Neuron 85(2): 429-438.

Larsen, B. and B. Luna (2018). "Adolescence as a neurobiological critical period for the development of higher-order cognition." Neurosci Biobehav Rev 94: 179-195.

Lewis, B. L. and P. O'Donnell (2000). "Ventral tegmental area afferents to the prefrontal cortex maintain membrane potential 'up' states in pyramidal neurons via D(1) dopamine receptors." Cereb Cortex 10(12): 1168-1175.

Li, L., B. Tasic, K. D. Micheva, V. M. Ivanov, M. L. Spletter, S. J. Smith and L. Luo (2010). "Visualizing the distribution of synapses from individual neurons in the mouse brain." PLoS One 5(7): e11503.

Li, X., J. Qi, T. Yamaguchi, H. L. Wang and M. Morales (2013). "Heterogeneous composition of dopamine neurons of the rat A10 region: molecular evidence for diverse signaling properties." Brain Struct Funct 218(5): 1159-1176.

Liu, C., P. Goel and P. S. Kaeser (2021). "Spatial and temporal scales of dopamine transmission." Nat Rev Neurosci 22(6): 345-358.

Lohani, S., A. K. Martig, K. Deisseroth, I. B. Witten and B. Moghaddam (2019). "Dopamine Modulation of Prefrontal Cortex Activity Is Manifold and Operates at Multiple Temporal and Spatial Scales." Cell Rep 27(1): 99-114 e116.

Manago, F., M. Mereu, S. Mastwal, R. Mastrogiacomo, D. Scheggia, M. Emanuele, M. A. De Luca, D. R. Weinberger, K. H. Wang and F. Papaleo (2016). "Genetic Disruption of Arc/Arg3.1 in Mice Causes Alterations in Dopamine and Neurobehavioral Phenotypes Related to Schizophrenia." Cell Rep 16(8): 2116-2128.

Mastwal, S., Y. Ye, M. Ren, D. V. Jimenez, K. Martinowich, C. R. Gerfen and K. H. Wang (2014). "Phasic dopamine neuron activity elicits unique mesofrontal plasticity in adolescence." J Neurosci 34(29): 9484-9496.

Morales, M. and E. B. Margolis (2017). "Ventral tegmental area: cellular heterogeneity, connectivity and behaviour." Nat Rev Neurosci 18(2): 73-85.

Mukherjee, A., F. Carvalho, S. Eliez and P. Caroni (2019). "Long-Lasting Rescue of Network and Cognitive Dysfunction in a Genetic Schizophrenia Model." Cell 178(6): 1387-1402 e1314. Murty, V. P., F. Calabro and B. Luna (2016). "The role of experience in adolescent cognitive development: Integration of executive, memory, and mesolimbic systems." Neurosci Biobehav Rev 70: 46-58.

Naneix, F., A. R. Marchand, G. Di Scala, J. R. Pape and E. Coutureau (2012). "Parallel maturation of goal-directed behavior and dopaminergic systems during adolescence." J Neurosci 32(46): 16223-16232.

Niwa, M., H. Jaaro-Peled, S. Tankou, S. Seshadri, T. Hikida, Y. Matsumoto, N. G. Cascella, S. Kano, N. Ozaki, T. Nabeshima and A. Sawa (2013). "Adolescent stress-induced epigenetic control of dopaminergic neurons via glucocorticoids." Science 339(6117): 335-339.

Niwa, M., A. Kamiya, R. Murai, K. Kubo, A. J. Gruber, K. Tomita, L. Lu, S. Tomisato, H. Jaaro-Peled, S. Seshadri, H. Hiyama, B. Huang, K. Kohda, Y. Noda, P. O'Donnell, K. Nakajima, A. Sawa and T. Nabeshima (2010). "Knockdown of DISC1 by in utero gene transfer disturbs postnatal dopaminergic maturation in the frontal cortex and leads to adult behavioral deficits." Neuron 65(4): 480-489.

O'Donnell, P. (2010). "Adolescent maturation of cortical dopamine." Neurotox Res 18(3-4): 306-312.

Oh, S. W., J. A. Harris, L. Ng, B. Winslow, N. Cain, S. Mihalas, Q. Wang, C. Lau, L. Kuan, A. M. Henry, M. T. Mortrud, B. Ouellette, T. N. Nguyen, S. A. Sorensen, C. R. Slaughterbeck, W. Wakeman, Y. Li, D. Feng, A. Ho, E. Nicholas, K. E. Hirokawa, P. Bohn, K. M. Joines, H. Peng, M. J. Hawrylycz, J. W. Phillips, J. G. Hohmann, P. Wohnoutka, C. R. Gerfen, C. Koch, A. Bernard, C. Dang, A. R. Jones and H. Zeng (2014). "A mesoscale connectome of the mouse brain." Nature 508(7495): 207-214.

Papathanou, M., S. Dumas, H. Pettersson, L. Olson and A. Wallen-Mackenzie (2019). "Off-Target Effects in Transgenic Mice: Characterization of Dopamine Transporter (DAT)-Cre Transgenic Mouse Lines Exposes Multiple Non-Dopaminergic Neuronal Clusters Available for Selective Targeting within Limbic Neurocircuitry." eNeuro 6(5).

Patriarchi, T., J. R. Cho, K. Merten, M. W. Howe, A. Marley, W. H. Xiong, R. W. Folk, G. J. Broussard, R. Liang, M. J. Jang, H. Zhong, D. Dombeck, M. von Zastrow, A. Nimmerjahn, V. Gradinaru, J. T. Williams and L. Tian (2018). "Ultrafast neuronal imaging of dopamine dynamics with designed genetically encoded sensors." Science 360(6396): 1420-+.

Porter, L. L., E. Rizzo and J. P. Hornung (1999). "Dopamine affects parvalbumin expression during cortical development in vitro." J Neurosci 19(20): 8990-9003.

Purcell, S. M., J. L. Moran, M. Fromer, D. Ruderfer, N. Solovieff, P. Roussos, C. O'Dushlaine, K. Chambert, S. E. Bergen, A. Kahler, L. Duncan, E. Stahl, G. Genovese, E. Fernandez, M. O. Collins, N. H. Komiyama, J. S. Choudhary, P. K. Magnusson, E. Banks, K. Shakir, K. Garimella, T. Fennell, M. DePristo, S. G. Grant, S. J. Haggarty, S. Gabriel, E. M. Scolnick, E. S. Lander, C. M. Hultman, P. F. Sullivan, S. A. McCarroll and P. Sklar (2014). "A polygenic burden of rare disruptive mutations in schizophrenia." Nature 506(7487): 185-190.

Robbins, T. W. (2000). "Chemical neuromodulation of frontal-executive functions in humans and other animals." Exp Brain Res 133(1): 130-138.

Rolls, E. T., M. Loh, G. Deco and G. Winterer (2008). "Computational models of schizophrenia and dopamine modulation in the prefrontal cortex." Nat Rev Neurosci 9(9): 696-709.

Seamans, J. K. and C. R. Yang (2004). "The principal features and mechanisms of dopamine modulation in the prefrontal cortex." Prog Neurobiol 74(1): 1-58.

Sesack, S. R., V. A. Hawrylak, C. Matus, M. A. Guido and A. I. Levey (1998). "Dopamine axon varicosities in the prelimbic division of the rat prefrontal cortex exhibit sparse immunoreactivity for the dopamine transporter." J Neurosci 18(7): 2697-2708.

Soden, M. E., S. M. Miller, L. M. Burgeno, P. E. M. Phillips, T. S. Hnasko and L. S. Zweifel (2016). "Genetic Isolation of Hypothalamic Neurons that Regulate Context-Specific Male Social Behavior." Cell Rep 16(2): 304-313.

Spear, L. (2000). "Modeling adolescent development and alcohol use in animals." Alcohol Res Health 24(2): 115-123.

Stagkourakis, S., G. Spigolon, P. Williams, J. Protzmann, G. Fisone and C. Broberger (2018). "A neural network for intermale aggression to establish social hierarchy." Nat Neurosci 21(6): 834-842. Sul, J. H., S. Jo, D. Lee and M. W. Jung (2011). "Role of rodent secondary motor cortex in value-based action selection." Nat Neurosci 14(9): 1202-1208.

Sul, J. H., H. Kim, N. Huh, D. Lee and M. W. Jung (2010). "Distinct roles of rodent orbitofrontal and medial prefrontal cortex in decision making." Neuron 66(3): 449-460.

Sulzer, D., S. J. Cragg and M. E. Rice (2016). "Striatal dopamine neurotransmission: regulation of release and uptake." Basal Ganglia 6(3): 123-148.

Tseng, K. Y. and P. O'Donnell (2007). "Dopamine modulation of prefrontal cortical interneurons changes during adolescence." Cereb Cortex 17(5): 1235-1240.

Vander Weele, C. M., C. A. Siciliano, G. A. Matthews, P. Namburi, E. M. Izadmehr, I. C. Espinel, E. H. Nieh, E. H. S. Schut, N. Padilla-Coreano, A. Burgos-Robles, C. J. Chang, E. Y. Kimchi, A. Beyeler, R. Wichmann, C. P. Wildes and K. M. Tye (2018). "Dopamine enhances signal-to-noise ratio in cortical-brainstem encoding of aversive stimuli." Nature 563(7731): 397-401.

Vijayraghavan, S., M. Wang, S. G. Birnbaum, G. V. Williams and A. F. Arnsten (2007). "Inverted-U dopamine D1 receptor actions on prefrontal neurons engaged in working memory." Nat Neurosci 10(3): 376-384.

Wen, Z., H. N. Nguyen, Z. Guo, M. A. Lalli, X. Wang, Y. Su, N. S. Kim, K. J. Yoon, J. Shin, C. Zhang, G. Makri, D. Nauen, H. Yu, E. Guzman, C. H. Chiang, N. Yoritomo, K. Kaibuchi, J. Zou, K. M. Christian, L. Cheng, C. A. Ross, R. L. Margolis, G. Chen, K. S. Kosik, H. Song and G. L. Ming (2014). "Synaptic dysregulation in a human iPS cell model of mental disorders." Nature 515(7527): 414-418.

Wood, J., Y. Kim and B. Moghaddam (2012). "Disruption of prefrontal cortex large scale neuronal activity by different classes of psychotomimetic drugs." J Neurosci 32(9): 3022-3031.

Ye, Y., S. Mastwal, V. Y. Cao, M. Ren, Q. Liu, W. Zhang, A. G. Elkahloun and K. H. Wang (2017). "Dopamine is Required for Activity-Dependent Amplification of Arc mRNA in Developing Postnatal Frontal Cortex." Cereb Cortex 27(7): 3600-3608.

Author response:

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

Reviewer 1:

The main weaknesses of the paper are a lack of significance in key findings, and relatedly, concluding effects from insignificant findings. Additional elements could be improved to help strengthen this overall well-rounded and intriguing set of results.

In the original manuscript, we reported that chemogenetic silencing of POA-social neurons (previously called POA-iso neurons; more details on rationale for renaming below in our responses to reviewer recommendations) tended to reduce mounting in both single-housed female and single-housed male mice, although these effects were non-significant. We have added samples to both datasets and now report that chemogenetic silencing of POA-social neurons significantly reduces the proportion of trials with mounting in both sexes (Fig. 2C and Fig. 6G). 

We have also included new analyses to test whether optogenetic activation of POAsocial neurons in group-housed females promotes social investigation (in addition to USV production, as reported in the original manuscript). We now report that optogenetic activation of POA-social neurons significantly increases the probability of social investigation (Fig. 4E-F) and significantly increases the duration of social investigation bouts (Fig. 4G). 

Additional recommendations from the reviewer are addressed in detail below. Thank you for your critical and insightful feedback.

Reviewer 2:

All the activity-dependent labeling experiments with TRAP mice, including the subsequent neural activity manipulation experiments (Figures 2, 3, 4, 5E-F), were conducted by labeling neurons only in socially isolated animals, not group-housed animals. The authors labeled neurons after 30-minute social interactions, raising the possibility that the labeled neurons simply represent a "social interaction/behavior population" (mediating mounting and USVs in females and males) rather than a set of neurons specific to social isolation.

I strongly recommend including experimental groups that involve labeling neurons after 30minute social interactions in group-housed female or male mice and inhibit TRAPed neurons after social isolation or activate TRAPed neurons after group housing. If manipulating the grouphoused TRAP neurons has similar effects to manipulating the isolated TRAP neurons, it would suggest the current labeling paradigm is not isolating neurons specific to the effect of social isolation per se. Rather, the neurons may mediate more general social interaction or motivationrelated activities. Given the known role of POA in male mating behavior, a group-housed TRAP experiment in males with a female visitor is especially important for understanding the selectivity of the labeled cells.

Without proper controls, referring to the labeled neurons as "POAiso" neurons is potentially misleading. The data thus far suggests these neurons may predominantly reflect a "POA social behavior" population rather than a set of cells distinctly responsive to isolated housing.

We agree with the reviewer that the POA neurons we are studying regulate the production of social behaviors in females and males, rather than representing a set of cells distinctly responsive to single housing. To more clearly reflect our thinking, we have changed the name of the neurons from “POA-iso neurons” to “POA-social neurons”. Thank you for this helpful criticism.

Our Fos data are consistent with the idea that the POA may regulate social behaviors in group-housed females (not just single-housed females). Namely, we found that counts of Fospositive POA neurons are significantly related to rates of social investigation (p = 0.01) and tend to be related to USV rates (p = 0.05) in group-housed females that engaged in same-sex interactions (Fig. S1C). We now include two new sets of experiments aimed at further testing the idea this idea. 

First, we include 2 control groups in which TRAPing sessions were performed in grouphoused females following same-sex interactions. We find that chemogenetic silencing of grouphoused-TRAPed POA neurons fails to reduce social behaviors in females that are subsequently single-housed and given a same-sex social interaction (Fig. 5A-D), and that optogenetic activation of group-housed-TRAPed POA neurons fails to promote female social behavior (Fig. 5E-H). At face value, these findings do not support the idea that the POA contains neurons that regulate social behaviors in group-housed females.

However, one important caveat is that group-housed females engage in low rates of social behaviors (low investigation time, no mounting, and few USVs), and thus TRAP-based labeling may not work efficaciously in these mice. There may be POA neurons that regulate social behaviors in group-housed females but that do not upregulate Fos following production of relatively low rates of social behaviors. To test this idea, we also include females in which POA neurons are chemogenetically silenced using a viral strategy that does not depend on activitydependent labeling. In this new experiment, we report that silencing of POA neurons significantly reduces USV production in group-housed females (Fig. 5J-L) and significantly reduces social investigation, mounting, and USV production when these same females are retested following single-housing (Fig. 5M-O). Together, these experiments suggest that the POA may regulate the production of social behaviors during same-sex interactions in group-housed females, but that these effects may be difficult to detect in some cases given the low rates at which group-housed females engage in social behaviors during same-sex interactions relative to single-housed females.

Finally, we want to highlight an additional new dataset that supports the idea that POAsocial neurons regulate social behaviors, rather than encoding the “state” of social isolation. We now include a control group for the chemogenetic silencing of female POA-social neurons, in which females were single-housed but were not given a social interaction prior to 4-OHT treatment (N = 5 non-social controls). Rates of social behaviors were subsequently unaffected following CNO delivery in these females (Fig. S2D-G). These new data support the conclusion that POA-social neurons regulate the production of social behaviors, rather than encoding the state of social isolation. 

Reviewer 3:

While the authors should be commended for performing and reporting multiple circuit perturbation experiments (e.g., chemogenetics, ablation), the conflicting effects on behavior are hard to interpret without additional experiments. For example, chemogenetic silencing of the POA neurons (using DREADDs) attenuated all three behavioral measures but the ablation of the same POA neurons (using CASPACE) decreased mounting duration without impacting social investigation or USV production. Similarly, optogenetic activation of POA neurons was sufficient to generate USV production as reported in earlier studies but mounting or social investigation remained unaffected. 

Do these discrepancies arise due to the efficiency differences between DREADD-mediated silencing vs. Casp3 ablation? Or does the chemogenetic result reflect off-manifold effects on downstream circuitry whereas a more permanent ablation strategy allows other brain regions to compensate due to redundancy? It is important to resolve whether these arise due to technical reasons or whether these reflect the underlying (perhaps messy) logic of neural circuitry. Therefore, while it is clear that POA neurons likely contribute to multiple behavioral readouts of social isolation, understanding their exact roles in any greater detail will require further experiments.

We have added new analyses to consider the possibility that optogenetic activation of female POA-social neurons promotes social investigation. In the original manuscript, we analyzed the duration of social investigation bouts in POA-social-ChR2 females according to whether they overlapped with laser stimulation or whether they did not overlap. We realized that we made an error in this first analysis and inadvertently included social investigation bouts that occurred during the first 5 minutes of the social sessions, prior to any laser stimulation. Because these earlier bouts tend to be longer duration than later bouts, this mistake washed out the effect of laser stimulation on social bout duration. After correcting that error, we now report that optogenetic activation of female POA-social neurons lengthens social investigation bout duration (Fig. 4G). Inspired by this interesting finding, we also included analyses of the probability of social investigation following laser stimulation (Fig. 4E-F; excluding laser stimulations that were preceded by social investigation in the pre-laser baseline period). These analyses support the conclusion that optogenetic activation of POA-social neurons promotes both USV production and social investigation in group-housed females.  

The majority of the females that we used in our TRAP2-based ablation experiments were heterozygous for TRAP2 (N = 11 of 15 POA-social-caspase subjects were TRAP2;Ai14 females), whereas all females used in our chemogenetic silencing experiments were homozygous for TRAP2. To test whether a more effective ablation of POA-social neurons might drive decreases in social investigation and USV production, we set up additional TRAP2 homozygous POA-social-caspase females and directly compare the effects of ablation between the two genotypes (Fig. S3; N = 11 hets in total and N = 9 homozygotes in total). These experiments revealed that effects on mounting were more pronounced following POA-social ablation in TRAP2 homozygotes vs. heterozygotes, but that neither group exhibited decreased social investigation or USV production following 4-OHT treatment.

To ask whether caspase-mediated ablation in TRAP2 homozygotes was effective in eliminating neural activity associated with social behaviors in females, we performed Fos immunostaining in a subset of the POA-social-caspase TRAP2 homozygotes following a samesex interaction. We found that POA Fos expression was robustly reduced in these females relative to control group-housed and control single-housed females that also engaged in samesex interactions, down to levels seen in group-housed and single-housed females that did not engage in a social interaction (comparison shown in Fig. S3D; control female data same as in Fig. 1). Moreover, the remaining POA Fos in these TRAP2 homozygotes was no longer positively correlated to social investigation or USV production (Fig. S3E-F). Together, these findings lead us to favor the interpretation suggested by the reviewer below, that permanent ablation of POA-social neurons leads to compensation from other brain regions due to redundancy. In addition, our finding that optogenetic activation of POA-social neurons promotes both USV production and social investigation supports the idea that POA-social neurons directly regulate these behaviors. We agree with the reviewer that additional work is needed to understand the complex sex- and context-dependent role played by the POA in the regulation of mouse social behaviors.

Recommendations for the Authors:

Reviewer 1 Recommendations:

(1) The largest issue is that many of the stated "key" behavioral findings are not statistically significant.

(1a) Figure 2C is not significant and Figure 5G is not significant

We have added N = 5 POA-social-hM4Di females, N = 3 POA-social-hM4Di males, and N = 3 POA-social-GFP males to the dataset. The decrease in mounting following chemogenetic silencing of POA-social neurons is now statistically significant in both sexes (p < 0.05 for both; see current Figs. 2C and 6G). We also simplified our statistical analysis of mounting in these experiments to consider the proportion of trials with and without resident-initiated mounting on saline vs. CNO days, using McNemar’s test for paired proportions. 

(1b) Mounting graphs are completely omitted in Figure 4. 

Given that mounting was only observed infrequently in POA-social-ChR2 females, we simply report this information in the Results text (lines 382-388). In our prior summary of the mounting results, we reported that mounting was observed in a total of 3 trials from 2 females, but we inadvertently included information from a duplicate trial from one of the POA-socialChR2 females in this summary (all other analyses of the POA-social-ChR2 females included one trial per female). We have corrected that error and now report that we observed mounting following laser stimulation in 1 trial from 1 POA-social-ChR2 female. We have expanded our consideration of potential effects of optogenetic activation of POA-social neurons on social investigation and include these new analyses as part of Figure 4 (Fig. 4E-G), following the existing analyses of USV production.

(1c) Figure 3C shows a reduction of mounting following the ablation of POA (although no stats on the graph to denote significance), but this ablation approach can't resolve whether POA is required to encode the state produced by the short period of isolation, and/or whether it needs to be online at test.

We have now added an asterisk in Fig. 3C to denote a p value less than 0.05. Thank you for catching our oversight.

We designed our activity-dependent labeling experiments to TRAP and express viruses in POA neurons that increase their activity in conjunction with the production of social behaviors in single-housed females. We believe our findings our most consistent with the conclusion that these neurons regulate the production of social behaviors, rather than encoding the state of social isolation, and we have renamed these neurons as “POA-social” neurons to better reflect our thinking.

We also now include control experiments (albeit chemogenetic inhibition, not caspase ablation) in which the TRAP2 strategy is used to express hM4Di in the POA of single-housed females that do not experience a social interaction prior to 4-OHT delivery (non-social controls, Fig. S2D-G). We report that chemogenetic inhibition of these neurons does not decrease social behavior in single-housed females during a subsequent same-sex interaction (p > 0.05 for saline vs. CNO rates of social investigation, mounting, and USVs). These additional findings support the idea that the activity of POA-social neurons is related to the production of social behaviors rather than to the state of social isolation. 

The reviewer is correct that our ablation approach cannot resolve the question of whether POA-social neuronal activity is required online during testing, but our reversible chemogenetic inhibition experiments provide evidence that the activity of POA-social neurons is required online at the time of testing to regulate social behavior.

(1d) A similar issue is seen regarding investigation (a general lack of significance with most of the LOF and GOF manipulations).

As reported in the original manuscript, we find that chemogenetic inhibition of POAsocial neurons reduces social investigation in females, while caspase-mediated ablation of female POA-social neurons does not. Our original caspase dataset used mostly but not all TRAP2 heterozygous females (N = 11 TRAP2 heterozygotes (TRAP2;Ai14), generated by crossing TRAP2 mice with Ai14 mice, for the purpose of visualizing the absence of tdTomato labeling to estimate spread of the caspase virus; and N = 4 TRAP2 homozygotes). By adding to the TRAP2 homozygous caspase dataset and comparing the effects on female social behavior of ablation of POA-social neurons in TRAP2 heterozygous vs. TRAP2 homozygous females, we

now provide evidence that the attenuation of mounting is more efficacious in TRAP2 homozygous females than in heterozygotes (Fig. S3B). Nonetheless, we fail to see effects on social investigation and USV production, even when caspase ablation of POA-social neurons is performed in TRAP2 homozygous females (Fig. S3A,C). 

In spite of the lack of effect on these behaviors, we show that caspase-mediated ablation of POA-social neurons in TRAP2 homozygous females leads to a dramatic reduction in social interaction-induced Fos expression in the POA. POA Fos expression in these caspase females is reduced to the levels seen in control group-housed and single-housed females that are not given social interactions and are significantly lower than Fos expression in group-housed and single-housed females that are given a same-sex interaction (Fig. S3D). Moreover, the remaining POA Fos expression in the caspase females is no longer related to rates of social investigation (Fig. S3E), as is normally the case in group-housed and single-housed control females (Fig. S1C, left). Together, these data support the idea that some type of neuronal compensation outside of the POA is occurring following ablation of POA-social neurons, and this compensation permits normal levels of USV production and social investigation.

As in the original manuscript, we report that chemogenetic inhibition of POA-social neurons in male mice reduces mounting but does not reduce social investigation (or USV production). We now include quantification of social behaviors produced by male and female POA-social-hM4Di mice in the TRAPing sessions that preceded 4-OHT delivery (Fig. S5). These measurements show that males spent significantly more time than females engaged in mounting, and we speculate that this bias in TRAPing session behavior might have led to a bias in TRAP-mediated viral labeling of male POA neurons that regulate mounting, at the expense of male POA neurons that regulate social investigation (or USV production).

We have added new analyses to consider the possibility that optogenetic activation of female POA-social neurons promotes social investigation. In the original manuscript, we analyzed the duration of social investigation bouts in POA-social-ChR2 females according to whether they overlapped with laser stimulation or whether they did not overlap. We realized that we made an error in this first analysis and inadvertently included social investigation bouts that occurred during the first 5 minutes of the social sessions, prior to any laser stimulation. Because these earlier bouts tend to be longer duration than later bouts, this mistake washed out the effect of laser stimulation on social bout duration. After correcting that error, we now report that optogenetic activation of female POA-social neurons lengthens social investigation bout duration (Fig. 4G). Inspired by this encouraging finding, we also included analyses of the probability of social investigation following laser stimulation (Fig. 4E-F; excluding laser stimulations that were preceded by social investigation in the pre-laser baseline period). These analyses support the conclusion that optogenetic activation of POA-social neurons promotes both USV production and social investigation in group-housed females.

(2) In Figure 1 and elsewhere, the authors use a Mann-Whitney U test, which should be used for non-parametric data, but in other places, they use statistical tests for normally distributed data. Why? How was the normality of distributions tested?

We tested the normality of data distributions using the Shapiro-Wilk test. Parametric tests were used for analyses that contained normally distributed data, and non-parametric tests were used for analyses that contained non-normally distributed data. This information is included in the Methods (lines 997-1000), and full details of statistical analyses can be found in Table S1.

(3) The method for "trapping" neurons that are part of the short-term isolation ensemble has some caveats that have not been adequately addressed. First, 4-OHT was administered after social interaction, but before 24 hours of isolation, making it unclear exactly WHAT is being trapped.

i) Is it neurons that encode the recent 3-day iso experience? (seems unlikely, as this would have been hours after the end of that iso window)

We now include a group of control females to directly test this possibility (Fig. S2D-G). These TRAP2 females were single-housed for 3 days but were not given a social interaction prior to 4-OHT treatment (N = 5 non-social controls). Presumably, POA neurons TRAPed in these females might encode the experience of short-term isolation. However, we found that chemogenetic inactivation of these TRAPed neurons during a subsequent same-sex interaction failed to decrease social behaviors in single-housed females (Fig. S2E-G; p > 0.05 for CNO vs. saline rates of social investigation, mounting, and USV production). These control experiments support the idea that we are TRAPing neurons whose activity is related to the production of social behaviors, and we have renamed the neurons as “POA-social” neurons to reflect this thinking.

ii) Is it neurons that encode the recent behavior impacted by the 3-day iso? (this seems to be the goal, but the authors do not provide evidence that the time course of their injection is efficient enough to recruit the recently activated neurons, nor do they provide evidence that opening the trapping window directly after the behavior is better than directly before)

We opted to perform IP injections of 4-OHT immediately following the behavior session, rather than behavior, due to concern that handling the mice and delivering IP injections prior to behavior sessions would stress the mice, leading to lower rates of social behaviors. The nonsocial female hM4Di experiments described above support the idea that we are TRAPing neurons related to the production of social behaviors, as the reviewer suggests. 

iii) Is it trapping neurons active during the subsequent 24 hours of isolation? (seems possible, but this would mean that the authors are looking at a different population of neurons than they claim).

If chemogenetic silencing of POA neurons that were TRAPed following 3-days of social isolation but in the absence of a social interaction (N = 5 non-social controls, Fig. S2D-G) does not alter social behaviors, there is no compelling reason to hypothesize that TRAPing POA neurons activated following the 24 hours of social isolation that follow a social interaction would do so. Moreover, in the original study characterizing the TRAP2 mice (DeNardo et al., 2019), the authors performed experiments to characterize the time course of TRAPing relative to 4-OHT treatment and concluded that the majority of TRAPing occurs within a 6-hour window centered around the 4-OHT injection.

(4) Relatedly, the authors seem to find a fair bit of variability in their TRAP-mediated experiments. This begs the question - are the effects of their GOF and LOF approaches

i) dependent on the iso-behaviors that were "trapped" for each animal (in other words, how does behavior at test 1 correlate with behavior at test 2)? 

To test the reviewer’s idea, we compared rates of TRAPing session behaviors for the POA-social-hM4Di females to the subsequent effects of neuronal silencing on these behaviors (calculated as (CNO behavior – saline behavior). These correlations are shown in Fig. S2A-C and are all non-significant. We also include below for the reviewer the same types of correlations for the other datasets in our study (loss-of-function experiments: female POAsocial-caspase, male POA-social-hM4Di; and gain-of-function experiments: female POA-socialChR2).

Author response image 1.

The only loss-of-function experiment comparison in the above figure that reveals a negative and significant correlation is the mounting comparison for the POA-social-hM4Di males (time spent mounting during TRAPing session vs. (CNO time spent mounting -saline time spent mounting). This significant correlation likely reflects that fact that (1) no males mounted in the CNO session and (2) that mounting rates for individual males are relatively consistent over time (in comparison to female mounting, which is more variable; see Author response image 2 below of TRAPing session vs. saline mounting in male vs. female POA-social-hM4Di experiments). The correlation between TRAPing session and testing session mounting is significant for the POA-social-ChR2 females, but despite the significant correlation, we would want to see more instances of optogenetically-elicited mounting to make any claim about its relationship to TRAPing session behavior.

Author response image 2.

Nonetheless, we agree with the reviewer’s intuition that one would expect the effects of POA activity manipulations on different behaviors to scale with rates at which these behaviors were performed during the TRAPing session. We speculate that variability in the TRAPing process might have obscured such a relationship. There is inevitable variability in the exact body cavity placement of IP injections, which can affect drug absorption, and another point is that we delivered a fixed volume of 4-OHT (10 mg/mL 4-OHT in 150 uL filtered corn oil) to all mice in the study, regardless of their weight, which likely added variability in TRAPing efficacy from animal to animal. This detail was reported inaccurately in the Methods, and that error has been corrected (line 920). With regard to our male POA-social-hM4Di dataset, we find that these males spend more time mounting during their TRAPing sessions than female POA-socialhM4Di (Fig. S5; males also spent less time investigating and tended to produce fewer USVs than females), a fact that we hypothesize may have led to a bias toward TRAPing mountingrelated POA neurons in male subjects. In addition, however, the fact that male mice typically weigh more than females and would have received a slightly lower effective dosage of 4-OHT may also have contributed to the weaker effects on behavior in the male POA-social-hM4Di experiments relative to the female POA-social-hM4Di experiments.

We also want to highlight that interpreting correlations for females between time spent mounting during the TRAPing session and time spent mounting during the test sessions can be complicated. For example, we see 2 cases in the female POA-social-hM4Di dataset in which the female did not mount in the TRAPing session, and then mounted on the saline day (12s and 10s total mounting for those 2 females) but not on the CNO day. One interpretation of the data from these 2 females is that mounting on the TRAPing day is not required to attenuate mounting on the later test days. However, female mounting behavior itself is variable, both across different females and across different tests of a given female, as noted above. If we consider all singlehoused females included in our dataset for which we quantified control behavioral data (i.e., behavior trials from unmanipulated females and TRAPing sessions from females that were later manipulated), we find that mounting is not observed in ~30% of the females (24 of 83). In ongoing behavioral experiments not included in this manuscript, we are investigating factors that regulate female mounting following single-housing. In that dataset, we also see little evidence that female mounting in one social interaction predicts mounting in a subsequent interaction

(i.e., there don’t appear to stable “high mounters” and “low mounters” following single housing). Thus, the small number of cases in which females did not mount in the TRAPing session and then displayed mounting on the CNO only day are difficult to interpret. 

Two additional considerations are that TRAPing may not be equally efficacious for POA neurons that regulate different behaviors, and that different behaviors may be differentially sensitive to perturbations of the POA. Previous elegant calcium imaging work has shown that different subsets of Esr1+ POA neurons exhibit activity that is “tuned” to specific behaviors (sniffing vs. mounting in males interacting with females; Yang et al., 2023). However, it is possible that these subsets of neurons display differential levels of Fos expression following the production of their preferred behavior and that some behavior-related subsets may thus be more easily TRAPed than others. It may also be the case that some behaviors are more easily disrupted by POA activity manipulations than others (e.g., perturbation in a smaller percentage of behavior-related POA neurons may be required to disrupt some behaviors relative to others). 

Despite these caveats, we have two lines of evidence that the effects of chemogenetic silencing of POA-social neurons depends on the behaviors produced during the TRAPing sessions.

(1) Social behavior is required during the TRAPing session to see subsequent effects on social behavior following chemogenetic silencing of TRAPed POA neurons. In control females that were single-housed but were not given a social interaction prior to 4OHT treatment, social behaviors are not reduced by chemogenetic silencing of TRAPed POA neurons (Figs. S2D-G).

(2) To directly test whether mounting in the TRAPing session is required to see attenuation of mounting during subsequent chemogenetic silencing of POA-social neurons, we performed control experiments in which single-housed females interacted with a female visitor that was placed under a cup during the TRAPing session prior to 4-OHT treatment. Mounting was not possible in this context, and we also found that females produced lower rates of USVs during the TRAPing session relative to single-housed females engaged in free social interaction. However, subject females spent more time engaged in social investigation of the visitor relative to single-housed females engaged in free social interactions (see Author response image 3 below).

Author response image 3.

Unfortunately, none of the experimental females in this cohort displayed mounting in the CNO or saline sessions. Given that we could use this dataset to address the intended question, we did not include it in the manuscript. However, it is quite interesting that female subjects displayed higher than normal social investigation and lower than normal USV production in their TRAPing sessions (relative to single-housed females engaged in free interactions), and subsequently, chemogenetic inhibition of TRAPed POA neurons decreased social investigation but did not decrease USV production (Author response image 4 below). 

Author response image 4.

Together, we think our data support the idea that the POA neurons that are TRAPed are related to the social behaviors performed by the animals, but these relationships may be complex and difficult to detect from comparisons across animals within a single experimental group.

And/or are they

ii) influenced by the spread or amount of virus for each animal? These correlations could help shed light on what exactly is being trapped - is it specific behaviors or is it the "state" of shortterm isolation?

Our control experiments with females that were single-housed but did not receive a social interaction prior to 4-OHT treatment provide evidence that the production of social behaviors is required to see subsequent effects on behavior following chemogenetic inhibition of TRAPed POA neurons (Figs. S2D-G).

The same volume of virus was injected across all activity manipulation experiments (200 nL). Because of the trajectory of our POA viral injections (performed at a slight rostral angle relative to vertical), we did sometimes see viral labeling that spread into the AH caudal to the POA. For this reason, we included the AH TRAPed control group (Fig. 2), to rule out the possibility that viral spread into the AH could account for the effects of chemogenetic silencing of POA-social neurons on female social behaviors. Also because of the injection angle used, we don’t see substantial viral spread rostral to our injection coordinates. In short, there isn’t systematic variability in the targeting or spread of our POA viral injections that can account for variability in the effects on USV production and social investigation of our LOF and GOF manipulations (female hM4Di and female ChR2 experiments).

In older lesion studies in male rodents and birds, there is some support for the idea that rostral vs. caudal POA neurons differentially regulate appetitive vs. consummatory sexual behaviors (as reviewed in Balthazart and Ball, 2007). However, all of our viral injections were placed in what that review paper would have considered ‘caudal’ POA. We also note that more recent imaging studies have reported that subsets of POA neurons are differentially tuned to male sniffing vs. male mounting (Yang et al.,2023), and these subsets must be relatively co-localized given that they are imaged in the same field of view. Whether distinct subsets of POA neurons regulate the production of different female social behaviors, and if so, how these subsets are localized within the POA, remains an important question for future study.

(5) The authors label their region of interest as the "POA" but images throughout (e.g. their fos image, Figure 1E), look more like the MPO. Why label it POA?

The POA neurons in our study are found in a band that spans the medial POA, as well as a bit of the lateral POA. To avoid over-specifying, we call this region the POA more generally.

(6) In all the experiments, mice are isolated and then re-group housed with siblings. Do all the siblings in the group belong to the same experimental group, or are siblings naïve? This may be critical to help determine whether some of the effects observed may be "group" effects.

In general, multiple (although not always all) mice in a cage belonged to the same experimental group. In our inhibitory DREADDs experiments, it is unclear how that could drive our observed effects on behavior, given that home cage behavior would only be expected to differ for a given mouse in the time period following their CNO session. 

For the female POA-social-caspase mice, we cannot rule out the possibility that their home cage behaviors differed in the time period following 4-OHT treatment and re-grouphousing and prior to post-4-OHT behavior measurements. However, given that the only social behavior affected by ablation of POA-social neurons was mounting, and that rates of mounting would be expected to be very low in group-housed females within home cages, it is unclear how our experimental result could be attributed to group effects.

If by “group” effects the reviewer means “litter” effects, we include a plot below that shows the CNO vs. saline behaviors for the POA-social-hM4Di females, separated by cage ID. There is no evidence that the effects of chemogenetic silencing of POA-social-hM4Di females are being driven by only certain cages (only social investigation and USVs are shown, because mounting was uniformly low (1 of 17 females mounted) in the CNO session).

Author response image 5.

(7) For chemogenetic experiments, the authors state that CNO and Saline were given in a counterbalanced order (eg line 189). Did the authors see any order effects?

We did not see order effects, and we can include plots of those data below for the female and male POA-social-hM4Di groups, with mice plotted according to which treatment they received first.

Author response image 6.

(8) In the control experiments in Figure 2 where VMH or AH are chemogenetically silenced, it isn't clear whether these groups include mice that were subjected to 3 days of isolation. Please clarify.

Yes, these female groups were also subjected to 3 days of isolation (first prior to the TRAPing session, and for a second time prior to the onset of the CNO/saline testing sessions). That information has been clarified in the Results section (line 214) and in the Methods (lines 935-938).

(9) Line 312. The title for this section, "POA neurons increase their activity....." is somewhat misleading. It sounds like the authors imaged trapped neurons. I think what they mean is that more POA neurons are activated following opposite-sex interactions with males.

Thanks for this catch. We have modified the section title, as well as the title of the first results sub-section.

(10) Figure 5A, right panels. The authors fail to find an increase in the investigation of male-male pairs following the short-term isolation of one. This contrasts with the main finding in Matthews et al., 2016 Cell, where short periods of isolation are said to promote pro-social behaviors. The authors could comment on this discrepancy in their discussion (eg difference in testing apparatus/test type? Difference in the number of days of isolation? etc.).

In current Fig. 6A, there is no significant interaction between the two main effects, but each main effect is significant: single-housed males spend more time investigating partners than group-housed males, and males spend more time investigating female partners than male partners. The significant main effect of housing condition is consistent with the findings of Matthews et al., 2016 and is included within the Results (lines 486-492). 

(11) Figure 5F, the authors seem to have a main effect of virus (more overall investigation in dreadds mice). Nothing about this is addressed.

We sometimes see differences in social behavior between cohorts of males when they are tested at different times and, correspondingly, with different groups of female social partners. Our POA-social-hM4Di and POA-social-GFP males were set-up and tested at largely non-overlapping times. We have added a brief note to the Results section to include this information (lines 535-539).

Reviewer 2 Recommendations:

(1) (C)ritical control experiments are missing to support this claim (that a population of preoptic hypothalamic neurons contribute to the effects of short-term social isolation on the social behaviors of female mice).  

(1a) All the activity-dependent labeling experiments with TRAP mice, including the subsequent neural activity manipulation experiments (Figures 2, 3, 4, 5E-F), were conducted by labeling neurons only in socially isolated animals, not group-housed animals. The authors labeled neurons after 30-minute social interactions, raising the possibility that the labeled neurons simply represent a "social interaction/behavior population" (mediating mounting and USVs in females and males) rather than a set of neurons specific to social isolation behaviors of mice)… The data thus far suggests these neurons may predominantly reflect a "POA social behavior" population rather than a set of cells distinctly responsive to isolated housing.

We agree with the reviewer that the POA neurons we are studying regulate the production of social behaviors in females and males, rather than representing a set of cells distinctly responsive to single housing. To more clearly reflect our thinking, we have changed the name of the neurons from “POA-iso neurons” to “POA-social neurons”. Thank you for this helpful criticism.

Our Fos data are consistent with the idea that the POA may regulate social behaviors in group-housed females (not just single-housed females). Namely, we found that counts of Fospositive POA neurons are significantly related to rates of social investigation (p = 0.01) and tend to be related to USV rates (p = 0.05) in group-housed females that engaged in same-sex interactions (Fig. S1C). We now include two new sets of experiments aimed at further testing the idea this idea. 

First, we include 2 control groups in which TRAPing sessions were performed in grouphoused females following same-sex interactions. We find that chemogenetic silencing of these group-housed-TRAPed POA neurons fails to reduce social behaviors in females that are subsequently single-housed and given a same-sex social interaction (Fig. 5A-D; GH-TRAPed POA hM4Di females), and that optogenetic activation of group-housed-TRAPed POA neurons fails to promote female social behavior (Fig. 5E-H; GH-TRAPed POA ChR2 females). At face value, these findings do not support the idea that the POA contains neurons that regulate social behaviors in group-housed females.

However, one important caveat is that group-housed females engage in low rates of social behaviors (low investigation time, no mounting, and few USVs), and thus TRAP-based labeling may not work efficaciously in these mice. There may be POA neurons that regulate social behaviors in group-housed females but that do not upregulate Fos following production of relatively low rates of social behaviors. To test this idea, we also include females in which POA neurons are chemogenetically silenced using a viral strategy that does not depend on activitydependent labeling. In this new experiment, we report that silencing of POA neurons significantly reduces USV production in group-housed females (Fig. 5J-L) and significantly reduces social investigation, mounting, and USV production when these same females are retested following single-housing (Fig. 5M-O).

(2) Please add strain background information of subject animals in the methods.

This information has been added to the Animals section within the Methods (lines 788802).

Responses to Reviewer 3 Recommendations:

(1a) (T)he conflicting effects on behavior are hard to interpret without additional experiments….Similarly, optogenetic activation of POA neurons was sufficient to generate USV production as reported in earlier studies but mounting or social investigation remained unaffected. 

We have added new analyses to consider the possibility that optogenetic activation of female POA-social neurons promotes social investigation. In the original manuscript, we analyzed the duration of social investigation bouts in POA-social-ChR2 females according to whether they overlapped with laser stimulation or whether they did not overlap. We realized that we made an error in this first analysis and inadvertently included social investigation bouts that occurred during the first 5 minutes of the social sessions, prior to any laser stimulation. Because these earlier bouts tend to be longer duration than later bouts, this mistake washed out the effect of laser stimulation on social bout duration. After correcting that error, we now report that optogenetic activation of female POA-social neurons lengthens social investigation bout duration (Fig. 4G). Inspired by this interesting finding, we also included analyses of the probability of social investigation following laser stimulation (Fig. 4E-F; excluding laser stimulations that were preceded by social investigation in the pre-laser baseline period). These analyses support the conclusion that optogenetic activation of POA-social neurons promotes both USV production and social investigation in group-housed females.

(1b) Do these discrepancies (between hM4Di and caspase) arise due to the efficiency differences between DREADD-mediated silencing vs. Casp3 ablation? Or does the chemogenetic result reflect off-manifold effects on downstream circuitry whereas a more permanent ablation strategy allows other brain regions to compensate due to redundancy? It is important to resolve whether these arise due to technical reasons or whether these reflect the underlying (perhaps messy) logic of neural circuitry.  

The possibility that the difference in effects on behavior between chemogenetic silencing and caspase ablation at face value seems inconsistent with the findings of previous experiments, in which ablation of large numbers of POA neurons failed to reduce USV production in male mice (POA lesions in Bean et al., 1981; ablation of VGAT+ POA neurons by Gao et al., 2018). These findings stand in contrast to those using chemogenetic silencing of large numbers of POA neurons, which report reduced USV production in male mice (VGAT+/Esr1+ in Karigo et al., 2021; Esr1+ in Chen et al., 2021).

However, it is the case that the majority of the females that we used in our TRAP2-based ablation experiments were heterozygous for TRAP2 (N = 11 of 15 POA-social-caspase subjects were TRAP2;Ai14 females), whereas all females used in our chemogenetic silencing experiments were homozygous for TRAP2. To test whether a more effective ablation of POAsocial neurons might drive decreases in social investigation and USV production, we set up additional TRAP2 homozygous POA-social-caspase females and directly compare the effects of ablation between the two genotypes (Fig. S3; N = 11 hets in total and N = 9 homozygotes in total). These experiments revealed that effects on mounting were more pronounced following POA-social ablation in TRAP2 homozygotes vs. heterozygotes, but that neither group exhibited decreased social investigation or USV production following 4-OHT treatment.

To ask whether caspase-mediated ablation in TRAP2 homozygotes was effective in eliminating neural activity associated with social behaviors in females, we performed Fos immunostaining in a subset of the POA-social-caspase TRAP2 homozygotes following a samesex interaction. We found that POA Fos expression was robustly reduced in these females relative to control group-housed and control single-housed females that also engaged in samesex interactions, down to levels seen in group-housed and single-housed females that did not engage in a social interaction (comparison shown in Fig. S3D; control female data same as in Fig. 1). Moreover, the remaining POA Fos in these TRAP2 homozygotes was no longer positively correlated to social investigation or USV production (Fig. S3E-F). Together, these findings lead us to favor the interpretation suggested by the reviewer below, that permanent ablation of POA-social neurons leads to compensation from other brain regions due to redundancy.

Given the negative results above, we favor this possibility and indicate so in our Discussion. In addition, our finding that optogenetic activation of POA-social neurons promotes both USV production and social investigation supports the idea that POA-social neurons directly regulate these behaviors. We agree with the reviewer that additional work is needed to understand the complex sex- and context-dependent role played by the POA in the regulation of mouse social behaviors.

(2) L 49: Please define Mesolimbic circuitry the first time it is mentioned.

We have added a definition (lines 52-53).

(3) L 210: In Figure 2C, the mounting duration baseline (saline) distribution seems lower than the same experimental baseline in Figures 1C and 3C. Does this reflect natural variability in the behavioral assay and might this be mitigated by additional sampling of animals?

Yes, there is substantial variability in the display of mounting behavior by single-housed females, including in the proportion of trials with mounting as well as in the total duration of mounting. In the revised manuscript, we have simplified our analysis of mounting in our TRAPbased experiments to quantify the proportion of trials with mounting, rather than considering the total time spent mounting. After adding N = 5 additional females to the POA-social-hM4Di dataset, we now report a statistically significant decrease in the proportion of trials with mounting following chemogenetic silencing of POA-social neurons (Fig. 2C; McNemar’s test for paired proportions). 

(4) L 310: The authors claim that "These findings suggest that a subset of POAiso neurons overlap with GABAergic, PAG-projecting POA neurons that have been demonstrated in previous work to promote USVs via disinhibition of excitatory PAG neurons important to USV production (Chen et al., 2021; Michael et al., 2020)." I think the data reported suggests the opposite since only 18.3% of all POA->PAG neurons are cFos+. Perhaps better rephrased as "A subset (18.3%) of POA->PAG neurons are labelled by cFos and that is sufficient to drive the production of USVs". Is it surprising?

We modified the phrasing (lines 468-469), but a bit differently than suggested above, because although we suspect that optogenetic activation of the PAG-projecting neurons within the larger population of POA-social neurons is responsible for eliciting USV production, we did not technically demonstrate this to be the case in the current dataset. 

We do find it surprising that so few (only ~20%) of PAG-projecting POA neurons upregulate Fos following female-female interactions marked by high rates of USV production. Even though optogenetic activation of PAG-projecting POA neurons elicits USV production, our finding suggests that the majority of PAG-projecting POA neurons may not play a role in regulating vocalization. In future work, it may be useful to apply an intersectional approach to further understand how the POA regulates USV production (for example, measure or manipulate activity selectively in projection-defined subsets of POA-social neurons).

(5) Given the considerable prior evidence of POA->PAG circuit in promoting USVs, it is hard to understand why chemogenetic inactivation of POA neurons in males affects mounting but not USV production (Figures 5F-H). Any potential explanation for this discrepancy?

We have two ideas about this surprising result. First, we examined the TRAPing session social behaviors of female and male POA-social-hM4Di mice. We found that male POA-socialhM4Di mice spent more time than female subjects mounting during the TRAPing sessions, and conversely, males spent less time investigating visitors and tended to produce fewer USVs than female subjects (Fig. S5). Given that our labeling method is activity-dependent, one possibility is that this bias in behavior is reflected in a bias toward labeling of POA neurons related to mounting.  

Second, each mouse in the TRAP2-based hM4Di datasets received an IP injection of the same amount of 4-OHT (150 nL of 10 mg/mL 4-OHT in filtered corn oil) not adjusted for weight of the mouse. This information was not reported accurately in the Methods, and we have adjusted that section accordingly (line 920). As a result, because male mice typically weigh more than females and would have received a lower effective dosage of 4-OHT, another possibility is that TRAPing in males was less efficient than in females and accounts for the less complete effects on social behaviors. We have added language to the Results to discuss these possibilities (lines 540-560).

(6) L 472: Typo. "we found that short-term isolation exerts more robust on the effects of male behavior during subsequent interactions with females than during interactions with males."

Thank you for catching this mistake.

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

Summary:

In this study, the authors address whether the dorsal nucleus of the inferior colliculus (DCIC) in mice encodes sound source location within the front horizontal plane (i.e., azimuth). They do this using volumetric two-photon Ca2+ imaging and high-density silicon probes (Neuropixels) to collect single-unit data. Such recordings are beneficial because they allow large populations of simultaneous neural data to be collected. Their main results and the claims about those results are the following:

(1) DCIC single-unit responses have high trial-to-trial variability (i.e., neural noise);

(2) approximately 32% to 40% of DCIC single units have responses that are sensitive tosound source azimuth;

(3) single-trial population responses (i.e., the joint response across all sampled single unitsin an animal) encode sound source azimuth "effectively" (as stated in title) in that localization decoding error matches average mouse discrimination thresholds;

(4) DCIC can encode sound source azimuth in a similar format to that in the central nucleusof the inferior colliculus (as stated in Abstract);

(5) evidence of noise correlation between pairs of neurons exists;

and 6) noise correlations between responses of neurons help reduce population decoding error.

While simultaneous recordings are not necessary to demonstrate results #1, #2, and #4, they are necessary to demonstrate results #3, #5, and #6.

Strengths:

- Important research question to all researchers interested in sensory coding in the nervous system.

- State-of-the-art data collection: volumetric two-photon Ca2+ imaging and extracellularrecording using high-density probes. Large neuronal data sets.

- Confirmation of imaging results (lower temporal resolution) with more traditionalmicroelectrode results (higher temporal resolution).

- Clear and appropriate explanation of surgical and electrophysiological methods. I cannot comment on the appropriateness of the imaging methods.

Strength of evidence for claims of the study:

(1) DCIC single-unit responses have high trial-to-trial variability - The authors' data clearlyshows this.

(2) Approximately 32% to 40% of DCIC single units have responses that are sensitive tosound source azimuth - The sensitivity of each neuron's response to sound source azimuth was tested with a Kruskal-Wallis test, which is appropriate since response distributions were not normal. Using this statistical test, only 8% of neurons (median for imaging data) were found to be sensitive to azimuth, and the authors noted this was not significantly different than the false positive rate. The Kruskal-Wallis test was not performed on electrophysiological data. The authors suggested that low numbers of azimuth-sensitive units resulting from the statistical analysis may be due to the combination of high neural noise and relatively low number of trials, which would reduce statistical power of the test. This may be true, but if single-unit responses were moderately or strongly sensitive to azimuth, one would expect them to pass the test even with relatively low statistical power. At best, if their statistical test missed some azimuthsensitive units, they were likely only weakly sensitive to azimuth. The authors went on to perform a second test of azimuth sensitivity-a chi-squared test-and found 32% (imaging) and 40% (e-phys) of single units to have statistically significant sensitivity. This feels a bit like fishing for a lower p-value. The Kruskal-Wallis test should have been left as the only analysis. Moreover, the use of a chi-squared test is questionable because it is meant to be used between two categorical variables, and neural response had to be binned before applying the test.

The determination of what is a physiologically relevant “moderate or strong azimuth sensitivity” is not trivial, particularly when comparing tuning across different relays of the auditory pathway like the CNIC, auditory cortex, or in our case DCIC, where physiologically relevant azimuth sensitivities might be different. This is likely the reason why azimuth sensitivity has been defined in diverse ways across the bibliography (see Groh, Kelly & Underhill, 2003 for an early discussion of this issue). These diverse approaches include reaching a certain percentage of maximal response modulation, like used by Day et al. (2012, 2015, 2016) in CNIC, and ANOVA tests, like used by Panniello et al. (2018) and Groh, Kelly & Underhill (2003) in auditory cortex and IC respectively. Moreover, the influence of response variability and biases in response distribution estimation due to limited sampling has not been usually accounted for in the determination of azimuth sensitivity.

As Reviewer #1 points out, in our study we used an appropriate ANOVA test (KruskalWallis) as a starting point to study response sensitivity to stimulus azimuth at DCIC. Please note that the alpha = 0.05 used for this test is not based on experimental evidence about physiologically relevant azimuth sensitivity but instead is an arbitrary p-value threshold. Using this test on the electrophysiological data, we found that ~ 21% of the simultaneously recorded single units reached significance (n = 4 mice). Nevertheless these percentages, in our small sample size (n = 4) were not significantly different from our false positive detection rate (p = 0.0625, Mann-Whitney, See Author response image 1 below).  In consequence, for both our imaging (Fig. 3C) and electrophysiological data, we could not ascertain if the percentage of neurons reaching significance in these ANOVA tests were indeed meaningfully sensitive to azimuth or this was due to chance. 

Author response image 1.

Percentage of the neuropixels recorded DCIC single units across mice that showed significant median response tuning, compared to false positive detection rate (α = 0.05, chance level).

We reasoned that the observed markedly variable responses from DCIC units, which frequently failed to respond in many trials (Fig. 3D, 4A), in combination with the limited number of trial repetitions we could collect, results in under-sampled response distribution estimations. This under-sampling can bias the determination of stochastic dominance across azimuth response samples in Kruskal-Wallis tests. We would like to highlight that we decided not to implement resampling strategies to artificially increase the azimuth response sample sizes with “virtual trials”, in order to avoid “fishing for a smaller p-value”, when our collected samples might not accurately reflect the actual response population variability.

As an alternative to hypothesis testing based on ranking and determining stochastic dominance of one or more azimuth response samples (Kruskal-Wallis test), we evaluated the overall statistical dependency to stimulus azimuth of the collected responses.  To do this we implement the Chi-square test by binning neuronal responses into categories. Binning responses into categories can reduce the influence of response variability to some extent, which constitutes an advantage of the Chi-square approach, but we note the important consideration that these response categories are arbitrary.

Altogether, we acknowledge that our Chi-square approach to define azimuth sensitivity is not free of limitations and despite enabling the interrogation of azimuth sensitivity at DCIC, its interpretability might not extend to other brain regions like CNIC or auditory cortex. Nevertheless we hope the aforementioned arguments justify why the Kruskal-Wallis test simply could not “have been left as the only analysis”.

(3) Single-trial population responses encode sound source azimuth "effectively" in that localization decoding error matches average mouse discrimination thresholds - If only one neuron in a population had responses that were sensitive to azimuth, we would expect that decoding azimuth from observation of that one neuron's response would perform better than chance. By observing the responses of more than one neuron (if more than one were sensitive to azimuth), we would expect performance to increase. The authors found that decoding from the whole population response was no better than chance. They argue (reasonably) that this is because of overfitting of the decoder modeltoo few trials used to fit too many parameters-and provide evidence from decoding combined with principal components analysis which suggests that overfitting is occurring. What is troubling is the performance of the decoder when using only a handful of "topranked" neurons (in terms of azimuth sensitivity) (Fig. 4F and G). Decoder performance seems to increase when going from one to two neurons, then decreases when going from two to three neurons, and doesn't get much better for more neurons than for one neuron alone. It seems likely there is more information about azimuth in the population response, but decoder performance is not able to capture it because spike count distributions in the decoder model are not being accurately estimated due to too few stimulus trials (14, on average). In other words, it seems likely that decoder performance is underestimating the ability of the DCIC population to encode sound source azimuth.

To get a sense of how effective a neural population is at coding a particular stimulus parameter, it is useful to compare population decoder performance to psychophysical performance. Unfortunately, mouse behavioral localization data do not exist. Therefore, the authors compare decoder error to mouse left-right discrimination thresholds published previously by a different lab. However, this comparison is inappropriate because the decoder and the mice were performing different perceptual tasks. The decoder is classifying sound sources to 1 of 13 locations from left to right, whereas the mice were discriminating between left or right sources centered around zero degrees. The errors in these two tasks represent different things. The two data sets may potentially be more accurately compared by extracting information from the confusion matrices of population decoder performance. For example, when the stimulus was at -30 deg, how often did the decoder classify the stimulus to a lefthand azimuth? Likewise, when the stimulus was +30 deg, how often did the decoder classify the stimulus to a righthand azimuth?

The azimuth discrimination error reported by Lauer et al. (2011) comes from engaged and highly trained mice, which is a very different context to our experimental setting with untrained mice passively listening to stimuli from 13 random azimuths. Therefore we did not perform analyses or interpretations of our results based on the behavioral task from Lauer et al. (2011) and only made the qualitative observation that the errors match for discussion.

We believe it is further important to clarify that Lauer et al. (2011) tested the ability of mice to discriminate between a positively conditioned stimulus (reference speaker at 0º center azimuth associated to a liquid reward) and a negatively conditioned stimulus (coming from one of five comparison speakers positioned at 20º, 30º, 50º, 70 and 90º azimuth, associated to an electrified lickport) in a conditioned avoidance task. In this task, mice are not precisely “discriminating between left or right sources centered around zero degrees”, making further analyses to compare the experimental design of Lauer et al (2011) and ours even more challenging for valid interpretation.

(4) DCIC can encode sound source azimuth in a similar format to that in the central nucleusof the inferior colliculus - It is unclear what exactly the authors mean by this statement in the Abstract. There are major differences in the encoding of azimuth between the two neighboring brain areas: a large majority of neurons in the CNIC are sensitive to azimuth (and strongly so), whereas the present study shows a minority of azimuth-sensitive neurons in the DCIC. Furthermore, CNIC neurons fire reliably to sound stimuli (low neural noise), whereas the present study shows that DCIC neurons fire more erratically (high neural noise).

Since sound source azimuth is reported to be encoded by population activity patterns at CNIC (Day and Delgutte, 2013), we refer to a population activity pattern code as the “similar format” in which this information is encoded at DCIC. Please note that this is a qualitative comparison and we do not claim this is the “same format”, due to the differences the reviewer precisely describes in the encoding of azimuth at CNIC where a much larger majority of neurons show stronger azimuth sensitivity and response reliability with respect to our observations at DCIC. By this qualitative similarity of encoding format we specifically mean the similar occurrence of activity patterns from azimuth sensitive subpopulations of neurons in both CNIC and DCIC, which carry sufficient information about the stimulus azimuth for a sufficiently accurate prediction with regard to the behavioral discrimination ability.

(5) Evidence of noise correlation between pairs of neurons exists - The authors' data andanalyses seem appropriate and sufficient to justify this claim.

(6) Noise correlations between responses of neurons help reduce population decodingerror - The authors show convincing analysis that performance of their decoder increased when simultaneously measured responses were tested (which include noise correlation) than when scrambled-trial responses were tested (eliminating noise correlation). This makes it seem likely that noise correlation in the responses improved decoder performance. The authors mention that the naïve Bayesian classifier was used as their decoder for computational efficiency, presumably because it assumes no noise correlation and, therefore, assumes responses of individual neurons are independent of each other across trials to the same stimulus. The use of decoder that assumes independence seems key here in testing the hypothesis that noise correlation contains information about sound source azimuth. The logic of using this decoder could be more clearly spelled out to the reader. For example, if the null hypothesis is that noise correlations do not carry azimuth information, then a decoder that assumes independence should perform the same whether population responses are simultaneous or scrambled. The authors' analysis showing a difference in performance between these two cases provides evidence against this null hypothesis.

We sincerely thank the reviewer for this careful and detailed consideration of our analysis approach. Following the reviewer’s constructive suggestion, we justified the decoder choice in the results section at the last paragraph of page 18:

“To characterize how the observed positive noise correlations could affect the representation of stimulus azimuth by DCIC top ranked unit population responses, we compared the decoding performance obtained by classifying the single-trial response patterns from top ranked units in the modeled decorrelated datasets versus the acquired data (with noise correlations). With the intention to characterize this with a conservative approach that would be less likely to find a contribution of noise correlations as it assumes response independence, we relied on the naive Bayes classifier for decoding throughout the study. Using this classifier, we observed that the modeled decorrelated datasets produced stimulus azimuth prediction error distributions that were significantly shifted towards higher decoding errors (Fig. 5B, C) and, in our imaging datasets, were not significantly different from chance level (Fig. 5B). Altogether, these results suggest that the detected noise correlations in our simultaneously acquired datasets can help reduce the error of the IC population code for sound azimuth.”

Minor weakness:

- Most studies of neural encoding of sound source azimuth are done in a noise-free environment, but the experimental setup in the present study had substantial background noise. This complicates comparison of the azimuth tuning results in this study to those of other studies. One is left wondering if azimuth sensitivity would have been greater in the absence of background noise, particularly for the imaging data where the signal was only about 12 dB above the noise. The description of the noise level and signal + noise level in the Methods should be made clearer. Mice hear from about 2.5 - 80 kHz, so it is important to know the noise level within this band as well as specifically within the band overlapping with the signal.

We agree with the reviewer that this information is useful. In our study, the background R.M.S. SPL during imaging across the mouse hearing range (2.5-80kHz) was 44.53 dB and for neuropixels recordings 34.68 dB. We have added this information to the methods section of the revised manuscript.

Reviewer #2 (Public Review):

In the present study, Boffi et al. investigate the manner in which the dorsal cortex of the of the inferior colliculus (DCIC), an auditory midbrain area, encodes sound location azimuth in awake, passively listening mice. By employing volumetric calcium imaging (scanned temporal focusing or s-TeFo), complemented with high-density electrode electrophysiological recordings (neuropixels probes), they show that sound-evoked responses are exquisitely noisy, with only a small portion of neurons (units) exhibiting spatial sensitivity. Nevertheless, a naïve Bayesian classifier was able to predict the presented azimuth based on the responses from small populations of these spatially sensitive units. A portion of the spatial information was provided by correlated trial-to-trial response variability between individual units (noise correlations). The study presents a novel characterization of spatial auditory coding in a non-canonical structure, representing a noteworthy contribution specifically to the auditory field and generally to systems neuroscience, due to its implementation of state-of-the-art techniques in an experimentally challenging brain region. However, nuances in the calcium imaging dataset and the naïve Bayesian classifier warrant caution when interpreting some of the results.

Strengths:

The primary strength of the study lies in its methodological achievements, which allowed the authors to collect a comprehensive and novel dataset. While the DCIC is a dorsal structure, it extends up to a millimetre in depth, making it optically challenging to access in its entirety. It is also more highly myelinated and vascularised compared to e.g., the cerebral cortex, compounding the problem. The authors successfully overcame these challenges and present an impressive volumetric calcium imaging dataset. Furthermore, they corroborated this dataset with electrophysiological recordings, which produced overlapping results. This methodological combination ameliorates the natural concerns that arise from inferring neuronal activity from calcium signals alone, which are in essence an indirect measurement thereof.

Another strength of the study is its interdisciplinary relevance. For the auditory field, it represents a significant contribution to the question of how auditory space is represented in the mammalian brain. "Space" per se is not mapped onto the basilar membrane of the cochlea and must be computed entirely within the brain. For azimuth, this requires the comparison between miniscule differences between the timing and intensity of sounds arriving at each ear. It is now generally thought that azimuth is initially encoded in two, opposing hemispheric channels, but the extent to which this initial arrangement is maintained throughout the auditory system remains an open question. The authors observe only a slight contralateral bias in their data, suggesting that sound source azimuth in the DCIC is encoded in a more nuanced manner compared to earlier processing stages of the auditory hindbrain. This is interesting, because it is also known to be an auditory structure to receive more descending inputs from the cortex.

Systems neuroscience continues to strive for the perfection of imaging novel, less accessible brain regions. Volumetric calcium imaging is a promising emerging technique, allowing the simultaneous measurement of large populations of neurons in three dimensions. But this necessitates corroboration with other methods, such as electrophysiological recordings, which the authors achieve. The dataset moreover highlights the distinctive characteristics of neuronal auditory representations in the brain. Its signals can be exceptionally sparse and noisy, which provide an additional layer of complexity in the processing and analysis of such datasets. This will be undoubtedly useful for future studies of other less accessible structures with sparse responsiveness.

Weaknesses:

Although the primary finding that small populations of neurons carry enough spatial information for a naïve Bayesian classifier to reasonably decode the presented stimulus is not called into question, certain idiosyncrasies, in particular the calcium imaging dataset and model, complicate specific interpretations of the model output, and the readership is urged to interpret these aspects of the study's conclusions with caution.

I remain in favour of volumetric calcium imaging as a suitable technique for the study, but the presently constrained spatial resolution is insufficient to unequivocally identify regions of interest as cell bodies (and are instead referred to as "units" akin to those of electrophysiological recordings). It remains possible that the imaging set is inadvertently influenced by non-somatic structures (including neuropil), which could report neuronal activity differently than cell bodies. Due to the lack of a comprehensive ground-truth comparison in this regard (which to my knowledge is impossible to achieve with current technology), it is difficult to imagine how many informative such units might have been missed because their signals were influenced by spurious, non-somatic signals, which could have subsequently misled the models. The authors reference the original Nature Methods article (Prevedel et al., 2016) throughout the manuscript, presumably in order to avoid having to repeat previously published experimental metrics. But the DCIC is neither the cortex nor hippocampus (for which the method was originally developed) and may not have the same light scattering properties (not to mention neuronal noise levels). Although the corroborative electrophysiology data largely eleviates these concerns for this particular study, the readership should be cognisant of such caveats, in particular those who are interested in implementing the technique for their own research.

A related technical limitation of the calcium imaging dataset is the relatively low number of trials (14) given the inherently high level of noise (both neuronal and imaging). Volumetric calcium imaging, while offering a uniquely expansive field of view, requires relatively high average excitation laser power (in this case nearly 200 mW), a level of exposure the authors may have wanted to minimise by maintaining a low the number of repetitions, but I yield to them to explain.

We assumed that the levels of heating by excitation light measured at the neocortex in Prevedel et al. (2016), were representative for DCIC also. Nevertheless, we recognize this approximation might not be very accurate, due to the differences in tissue architecture and vascularization from these two brain areas, just to name a few factors. The limiting factor preventing us from collecting more trials in our imaging sessions was that we observed signs of discomfort or slight distress in some mice after ~30 min of imaging in our custom setup, which we established as a humane end point to prevent distress. In consequence imaging sessions were kept to 25 min in duration, limiting the number of trials collected. However we cannot rule out that with more extensive habituation prior to experiments the imaging sessions could be prolonged without these signs of discomfort or if indeed influence from our custom setup like potential heating of the brain by illumination light might be the causing factor of the observed distress. Nevertheless, we note that previous work has shown that ~200mW average power is a safe regime for imaging in the cortex by keeping brain heating minimal (Prevedel et al., 2016), without producing the lasting damages observed by immunohistochemisty against apoptosis markers above 250mW (Podgorski and Ranganathan 2016, https://doi.org/10.1152/jn.00275.2016).

Calcium imaging is also inherently slow, requiring relatively long inter-stimulus intervals (in this case 5 s). This unfortunately renders any model designed to predict a stimulus (in this case sound azimuth) from particularly noisy population neuronal data like these as highly prone to overfitting, to which the authors correctly admit after a model trained on the entire raw dataset failed to perform significantly above chance level. This prompted them to feed the model only with data from neurons with the highest spatial sensitivity. This ultimately produced reasonable performance (and was implemented throughout the rest of the study), but it remains possible that if the model was fed with more repetitions of imaging data, its performance would have been more stable across the number of units used to train it. (All models trained with imaging data eventually failed to converge.) However, I also see these limitations as an opportunity to improve the technology further, which I reiterate will be generally important for volume imaging of other sparse or noisy calcium signals in the brain.

Transitioning to the naïve Bayesian classifier itself, I first openly ask the authors to justify their choice of this specific model. There are countless types of classifiers for these data, each with their own pros and cons. Did they actually try other models (such as support vector machines), which ultimately failed? If so, these negative results (even if mentioned en passant) would be extremely valuable to the community, in my view. I ask this specifically because different methods assume correspondingly different statistical properties of the input data, and to my knowledge naïve Bayesian classifiers assume that predictors (neuronal responses) are assumed to be independent within a class (azimuth). As the authors show that noise correlations are informative in predicting azimuth, I wonder why they chose a model that doesn't take advantage of these statistical regularities. It could be because of technical considerations (they mention computing efficiency), but I am left generally uncertain about the specific logic that was used to guide the authors through their analytical journey.

One of the main reasons we chose the naïve Bayesian classifier is indeed because it assumes that the responses of the simultaneously recorded neurons are independent and therefore it does not assume a contribution of noise correlations to the estimation of the posterior probability of each azimuth. This model would represent the null hypothesis that noise correlations do not contribute to the encoding of stimulus azimuth, which would be verified by an equal decoding outcome from correlated or decorrelated datasets. Since we observed that this is not the case, the model supports the alternative hypothesis that noise correlations do indeed influence stimulus azimuth encoding. We wanted to test these hypotheses with the most conservative approach possible that would be least likely to find a contribution of noise correlations. Other relevant reasons that justify our choice of the naive Bayesian classifier are its robustness against the limited numbers of trials we could collect in comparison to other more “data hungry” classifiers like SVM, KNN, or artificial neuronal nets. We did perform preliminary tests with alternative classifiers but the obtained decoding errors were similar when decoding the whole population activity (Author response image 2A). Dimensionality reduction following the approach described in the manuscript showed a tendency towards smaller decoding errors observed with an alternative classifier like KNN, but these errors were still larger than the ones observed with the naive Bayesian classifier (median error 45º). Nevertheless, we also observe a similar tendency for slightly larger decoding errors in the absence of noise correlations (decorrelated, Author response image 2B). Sentences detailing the logic of classifier choice are now included in the results section at page 10 and at the last paragraph of page 18 (see responses to Reviewer 1).

Author response image 2.

A) Cumulative distribution plots of the absolute cross-validated single-trial prediction errors obtained using different classifiers (blue; KNN: K-nearest neighbors; SVM: support vector machine ensemble) and chance level distribution (gray) on the complete populations of imaged units. Cumulative distribution plots of the absolute cross-validated singletrial prediction errors obtained using a Bayes classifier (naive approximation for computation efficiency) to decode the single-trial response patterns from the 31 top ranked units in the simultaneously imaged datasets across mice (cyan), modeled decorrelated datasets (orange) and the chance level distribution associated with our stimulation paradigm (gray). Vertical dashed lines show the medians of cumulative distributions. K.S. w/Sidak: Kolmogorov-Smirnov with Sidak.

That aside, there remain other peculiarities in model performance that warrant further investigation. For example, what spurious features (or lack of informative features) in these additional units prevented the models of imaging data from converging?

Considering the amount of variability observed throughout the neuronal responses both in imaging and neuropixels datasets, it is easy to suspect that the information about stimulus azimuth carried in different amounts by individual DCIC neurons can be mixed up with information about other factors (Stringer et al., 2019). In an attempt to study the origin of these features that could confound stimulus azimuth decoding we explored their relation to face movement (Supplemental Figure 2), finding a correlation to snout movements, in line with previous work by Stringer et al. (2019).

In an orthogonal question, did the most spatially sensitive units share any detectable tuning features? A different model trained with electrophysiology data in contrast did not collapse in the range of top-ranked units plotted. Did this model collapse at some point after adding enough units, and how well did that correlate with the model for the imaging data?

Our electrophysiology datasets were much smaller in size (number of simultaneously recorded neurons) compared to our volumetric calcium imaging datasets, resulting in a much smaller total number of top ranked units detected per dataset. This precluded the determination of a collapse of decoder performance due to overfitting beyond the range plotted in Fig 4G.

How well did the form (and diversity) of the spatial tuning functions as recorded with electrophysiology resemble their calcium imaging counterparts? These fundamental questions could be addressed with more basic, but transparent analyses of the data (e.g., the diversity of spatial tuning functions of their recorded units across the population). Even if the model extracts features that are not obvious to the human eye in traditional visualisations, I would still find this interesting.

The diversity of the azimuth tuning curves recorded with calcium imaging (Fig. 3B) was qualitatively larger than the ones recorded with electrophysiology (Fig. 4B), potentially due to the larger sampling obtained with volumetric imaging. We did not perform a detailed comparison of the form and a more quantitative comparison of the diversity of these functions because the signals compared are quite different, as calcium indicator signal is subject to non linearities due to Ca2+ binding cooperativity and low pass filtering due to binding kinetics. We feared this could lead to misleading interpretations about the similarities or differences between the azimuth tuning functions in imaged and electrophysiology datasets. Our model uses statistical response dependency to stimulus azimuth, which does not rely on features from a descriptive statistic like mean response tuning. In this context, visualizing the trial-to-trial responses as a function of azimuth shows “features that are not obvious to the human eye in traditional visualizations” (Fig. 3D, left inset).

Finally, the readership is encouraged to interpret certain statements by the authors in the current version conservatively. How the brain ultimately extracts spatial neuronal data for perception is anyone's guess, but it is important to remember that this study only shows that a naïve Bayesian classifier could decode this information, and it remains entirely unclear whether the brain does this as well. For example, the model is able to achieve a prediction error that corresponds to the psychophysical threshold in mice performing a discrimination task (~30 {degree sign}). Although this is an interesting coincidental observation, it does not mean that the two metrics are necessarily related. The authors correctly do not explicitly claim this, but the manner in which the prose flows may lead a non-expert into drawing that conclusion.

To avoid misleading the non-expert readers, we have clarified in the manuscript that the observed correspondence between decoding error and psychophysical threshold is explicitly coincidental.

Page 13, end of middle paragraph:

“If we consider the median of the prediction error distribution as an overall measure of decoding performance, the single-trial response patterns from subsamples of at least the 7 top ranked units produced median decoding errors that coincidentally matched the reported azimuth discrimination ability of mice (Fig 4G, minimum audible angle = 31º) (Lauer et al., 2011).”

Page 14, bottom paragraph:

“Decoding analysis (Fig. 4F) of the population response patterns from azimuth dependent top ranked units simultaneously recorded with neuropixels probes showed that the 4 top ranked units are the smallest subsample necessary to produce a significant decoding performance that coincidentally matches the discrimination ability of mice (31° (Lauer et al., 2011)) (Fig. 5F, G).”

We also added to the Discussion sentences clarifying that a relationship between these two variables remains to be determined and it also remains to be determined if the DCIC indeed performs a bayesian decoding computation for sound localization.

Page 20, bottom:

“… Concretely, we show that sound location coding does indeed occur at DCIC on the single trial basis, and that this follows a comparable mechanism to the characterized population code at CNIC (Day and Delgutte, 2013). However, it remains to be determined if indeed the DCIC network is physiologically capable of Bayesian decoding computations. Interestingly, the small number of DCIC top ranked units necessary to effectively decode stimulus azimuth suggests that sound azimuth information is redundantly distributed across DCIC top ranked units, which points out that mechanisms beyond coding efficiency could be relevant for this population code.

While the decoding error observed from our DCIC datasets obtained in passively listening, untrained mice coincidentally matches the discrimination ability of highly trained, motivated mice (Lauer et al., 2011), a relationship between decoding error and psychophysical performance remains to be determined. Interestingly, a primary sensory representations should theoretically be even more precise than the behavioral performance as reported in the visual system (Stringer et al., 2021).”

Moreover, the concept of redundancy (of spatial information carried by units throughout the DCIC) is difficult for me to disentangle. One interpretation of this formulation could be that there are non-overlapping populations of neurons distributed across the DCIC that each could predict azimuth independently of each other, which is unlikely what the authors meant. If the authors meant generally that multiple neurons in the DCIC carry sufficient spatial information, then a single neuron would have been able to predict sound source azimuth, which was not the case. I have the feeling that they actually mean "complimentary", but I leave it to the authors to clarify my confusion, should they wish.

We observed that the response patterns from relatively small fractions of the azimuth sensitive DCIC units (4-7 top ranked units) are sufficient to generate an effective code for sound azimuth, while 32-40% of all simultaneously recorded DCIC units are azimuth sensitive. In light of this observation, we interpreted that the azimuth information carried by the population should be redundantly distributed across the complete subpopulation of azimuth sensitive DCIC units.

In summary, the present study represents a significant body of work that contributes substantially to the field of spatial auditory coding and systems neuroscience. However, limitations of the imaging dataset and model as applied in the study muddles concrete conclusions about how the DCIC precisely encodes sound source azimuth and even more so to sound localisation in a behaving animal. Nevertheless, it presents a novel and unique dataset, which, regardless of secondary interpretation, corroborates the general notion that auditory space is encoded in an extraordinarily complex manner in the mammalian brain.

Reviewer #3 (Public Review):

Summary:

Boffi and colleagues sought to quantify the single-trial, azimuthal information in the dorsal cortex of the inferior colliculus (DCIC), a relatively understudied subnucleus of the auditory midbrain. They used two complementary recording methods while mice passively listened to sounds at different locations: a large volume but slow sampling calcium-imaging method, and a smaller volume but temporally precise electrophysiology method. They found that neurons in the DCIC were variable in their activity, unreliably responding to sound presentation and responding during inter-sound intervals. Boffi and colleagues used a naïve Bayesian decoder to determine if the DCIC population encoded sound location on a single trial. The decoder failed to classify sound location better than chance when using the raw single-trial population response but performed significantly better than chance when using intermediate principal components of the population response. In line with this, when the most azimuth dependent neurons were used to decode azimuthal position, the decoder performed equivalently to the azimuthal localization abilities of mice. The top azimuthal units were not clustered in the DCIC, possessed a contralateral bias in response, and were correlated in their variability (e.g., positive noise correlations). Interestingly, when these noise correlations were perturbed by inter-trial shuffling decoding performance decreased. Although Boffi and colleagues display that azimuthal information can be extracted from DCIC responses, it remains unclear to what degree this information is used and what role noise correlations play in azimuthal encoding.

Strengths:

The authors should be commended for collection of this dataset. When done in isolation (which is typical), calcium imaging and linear array recordings have intrinsic weaknesses. However, those weaknesses are alleviated when done in conjunction with one another - especially when the data largely recapitulates the findings of the other recording methodology. In addition to the video of the head during the calcium imaging, this data set is extremely rich and will be of use to those interested in the information available in the DCIC, an understudied but likely important subnucleus in the auditory midbrain.

The DCIC neural responses are complex; the units unreliably respond to sound onset, and at the very least respond to some unknown input or internal state (e.g., large inter-sound interval responses). The authors do a decent job in wrangling these complex responses: using interpretable decoders to extract information available from population responses.

Weaknesses:

The authors observe that neurons with the most azimuthal sensitivity within the DCIC are positively correlated, but they use a Naïve Bayesian decoder which assume independence between units. Although this is a bit strange given their observation that some of the recorded units are correlated, it is unlikely to be a critical flaw. At one point the authors reduce the dimensionality of their data through PCA and use the loadings onto these components in their decoder. PCA incorporates the correlational structure when finding the principal components and constrains these components to be orthogonal and uncorrelated. This should alleviate some of the concern regarding the use of the naïve Bayesian decoder because the projections onto the different components are independent. Nevertheless, the decoding results are a bit strange, likely because there is not much linearly decodable azimuth information in the DCIC responses. Raw population responses failed to provide sufficient information concerning azimuth for the decoder to perform better than chance. Additionally, it only performed better than chance when certain principal components or top ranked units contributed to the decoder but not as more components or units were added. So, although there does appear to be some azimuthal information in the recoded DCIC populations - it is somewhat difficult to extract and likely not an 'effective' encoding of sound localization as their title suggests.

As described in the responses to reviewers 1 and 2, we chose the naïve Bayes classifier as a decoder to determine the influence of noise correlations through the most conservative approach possible, as this classifier would be least likely to find a contribution of correlated noise. Also, we chose this decoder due to its robustness against limited numbers of trials collected, in comparison to “data hungry” non linear classifiers like KNN or artificial neuronal nets. Lastly, we observed that small populations of noisy, unreliable (do not respond in every trial) DCIC neurons can encode stimulus azimuth in passively listening mice matching the discrimination error of trained mice. Therefore, while this encoding is definitely not efficient, it can still be considered effective.

Although this is quite a worthwhile dataset, the authors present relatively little about the characteristics of the units they've recorded. This may be due to the high variance in responses seen in their population. Nevertheless, the authors note that units do not respond on every trial but do not report what percent of trials that fail to evoke a response. Is it that neurons are noisy because they do not respond on every trial or is it also that when they do respond they have variable response distributions? It would be nice to gain some insight into the heterogeneity of the responses.

The limited number of azimuth trial repetitions that we could collect precluded us from making any quantification of the unreliability (failures to respond) and variability in the response distributions from the units we recorded, as we feared they could be misleading. In qualitative terms, “due to the high variance in responses seen” in the recordings and the limited trial sampling, it is hard to make any generalization. In consequence we referred to the observed response variance altogether as neuronal noise. Considering these points, our datasets are publicly available for exploration of the response characteristics.

Additionally, is there any clustering at all in response profiles or is each neuron they recorded in the DCIC unique?

We attempted to qualitatively visualize response clustering using dimensionality reduction, observing different degrees of clustering or lack thereof across the azimuth classes in the datasets collected from different mice. It is likely that the limited number of azimuth trials we could collect and the high response variance contribute to an inconsistent response clustering across datasets.

They also only report the noise correlations for their top ranked units, but it is possible that the noise correlations in the rest of the population are different.

For this study, since our aim was to interrogate the influence of noise correlations on stimulus azimuth encoding by DCIC populations, we focused on the noise correlations from the top ranked unit subpopulation, which likely carry the bulk of the sound location information.  Noise correlations can be defined as correlation in the trial to trial response variation of neurons. In this respect, it is hard to ascertain if the rest of the population, that is not in the top rank unit percentage, are really responding and showing response variation to evaluate this correlation, or are simply not responding at all and show unrelated activity altogether. This makes observations about noise correlations from “the rest of the population” potentially hard to interpret.

It would also be worth digging into the noise correlations more - are units positively correlated because they respond together (e.g., if unit x responds on trial 1 so does unit y) or are they also modulated around their mean rates on similar trials (e.g., unit x and y respond and both are responding more than their mean response rate). A large portion of trial with no response can occlude noise correlations. More transparency around the response properties of these populations would be welcome.

Due to the limited number of azimuth trial repetitions collected, to evaluate noise correlations we used the non parametric Kendall tau correlation coefficient which is a measure of pairwise rank correlation or ordinal association in the responses to each azimuth. Positive rank correlation would represent neurons more likely responding together. Evaluating response modulation “around their mean rates on similar trials” would require assumptions about the response distributions, which we avoided due to the potential biases associated with limited sample sizes.

It is largely unclear what the DCIC is encoding. Although the authors are interested in azimuth, sound location seems to be only a small part of DCIC responses. The authors report responses during inter-sound interval and unreliable sound-evoked responses. Although they have video of the head during recording, we only see a correlation to snout and ear movements (which are peculiar since in the example shown it seems the head movements predict the sound presentation). Additional correlates could be eye movements or pupil size. Eye movement are of particular interest due to their known interaction with IC responses - especially if the DCIC encodes sound location in relation to eye position instead of head position (though much of eye-position-IC work was done in primates and not rodent). Alternatively, much of the population may only encode sound location if an animal is engaged in a localization task. Ideally, the authors could perform more substantive analyses to determine if this population is truly noisy or if the DCIC is integrating un-analyzed signals.

We unsuccessfully attempted eye tracking and pupillometry in our videos. We suspect that the reason behind this is a generally overly dilated pupil due to the low visible light illumination conditions we used which were necessary to protect the PMT of our custom scope.

It is likely that DCIC population activity is integrating un-analyzed signals, like the signal associated with spontaneous behaviors including face movements (Stringer et al., 2019), which we observed at the level of spontaneous snout movements. However investigating if and how these signals are integrated to stimulus azimuth coding requires extensive behavioral testing and experimentation which is out of the scope of this study. For the purpose of our study, we referred to trial-to-trial response variation as neuronal noise. We note that this definition of neuronal noise can, and likely does, include an influence from un-analyzed signals like the ones from spontaneous behaviors.

Although this critique is ubiquitous among decoding papers in the absence of behavioral or causal perturbations, it is unclear what - if any - role the decoded information may play in neuronal computations. The interpretation of the decoder means that there is some extractable information concerning sound azimuth - but not if it is functional. This information may just be epiphenomenal, leaking in from inputs, and not used in computation or relayed to downstream structures. This should be kept in mind when the authors suggest their findings implicate the DCIC functionally in sound localization.

Our study builds upon previous reports by other independent groups relying on “causal and behavioral perturbations” and implicating DCIC in sound location learning induced experience dependent plasticity (Bajo et al., 2019, 2010; Bajo and King, 2012), which altogether argues in favor of DCIC functionality in sound localization.

Nevertheless, we clarified in the discussion of the revised manuscript that a relationship between the observed decoding error and the psychophysical performance, or the ability of the DCIC network to perform Bayesian decoding computations, both remain to be determined (please see responses to Reviewer #2).

It is unclear why positive noise correlations amongst similarly tuned neurons would improve decoding. A toy model exploring how positive noise correlations in conjunction with unreliable units that inconsistently respond may anchor these findings in an interpretable way. It seems plausible that inconsistent responses would benefit from strong noise correlations, simply by units responding together. This would predict that shuffling would impair performance because you would then be sampling from trials in which some units respond, and trials in which some units do not respond - and may predict a bimodal performance distribution in which some trials decode well (when the units respond) and poor performance (when the units do not respond).

In samples with more that 2 dimensions, the relationship between signal and noise correlations is more complex than in two dimensional samples (Montijn et al., 2016) which makes constructing interpretable and simple toy models of this challenging. Montijn et al. (2016) provide a detailed characterization and model describing how the accuracy of a multidimensional population code can improve when including “positive noise correlations amongst similarly tuned neurons”. Unfortunately we could not successfully test their model based on Mahalanobis distances as we could not verify that the recorded DCIC population responses followed a multivariate gaussian distribution, due to the limited azimuth trial repetitions we could sample.

Significance:

Boffi and colleagues set out to parse the azimuthal information available in the DCIC on a single trial. They largely accomplish this goal and are able to extract this information when allowing the units that contain more information about sound location to contribute to their decoding (e.g., through PCA or decoding on top unit activity specifically). The dataset will be of value to those interested in the DCIC and also to anyone interested in the role of noise correlations in population coding. Although this work is first step into parsing the information available in the DCIC, it remains difficult to interpret if/how this azimuthal information is used in localization behaviors of engaged mice.

Recommendations for the authors:

Reviewer #2 (Recommendations For The Authors):

General:

The manuscript is generally well written, but could benefit from a quick proof by a native English speaker (e.g., "the" inferior colliculus is conventionally used with its article). The flow of arguments is also generally easy to follow, but I would kindly ask the authors to consider elaborating or clarifying the following points (including those already mentioned in my public review).

(1) Choice of model:

There are countless ways one can construct a decoder or classifier that can predict a presented sensory stimulus based on a population neuronal response. Given the assumptions of independence as mentioned in my public review, I would ask the authors to explicitly justify their choice of a naïve Bayesian classifier.

A section detailing the logic of classifier choice is now included in the results section at page 10 and the last paragraph of page 18 from the revised version of the manuscript.

(2) Number of imaging repetitions:

For particularly noisy datasets, 14 repetitions is indeed quite few. I reckon this was not the choice of the authors, but rather limited by the inherent experimental conditions. Despite minimisation of required average laser power during the development of s-TeFo imaging, the authors still required almost 200 mW (which is still quite a lot of exposure). Although 14 repetitions for 13 azimuthal locations every 5 s is at face value a relatively short imaging session (~15 min.), at 191 mW, with the desire to image mice multiple times, I could imagine that this is a practical limitation the authors faced (to avoid excessive tissue heating or photodamage, which was assessed in the original Nature Methods article, but not here). Nevertheless, this logic (or whatever logic they had) should be explained for non-imaging experts in the readership.

This is now addressed in the answers to the public reviews.

(3) Redundancy:

It is honestly unclear to me what the authors mean by this. I don't speculate that they mean there are "redundant" (small) populations of neurons that sufficiently encode azimuth, but I'm actually not certain. If that were the case, I believe this would need further clarification, since redundant representations would be both inconsistent with the general (perhaps surprising) finding that large populations are not required in the DCIC, which is thought to be the case at earlier processing stages.

In the text we are referring to the azimuth information being redundantly distributed across DCIC top ranked units. We do not mention redundant “populations of neurons”.

(4) Correspondence of decoding accuracy with psychometric functions in mice: While this is an interesting coincidental observation, it should not be interpreted that the neuronal detection threshold in the DCIC somehow is somehow responsible its psychometric counterpart (which is an interesting yet exceedingly complex question). Although I do not believe the authors intended to suggest this, I would personally be cautious in the way I describe this correspondence. I mention this because the authors point it out multiple times in the manuscript (whereas I would have just mentioned it once in passing).

This is now clarified in the revised manuscript.

(5) Noisy vs. sparse:

I'm confident that the authors understand the differences between these terms, both in concept (stochastic vs. scattered) and in context (neuronal vs. experimental), but I personally would be cautious in the way I use them in the description of the study. Indeed, auditory neuronal signals are to my knowledge generally thought to be both sparse and noisy, which is in itself interesting, but the study also deals with substantial experimental (recording) noise, and I think it's important for the readership to understand when "noise" refers to the recordings (in particular the imaging data) and to neuronal activity. I mention this specifically because "noisy" appears in the title.

We have clarified this issue at the bottom of page 5 by adding the following sentences to the revised manuscript:

“In this section we used the word “noise” to refer to the sound stimuli used and recording setup background sound levels or recording noise in the acquired signals. To avoid confusion, from now on in the manuscript the word “noise” will be used in the context of neuronal noise, which is the trial-to-trial variation in neuronal responses unrelated to stimuli, unless otherwise noted.”

(6)  More details in the Methods:

The Methods section is perhaps the least-well structured part of the present manuscript in my view, and I encourage the authors to carefully go through it and add the following information (in case I somehow missed it).

a. Please also indicate the number of animals used here.

Added.

b. How many sessions were performed on each mouse?

This is already specified in the methods section in page 25:

“mice were imaged a total of 2-11 times (sessions), one to three times a week.”

We added for clarification:

“Datasets here analyzed and reported come from the imaging session in which we observed maximal calcium sensor signal (peak AAV expression) and maximum number of detected units.”

c. For the imaging experiments, was it possible to image the same units from session tosession?

This is not possible for sTeFo 2P data due to low spatial resolution which makes precisely matching neuron ROIs across sessions challenging.

d. Could the authors please add more detail to the analyses of the videos (to track facialmovements) or provide a reference?

Added citation.

e. The same goes for the selection of subcellular regions of interest that were used as"units."

Added to page 25:

“We used the CaImAn package (Giovannucci et al., 2019) for automatic ROI segmentation through constrained non negative matrix factorization and selected ROIs (Units) showing clear Ca transients consistent with neuronal activity, and IC neuron somatic shape and size (Schofield and Beebe, 2019).”

Specific: In order to maximise the efficiency of my comments and suggestions (as there are no line numbers), my numerated points are organised in sequential order.

(1) Abstract: I wouldn't personally motivate the study with the central nucleus of the IC (i.e. Idon't think this is necessary). I think the authors can motivate it simply with the knowledge gaps in spatial coding throughout the auditory system, in which such large data sets such as the ones presented here are of general value.

(2) Page 4: 15-50 kHz "white" noise is incorrect. It should be "band-passed" noise.

Changed.

(3) Supplemental figure 1, panel A: Since the authors could not identify cell bodiesunequivocally from their averaged volume timeseries data, it would be clearer to the readership if larger images are shown, so that they can evaluate (speculate) for themselves what subcellular structures were identified as units. Even better would be to include a planar image through a cross-section. As mentioned above, not everything determined for the cortex or hippocampus can be assumed to be true for the DCIC.

The raw images and segmentations are publicly available for detailed inspections.

(4) Supplemental figure 2, panel A: This panel requires further explanation, in particular thepanel on the right. I assume that to be a simple subtraction of sequential frames, but I'm thrown off by the "d(Grey)" colour bar. Also, if "grey" refers to the neutral colour, it is conventionally spelled "gray" in US-American English.

Changed.

(5) Supplemental figure 2, panel B: I'm personally curious why the animals exhibitedmovement just prior to a stimulus. Did they learn to anticipate the presentation of a sound after some habituation? Is that somehow a pre-emptive startle response? We observe that in our own experiments (but as we stochastically vary the inter-trial-intervals, the movement typically occurs directly after the stimulus). I don't suggest the authors dwell on this, but I find it an interesting observation.

It is indeed interesting, but we can’t conclude much about it without comparing it to random inter-trial-intervals.

(6) Supplemental figure 3: I personally find these data (decoding of all electrophysiologicaldata) of central relevance to the study, since it mirrors the analyses presented for its imaging data counterpart and encourage the authors to move it to the main text.

Changed.

(7) Page 12: Do the authors have any further analyses of spatial tuning functions? We allknow they can parametrically obscure (i.e., bi-lobed, non-monotonic, etc.), but having these parameters (even if just in a supplemental figure) would be informative for the spatial auditory community.

We dedicated significant effort to attempt to parametrize and classify the azimuth response dependency functions from the recorded DCIC cells in an unbiased way. Nevertheless, given the observed response noise and the “obscure” properties of spatial tuning functions mentioned by the reviewer, we could only reach the general qualitative observation of having a more frequent contralateral selectivity.

(8) Page 14 (end): Here, psychometric correspondence is referenced. Please add theLauer et al., (2011) reference, or, as I would, remove the statement entirely and save it for the discussion (where it is also mentioned and referenced).

Changed.

(9) Figure 5, Panels B and C: Why don't the authors report the Kruskal-Wallis tests (forincreasing number of units training the model), akin to e.g., Panel G of Figure 4? I think that would be interesting to see (e.g., if the number of required units to achieve statistical significance is the same).

Within class randomization produced a moderate effect on decoder performance, achieving statistical significance at similar numbers of units, as seen in figure 5 panels B and C. We did not include these plots for the sake of not cluttering the figure with dense distributions and fuzzing the visualization of the differences between the distributions shown.

(10) Figure 5, Panels B and C (histograms): I see a bit of skewedness in the distributions(even after randomisation). Where does this come from? This is just a small talking point.

We believe this is potentially due to more than one distribution of pairwise correlations combined into one histogram (like in a Gaussian mixture model).

(11) Page 21: Could the authors please specify that the Day and Delgutte (2013) study wasperformed on rabbits? Since rabbits have an entirely different spectral hearing range compared to mice, spatial coding principles could very well be different in those animals (and I'm fairly certain such a study has not yet been published for mice).

Specified.

(12) Page 22: I'd encourage the authors to remove the reference to Rayleigh's duplextheory, since mice hardly (if at all) use interaural time differences for azimuthal sound localisation, given their generally high-frequency hearing range.

That sentence is meant to discuss beyond the mouse model an exciting outlook of our findings in light of previous reports, which is a hypothetical functional relationship between the tonotopy in DCIC and the spatial distribution of azimuth sensitive DCIC neurons. We have clarified this now in the text.

(13) Page 23: I believe the conventional verb for gene delivery with viruses is still"transduce" (or "infect", but not "induce"). What was the specific "syringe" used for stereotactic injections? Also, why were mice housed separately after surgery? This question pertains to animal welfare.

Changed. The syringe was a 10ml syringe to generate positive or negative pressure, coupled to the glass needle through a silicon tubing via a luer 3-way T valve. Single housing was chosen to avoid mice compromising each other’s implantations. Therefore this can be seen as a refinement of our method to maximize the chances of successful imaging per implanted mouse.

(14) Page 25: Could the authors please indicate the refractory period violation time windowhere? I had to find it buried in the figure caption of Supplementary figure 1.

Added.

(15) Page 27: What version of MATLAB was used? This could be important for reproductionof the analyses, since The Mathworks is infamously known to add (or even more deplorably, modify) functions in particular versions (and not update older ones accordingly).

Added.

Reviewer #3 (Recommendations For The Authors):

Overall I thought this was a nice manuscript and a very interesting dataset. Here are some suggestions and minor corrections:

You may find this work of interest - 'A monotonic code for sound azimuth in primate inferior colliculus' 2003, Groh, Kelly & Underhill.

We thank the reviewer for pointing out this extremely relevant reference, which we regrettably failed to cite. It is now included in the revised version of the manuscript.

In your introduction, you state "our findings point to a functional role of DCIC in sound location coding". Though your results show that there is azimuthal information contained in a subset of DCIC units there's no evidence in the manuscript that shows a functional link between this representation and sound localization.

This is now addressed in the answers to the public reviews.

I found the variability in your DCIC population quite striking - especially during the intersound intervals. The entrainment of the population in the imaging datatset suggests some type of input activating the populations - maybe these are avenues for further probing the variability here:

(1) I'm curious if you can extract eye movements from your video. Work from Jennifer Grohshows that some cells in the primate inferior colliculus are sensitive to different eye positions (Groh et. al., 2001). With recent work showing eye movements in rodents, it may explain some of the variance in the DCIC responses.

This is now addressed in the answers to the public reviews.

(2) I was also curious if the motor that moves the speaker made noise It could be possiblesome of the 'on going' activity could be some sound-evoked response.

We were careful to set the stepper motor speed so that it produced low frequency noise, within a band mostly outside of the hearing range of mice (<4kHz). Nevertheless, we cannot fully rule out that a very quiet but perhaps very salient component of the motor noise could influence the activity during the inter trial periods. The motor was stationary and quiet for a period of at least one stimulus duration before and during stimulus presentation.  

(3) Was the sound you present frozen or randomly generated on each trial? Could therebe some type of structure in the noise you presented that sometimes led cells to respond to a particular azimuth location but not others?

The sound presented was frozen noise. This is now clarified in the methods section.

It may be useful to quantify the number of your units that had refractory period violations.

Our manual curation of sorted units was very stringent to avoid mixing differently tuned neurons. The single units analyzed had very infrequent refractory period violations, in less than ~5% of the spikes, considering a 2 ms refractory period.

Was the video recording contralateral or ipsilateral to the recording?

The side of the face ipsilateral to the imaged IC was recorded. Added to methods.

I was struck by the snout and ear movements - in the example shown in Supplementary Figure 2B it appears as they are almost predicting sound onset. Was there any difference in ear movements in the habituated and non-habituated animals? Also, does the placement of the cranial window disturb any of the muscles used in ear movement?

Mouse snout movements appear to be quite active perhaps reflecting arousal (Stringer et al., 2019). We cannot rule out that the cranial window implantation disturbed ear movement but while moving the mouse headfixed we observed what could be considered normal ear movements.

Did you correlate time-point by time-point in the average population activity and movement or did you try different temporal labs/leads in case the effect of the movements was delayed in some way?

Point by point due to 250ms time resolution of imaging.

Are the video recordings only available during the imaging? It would be nice to see the same type of correlations in the neuropixel-acquired data as well.

Only imaging. For neuropixels recordings, we were skeptical about face videography as we suspected that face movements were likely influenced by the acute nature of the preparation procedure. Our cranial window preparation in the other hand involved a recovery period of at least 4 weeks. Therefore we were inclined to perform videographical interrogation of face movements on these mice instead.

If you left out more than 1 trial do you think this would help your overfitting issue (e.g. leaving out 20% of the data).

Due to the relatively small number of trial repetitions collected, fitting the model with an even smaller training dataset is unlikely to help overfitting and will likely decrease decoder performance.

It would be nice to see a confusion matrix - even though azimuthal error and cumulative distribution of error are a fine way to present the data - a confusion matrix would tell us which actual sounds the decoder is confusing. Just looking at errors could result in some funky things where you reduce the error generally but never actually estimate the correct location.

We considered confusion matrices early on in our study but they were not easily interpretable or insightful, likely due to the relatively low discrimination ability of the mouse model with +/- 30º error after extensive training. Therefore, we reasoned that in passively listening mice (and likely trained mice too) with limited trial repetitions, an undersampled and diffuse confusion matrix is expected which is not an ideal means of visualizing and comparing decoding errors. Hence we relied on cumulative error distributions.

Do your top-ranked units have stronger projections onto your 10-40 principal components?

It would be interesting to know if the components are mostly taking into account those 30ish percent of the population that is dependent upon azimuth.

Inspection of PC loadings across units ranked based on response dependency to stimulus azimuth does not show a consistent stronger projection of top ranked units onto the first 10-40 principal components (Author response image 3).

Author response image 3.

PC loading matrices for each recorded mouse. The units recorded in each mouse are ranked in descending order of response dependency to stimulus azimuth based on  the p value of the chi square test. Units above the red dotted line display a chi square p value < 0.05, units below this line have p values >= 0.05.

How much overlap is there in the tuning of the top-ranked units?

This is quite varying from mouse to mouse and imaging vs electrophysiology, which makes it hard to make a generalization since this might depend on the unique DCIC population sampled in each mouse.

I'm not really sure I follow what the nS/N adds - it doesn't really measure tuning but it seems to be introduced to discuss/extract some measure of tuning.

nS/N is used to quantify how noisy neurons are, independent of how sensitive their responses are to the stimulus azimuth.

Is the noise correlation - observed to become more positive - for more contralateral stimuli a product of higher firing rates due to a more preferred stimulus presentation or a real effect in the data? Was there any relationship between distance and strength of observed noise correlation in the DCIC?

We observed a consistent and homogeneous trend of pairwise noise correlation distributions either shifted or tailed towards more positive values across stimulus azimuths, for imaging and electrophysiology datasets (Author response image 3). The lower firing frequency observed in neuropixels recordings in response to ipsilateral azimuths could have affected the statistical power of the comparison between the pairwise noise correlation coefficient distribution to its randomized chance level, but the overall histogram shapes qualitatively support this consistent trend across azimuths (Author response image 4).

Author response image 4.

Distribution histograms for the pairwise correlation coefficients (Kendall tau) from pairs of simultaneously recorded top ranked units across mice (blue) compared to the chance level distribution obtained through randomization of the temporal structure of each unit’s activity to break correlations (purple). Vertical lines show the medians of these distributions. Imaging data comes from n = 12 mice and neuropixels data comes from n = 4 mice.

Typos:

'a population code consisting on the simultaneous" > should on be of?

'half of the trails' > trails should be trials?

'referncing the demuxed channels' > should it be demixed?

Corrected.

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public Review):

Summary:

Tateishi et al. report a Tn-seq-based analysis of genetic requirements for growth and fitness in 8 clinical strains of Mycobacterium intracellulare Mi), and compare the findings with a type strain ATCC13950. The study finds a core set of 131 genes that are essential in all nine strains, and therefore are reasonably argued as potential drug targets. Multiple other genes required for fitness in clinical isolates have been found to be important for hypoxic growth in the type strain.

Strengths:

The study has generated a large volume of Tn-seq datasets of multiple clinical strains of Mi from multiple growth conditions, including from mouse lungs. The dataset can serve as an important resource for future studies on Mi, which despite being clinically significant remains a relatively understudied species of mycobacteria.

Thank you for reviewing our manuscript and finding the significance of our data.

Weaknesses:

The paper lacks clarity in data presentation and organization. For example, some of the key data on cfu counts of clinical Mi strains in a mouse model can be presented along with the Tn-seq dataset in Figure 6, the visualization of which can be improved with volcano plots. etc. Improvement in data visualization is perhaps necessary throughout the paper.

Thank you for the comment on the data presentation of in vivo studies. We previously revealed the time-course data on CFUs, animal survival, and tissue pathology from the pure strains (Tateishi Y. BMC Microbiol. 2023; new Ref #22) . Based on these data, we assumed that we would be able to harvest sufficient number of colonies from mice infected with M.i.27 or M.i.198, and we performed in vivo TnSeq studies using these two strains. We have referred to our previous publication (new Ref #22) on the virulence of MAC-PD strains used in this study for mice in the revised manuscript (page12, line 212).

The data of CFU counts were shown in new Supplementary Fig. 3b. In the manuscript text, we explained as follows (page 12, lines 212-216): “The time course of the changes in the bacterial burden showed a pattern similar to those of the wild-type strains M.i.198 and M.i.27, respectively, except that it was not possible to harvest sufficient colonies (as few as 104/mouse) in the few mice infected with the M.i.27 Tn mutant strain in week 8 and week 16 (page 12, lines 212-216; new Supplementary Fig, 3b, new Supplementary Table 8)”.

Regarding the suggestion to include volcano plots, we appreciate the proposal but chose not to adopt this format, as the main aim of this study was to identify genes commonly required for in vitro and in vivo fitness across multiple M. intracellulare strains, rather than to highlight differential genetic requirements within a single strain. Volcano plots are useful for visualizing differential values and significance for a single dataset but are less suited for cross-strain comparisons of shared gene sets. Our approach is aligned with the methodology used by Cary et al. (PLoS Pathog. 2018; new Ref#8), who similarly focused on identifying conserved genetic requirements across M. tuberculosis genotypes without employing volcano plots.

[References]

Tateishi, Y. et al. Virulence of Mycobacterium intracellulare clinical strains in a mouse model of lung infection - role of neutrophilic inflammation in disease severity. BMC Microbiol 23, 94 (2023).

Carey, A.F. et al. TnSeq of Mycobacterium tuberculosis clinical isolates reveals strain-specific antibiotic liabilities. PLoS Pathog 14, e1006939 (2018).

The primary claim of the study that the clinical strains are better adapted for hypoxic growth is not well-supported by the data presented in Figure 7.

Thank you for the comments on the difference of adaptation for hypoxic growth between ATCC13950 and clinical MAC-PD strains. To clarify, growth rates shown in Figure 7 were calculated at the inflection point (midpoint) of the growth curves, which were modeled using a four-parameter logistic (4P logistic) model. As described in the Discussion, we found the pattern of hypoxic adaptation characteristics of the clinical MAC-PD strains from the growth curve forms. Taking into consideration the impact of growing bacteria on the disease progression of MAC-PD, the slow growth with early entry to log phase implicates the continuous impact on the infected hosts during logarithmic bacterial growth, which may be involved in the persistent and steadily progressive illness of MAC-PD for years in humans.

Unlike time-lapse imaging assay, the completely seamless sampling of culture for CFU assay is impossible. Nevertheless, we collected sufficient timepoints to allow reliable curve fitting with the 4P logistic model and thus consider our growth data to represent a valid approximation of continuous growth dynamics.

Regarding the suggestion of mixed culture experiments, we agree that such studies could be informative. However, co-culture conditions introduce additional variables, including inter-strain competition or synergy, which can obscure the specific contributions of hypoxic adaptation in each strain. Therefore, we believe that the current approach using monoculture growth curves under defined oxygen conditions offers a clearer interpretation of strain-specific hypoxic responses.

The title of the paper is misleading as the study doesn't provide any mechanistic aspect of hypoxic adaptation in Mi.

Thank you for the comment on the article title. We admit that this paper does not directly reveal the mechanism of hypoxic adaptation in M. intracellulare strains but provides the data on the different pattern of hypoxic adaptation between M. intracellulare strains in relation to the difference of genetic requirements. Therefore, we revised the title as ”Functional genomics reveals strain-specific genetic requirements conferring hypoxic growth in Mycobacterium intracellulare”

Reviewer #2 (Public Review):

Summary:

In the study titled "Functional genomics reveals the mechanism of hypoxic adaptation in nontuberculous mycobacteria" by Tateishi et al., the authors have used TnSeq to identify the common essential and growth-defect-associated genes that represent the genomic diversity of clinical M. intracellulare strains in comparison to the reference type strain. By estimating the frequency of Tn insertion, the authors speculate that genes involved in gluconeogenesis, the type VII secretion system, and cysteine desulfurase are relatively critical in the clinical MAC-PD strains than in the type strain, both for the extracellular survival and in a mouse lung infection model.

Based on their analysis, the authors proposed to identify the mechanism of hypoxic adaptation in nontuberculous mycobacteria (NTM) which offer promising drug targets in the strains causing clinical Mycobacterium avium-intracellulare complex pulmonary disease (MAC-PD).

Strengths:

A major strength of the manuscript is the performance of the exhaustive set of TnSeq experiments with multiple strains of M. intracellulare during in vitro growth and animal infection.

Thank you for reviewing our manuscript and acknowledging the performance of producing datasets in this study.

Weaknesses:

(1) The study suffers from the authors' preconceived bias toward a small subset of genes involved in hypoxic pellicle formation in ATCC13950.

Thank you for the comment regarding a potential bias toward a small subset of genes involved in hypoxic pellicle formation in ATCC13950. The rationale for the importance of hypoxic pellicle genes in clinical MAC-PD strains is that the profiles of genetic requirements in each bacterial strain reflect the adaptation to the environment in which each strain lives. When the strains are placed in a special environment, they can adapt to the situation by altering the profiles of genetic requirements, resulting in the remodeling of metabolic pathways.

In this study, we found that several of these pellicle-associated genes also showed increased genetic requirement in the clinical MAC-PD strains, suggesting a possible overlap in hypoxic adaptation mechanisms. We did not insist that clinical MAC-PD strains showed an increase of genetic requirements in all genes of hypoxic pellicle formation. Except for the gene sets involved in hypoxic pellicle formation in ATCC13950, almost no global information has been revealed on the pathogenesis of nontuberculous mycobacterial disease, which differs from the case of tuberculosis. Along with this finding, we investigated the effect of gene silencing on bacterial growth and preferential hypoxic adaptation observed by growth kinetics in clinical MAC-PD strains compared to ATCC13950. At first glance, to focus on the gene sets of hypoxic pellicle formation seems to be “biased”, but we proceeded this research step by step based on our achievements. We consider these data provide valuable information on the pathogenesis of MAC-PD by clinical MAC-PD strains.

We have added the description of the rationale for the importance of hypoxic pellicle genes in clinical MAC-PD strains in the revised manuscript (page 9, lines 148-155).

(2) An important set of data with the ATCC13950 reference strain is missing in the mouse infection study. In the absence of this, it is difficult to establish whether the identified genes are critical for infection/intracellular proliferation, specifically in the clinical isolates that are relatively more adapted for hypoxia.

Thank you for the comment on the necessity of setting ATCC13950 as a control strain of mouse TnSeq experiment. To set ATCC13950 as a control strain in mouse infection experiments would be ideal. However, we proved that ATCC13950 is eliminated within 4 weeks of infection (Tateishi Y. BMC Microbiol. 2023; new Ref#22). That means, it is impossible to perform in vivo TnSeq study due to the inability to harvest sufficient number of colonies.

[Reference]

Tateishi, Y. et al. Virulence of Mycobacterium intracellulare clinical strains in a mouse model of lung infection - role of neutrophilic inflammation in disease severity. BMC Microbiol 23, 94 (2023).

(3) Statistical enrichment analysis of gene sets by GSEA wrongly involves genes required for hypoxic pellicle formation in ATCC13950 together with the gene sets found essential in the clinical MAC-PD strains, to claim that a significant % of genes belong to hypoxia-adaptation pathways. It could be factually incorrect because a majority of these might overlap with those found critical for the in vitro survival of MAC-PD strains (and may not be related to hypoxia).

Thank you for the suggestion on the re-analysis of gene enrichment analysis of genes required for M.i.27 and M.i.198 in vivo infection, individually with genes involved in hypoxic pellicle formation in ATCC13950 and with those showing increased genetic requirements in clinical MAC-PD strains compared to ATCC13950.

About 50% (92 and 94 out of 181 genes through Day 1 to Week 16 and Week4 to Week16 of infection) and 40% (70 and 79 genes out of 179 through Day 1 to Week 16 and Week 4 to Week 16 of infection) of genes required for hypoxic pellicle formation in ATCC13950 were listed as enriched in genes required for mouse lung infection in M.i.27 and M.i.198, respectively. In addition, about 42% (54 and 56 out of 128 genes through Day 1 to Week 16 and thorough Week 4 to Week 16 of infection) and 40% (79 and 68 out of 179 genes through Day 1 to Week 16 and through Week 4 to Week 16 of infection) of genes showing increased requirements in clinical MAC-PD strains compared to ATCC13950 were listed as enriched in genes required for mouse lung infection in M.i.27 and M.i.198, respectively.

These data indicate that about 40-50% genes required for in vitro hypoxic pellicle formation are shared with the genes required for in vivo bacterial growth, and that about 40% strain-dependent/accessory essential genes are shared with the genes required for in vivo bacterial growth. Thus, the genes required for the growth of M.i.27 and M.i.198 in mouse lungs are enriched individually with those involved in hypoxic pellicle formation in ATCC13950, and with the gene sets found critical for in vitro growth. We have added the description of the reanalyzed data of GSEA in the manuscript (pages 16-17, lines 287-290). And the details of reanalyzed data of GSEA have been shown in Supplementary Fig. 5 and 6 as well as Supplementary Tables 15 and 16.

(4) Validation of mouse infection experiments with individual mutants is missing.

Thank you for the suggestion on the validation of the TnSeq hit genes on the in vivo survival. We acknowledge the importance of validating the TnSeq-hit genes by constructing knockout mutants. We have recently succeeded in constructing the vectors for making knockout strains of M. intracellulare (Tateishi. Microbiol Immunol. 2024). We will proceed to the infection experiment of knockout mutants by using our system for constructing them.

[Reference]

Tateishi Y. et al. Construction of knockout mutants in Mycobacterium intracellulare ATCC13950 strain using a thermosensitive plasmid containing negative selection marker rpsL. Microbiol Immunol 68, 339-347 (2024).

(5) Phenotypes with TnSeq and CRISPRi-based KD exhibit poor correlation with misleading justifications by the authors.

Thank you for the comment on the issue of inconsistent results between TnSeq and CRISPR-i based knockdown. We acknowledge that some inconsistencies were observed, particularly among strain-dependent/accessory essential or growth-defect-associated genes. By contrast, we found consistent data between TnSeq and CRISPR-i based knockdown results among universal essential genes such as glcB, inhA, gyrB and embB. Although the mechanism has not been fully proven yet, we consider that such inconsistent phenotypes with TnSeq and CRISPR- based knockdown may be related to the recently revealed bypass mechanism of gene essentiality which is characteristically observed in strain-specific/accessory essential genes (Rosconi F. Nat Micorbiol. 2022; new Ref#14). They suggested this bypass mechanism of gene essentiality in strain-dependent/accessory essential or growth-defect-associated genes from the ‘forced-evolution experiments’ of 36 clinical Streptococcus pneumoniae strains. For example, knockout mutants are successfully recovered from transformation experiments targeting strain-specific/accessory essential genes in TnSeq such as cytidine monophosphate kinase cmk, formate tetrahydrofolate ligase fhs and farnesyl-diphosphate synthase fpp. The bypassing of gene essentiality can be suggested by observing suppressor mutations and synthetic lethality in knockout strains. By contrast, universal essential genes fulfill the following three categories: i) high levels of conservation within and often across species, iii) limited genetic diversity, and iii) high and stable expression levels. Consequently, the universal essential genes are rigid, largely immutable key components to an organism’s survival. In the universal essential genes, the knockout recovery fails as shown by no colonies or only appearance of merodiploids. Taking into consideration such bypass mechanism of gene essentiality in strain-dependent/accessory essential genes, the lower effect of gene silencing of strain-dependent/accessory essential genes on bacterial growth may reflect pathway rewiring that helps the bacterial growth under suppression of the target gene expression.

We have added the description of the possible reason for inconsistency between TnSeq and CRISPR-i results in the Result and Discussion in the revised manuscript (page 21, lines 367-376; pages 28-29, lines 489-519).

[Reference]

Rosconi, F. et al. A bacterial pan-genome makes gene essentiality strain-dependent and evolvable. Nat Microbiol 7, 1580–1592 (2022).

In summary, this study is unable to provide mechanistic insights into why and how different MAC-PD mutant strains exhibit differential survival (in vitro and in animals) and adaptation to hypoxia. It remains to understand why the clinical strains show better adaptation to hypoxia and what is the impact of other stresses on their growth rates.

Thank you for the comments on the issue of being unable to prove the mechanism of MAC-PD pathogenesis and adaptation to hypoxia. We admit that the original manuscript did not provide the apparent reason and mechanism of MAC-PD pathogenesis and adaptation to hypoxia. Following the comment, we have modified the manuscript tile as “Functional genomics reveals strain-specific genetic requirements conferring hypoxic growth in Mycobacterium intracellulare”.

However, we revealed the diversity of genetic requirements among the genus M. intracellulare including the type strain ATCC13950 and clinical MAC-PD strains. We revealed the characteristics of genetic requirements in clinical MAC-PD strains as increased genetic requirements of gluconeogenesis, type VII secretion system and cysteine desulfurase, the former two of which are also required in hypoxic pellicle formation in ATCC13950. Along with this, we demonstrated the difference of growth behavior under hypoxia between clinical MAC-PD strains and ATCC13950. Overall, we consider that we could provide the basic information suggesting the involvement of difference of genetic requirements among strains in the pathogenesis of MAC-PD.

Reviewer #3 (Public Review):

Summary:

The study by Tateishi et al. utilized TnSeq in nine genetically diverse M. intracellulare strains, identifying 131 common essential and growth-defect-associated genes across those strains, which could serve as potential drug targets. The authors also provided an overview of the differences in gene essentiality required for hypoxic growth between the reference strain and the clinical strains. Furthermore, they validated the universal and accessory/strain-dependent essential genes by knocking down their expression using CRISPRi technique. Overall, this study offers a comprehensive assessment of gene requirements in different clinical strains of M. intracellular.

Thank you for reviewing our manuscript and finding the significance of our data.

(1) The rationale for using ATCC13950 versus clinical strains needs to be clarified. The reference strain ATCC13950 was obtained from the abdominal lymph node of a patient around 10 years ago and is therefore considered a clinical strain that has undergone passages in vitro. How many mutations have accumulated during these in vitro passages? Are these mutations significant enough to cause the behavior of ATCC13950 to differ from other recently sampled clinical strains? From the phylogenetic tree, ATCC13950 is located between M018 and M.i.27. Did the authors observe a similarity in gene essentiality between ATCC13950 and its neighbor strains? What is the key feature that separates ATCC13950 from these clinical strains? The authors should provide a strong rationale for how to interpret the results of this comparison in a clinical or biological context.

Thank you for the comments on the rationale for using ATCC13950 versus clinical strains and the key feature that separates ATCC13950 from clinical MAC-PD strains.

ATCC13950 was isolated in 1949, (not around 10 years ago) from 34-month-old female abdominal lymph node (Cuttino. Am J Pathol 1949; new Ref#11). Of note, the clinical background of the patient infected with ATCC13950 is quite different from the patients with MAC-pulmonary disease (MAC-PD), the incidence rate of which has been increasing worldwide without predisposing immunological disorders. ATCC13950 has been regarded as a type strain of genus M. intracellulare historically. And ATCC13950 is the first M. intracellulare strain to be sequenced in 2012 (Kim. J Bacteriol 2012; new Ref#59).

The rationale for using ATCC13950 versus clinical MAC-PD strains is to find the answer to the question whether the essential genes and genetic requirements are similar or different between clinical MAC-PD strains and historical type strain ATCC13950. So far, there are two reports on TnSeq that compare genetic requirements between clinical mycobacterial strains and the type strains, one of which is M. tuberculosis (Carey AF. PLoS Pathogens. 2018; new Ref#8) and the other is M. abscessus (Akusobi C. mBio. 2025; new Ref#9, published after submission of our manuscript). They reported the difference and diversity of genetic requirements between clinical strain and type strains such as M. tuberculosis H37Rv and M. abscessus ATCC19977. We have added the mention of these previous reports to explain the rationale for setting the type strain ATCC13950 as a referential control strain. (page 5, lines 83-87)

The genetic and functional analysis of clinical MAC-PD strains has not been conducted for a long time. In 2021, we have revealed the genomic diversity between clinical MAC-PD and ATCC13950 by comparative genomic analysis (Tateishi BMC Microbiol. 2021; new Ref#5). Except for our TnSeq study of ATCC13950 (Tateishi Sci Rep 2020; new Ref#10), no functional analysis has been conducted in clinical M. intracellulare strains. On our research stream of clinical MAC-PD strains, we expected that we could reveal the functional genomic characteristics of clinical MAC-PD strains by setting ATCC13950 as a referential control strain for analyzing TnSeq data.

It seems an interesting viewpoint to consider the relationship between accumulation of mutations by in vitro passages during prolonged periods from first isolation in ATCC13950, and the difference of phenotypes between ATCC13950 and recently sampled clinical MAC-PD strains. However, there are no time-course samples of ATCC13950 isolates available. Therefore, we can neither investigate how many mutations have accumulated in a time-course manner, nor evaluate how much the accumulated mutations influence the phenotype in ATCC13950. It can be expected that the accumulation of in vitro mutations may cause the behavior of ATCC13950 different from clinical MAC-PD strains. However, it is to be elucidated yet which kinds of factors contribute to the characteristics of ATCC13950 that differ from clinical MAC-PD strains specifically.

It seems an interesting viewpoint to investigate the similarity of gene essentiality between genetical neighbor strains. However, we focused on the overview of the profiles of gene essentiality in clinical MAC-PD strains compared to ATCC13950. Thus, it was out of scope to elucidate the details of gene essentiality in each genetic phylogeny that clinical MAC-PD strains belong. The overview of phylogenetic trees should be referred to our previous publication on the comparative genomic analysis of 55 strains (Tateishi Y. BMC Microbiol. 2021; new Ref#5, new Supplementary Fig. 1), and we have shown Fig. 1 as the extracted phylogenetic tree of subject strains. To elucidate the details of gene essentiality in each genetic clade, it would be necessary to include a considerable number of strains that we used for comparative genomic analysis in 2021 (Tateishi Y. BMC Microbiol. 2021; new Ref#5). Furthermore, it would be necessary to set a referential control strain other than ATCC13950 for comparing gene essentiality. So far, it is not the highest priority for us to elucidate the similarity of gene essentiality between phylogenetic clades in detail, and such investigation will be planned as a future study.

The key features that separate ATCC13950 and clinical MAC-PD strains have not been proved yet, in contrast to the case of M. tuberculosis such as mutations in the gene of the response regulator PhoPR in the type strain H37Rv vs most clinical strains. However, the features that separate ATCC13950 and clinical MAC-PD strains may not be explained by a single genetic factor but may be explained by complicated factors such as epigenetic and/or regulatory factors. For example, the reason for the weakened virulence of H37Ra compared to H37Rv has not been able to be explained by simple genetic differences (Brosch R. Infect Immun. 1999).

In summary, we set the historical type strain ATCC13950 which is derived from infant abdominal lymphadenitis as a referential control strain for TnSeq analysis, because we intended to reveal the characteristics of clinical MAC-PD strains in terms of the gene essentiality and genetic requirements by comparing the clinical MAC-PD strains with the non-MAC-PD reference strain. We consider that the profiles of gene essentiality and genetic requirements specific to clinical MAC-PD strains confer the pathogenesis in an increasing number of MAC-PD patients worldwide without predisposing immunological disorders.

[References]

Cuttino, J.T. & Mc, C.A. Pure granulomatous nocardiosis, a new fungus disease distinguished by intracellular parasitism; a description of a new disease in man due to a hitherto undescribed organism, Nocardia intracellularis, n. sp., including a study of the biologic and pathogenic properties of this species. Am J Pathol 25, 1-47 (1949).

Kim, B.J. et al. Complete genome sequence of Mycobacterium intracellulare clinical strain MOTT-64, belonging to the INT1 genotype. J Bacteriol 194, 3268 (2012).

Carey, A.F. et al. TnSeq of Mycobacterium tuberculosis clinical isolates reveals strain-specific antibiotic liabilities. PLoS Pathog 14, e1006939 (2018).

Akusobi. C. et al.. Transposon-sequencing across multiple Mycobacterium abscessus isolates reveals significant functional genomic diversity among strains. mBio 6, e0337624 (2025).

Tateishi, Y. et al. Comparative genomic analysis of Mycobacterium intracellulare: implications for clinical taxonomic classification in pulmonary Mycobacterium avium-intracellulare complex disease. BMC Microbiol 21, 103 (2021).

Tateishi, Y. et al. Genome-wide identification of essential genes in Mycobacterium intracellulare by transposon sequencing - Implication for metabolic remodeling. Sci Rep 10, 5449 (2020)

Brosch R. et al. Genomic analysis reveals variation between Mycobacterium tuberculosis H37Rv and the attenuated M. tuberculosis H37Ra strain. Infect Immun. 67, 5768-74 (1999).

(2) Regarding the 'nine representative strains of M. intracellulare with diverse genotypes in this study,' how were these nine strains selected? To what extent do they represent the genetic diversity of the M. intracellulare population? A phylogenetic tree illustrating the global genetic diversity of the M. intracellulare population, with these strains marked on it, would be important to demonstrate their genetic representativeness.

Thank you for the comments on the selection of 9 subject strains. We selected the 9 strains based on the phylogenetic tree we published in 2021 (BMC Microbiol 2021; new Ref#5). We have shown the global phylogenetic tree of the M. intracellulare population in new supplementary Fig. 1. We have selected 4 or 5 strains from the major two groups (typical M. intracellulare group and M. paraintracellulare-M. indicus pranii group) for this TnSeq study, respectively.

[Reference]

Tateishi, Y. et al. Comparative genomic analysis of Mycobacterium intracellulare: implications for clinical taxonomic classification in pulmonary Mycobacterium avium-intracellulare complex disease. BMC Microbiol 21, 103 (2021).

(3) The authors observed a considerable amount of differential gene requirements in clinical strains. However, the genetic underpinning underlying the differential requirement of genes in clinical strains was not investigated or discussed. Because M. intracellulare has a huge number of accessory genes, the authors should at least check whether the differential requirement could be explained by the existence of a second copy of functional analogous genes or duplications.

Thank you for the comments on the effect of gene duplication on the change of genetic requirements between strains. Following the comments, we conducted blast search for the 162 genes showing increased Tn insertion reads in each subject strain. We found that M019 has duplicate genes of OCU_RS44705 coding adenosylhomocysteinase (LOCUS_42940: ahcY_1, LOCUS_21000: ahcY_2). However, there were no duplicate genes found in the remaining 161 genes showing increased Tn insertion reads.

From these results, we consider that gene duplication has minor effects on the change of genetic requirements between strains. Rather, sequence differences and accessory genes may play a key role in determining the difference of genetic requirements.

We have added a description of the above-mentioned result in the Result section (pages11-12, lines 191-199).

(4) Growth in aerobic and hypoxic conditions: The authors concluded that clinical strains are better adapted to hypoxia, as reflected by their earlier entry into the log phase. They presented the 'Time at midpoint' and 'Growth rate at midpoint.' However, after reviewing the growth curves, I noticed that ATCC13950 had a longer lag phase compared to other strains under hypoxic conditions, and its phylogenetic neighbor M018 also had a longer lag phase. Hence, I do not believe a conclusion can be drawn that clinical strains are better adapted to hypoxia, as this behavior could be specific to a particular clade. It's also possible that the ATCC13950 strain has adapted to aerobic growth. I would suggest that the authors include growth curves in the main figures. The difference in 'Time at midpoint' could be attributed to several factors, and visualizing the growth curves would provide additional context and clarity.

Thank you for the comments on the possibility of genotypes as a determinant of growth pattern in M. intracelulare. Following the comments, we performed aerobic and hypoxic growth assay in the two strains (M005 and M016) that neighbor ATCC13950.

Author response image 1.

The phylogenetic relationship between M005, M016 and ATCC13950. The former two strains are squared in blue.

M005 reached midpoint later than ATCC13950 both in aerobic and hypoxic conditions. By contrast, M016 reached midpoint three quarters earlier than ATCC13950 under hypoxic conditions. The growth rate was not significantly different between M005, M016 or ATCC13950 under either aerobic or hypoxic conditions, although P-value of M005 vs ATCC13950 was 0.0512 under aerobic conditions on Steel’s multiple comparisons test.

From the data of growth pattern in M005 and M016, we suggest that in addition to gene essentiality, genotypes may have some impact on the bacterial growth pattern under hypoxia; however, since there was a significant difference in the timing of hypoxic adaptation among ATCC13950 and its neighbor strains, bacterial growth pattern under hypoxia is considered to be determined by multiple factors such as genetic requirements and unproven regulatory systems. Taking into consideration that there are lots of genetically diverse strains other than ATCC13950 clade, many clinical MAC-PD strains are possibly better adapted to hypoxia.

Responding to the reviewer’s recommendation, we have added the description of the above-mentioned result in the revised manuscript (page 18, lines 313-322). And we have shown the data of growth curves of the original 9 subject strains in the new Fig 7. And we have added the data of the growth curves of M005 and M016 in new Supplementary Fig 7. Additionally, we have corrected the label of y-axis in new Fig. 7a and new Supplementary Fig. 7a and added the description as “Data are represented as CFUs in 4 μl sample at each timepoint.” in the Figure legends. (page 58, lines 1027-1028 and Supplementary Fig. 7a legend)

(5) Lack of statistical statement: The authors emphasized the role of pellicle-formation-associated genes in strain-dependent essential and accessory essential genes. Additionally, the authors observed that 10% of the genes required for mouse infection are also required for hypoxic pellicle formation. However, these are merely descriptive statements. There is no enrichment analysis to justify whether pellicle-formation-associated genes are significantly enriched in these groups.

Thank you for the comments on the enrichment of pellicle-formation associated genes in strain-dependent essential and accessory essential genes. We performed GSEA and found that 9.1% (16 out of 175) genes were hit as core enrichment. Of them, 4 genes were hit commonly as genes showing increased genetic requirements analyzed by resampling plus HMM analyses including genes of phosphoenolpyruvate carboxykinase pckA (OCU_RS48660), type VII secretion-associated serine protease mycP5 (OCU_RS38275), type VII secretion protein eccC5 (OCU_RS38345) and glycine cleavage system amino-methyltransferase gcvT (OCU_RS35955).

Author response image 2.

We have added the description of GSEA result in the revised manuscript (page 8, lines 137-144; Supplementary Fig. 2; Supplementary Table 5).

Reviewer #1 (Recommendations For The Authors):

Tn-seq and hypoxia adaption in clinical isolates of M. intracellulare (Mi): The authors claim that clinical strains are better adapted to hypoxia because their genetic requirements for optimum fitness overlap with genetic requirements for fitness of the type strain under hypoxia. This is a reasonable hypothesis, but it has not been well-supported by the data presented in Figure 7. The growth rates (Figure 7b) of most of the clinical strains under hypoxia appear to be less than the type strain, although they all seem to grow better than the type strain under normoxia. Perhaps a continuous growth curve of each strain, both as pure and mixed cultures under these conditions will provide a clearer picture.

Thank you for the comments on the difference of adaptation of hypoxic growth between ATCC13950 and MAC-PD strains. To clarify, growth rates shown in Figure 7 were calculated at the inflection point (midpoint) of the growth curves, which were modeled using a four-parameter logistic (4P logistic) model. As described in the Discussion, we found the pattern of hypoxic adaptation characteristics of the clinical MAC-PD strains from the growth curve forms. Taking into consideration the impact of growing bacteria on the disease progression of MAC-PD, the slow growth with early entry to log phase implicates the continuous impact on the infected hosts during logarithmic bacterial growth, which may be involved in the persistent and steadily progressive illness of MAC-PD for years in humans.

Unlike time-lapse imaging assay, the completely seamless sampling of cultures for CFU assay is impossible. Nevertheless, we collected sufficient timepoints to allow reliable curve fitting with the 4P logistic model, and thus consider our growth data to represent a valid approximation of continuous growth dynamics.

Regarding the suggestion of mixed culture experiments, we agree that such studies could be informative. However, co-culture conditions introduce additional variables, including inter-strain competition or synergy, which can obscure the specific contributions of hypoxic adaptation in each strain. Therefore, we believe that the current approach using monoculture growth curves under defined oxygen conditions offers a clearer interpretation of strain-specific hypoxic responses.

In vivo studies: It is unclear how virulent the two clinical strains, Mi27 and Mi198 are in the mouse model. The CFU data in Figure S1b reports the bacterial burden of the Tn libraries of the two strains, of which the overall population of Mi27 library seems to be declining. Without any information on the CFU, animal survival, and tissue pathology from the pure strains, data from the library will have limited implications.

Thank you for the comments on the data presentation of in vivo studies. We previously revealed the time-course data on CFUs, animal survival, and tissue pathology from the pure strains (Tateishi Y. BMC Microbiol. 2023; new Ref#22). Based on these data, we assumed that we would be able to harvest sufficient number of colonies from mice infected with M.i.27 or M.i.198, and we performed in vivo TnSeq studies using these two strains. We have referred to our previous publication on the virulence of MAC-PD pure strains used in this study for mice in the revised manuscript (page 12, line 212; new Ref #22).

The data of CFU counts were shown in new Supplementary Figure 3b. In the manuscript text, we explained as follows (page 12, lines 212-216): “The time course of the changes in the bacterial burden showed a pattern similar to those of the wild-type strains M.i.198 and M.i.27, respectively (Tateishi Y. BMC Microbiol. 2023; new Ref#22), except that it was not possible to harvest sufficient colonies (as few as 104/mouse) in the few mice infected with the M.i.27 Tn mutant strain in week 8 and week 16 (new Supplementary Fig, 3b, new Supplementary Table 8)”. The decline of overall population of M.i.27 Tn mutant library strains in the infected lungs can be explained by the lower virulence of M.i.27 pure strain that shows intermediate virulence phenotype than M.i.198 that shows high virulence phenotype.

[References]

Tateishi, Y. et al. Virulence of Mycobacterium intracellulare clinical strains in a mouse model of lung infection - role of neutrophilic inflammation in disease severity. BMC Microbiol 23, 94 (2023).

Data presentation: The manuscript suffers from a lack of clarity in data visualization and presentation, especially the Tn-Seq datasets. Panels describe the experimental workflow with a densely-worded paragraph, making it difficult to navigate through the major findings.

Thank you for the comments on the issue of Fig. 1b. Following the suggestion, we have modified the new Fig. 1b entitled “Strategy of the procedure of TnSeq analyses”.

Language: The paper should be extensively revised for language. Often the authors have mixed-up the terms like 'core' and 'accessory' 'genes' in lines 116-119 with 'core and accessory genomes' in Figure 2c, which is not even mentioned in the paper. It is further unclear how they identified 3153 and 5824 core and accessory genes, respectively, from 55 different strains of Mi. Line 251: ..."involved by confer..." needs revision. The terms "increased gene essentiality" and 'growth-defected associated genes" are very confusing. The essentiality of a gene is either absolute or conditional but is not quantitative. Similarly, 'growth-defect associated' can be replaced with a better phrase that alludes to fitness loss in the clone. Additional typos were found throughout the paper that need to be fixed.

Thank you for the comments on the issue of scientific words including “'core and accessory genomes” and “gene essentiality” used in this study.

Based on Rosconi’s paper (Panel C of Fig. 1 in Nat Microbiol. 2022; new Ref#14), we used the phrases “accessory genome and core genome” as a meaning of a whole set of genes belonging to accessory and core genes. To avoid the confusion and keep consistency, we replaced the term “genomes” to “genes” in the revised manuscript.

In our previous comparative genomic analysis, we identified 3153 and 5824 core and accessory genes, respectively, from 55 different strains of M. intracellulare (Tateishi Y. BMC Microbiol. 2021; new Ref #5). To perform pangenomic analysis, we used the software Bacterial Pan-Genome Analysis tool (BGAP) (Narendrakumar NM. Sci Rep 2016).

We admit that gene essentiality is a qualitative but not a quantitative trait. We have corrected the term "increased gene essentiality" as "increased genetic requirements" throughout the manuscript.

We have used the term “growth-defect (GD)” based on the classification of gene essentiality calculated by the Hidden Markov Model (HMM) complied by TRANSIT software (DeJesus. PLoS Comput Biol. 2015; new Ref#12). The HMM classifies genes as essential (ES), GD, non-essential (NE), growth-advantage (GA). GD means difficulties of growth (growth deficiency) in aerobic conditions in vitro, because Tn insertions are less frequent. The suggested phrases “fitness loss” or “less fit” may include the meaning of the comparison of two different conditions such as culture conditions exposed to a single bacterial strain. Since the HMM analysis is performed in data of a single strain in a specific bacterial condition, we consider that the phrase including “fitness” is somewhat unsuitable for describing the classification of gene essentiality. Thus, it is difficult for us to rephrase GD to the word that implies fitness levels between two conditions in a single bacterial strain.

[References]

Rosconi, F. et al. A bacterial pan-genome makes gene essentiality strain-dependent and evolvable. Nat Microbiol 7, 1580–1592 (2022).

Tateishi, Y. et al. Comparative genomic analysis of Mycobacterium intracellulare: implications for clinical taxonomic classification in pulmonary Mycobacterium avium-intracellulare complex disease. BMC Microbiol 21, 103 (2021).

Narendrakumar NM et al. BPGA- an ultra-fast pan-genome analysis pipeline. Sci Rep 2016 6, 24373 (2016).

DeJesus, M.A. et al. TRANSIT--A Software Tool for Himar1 TnSeq Analysis. PLoS Comput Biol 11, e1004401 (2015).

Reviewer #2 (Recommendations For The Authors):

Major Comments:

(1) Result 1 (Page 6-7): Common essential and growth-defect-associated genes representing the genomic diversity of M. intracellulare strains.

(1a) From Table S1, it is observed that the numbers of Tn-inserted TA sites significantly vary (p >0.05) among biological replicates for each strain when compared with the reference strain ATCC13950. the authors should provide an explanation of how they overcame this variation in their analysis.

Thank you for the comment on the issue of a variable number of Tn-inserted TA sites among biological replicates for each strain of MAC-PD.

On TRANSIT software, we set the replicate option as Sum to combine read counts. And we used Beta-Geometric correction (BGC) to normalize the datasets to fit an “ideal” geometric distribution with a variable probability parameter ρ.

Following the comment, we have added the description on which option we used for handling the replicate data and normalization (page 36, lines 640-643).

(1b) Importantly, saturation in most of the strains is only ~50-60%. In such a case, there will be a high probability that Tn will not hit a nonessential region due to chance instead of selection (See DeJasus et al., mBio, 2017). It has been observed that the sequence pattern (GC)GNTANC(GC) is strongly associated with non-permissiveness. As shown earlier, the authors need to carefully look for the potential non-permissive sites before concluding the fate of a gene. Also, they should acknowledge the potential limitations of their approach due to the suboptimal level of saturation.

Thank you for the comment on the saturation of Tn mutant libraries. Our method of comparison of genetic requirements between strains are the same as a previous report that used duplicate Tn mutant libraries of clinical Mtb strains of different genotypes and triplicate Tn mutant libraries of H37Rv for identifying increased genetic requirements of clinical Mtb strains (Carey. PLoS Pathog 2018; new Ref#8). Our method is also based on the coauthor’s TnSeq study on H37Rv (Minato Y. mSystems 2019; new Ref#61). Moreover, by combining replicates, the saturation of our Tn mutant libraries became 62-79% as follows: ATCC13950: 67.6%, M001: 72.9%, M003: 63.0%, M018: 62.4%, M019: 74.5%, M.i.27: 76.6%, M.i.198: 68.0%, MOTT64: 77.6%, M021: 79.9%. That is, we calculated gene essentiality from the Tn mutant libraries with 62-79% saturation in each strain. The levels of saturation of transposon libraries in our study is similar to the very recent TnSeq anlaysis by Akusobi where 52-80% saturation libraries (so-called “high-density” transposon libraries) are used for HMM and resampling analyses (Supplemental Methods Table 1[merged saturation] in Akusobi. ｍBio. 2025; new Ref#9). The saturation of Tn insertion in individual replicates of our libraries is also comparable to that reported by DeJesus (Table S1 in mBio 2017; new Ref#57). Thus, we consider that our TnSeq method of identifying essential genes and detecting the difference of genetic requirements between clinical MAC-PD strains and ATCC13950 is acceptable.

As the Reviewer indicates, there is non-permissive sequence pattern in mariner transposon mutagenesis. Using more than 10 replicates of Tn mutant libraries is quite an accurate method for detecting essential genes in nonstructural small genes such as small regulatory RNAs. However, as DeJesus shows, the number essential genes identified by TnSeq are comparable in large genes possessing more than 10 TA sites between 2 and 14 TnSeq datasets, most of which seem to be structural genes (Supplementary Fig 2 in mBio 2017; new Ref#57). Thus, we do not consider that we made a serious mistake for the classification of essentiality in most of the structural genes that encode proteins. With respect to the coverage of non-permissive sites, our TnSeq method might need to be improved if it is intended to classify the gene essentiality quite accurately on the small genes including small regulatory RNAs.

We investigated the non-permissive TA sites in ATCC13950. There are 4136 (6.43% of total ORFs) nonpermissive TA sites in ATCC13950, which is less than in H37Rv (9% of total ORFs) (DeJesus MA. mBio 20171; new Ref#57) and in M. abscessus ATCC19977 (8.1% of total ORFs)(Rifat D. mBio. 2021; new Ref#58). As for larger ORFs (TA sites > = 10), there are nonpermissive TA sites in 89 genes (ORFs) of common “essential (ES)” or “growth-defect-associated (GD)” (4.82% of a total of 1844 larger ORFs in ATCC13950). As for small ORFs (2-9 TA sites), there are nonpermissive TA sites in 41 genes (ORFs) of common ES or GD (1.35% of a total of 3021 smaller ORFs in ATCC13950).

We appreciate the idea of concluding the fate of gene essentiality by the presence/absence of non-permissive TA sites. However, we cannot conclude the fate of gene essentiality classification only by the presence/absence of potential non-permissive sites. Because, strictly to say, it is impossible to conclude the scientific truth of gene essentiality without functional analysis using gene manipulation. In accurate, TnSeq can “predict” the gene essentiality but cannot perfectly guarantee the functional significance. However, in the current situation, most of the recent TnSeq studies have been published only by the TnSeq analysis without functional analysis that uses gene manipulation strains of all targets they identified. Taking such limitations of TnSeq including non-permissive sites into consideration, we consider that ｔhe essentiality of the detected genes should be determined in further studies, mainly including biological experiments such as functional studies using gene manipulation strains.

We have added the above-mentioned contents in the revised manuscript (pages 32-33, lines 559-580).

[References]

Carey, A.F. et al. TnSeq of Mycobacterium tuberculosis clinical isolates reveals strain-specific antibiotic liabilities. PLoS Pathog 14, e1006939 (2018).

Minato, Y., et al. Genomewide assessment of Mycobacterium tuberculosis conditionally essential metabolic pathways. mSystems. 4, e00070-192019 (2019).

Akusobi. C. et al. Transposon-sequencing across multiple Mycobacterium abscessus isolates reveals significant functional genomic diversity among strains. mBio 6, e0337624 (2025).

DeJesus, M.A. et al. Comprehensive essentiality analysis of the Mycobacterium tuberculosis genome via saturating transposon mutagenesis. mBio 8, e02133-16 (2017).

Rifat, D., Chen L., Kreiswirth, B.N. & Nuermberger, E.L.. Genome-wide essentiality analysis of Mycobacterium abscessus by saturated transposon mutagenesis and deep sequencing. mBio 12, e0104921 (2021).

(1c) Line 100: Authors report a total of 131 genes identified as essential or growth-defect-associated with the HMM analysis across all M. intracellulare strains. It should be explained in more detail how gene essentiality was determined (see above comment in (1b)). Furthermore, in Table S3 authors should mention the essential and growth defective trait of each of the 131 genes.

Thank you for the comment on how to classify the 131 genes as essential or growth-defect-associated with the HMM analysis across all M. intracellulare strains. As replied in (1b), the average saturation of Tn insertion of our libraries became 62-79% when combining duplicate or triplicate data in each strain. The levels of saturation of transposon libraries in our study is similar to the very recent TnSeq analysis by Akusobi where 52-80% saturation libraries (so-called “high-density” transposon libraries) were used for HMM and resampling analyses, and most of triplicate libraries ranges 70-79% saturation (Supplemental Methods Table 1[merged saturation] in Akusobi. ｍBio. 2025; new Ref#9). The saturation of Tn insertion in individual replicates of our libraries is also comparable to those with DeJesus (Table S1 in mBio 2017; new Ref#57). Thus, we consider that our TnSeq libraries are acceptable for identifying essential genes and growth-defect-associated genes by the HMM method.

We used the HMM method as reported by DeJesus (DeJesus. PLoS Comput Biol. 2015; new Ref#12). HMM method can categorize the gene essentiality throughout the genome including “Essential”, “Growth-defect”, “Non-essential” and “Growth-advantage”. “Essential” genes are defined as no insertions in all or most of their TA sites. “Non-essential” genes are defined as regions that have usual read counts. “Growth-defect” genes are defined as regions that have unusually low read counts. “Growth-advantage” genes are defined as regions that have unusually high low read counts.

Following the previous report (Carey AF. PLos Pathog 2018; new Ref#8), the annotation for the clinical MAC-PD strains was adapted from that of ATCC13950 by adjusting the START and END coordinates of each ORF in the clinical MAC-PD strains according to their alignment with the corresponding ORFs of ATCC13950. By using an adjusted annotation table, gene essentiality was classified by the HMM analysis.

We have added the explanation of how we identified essential and growth-defect-associated genes in the Methods (pages 35-36, lines 620-632). And following the comment, we have added the data of classification of gene essentiality in the 131 genes in the new Supplementary Table 3 in the revised manuscript.

[Reference]

DeJesus, M.A. et al. TRANSIT--A Software Tool for Himar1 TnSeq Analysis. PLoS Comput Biol 11, e1004401 (2015).

Carey, A.F. et al. TnSeq of Mycobacterium tuberculosis clinical isolates reveals strain-specific antibiotic liabilities. PLoS Pathog 14, e1006939 (2018).

Akusobi. C. et al. Transposon-sequencing across multiple Mycobacterium abscessus isolates reveals significant functional genomic diversity among strains. mBio 6, e0337624 (2025).

DeJesus, M.A. et al. Comprehensive essentiality analysis of the Mycobacterium tuberculosis genome via saturating transposon mutagenesis. mBio 8, e02133-16 (2017).

(1d) In Table S4, the authors show strain-specific putative essential genes from the core and accessory gene sets. For the sake of clarity, it is important to have the name of all the strains against each gene in which it is predicted essential or growth defective.

Thank you for the comment on the hit strains on the genes classified as strain-specific and accessory putative essential of growth-defect associated. Following the comment, we have added the data of hit strains in the new Supplementary Table 4 in the revised manuscript.

(1e) Lines 123-126: It is not clear what is the relevance of highlighting genes involved in hypoxic pellicle formation in ATCC13950. These appear to be randomly distributed across different clinical isolates and is not clear whether they correlate with differential susceptibility of the reference strain and clinical isolates to hypoxia.

Thank you for the comment on the relevance of highlighting genes involved in hypoxic pellicle formation in ATCC13950. The rationale for the importance of hypoxic pellicle genes in clinical MAC-PD strains is that the profiles of genetic requirements in each bacterial strain reflect the adaptation to the environment in which each strain lives. When the strains are placed in a special environment, they can adapt to the situation by altering the profiles of genetic requirements, resulting in the remodeling of metabolic pathways. We indeed found that the genetic requirements of several hypoxic pellicle genes were increased in clinical MAC-PD strains in vitro situations. These data suggest the hypoxic pellicle genes become more important in clinical MAC-PD strains for in vitro growth than in ATCC13950.

Moreover, hypoxia is known to be one of the characteristic conditions in vivo including clinical lesions (McKeown. Br Br J Radiol. 2014). We consider it reasonable to expect that the strains derived from MAC-PD patients without predisposing immunological disorders may adapt under hypoxic conditions for maintaining bacterial survival. Therefore, we highlighted the genes involved in hypoxic pellicle formation in ATCC13950.

We have added the description of the rationale for the importance of hypoxic pellicle genes in clinical MAC-PD strains in the revised manuscript (page 9, lines 148-155).<br /> [Reference]

McKeown, S.R., et al. Defining normoxia, physoxia and hypoxia in tumours-implications for treatment response. Br Br J Radiol 87,: 20130676 (2014).

(2) Result 2 (pages 8-10): Genes with increased gene essentiality in clinical MAC-PD strains are also required for hypoxic pellicle formation in the type strain.

(2a) As reported by authors (lines 123-126), only a small fraction of genes showing essentiality in clinical MAC-PD strains are required for hypoxic pellicle formation in the reference strain, which might be due to random distribution. Authors should avoid making such a generalised statement that reflects the association of the entire essential gene pool in clinical MAC-PD strains with hypoxic pellicle formation.

Thank you for the comment on the issue of a small fraction of genes showing increased genetic requirements in clinical MAC-PD strains that is shared with genes required for hypoxic pellicle formation in the type strain ATCC13950. We admit that the section title may mislead that the genes required for hypoxic pellicle formation confer the entire essential gene pool of clinical MAC-PD strains. Following the comment, we have revised the section title as “Partial overlap of the genes showing increased genetic requirements in clinical MAC-PD strains with those required for hypoxic pellicle formation in ATCC13950” (page 9, lines 146-147).

We consider that it cannot be explained by a mere coincidence that we obtained the data of partial overlap of genes showing essentiality in clinical MAC-PD strains with genes required for hypoxic pellicle formation in ATCC13950, because we demonstrated the supporting data such as the pattern of genetic requirements suggesting gluconeogenic metabolic shift (Fig. 5) and the different pattern of hypoxic growth curves between clinical MAC-PD strains and ATCC13950 (Fig. 7).

(2b) I fail to understand how the number of Tn insertions determines "more" or "less" essentiality of a gene particularly with 50-60% saturation. To my understanding, essentiality is a qualitative trait. Either a gene will be essential (based on no Tn insertion despite having the permissive sites), critical (poor representation of Tn insertions at the permissive sites due to growth defect of the strain in the pool), non-essential (expected frequency of insertion) or growth-advantageous (higher representation of Tn insertions at the permissive sites due to growth advantage of the strain in the pool). Hence, authors should avoid quantifying the essentiality of a gene.

Thank you for the comments on the trait of gene essentiality. We realize that essentiality is a qualitative trait, not a quantitative trait. Taking into consideration the number of Tn insertions determines "more" or "less" requirements of a gene, we have corrected the manuscript by using the phrase “genetic requirements” instead of “gene essentiality”.

As mentioned earlier, our method of comparison of genetic requirements between strains are the same as a previous report that used duplicate Tn mutant libraries of clinical Mtb strains of different genotypes and triplicate Tn mutant libraries of H37Rv for identifying increased genetic requirements of clinical Mtb strains (Carey AF. PLoS Pathog 2018; new Ref#8). Moreover, as described in rebuttal (1b), the saturation of our Tn mutant libraries by combining replicates are 62-79% as follows: ATCC13950: 67.6%, M001: 72.9%, M003: 63.0%, M018: 62.4%, M019: 74.5%, M.i.27: 76.6%, M.i.198: 68.0%, MOTT64: 77.6%, M021: 79.9%. That is, we calculated gene essentiality from the Tn mutant libraries with 62-79% saturation in each strain. The levels of saturation of transposon libraries in our study is similar to the recent TnSeq analysis by Akusobi where 52-80% saturation libraries (“high-density” transposon libraries) were used for HMM and resampling analyses (Supplemental Methods Table 1[merged saturation] in Akusobi C. ｍBio. 2025; new Ref#9).

Thus, we consider that our data of the difference of genetic requirements between clinical MAC-PD strains and ATCC13950 are acceptable.

[Reference]

Akusobi. C. et al. Transposon-sequencing across multiple Mycobacterium abscessus isolates reveals significant functional genomic diversity among strains. mBio 6, e0337624 (2025).

(2c) From Figures 3-4, it seems the authors intend to highlight the insertion frequencies of certain genes in the clinical isolates compared to those in the reference strain to conclude whether a gene has become more critical and its disruption results in the growth defective phenotype (poor representation) in the clinical isolates, or a critical/essential gene has become dispensable in these strains.

Based on these arguments, I suggest that the authors modify the title of the result such as "Tn insertion reveals differential requirement of genes for in vitro growth of clinical MAC-PD strains" or "Identification of genes differentially required for in vitro growth of clinical MAC-PD strains" as this is precisely the information we gain from this section of the study. Also, it is suggested to re-draft the rationale of this section as only 4 genes associated with hypoxic pellicle formation, were found to exhibit reduced insertion frequencies in the clinical isolates out of total of 283 genes. Hypoxia-related genes can be highlighted in the next section (see below).

Thank you for the suggestion to modify the section title and to re-draft the rationale of the section. Following the comment, we modified the section title as “Partial overlap of the genes showing increased genetic requirements in clinical MAC-PD strains with those required for hypoxic pellicle formation in ATCC13950 (page 9, lines 146-147)

Following the suggestion, we have revised the rationale of this section as follows: “The sharing of strain-dependent and accessory essential or growth-defect-associated genes with genes required for hypoxic pellicle formation in ATCC13950 prompts us to consider that the profiles of gene essentiality in clinical MAC-PD strains may be associated with the genes required for hypoxic pellicle formation in ATCC13950.” (page 9, lines 151-155)

The reviewer points out that only 4 genes associated with hypoxic pellicle formation were found to exhibit reduced insertion frequencies in the clinical isolates out of total of 283 genes. However, to discuss how much proportion of the genes were detected to be increasingly required in clinical MAC-PD strains compared to ATCC13950, we should focus on the 121 genes showing increased requirements in clinical MAC-PD strains compared to ATCC13950, excluding the 162 genes indispensable for clinical MAC-PD strains. Thus, we described that 4 genes associated with hypoxic pellicle formation, were found to exhibit reduced insertion frequencies in the clinical isolates out of the 121 genes having significantly fewer Tn insertions than ATCC13950 in the manuscript (Fig. 3).

(3) Result 3 (Page 10-14): Requirement of genes with increased gene essentiality in the clinical MAC-PD strains for mouse lung infection.

(3a) The title should be modified to "Identification of genes in the clinical MAC-PD strains required for mouse lung infection".

Following the comment, we have modified the section title as "Identification of genes in the clinical MAC-PD strains required for mouse lung infection". (page 12, lines 201-202).

(3b) Further, the rationale of this experiment needs to be modified. As mentioned above, up until now the impact of hypoxic pellicle formation genes in the growth of MAC-PD strains remains unconvincing. The rationale of mouse infection experiments could be straightforward- "to identify genes critical for animal infection of the clinical isolates".

Thank you for the comment on the rationale of the in vivo TnSeq experiment. Following the comment, we have revised the rationale as “The impact of hypoxia on mycobacteria under various ecological circumstances implies that the genes required for pathogenesis of MAC-PD may be in some degrees, overlapped with the genes with increased requirements in the clinical MAC-PD strains compared to ATCC13950, and also with the genes required for hypoxic pellicle formation in ATCC13950. To identify genes required for in vivo infection of clinical MAC-PD strains,” in the revised manuscript (page 12, lines 204-210).

(3c) The authors should avoid using the term "genes with increased essentiality" for the reasons explained above in point #2b.

Following the comment, we have corrected the term as “genes with increased requirements” in the revised manuscript (page 12, line 207).

(3d) From Tables S8 and S9, I can find 93 genes in Mi198Tn and 74 genes in Mi27Tn for which Tn insertion mutants are under-represented in TnSeq at all time points from Day 1 to Wk 16 in comparison to input. Importantly, excluding results from Day 1 when the infection has just settled, I find 172 and 121 genes in Mi198Tn and Mi27Tn, respectively, under-represented in lungs between Wk 4-16. My suggestion is that authors should focus more on such genes and identify the characteristics of these genes and what fraction belongs to those involved in hypoxic pellicle formation in the reference strain. I am perplexed why authors have categorically ignored other genes and only focused on a set of genes that correspond to ~10-12% of entire differentially abundant mutant pool.

Thank you for the suggestion on the genes that Tn insertion mutants are under-represented in TnSeq from Weeks 4-16 in the infected mouse lungs be analyzed for overlapping the genes involved in hypoxic pellicle formation in the type strain ATCC13950. We found that at all timepoints from Day1 to Week 16, 74 genes and 99 genes were under-represented in lungs infected with M.i.27Tn and M.i.198Tn, respectively. Of them, 21 (28.3%) and 12 (12.1%) genes belonged to the genes involved in the genes required for hypoxic pellicle formation in the type strain. We found that at timepoints from Week 4 to Week 16, 121 genes and 172 genes were under-represented in lungs infected with M.i.27Tn and M.i.198Tn, respectively. Of them, 21 (23.1%) and 30 (18.0%) genes belonged to genes involved in hypoxic pellicle formation in the type strain. These hypoxic pellicle-associated genes detected both in M.i.27 and M.i.198 encoded methionine synthesis, acyl-CoA dehydrogenase, isocitrate lyase, MMPL family transporter at all time points (from Day1 to Week 16). And additionally, multifunctional oxoglutarate decarboxylase/dehydrogenase, proteasome subunits, ABC transporter ATP-binding protein/permease, lipase chaperone at all time points (from Week 4 to Week 16). We have described these results in the Result section (page 14 lines 236-248) and new Supplementary Tables 12 and 13.

As for M. intracellulare, conditionally essential genes have not been revealed except for those required for hypoxic pellicle formation in ATCC13950 revealed by us (Tateishi Y. Sci Rep. 2020; new Ref#10). This study is the first to focus on the relationship between the difference of genetic requirements among strains and hypoxic adaptation. We found a certain proportion of overlapped genes required for mouse lung infection and ATCC13950’s hypoxic pellicle formation. We consider it reasonable to focus on the category of genes required for hypoxic pellicle formation to analyze the datasets of TnSeq in mice.

[Reference]

Tateishi, Y. et al. Genome-wide identification of essential genes in Mycobacterium intracellulare by transposon sequencing - Implication for metabolic remodeling. Sci Rep 10, 5449 (2020).

(3e) Page 13, lines 224-227: "Despite the differences in the profiles of the genes required for in vivo infection between strains, these data suggest that increased gene essentiality for hypoxic growth confers advantages for pathogenesis in vivo."

For the reason described above, I find it a misleading hypothesis that hypoxic growth confers advantages for pathogenesis in vivo. How come only 10-12% of the entire gene sets which include genes of varying functions, can be the sole contributors to bacterial survival in host organelles during infection?

More importantly, the mouse is not considered a good model for hypoxia as mouse infection does not lead to the formation of solid granuloma with a hypoxic core Though I am not convinced with the authors' bias toward hypoxia-related genes, however, if at all they aim to investigate the role of such genes by an unbiased enrichment of TnSeq mutant, they should have used C3HeJ mice which are known to form granulomas (Boute et al., 2017 (doi: 10.1186/s13567-017-0477-7)).

Thank you for the comments on the issue of the contribution of genes required for hypoxic growth and on the difference of hypoxic levels between mouse lineages. We did not intend to mention that a set of genes required for hypoxic growth is the sole contributor to bacterial survival in host organs during infection. As we discussed in the Discussion section, we acknowledge that the adaptation to the difference of carbon source between in vitr_o and _in vivo infection (i.e. preferential usage of lipid carbon source in vivo) is involved in the pathogenesis of mycobacterial diseases (Yang. Front Microbiol 2018; new Ref#33, Gouzy. Proc Natl Acad Sci U S A 2021; new Ref#29, Quinonez. mBio 2022; new Ref#40, Pandey. Proc Natl Acad Sci U S A 2008; new Ref#41). We consider that not only the genes required for hypoxic pellicle formation but also strain-dependent/accessory genes conferring kinds of metabolism other than hypoxic pellicle formation can be estimated to be involved in the in vivo mouse lung infection.

We have modified the sentence to clearly express our intention as follows: “These in vivo TnSeq data suggest that, despite the differences in the profiles of the genes required for in vivo infection between strains, increase of genetic requirements for hypoxic growth in part contribute to the pathogenesis in vivo.” (pages 15-16, lines 269-271)

It seems to be an interesting idea to perform TnSeq by using C3HeJ mice. The granuloma formed in C3HeJ mice becomes extremely hypoxic (less than 1%, corresponding the level of “pathological” hypoxia) which is as severe as the detection range by pimonidazole. In our model, the effect of such pathological levels of hypoxia on granuloma formation might not be detected. However, the lesion formed in C57BL/6 mice becomes a “physiological” level of hypoxia (5% O2) (McKeown SR. Br Br J Radiol. 2014) which is the same O2 level for M. intracellulare to form pellicles. In principle, oxygen levels inside human bodies are physiologically hypoxic, and many biological events are experimentally investigated in this condition. Thus, we consider that we were able to observe the effect of physiological hypoxia on M. intracellulre growth both in vitro (hypoxic pellicles) and in vivo (infected C57BL/6 mice).

[Reference]

Yang, T. et al. Pan-genomic study of Mycobacterium tuberculosis reflecting the primary/secondary genes, generality/individuality, and the interconversion through copy number variations. Front Microbiol 9, 1886 (2018).

Gouzy, A., Healy, C., Black, K.A., Rhee, K.Y. & Ehrt, S. Growth of Mycobacterium tuberculosis at acidic pH depends on lipid assimilation and is accompanied by reduced GAPDH activity. Proc Natl Acad Sci U S A 118, e2024571118 (2021).

Quinonez, C.G. et al. The role of fatty acid metabolism in drug tolerance of Mycobacterium tuberculosis. mBio 13, e0355921 (2022).

Pandey, A.K. & Sassetti, C.M. Mycobacterial persistence requires the utilization of host cholesterol. Proc Natl Acad Sci U S A 105, 4376-4380 (2008).

McKeown., S.R. et al. Defining normoxia, physoxia and hypoxia in tumours-implications for treatment response. Br Br J Radiol 87, 20130676 (2014).

(3f) An important set of data with the ATCC13950 reference strain is missing here. It is suggested that authors perform this study with the reference strain to identify whether the enrichment of genes is similar across all strains or is specific to the clinical isolates.

Thank you for the comment on the setting of ATCC13950 as a control strain in the mouse infection experiment. However, we proved that bacterial burden of ATCC13950 is reduced continuously from 4 weeks of infection, and that ATCC13950 is almost completely eliminated from 8 to 16 weeks of infection (BMC Microiol 2023; new Ref#22). Therefore, it is impossible to perform TnSeq to detect the genes required for persistent infection in mice infected with ATCC13950.

[Reference]

Tateishi, Y. et al. Virulence of Mycobacterium intracellulare clinical strains in a mouse model of lung infection - role of neutrophilic inflammation in disease severity. BMC Microbiol 23, 94 (2023).

(3g) Pages 13-14, lines 228-245: "We have performed a statistical enrichment analysis of gene sets by GSEA...".

The comparison made here is not clear to me. It seems the authors do compare genes required for the growth of M.i.27 and M.i.198 in mouse lungs with the gene sets required for hypoxic pellicle formation in ATCC13950 together with the gene sets showing increased gene essentiality observed in the clinical MAC-PD strains, and claim that a significant % of genes belong to hypoxia-adaptation pathways. It is factually incorrect because a majority of these might overlap with those found critical for the in vitro survival of MAC-PD strains. It is suggested that authors re-analyze their data by comparing genes required for the growth of M.i.27 and M.i.198 in mouse lungs individually with those involved in hypoxic pellicle formation in ATCC13950, and with the gene sets found critical for in vitro growth, and present accordingly.

Thank you for the suggestion on the re-analysis of gene enrichment analysis of genes required for M.i.27 and M.i.198 in vivo infection, individually with genes involved in hypoxic pellicle formation in ATCC13950 and with those showing genetic requirements in clinical MAC-PD strains compared to ATCC13950.

About 50% (92 and 94 out of 181 genes through Day 1 to Week 16 and through Week4 to Week16 of infection) and 40% (70 and 79 out of 179 genes through Day 1 to Week 16 and through Week 4 to Week 16 of infection) of genes required for hypoxic pellicle formation in ATCC13950 were listed as enriched in genes required for mouse lung infection in M.i.27 and M.i.198, respectively. In addition, about 42% (54 and 56 out of 128 genes through Day 1 to Week 16 and through Week 4 to Week 16 of infection) and 40% (79 and 68 out of 179 genes through Day 1 to Week 16 and through Week 4 to Week 16 of infection) of genes showing increased requirements in clinical MAC-PD strains compared to ATCC13950 were listed as enriched in genes required for mouse lung infection in M.i.27 and M.i.198, respectively.

The tables and graphs of GSEA results are shown in Supplementary Figs. 5, 6.

These data indicate that 40-50% of the genes required for in vitro hypoxic pellicle formation and the strain-dependent/accessory essential genes are significantly enriched individually with in vivo bacterial growth. We have added the result of reanalyzed data of GSEA in the Result (pages 16-17, lines 287-290). We have shown the detail of reanalyzed data of GSEA in Supplementary Figs. 5, 6 and Supplementary Tables 15, 16.

(3h) Since authors have used Tnseq of pooled mutants, which often yields misleading information, it is important to validate some of their findings upon mouse infection with individual mutants that yield prominent as well as baseline reduction at different time points. In the absence of validation, it remains a mere speculation of the role of these genes in the infection of these strains to animals.

Thank you for the suggestion on the validation of the TnSeq hit genes on the in vivo survival. We acknowledge the importance of validating the TnSeq-hit genes by constructing knockout mutants. We have recently succeeded in constructing the vectors for making knockout strains of M. intracellulare (Tateishi Y. Microbiol Immunol. 2024). We will proceed to the infection experiment of knockout mutants by using our system for constructing them.

[Reference]

Tateishi Y. et al. Construction of knockout mutants in Mycobacterium intracellulare ATCC13950 strain using a thermosensitive plasmid containing negative selection marker rpsL+. Microbiol Immunol 68, 339-347 (2024).

(4) Result 4 (Page 14-15): Preferential hypoxic adaptation of clinical MAC-PD strains evaluated with bacterial growth kinetics.

(4a) "The metabolic remodeling, such as the increased gene essentiality of gluconeogenesis and the type VII secretion system..". As stated above, the essentiality of a gene, being a qualitative trait, should not be presented in quantitative terms. The authors should re-phrase this statement.

Following the comment, we have corrected the term as “The metabolic remodeling, such as the increased genetic requirements of gluconeogenesis and the type VII secretion system.” (page 17, lines 296-297)

(4b) "overlap of the genes required for mouse lung infection and those required for hypoxic pellicle formation involved by conferring these metabolic pathways..". There is a syntax error in this statement and needs revision.

Following the comment, we have corrected the phrase as “overlap of the genes required for mouse lung infection and those required for hypoxic pellicle formation involved by these metabolic pathways”. (page 17, lines 297-299)

(4c) The altered requirement of genes in different clinical strains for survival provides only circumstantial evidence of metabolic remodeling. Authors are suggested to perform metabolic profiling of representative clinical and reference strains, as it is important to examine whether these bacteria indeed undergo metabolic shift.

Thank you for the comment on the metabolic profiling of the representative clinical and reference strains. We previously published the TnSeq result of ATCC13950 and we produced the current data by organizing with our previous findings (Fig. 4 in Tateishi Y. Sci Rep 2020; new Ref#10). The priority of the current study was to elucidate the difference and diversity of genetic requirements between clinical MAC-PD strains and ATCC13950. We consider that it is of some value to show the even circumstantial evidence of metabolic remodeling by TnSeq, because it provides a strong rationale for proceeding to the next study including metabolomic analysis.

[Reference]

Tateishi, Y. et al. Genome-wide identification of essential genes in Mycobacterium intracellulare by transposon sequencing - Implication for metabolic remodeling. Sci Rep 10, 5449 (2020).

(5) Result 5 (Page 16-18): Effects of knockdown of universal and accessory/strain-dependent essential or growth-defect-associate genes in clinical MAC-PD strains.

(5a) Lines 273-277: The rationale of using CRISPRi should be correctly presented to evaluate the effect of individual genes' suppression on the downstream phenotype and not to establish the CRISPRi silencing tool in MAC.

Thank you for the comment on the rationale of the section of the CRISPR-i experiment. Following the comment, we have modified the sentence as follows: “With an intention to evaluate the effect of suppressing TnSeq-hit genes on bacterial growth.” (page 19, lines 333-334 in the revised manuscript).

(5b) Line 278: pRH2052/pRH2521 are the plasmids and not the CRISPRi system.

Following the comment, we have corrected the phrase as “pRH2052/pRH2521 clustered regularly interspaced short palindromic repeats interference (CRISPR-i) plasmids.” (page 19, lines 334-335 in the revised manuscript).

(5c) Line 280: Other pioneering studies on the use of CRISPRi for gene silencing in mycobacteria (Chaudhary et al., Nat Comm, Rock et al., Nat Microbio) should also be cited.

Thank you for the comment on adding the reference papers on CRISPR-i in mycobacteria. We have added the two suggested papers in the revised manuscript as new Ref #30 and #31. (page 19, line 336).

(5d) Lines 282-283: It is not clear why M001 and MOTT64 could not be transformed. Did the authors use any control plasmid to evaluate the transformation efficiency of these strains?

Thank you for the comment on the failure of transformation in M001 and MOTT64.

Following the comment, we have performed the experiment for evaluating the efficiency of transformation in the 9 M. intracellulare strains we used in this study. We have used an E. coli-mycobacteria shuttle vector pSO246KM-Prhsp65-luc that expresses firefly luciferase as a control plasmid (Aoki K. J Biol Chem 2004). For obtaining transformed colonies, we used 7H10/OADC agar plates containing the same concentration of kanamycin that we used for preparing Tn mutant libraries and for obtaining CRSISPR-i knockdown strains.

We have observed no colonies grown on agar plates in MOTT64 after electroporation of the pSO246KM-Prhsp65-luc plasmid. In most of the remaining strains, the transformed colonies have emerged fully on day 10 of culture after electroporation of the plasmid. However, we have observed that M001 needs twice as long as a period for the emergence of transformed colonies. On day 21, the number of colonies in M 001 have finally become comparable to that of the other strains. We have checked the luciferase activity of 6-12 colonies in each strain except for MOTT64, and we have confirmed the transformation of the plasmid by the data of higher luciferase activity in the colonies undergoing electroporation of the plasmid than in those not undergoing electroporation.

The possible reason for the incapability of obtaining transformants of CRISPR-i vectors in MOTT64 may be due to the extremely low efficiency of acquiring foreign DNA. And the possible reason for the incapability of obtaining transformants of CRISPR-i vectors in M001 may be intolerable to the stress caused by transformation of plasmids compared to other M. intracellulare strains. For M001, pSO246KM-Prhsp65-luc plasmid may cause tolerable stress for transformation, resulting in the delayed emergence of transformed colonies. By contrast, the CRIPSR-i plasmids may cause greater stress for M001 than pSO246KM-Prhsp65-luc plasmid, resulting in being intolerable for transformation.

Author response table 1.

Author response image 3.

Result of luciferase activities before and after transformation of pS0246KM-Prhsp65-luc plasmid. Fifty microliter of cultures were mixed with 50 u L of assay reagents (Luciferase assay system E1500, Promega) and luciferase activity was measured by the luminometer (FilterMax F5, Molecular Devices). Data are shown as mean ± SD of 6-12 colonies

[Reference]

Aoki K. Extracellular mycobacterial DNA-binding protein 1 participates in Mycobacterium-lung epithelial cell interaction through hyaluronic acid. J Biol Chem 279, 39798–39806 (2004).

(5e) Lines 283-186: "To confirm the gene essentiality detected with the HMM analysis, we evaluated the consequent growth inhibition in the knockdown strains of representative universal essential or growth-defect-associated genes, including glcB, inhA, gyrB, and embB.." It is not clear what was the level of suppression of these genes in the respective KD strains. Authors should include the level of suppression of these genes also by qRT-PCR.

Thank you for the comment on the suppression levels of gene expression in knockdown strains of universal essential genes. Following the comment, we have evaluated them by qRT-PCR and we observed comparable levels of knockdown efficiency in the knockdown strains between universally essential genes and strain-specific/accessory essential genes (new Supplementary Fig. 9). Overall, the gene expression was suppressed to 20 - 70% in the knockdown strains compared to the vector control strains that do not express sgRNA.

We have added the data of qRT-PCR of knockdown strains of universal essential genes such as glcB, inhA, gyrB, and embB (new Supplementary Fig. 9). We have revised the Result and Discussion in the manuscript (page 21, lines 367-376; page28, lines 490-497).

(5f) Lines 293-: I am unable to establish any correlation between the growth of the knockdown with Tn insertion reads in the respective genes. For instance, pckA exhibits reduced Tn insertion reads in almost all the strains except in M.i.27, but the effect of its KD on growth is seen only in M.i.198 and M003; glpX exhibits reduced Tn insertion reads in M003, M019, M021 but the effect of its KD on growth is seen only in M003; csd exhibits reduced Tn insertion reads in M.i.198, M003, M019 but the effect of its KD on growth is seen only in M.i.198 and M003. The authors argue that these contradictory phenotypes are due to difficulties in the effective operation of genetically modified systems using foreign genes from different bacterial species in MAC-PD strains (Lines 310-312) or the desired effect on growth could not be observed due to the inability of CRISPRi to yield >99% suppression (Line 314) are not the valid justifications. Indeed, a close look at the RT-PCR data (Figure S5) reveals that pckA levels are ~0.22, 0.5, 0.2, 0.22, 0.2, 0.5, and 0.3 fold relative to sigA in M.i.198, M.i.27, ATCC13950, M018, M019, M003 and M021, respectively, but the effect of its suppression on growth by CRISPRi is seen only in M.i.198 and M003. Secondly, >99% suppression is not a universal prerequisite for all the genes to show growth defect (as might be the case with glcB, inhA, gyrB, and embB genes in this study). Hence, it remains unclear why contrasting results are obtained for most of the genes by TnSeq and CRISPRi.

Thank you for the comments on the issue of inconsistent results between TnSeq and CRISPR-i based knockdown. We acknowledge that some inconsistencies were observed, particularly among strain-dependent/accessory essential or growth-defect associated genes. By contrast, we found consistent data between TnSeq and CRISPR-i based knockdown results of universal essential genes. By obtaining the data of suppression levels of gene expression in the knockdown strains of universal essential genes, we have acknowledged that the low efficiency of knockdown does not explain the reason of the discrepancy between TnSeq and CRISPR-i results because the levels of knockdown efficiency were comparable between strain-dependent/accessory essential genes and universally essential genes.  

Although the mechanism has not been fully proven yet only from the current study, we consider that such inconsistent phenotypes with TnSeq and CRISPR-i based knockdown may be related to the recently revealed the bypass mechanism of gene essentiality which is characteristically observed in strain-dependent/accessory essential or growth-defect-associated genes. According to the publication by Rosconi (Nat Microbiol. 2022: new Ref#14) reporting the ‘forced-evolution experiments’ of 36 clinical Streptococcus pneumoniae strains, gene essentiality can be bypassed by several mechanisms including the composition of the accessory genome and pathway rewiring. They recovered successfully knockout mutants from transformation experiments in strain-specific/accessory essential genes such as cytidine monophosphate kinase, a folate pathway enzyme formate tetrahydrofolate ligase and an undecaprenyl phosphate-biosynthesis pathway enzyme farnesyl-diphosphate synthase. The bypassing of gene essentiality could be suggested by observing suppressor mutations and synthetic lethality in knockout strains. By contrast, universal essential genes were reported to fulfill the three categories including high levels of conservation within and often across species, limited genetic diversity, and high and stable expression levels. Consequently, universal essential genes are estimated to be rigid, largely immutable key components to an organism’s survival.

We consider that this is the case with our study on NTM because NTM is pangenomic. The knockdown of universal essential genes resulted in the clear growth suppression; however, the knockdown of strain-dependent/accessory essential genes did not show the consistent growth suppression. We consider that the bypass mechanism of gene essentiality can explain the inconsistent effect of gene silencing of strain-dependent/accessory genes on bacterial growth suppression.

We have added the above-mentioned description in the Discussion (pages 28-29, lines 497-519).

[Reference]

Rosconi, F. et al. A bacterial pan-genome makes gene essentiality strain-dependent and evolvable. Nat Microbiol 7, 1580–1592 (2022).

Minor Comments:

(1) The authors should mention the cut-off of fold-change for all the experiments in the methods section.

Thank you for the comment on the cut-off of fold-change. We set the cut-off of fold-change as adjusted P-value < 0.05. We added the description in the Methods section. (page 41, lines 724-725)

(2) Figure 7 legend (Lines 888-889): "Data are shown as the means {plus minus} SD of triplicate experiments. Data from one experiment representative of three independent experiments (N = 3) are shown."

Figure S3 legend: Data on the growth curves are the means of triplicate experiments. Data from one experiment representative of three independent experiments (N = 3) are shown.

Figure S4 legend: Data are shown as the means {plus minus} SD of triplicate experiments. Data from one experiment representative of two independent experiments (N = 2) are shown.

Figure S5 legend: Gene expression data are the means {plus minus} SD of triplicate experiments. Data from one experiment representative of two independent experiments (N = 2) are shown.

These statements need clarification. Whether multiple independent experiments (biological repeats), each with 2-3 technical replicates performed and the data shown represent one of the multiple biological repeats?

Thank you for the comments on the number of experiments performed and the number of replicates. We have performed two or three independent experiments with 2-3 technical replicates. The data shown represent one of the independent experiments.

(3) Figure 7b: Statistics are missing in the bar graph for growth rate under aerobic conditions.

Thank you for the comment on the statistics of the data regarding growth rate under aerobic conditions. We have added the statistics in the new Fig. 7c.

(4) The authors should check the y-axis in Figure 7b, as it is not clear whether bacteria indeed show a growth rate of 1-3 CFUs/day.

Thank you for the comment on the y-axis in Figure 7b. We have corrected the label of y-axis as “log10[CFUs]/day” in the new Fig. 7c. Additionally, we have corrected the label of y-axis in new Fig. 7a and added the description as “Data are represented as CFUs in 4 μl sample at each timepoint.” in the Fig. 7a legend.

Reviewer #3 (Recommendations For The Authors):

(1) It's notable that strains M001 and MOTT64 failed to undergo a transformation, while seven other strains did. Given that M001, MOTT64, and M019 belong to the same phylogenetic clade, it raises questions about why particular strains within this clade showed different transformation outcomes. It might be valuable for them to discuss this discrepancy in their study.

Thank you for the comment on the difference in capacity of transformation between strains belonging to the same genomic subgroup. Although the direct mechanism determining the competency for foreign DNA has not been elucidated in M. intracellulare and other pathogenic NTM species, several studies on general bacteria suggest the difficulties of introducing foreign DNA into clinical strains compared to the laboratory strains. As suggested in Staphylococcus aureus (Covaglia AR. PNAS. 2010; new ref#55), some clinical strains develop elimination system of foreign nucleic acids such as a type III-like restriction restriction endonuclease. As suggested in gran-negative bacteria (Qin J. Sci Rep. 2022; new Ref#56), there may be some difference in cell surface structures between strains, resulting in the necessity of polymyxin B nonapeptide targeting cell membrane for transforming clinical strains. The efficiency of eliminating foreign DNA may be attributed to various kinds of strain-specific factors including restriction endonuclease, natural CRISPR-interference system and cell wall structures rather than a simple genotypic factor.

We have added the description on the difference of capability in transformation in the Discussion. (page 31, lines 546-558)

[References]

Corvaglia, A.R., François, P., Hernandez, D., Perron, K., Linder, P. & Schrenzel, J. A type III-like restriction endonuclease functions as a major barrier to horizontal gene transfer in clinical Staphylococcus aureus strains. Proc Natl Acad Sci U S A 107, 11954-11958 (2010).

Qin, J., Hong, Y., Pullela, K., Morona, R., Henderson, I.R. & Totsika, M. A method for increasing electroporation competence of Gram-negative clinical isolates by polymyxin B nonapeptide. Sci Rep 12,:11629 (2022).

(2) The authors should consider specifying M. intracellulare in their title.

Thank you for the comment on the manuscript title. Following the comments from all Reviewers, we have modified the title as “Functional genomics reveals strain-specific genetic requirements conferring hypoxic growth in Mycobacterium intracellulare”.

Author Response

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

eLife assessment

This study presents a valuable finding on the immunophenotypes of cancer treatment-related pneumonitis. The evidence supporting the claims of the authors is solid, although the inclusion of controls, as suggested by one of the reviewers, strengthened the study. The work will be of interest to cancer immunologists.

Response: We are thankful for the editor's recognition of the contribution our study makes to understanding the immunophenotypes associated with cancer treatment-related pneumonitis. We agree that the inclusion of control data is pivotal for benchmarking biomarkers. While our initial study design was constrained by the availability of BALF from healthy individuals within clinical settings, we addressed this limitation by incorporating scRNA-seq data from healthy control and COVID-19 BALF cells sourced from the GSE145926 dataset. This additional analysis has provided a baseline for comparison, revealing that CD16 is expressed in a minority of T cells in healthy BALF, specifically 1.0% of CD4+ T cells and 1.6% of CD8+ T cells. The inclusion of this data as Figures 6H and 6I in our manuscript offers a robust context for the significant increase in CD16-expressing T cells observed in patients with PCP, thus enhancing the robustness of our study's conclusions.

Author response image 1.

Reviewer #1 (Recommendations For The Authors):

Many thanks for giving me the opportunity to review your paper. I really enjoyed the way you carried out this work - for example, your use of a wide panel of markers and the use of two analytical methods - you have clearly given great thought to bias avoidance. I also greatly appreciated your paragraph on the limitations, as there are several, but you do not 'over-sell' your conclusions so there is no issue here for me.

To improve the piece, there are a few typos (eg 318 - specific to alpha-myosin) and I was briefly confused about the highlighted clusters in Figure 4. Perhaps mention why they are highlighted when they first appear in 4D instead of E?

Response: We have corrected the typos, and we have rearranged the sequence of Figures 3E and 3F, as well as 4D and 4E, to ensure a logical flow. Citrus-generated violin plots are now presented prior to the heatmap of the clusters, which better illustrates the progression of our analysis and the derivation of the clusters.

In terms of improvements to the data, obviously it would have been ideal if you had had some sort of healthy control as a point of reference for all cohorts, but working in the field I understand the difficulties in getting healthy BAL. It would be worth your while however trying to find more supportive data in the literature in general. There are studies which assess various immune markers in healthy BAL eg https://journal-inflammation.biomedcentral.com/articles/10.1186/1476-9255-11-9. and so I think it is worth looking wrt the main findings. For example, are CD16+ T cells seen in healthy BAL or any other conditions (at present the COVID study is being over-relied on)? Could these cells be gamma deltas? (gamma deltas frequently express CD8 and CD16, and can switch to APC like phenotypes).

Response: We are grateful for the reviewer's consideration of the practical challenges associated with collecting BALF from healthy individuals. Alternatively, we have supplemented our analysis with single-cell RNA sequencing data from BALF cells of healthy controls, as found in existing literature (Nature Medicine 2020; 26: 842-844). We have accessed to GSE145926 and downloaded data of BALF cells from healthy control (n=3) and severe COVID19 (n=6). The filtered gene-barcode matrix was first normalized using ‘NormalizeData’ methods in Seurat v.4 with default parameters. The top 2,000 variable genes were then identified using the ‘vst’ method in Seurat FindVariableFeatures function. Then PCA and UMAP was performed. T cells were identified as CD2 >1 and CD3E >1, and FCGR3A expression was explored using an expression threshold of 0.5. Violin plots and bar plots were generated by ggplot function.

Regarding the pivotal finding of increased CD16-expressing T cells in patients with PCP, the scRNA-seq data mining indicates that CD16 is expressed by a minority of T cells in healthy BALF—1.0% of CD4+ T cells and 1.6% of CD8+ T cells. These figures, now incorporated into our revised manuscript as Figures 6H and 6I, substantiate our findings. These cells could be gamma delta T cells, but we could not confirm it with the limited data. We will investigate in the future study. The main text has been updated to reflect these findings.

Author response image 2.

I would agree with your approach of not going down the transcript route, so just focus on protein expression.

I think you need to mention more about the impact of ICI on PD1 expression - in the methods you lose one approach owing to low T cell expression (132) but in the discussion you mention ICI induced high expression (311) as previously reported. This apparent contradiction needs an explanation.

Response: We acknowledge the need for clarification regarding the impact of ICIs on PD-1 expression. In the methods section, the low detection of PD-1 expression on T cells in patients treated with nivolumab was indeed noted; this was due to the competitive nature of the PD-1 detection antibody EH12.2 with nivolumab. As reported by Suzuki et al. (International Immunology 2020; 32: 547-557), T cells from patients with ICI-induced ILD, including those treated with nivolumab, exhibit upregulated PD-1 expression, where the PD-1 detection antibody (clone: MIH4). Conversely, as outlined by Yanagihara et al. (BBRC 2020; 527: 213-217), the PD-1 detection antibody clone EH12.2 conjugated with 155Gd (#3155009B) used in our study is unable to detect PD-1 when patients are under nivolumab treatment due to competitive inhibition. The absence of a metal-conjugated PD-1 antibody with the MIH4 clone presented a limitation in our study. Ideally, we would have conjugated the MIH4 antibody with 155Gd for our analysis, which is a refinement we aim to incorporate in future research. We have now included this discussion in our manuscript to clarify the contradiction between the methodological limitations and the high PD-1 expression induced by ICIs, as reported in the literature. This addition will guide readers through the nuances of antibody selection and its implications for detecting PD-1 expression in the context of ICI treatment.

Finally, since you have the severity data, it would be good to assess all the significantly different clusters against this metric, as you have done for CD16+ T cells. Not only may this reveal more wrt the impact of other immune populations, but it'll also give a point of reference for the CD16+ T cell data.

Response: Thank you for the suggestion to assess all significantly different clusters against the disease severity metric. We have expanded our analysis to include a thorough correlation study between the disease severity and intensity of various T-cell markers. Notably, we observed that intensity of CCR7 expression correlates with the disease severity. Although the precise biological significance of this correlation remains to be elucidated, it may suggest a role for CCR7+ T cells in the pathogenesis or progression of the disease. We have considered the potential implications of this finding and included it as Supplementary Figure 5. We have also discussed this observation in the discussion section.

Author response image 3.

Overall though I think this is a really nice study, with a potentially very significant finding in linking CD16+ T cells with severity. Congratulations.

Response: We would like to thank the reviewer’s heartful comments on our manuscript.

Reviewer #2 (Recommendations For The Authors):

General:

1) The fact that this is a retrospective study should be indicated earlier in the paper.

Response: Now we have mentioned the retrospective nature of the study in the method section as follows: In this retrospective study, patients who were newly diagnosed with PCP, DI-ILD, and ICI-ILD and had undergone BALF collection at Kyushu University Hospital from January 2017 to April 2022 were included. The retrospective study was approved by the Ethics Committee of Kyushu University Hospital (reference number 22117-00).

2) tSNE and UMAP are dimensionality reduction techniques that don't cluster the cells, the authors should specify what clustering algorithm was used subsequently (e.g FlowSOM)

Response: The cluster was determined manually by their expression pattern.

3) With regards to the role of CD16 in a potential exacerbated cytotoxicity in the fatal PCP case, the authors could measure the levels of C3a related proteins in patient serum to link to a common immunopathogenic pathway with COVID.

Response: We did not collect serum from the patients in this study as our research protocol was approved by the Ethics committee for the use of BALF only. However, we agree with your assessment that the measurement of serum C3a levels would be informative. In future studies, we will incorporate the measurement of serum C3a levels to provide more comprehensive insights into the impact of C3a on immune function. Thank you for your valuable feedback and for helping us to improve the quality of our research.

Line-specific:

101 The authors should provide some information on how the cryopreservation of the BALF was carried out.

Response: Upon collection, BALF samples were immediately centrifuged at 300 g for 5 minutes to pellet the cells. The resultant cell pellets were then resuspended in Cellbanker 1 cryopreservation solution (Takara, catalog #210409). This suspension was aliquoted into cryovials and gradually frozen to –80ºC using a controlled rate freezing method to ensure cell viability. The samples were stored at –80ºC until required for experimental analysis. We have added the information in the method section.

Fig 3B: It would be very helpful if the authors could add a supplementary figure with marker expression on the UMAP projection.

Response: We have added Supplementary Figure 4 with marker expression on the UMAP projection in Figure 3B.

Fig 4A: Same as Fig 3B

Response: We have added Supplementary Figure 5 with marker expression on the UMAP projection in Figure 4A.

Fig 5B: Same as Fig 3B

Response: We have added Supplementary Figure 6 with marker expression on the tSNE projection in Figure 5B.

266 Authors should state if the data is not shown with regards to differences in myeloid cell fractions

430 Marker intensity is not shown in panel D

Re: Corrected as follows: “Citrus network tree visualizing the hierarchical relationship of each marker between identified T cell ~”

446 The legend says patients have IPF, CTD-ILD, sarcoidosis but the figure shows PCP, DI-ILD, ICI-ILD.

Re: Corrected.

451 What do the authors mean in "Graphical plots represent individual samples"? Panel B is a dot plot of all samples.

Response: Corrected as “Dot plots represent ~”.

472 What do the authors mean in "Graphical plots represent individual samples"? Panel C is a dot plot of all samples.

Response: Corrected as “Dot plots represent ~”.

Reviewer #3 (Recommendations For The Authors):

An important thing is to add comparisons against healthy donors, at least. A common baseline is needed to firmly establish any biomarkers.

Response: We acknowledge the reviewer's concern regarding the comparison with healthy donors. Although our study did not initially include BALF collection from healthy controls due to the constraints of clinical practice, we recognize the importance of a control baseline to validate biomarkers. To address this, we have integrated scRNA-seq data from healthy control BALF cells available in public datasets (Nature Medicine 2020; 26: 842-844), accessed from GSE145926. This dataset includes BALF cells from healthy controls (n=3) alongside severe COVID-19 patients (n=6). Data mining confirmed that CD16 expression is in a minority of T cells in healthy BALF—1.0% of CD4+ T cells and 1.6% of CD8+ T cells. We have included this comparative data in our manuscript as Figures 6H and 6I to provide context for the observed increase in CD16-expressing T cells in PCP patients, which substantiates our findings.

Author response image 4.

Data analysis needs to go deeper. There are several other tools on Cytobank alone that would allow a more quantitative analysis of the data. Fold changes in marker expressions would be very important as measurements of phenotypic changes.

Response: We thank the reviewer for their constructive feedback on the depth of our data analysis. We acknowledge the value of a more quantitative approach, including the use of fold change measurements to assess phenotypic alterations, and recognize the potential insights such tools on Cytobank could provide. Due to the scope and limited space of the current study, we have focused our analysis on the most pertinent findings relevant to our research questions. We believe the present analysis serves the immediate objectives of this study. However, we agree that further quantitative analysis would enhance the understanding of the data. We have expanded our analysis to include a thorough correlation study between the disease severity of PCP and intensity of various T-cell markers. Notably, we observed that intensity of CCR7 expression correlates with the disease severity of PCP. Although the precise biological significance of this correlation remains to be elucidated, it may suggest a role for CCR7+ T cells in the pathogenesis or progression of the disease. We have considered the potential implications of this finding and included it as Supplementary Figure 5. We have also discussed this observation in the discussion section. We aim to consider these approaches in future work to build upon the foundation laid by this study. Your suggestions are invaluable and will be kept at the forefront as we plan subsequent research phases.

Author response image 5.

Reviewer #1 (Public Review):

Cytotoxic agents and immune checkpoint inhibitors are the most commonly used and efficacious treatments for lung cancers. However their use brings two significant pulmonary side-effects; namely Pneumocystis jirovecii infection and resultant pneumonia (PCP), and interstitial lung disease (ILD). To observe the potential immunological drivers of these adverse events, Yanagihara et al. analysed and compared cells present in the bronchoalveolar lavage of three patient groups (PCP, cytotoxic drug-induced ILD [DI-ILD], and ICI-associated ILD [ICI-ILD]) using mass cytometry (64 markers). In PCP, they observed an expansion of the CD16+ T cell population, with the highest CD16+ T proportion (97.5%) in a fatal case, whilst in ICI-ILD, they found an increase in CD57+ CD8+ T cells expressing immune checkpoints (TIGIT+ LAG3+ TIM-3+ PD-1+), FCRL5+ B cells, and CCR2+ CCR5+ CD14+ monocytes. Given the fatal case, the authors also assessed for, and found, a correlation between CD16+ T cells and disease severity in PCP, postulating that this may be owing to endothelial destruction. Although n numbers are relatively small (n=7-9 in each cohort; common numbers for CyTOF papers), the authors use a wide panel (n=65) and two clustering methodologies giving greater strength to the conclusions. The differential populations discovered using one or two of the analytical methods are robust: whole population shifts with clear and significant clustering. These data are an excellent resource for clinical disease specialists and pan-disease immunologists, with a broad and engaging contextual discussion about what they could mean.

Strengths:

• The differences in immune cells in BAL in these specific patient subgroups is relatively unexplored.

• This is an observational study, with no starting hypothesis being tested.

• Two analytical methods are used to cluster the data.

• A relatively wide panel was used (64 markers), with particular strength in the alpha beta T cells and B cells.

• Relevant biomarkers, beta-D-glucan and KL-6 were also analysed

• Appropriate statistics were used throughout.

• Numbers are low (7 cases of PCP, 9 of DI-ILD, and 9 of ICI-ILD) but these are difficult samples to collect and so in relative terms, and considering the use of CyTOF, these are good numbers.

• Beta-D-glucan shows potential as a biomarker for PCP (as previously reported) whilst KL-6 shows potential as a biomarker for ICI-ILD (not reported before). Interestingly, KL-6 was not seen to be increased in DI-ILD patients.

• Despite the relatively low n numbers and lack of matching there are some clear differentials. The CD4/CD8+CD16+HLA-DR+CXCR3+CD14- T cell result is striking - up in PCP (with EM CD4s significantly down) - whilst the CD8 EMRA population is clear in ICI-ILD and 'non-exhausted' CD4s, with lower numbers of EMRA CD8s in DI-ILD.

• The authors identify 17/31 significantly differentiated clusters of myeloid cells, eg CD11bhi CD11chi CD64+ CD206+ alveolar macrophages with HLA-DRhi in PCP.

• With respect to B cells, the authors found that FCRL5+ B cells were more abundant in patients with ICI-ILD compared to those with PCP and DI-ILD, suggesting these FCRL5+ B cells may have a role in irAE.

• One patient's extreme CD16+ T cell (97.5% positive) and death, led the authors to consider CD16+ T cells as an indicator of disease severity in PCP. This was then tested and found to be correct.

• Authors discuss results in context of literature leading them to suggest that CD16+ T cells may target endothelial cells and wonder if anti-complement therapy may be efficacious in PCP.

• Great discussion on auto-reactive T cell clones where the authors suggest that in ICI-ILD CD8s may react against healthy lung, driving ILD.

• An observation of CXCR3 in different CD8 populations in ICI-ILD and PCP lead the authors to hypothesise on the chemoattractants in the microenvironment.

• Excellent point suggesting CD57 may not always be a marker of senescence on T cells - reflective of growing change within the community.

• Well considered suggestion that FCRL5+ B cells may be involved in ICI-ILD driven autoimmunity.

• The authors discuss the main weaknesses in the discussion and stress that the findings detailed in the paper "demonstrate a correlation rather than proof of causation".

• Figures and legends are clear and pleasing to the eye.

Weaknesses:

• This is an observational study, with no starting hypothesis being tested.

• Only patients who were able to have a lavage taken have been recruited.

• One set of analysis wasn't carried out for one subgroup (ICI-ILD) as PD1 expression was negative owing to the use of nivolumab.

• Some immune cell subsets wouldn't be picked up with the markers and gating strategies used; e.g. NK cells.

• Some immune cells would be disproportionately damaged by the storage, thawing and preparation of the samples; e.g. granulocytes.

• Numbers are low (7 cases of PCP, 9 of DI-ILD, and 9 of ICI-ILD), sex, age and adverse event matching wasn't performed, and treatment regimen are varied and 'suspected' (suggesting incomplete clinical data) - but these are difficult samples to collect. These numbers drop further for some analyses e.g. T cell clustering owing to factors such as low cell number.

• The disease comparisons are with each other, there is no healthy control.

• Samples are taken at one time point.

• The discussion on probably the stand out result - the CD16+ T cells in PCP - relies on two papers - leading to a slightly skewed emphasis on one paper on CD16+ cells in COVID. There are other papers out there that have observed CD16+ T cells in other conditions. It is also worth being in mind that given the markers used, these CD16+ T cell may be gamma deltas.

• The discussion on ICI patient consistently showing increased PD1, could have been greater, as given the ICI is targeting PD1, one would expect the opposite as commented on, and observed, in the methods section.

Reviewer #2 (Public Review):

Yanagihara and colleagues investigated the immune cell composition of bronchoalveolar lavage fluid (BALF) samples in a cohort of patients with malignancy undergoing chemotherapy and with with lung adverse reactions including Pneumocystis jirovecii pneumonia (PCP) and immune-checkpoint inhibitors (ICIs) or cytotoxic drug induced interstitial lung diseases (ILDs). Using mass cytometry, their aim was to characterize the cellular and molecular changes in BAL to improve our understanding of their pathogenesis and identify potential biomarkers and therapeutic targets. In this regard, the authors identify a correlation between CD16 expression in T cells and the severity of PCP and an increased infiltration of CD57+ CD8+ T cells expressing immune checkpoints and FCLR5+ B cells in ICI-ILD patients.

The conclusions of this paper are mostly well supported by data, but some aspects of the data analysis need to be clarified and extended.

1) The authors should elaborate on why different set of markers were selected for each analysis step. E.g., Different set of markers were used for UMAP, CITRUS and viSNE in the T cell and myeloid analysis.

2) The authors should state if a normality test for the distribution of the data was performed. If not, non-parametric tests should be used.

3) The authors should explore the correlation between CD16 intensity and the CTCAE grade in T cell subsets such as EMRA CD8 T cells, effector memory CD4, etc as identified in Figure 1B.

4) The authors could use CITRUS to better assess the B cell compartment.

Reviewer #3 (Public Review):

The authors collected BALF samples from lung cancer patients newly diagnosed with PCP, DI-ILD or ICI-ILD. CyTOF was performed on these samples, using two different panels (T-cell and B-cell/myeloid cell panels). Results were collected, cleaned-up, manually gated and pre-processed prior to visualisation with manifold learning approaches t-SNE (in the form of viSNE) or UMAP, and analysed by CITRUS (hierarchical clustering followed by feature selection and regression) for population identification - all using Cytobank implementation - in an attempt to identify possible biomarkers for these disease states. By comparing cell abundances from CITRUS results and qualitative inspection of a small number of marker expressions, the authors claimed to have identified an expansion of CD16+ T-cell population in PCP cases and an increase in CD57+ CD8+ T-cells, FCRL5+ B-cells and CCR2+ CCR5+ CD14+ monocytes in ICI-ILD cases.

By the authors' own admission, there is an absence of healthy donor samples and, perhaps as a result of retrospective experimental design, also an absence of pre-treatment samples. The entire analysis effectively compares three yet-established disease states with no common baseline - what really constitutes a "biomarker" in such cases? The introduction asserts that "y characterizing the cellular and molecular changes in BAL from patients with these complications, we aim to improve our understanding of their pathogenesis and identify potential therapeutic targets" (lines 82-84). Given these obvious omissions, no real "changes" have been studied in the paper. These are very limited comparisons among three, and only these three, states.

Even assuming more thorough experimental design, the data analysis is unfortunately too shallow and has not managed to explore the wealth of information that could potentially be extracted from the results. CITRUS is accessible and convenient, but also make a couple of big assumptions which could affect data analysis - 1) Is it justified to concatenate all FCS files to analyse the data in one batch / small batches? Could there be batch effects or otherwise other biological events that could confuse the algorithm? 2) With a relatively small number of samples, and after internal feature selection of CITRUS, is the regression model suitable for population identification or would it be too crude and miss out rare populations? There are plenty of other established methods that could be used instead. Have those methods been considered?

Colouring t-SNE or UMAP (e.g. Figure 6C) plots by marker expression is useful for quick identification of cell populations but it is not a quantitative analysis. In a CyTOF analysis like this, it is common to work out fold changes of marker expressions between conditions. It is inadequate to judge expression levels and infer differences simply by looking at colours.

The relatively small number of samples also mean that most results presented in the paper are not statistical significant. Whilst it is understandable that it is not always possible to collect a large number of patient samples for studies like this, having several entire major figures showing "n.s." (e.g. Figures 3A, 4B and 5C), together with limitations in the comparisons themselves and inadequate analysis, make the observations difficult to be convincing, and even less so for the single fatal PCP case where N = 1.

It would also be good scientific practice to show evidence of sample data quality control. Were individual FCS files examined? Did the staining work? Some indication of QC would also be great.

This dataset generated and studied by the authors have the potential to address the question they set out to answer and thus potentially be useful for the field. However, in the current state of presentation, more evidence and more thorough data analysis are needed to draw any conclusions, or correlations, as the authors would like to frame them.

Author response:

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

eLife assessment

This paper provides useful information about how the ionome of Arabidopsis thaliana adapts to very high CO2-levels, backed up by solid evidence and carefully designed studies. However, the broader claims of the paper about climate change and food security - heavily emphasized in the abstract, introduction, and discussion - are inappropriate, as there is no direct link to the presented work.

We sincerely thank you for the work you have done in reviewing our manuscript. We very much appreciate your overall positive assessment of the experimental work as a whole, its value and robustness.

In this revised version, we took on board the majority of your suggestions and your comments. In particular, we understood your critical point about overstating our objectives, which might in turn seem uncorrelated with our results. We fully agree with the comments that have been made on this point. Consequently, we have made substantial modifications and corrections in order to clarify our objectives and their implications: exploring in depth the natural variation of the shoot ionome response to elevated CO2, and generating a valuable resource allowing a better understanding of the genetic and molecular mechanisms involved in the regulation of plant mineral nutrition by the elevation of atmospheric CO2.

We also made modifications in response to the other suggestions, including a clarification of the functional experiments carried out around the function of TIP2;2 in response to elevated CO2. Figure 7 now comprises the comparison between both ambient and elevated CO2 conditions, which is much more informative that what appeared in the previous version.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

The study's abstract, introduction, and conclusions are not supported by the methods and results conducted. In fact, the results presented suggest that Arabidopsis could easily adapt to an extremely high CO2 environment.

We understand the reviewer’s comment. Although our work is considered useful, robust and well designed, we agree with the reviewer's point. We have certainly overemphasized the significance of our work to address the issue of food security in response to rising atmospheric CO2, at the expense of the factual description of the results of our fundamental study of the mechanisms at the interface between CO2 and mineral nutrition. We have clarified this focus by modifying the text of the introduction, objectives and discussion. We hope that these modifications will enable readers to better appreciate the core of this work.

Regarding the last part of the comment, our results do suggest that genetic variation could allow adaptation to rising atmospheric CO2, and our study does indeed aim to identify the extent and basis of this genetic variation.

This study offers good evidence pointing to a genetic basis for Arabidopsis thaliana's response to elevated CO2 (eCO2) levels and its subsequent impact on the leaf ionome. The natural variation analyses in the study support the hypothesis that genetic factors, rather than local adaptation, guide the influence of eCO2 on the ionome of rosette leaves in Arabidopsis. However, the manuscript's claim regarding its role in "the development of biofortified crops adapted to a high-CO2 world" (line 23) is overstated, especially given the absence of any analysis on the influence of eCO2 on the seed ionome and Arabidopsis is a poor model for harvest index for any crop. The manuscript, in its current form, necessitates massive revisions, particularly in clarifying its broader implications and in providing more substantial evidence for some of its assertions.

We thank the reviewer for this comment, and we would like to thank the reviewer for the positive appreciation for the identification of genetic basis for Arabidopsis thaliana's response to elevated CO2 and its subsequent impact on the leaf ionome. Nevertheless, it is true that the study of the leaf ionome is far from being able to lead to the development of biofortified plants. Some papers described that nutrient harvest index in Arabidopsis is a potential indicator of nutrient use efficiency (for instance, Masclaux-Daubresse and Chardon, Journal of Experimental Botany 2011 or Aranjuelo et al., Journal of Experimental Botany 2013). However, as we did not include any seed ionome data in the paper, we added clear mentions that our analyses were made on leaves (lines 56/57/250/319) and a comment in the discussion section to address this limitation (lines 325-328).

Major Drawbacks and Questions:

(1) Evidence for the Central Premise:

The foundational premise of the study is the assertion that rising atmospheric CO2 levels result in a decline in plant mineral content. This phenomenon is primarily observed in C3 plants, with C4 plants seemingly less affected. The evidence provided on this topic is scant and, in some instances, contradicts the authors' own references. The potential reduction of certain minerals, especially in grains, can be debated. For instance, reduced nitrogen (N) and phosphorus (P) content in grains might not necessarily be detrimental for human and animal consumption. In fact, it could potentially mitigate issues like nitrogen emissions and phosphorus leaching. Labeling this as a "major threat to food security" (line 30) is exaggerated. While the case for microelements might be more compelling, the introduction fails to articulate this adequately. Furthermore, the introduction lacks any discussion on how eCO2 might influence nutrient allocation to grains, which would be crucial in substantiating the claim that eCO2 poses a threat to food security. A more comprehensive introduction that clearly delineates the adverse effects of eCO2 and its implications for food security would greatly enhance the manuscript.

We partially agree with this comment. The decline in mineral status of C3 plants under conditions of elevated atmospheric CO2 has been widely described in the literature, and specifically documented for the cereal grains. While there are variations in this effect (depending on species, ecotype, cultivar), there is no debate about its acceptance. Here are just a few of the many works describing this effect, both on a global scale and at the level of the individual plant (Cotrufo MF (1998) Elevated CO2 reduces the nitrogen concentration of plant tissues. Global Change Biology 4: 43-54; Loladze I (2014) Hidden shift of the ionome of plants exposed to elevated CO(2)depletes minerals at the base of human nutrition. eLife 3: e02245; Myers SS (2014) Increasing CO2 threatens human nutrition. Nature 510: 139-142; Poorter H (1997) The effect of elevated CO2 on the chemical composition and construction costs of leaves of 27 C3 species. Plant, Cell & Environment 20: 472-482 ; Soares JC (2019) Preserving the nutritional quality of crop plants under a changing climate: importance and strategies. Plant and Soil 443: 1-26; Stitt] M (1999) The interaction between elevated carbon dioxide and nitrogen nutrition: the physiological and molecular background. Plant, Cell & Environment 22: 583-621; Uddling J (2018) Crop quality under rising atmospheric CO2. Curr Opin Plant Biol 45: 262-267).

In addition to this, the threat to food security posed by this alteration in plant mineral status has also been well described in the literature by several modeling approaches (Beach RH (2019) Combining the effects of increased atmospheric carbon dioxide on protein, iron, and zinc availability and projected climate change on global diets: a modelling study. Lancet Planet Health 3: e307-e317; Ebi KL (2019) Elevated atmospheric CO(2) concentrations and climate change will affect our food's quality and quantity. Lancet Planet Health 3: e283-e284; Medek DE (2017) Estimated Effects of Future Atmospheric CO2 Concentrations on Protein Intake and the Risk of Protein Deficiency by Country and Region. Environ Health Perspect 125: 087002; Smith MR (2018) Impact of anthropogenic CO2 emissions on global human nutrition. Nature Climate Change 8: 834-839; Weyant C (2018) Anticipated burden and mitigation of carbon-dioxide-induced nutritional deficiencies and related diseases: A simulation modeling study. PLoS Med 15: e1002586; Zhu C (2018) Carbon dioxide (CO2) levels this century will alter the protein, micronutrients, and vitamin content of rice grains with potential health consequences for the poorest rice-dependent countries. Sci Adv 4: eaaq1012). To reinforce this point, we have added a sentence and references (lines 30-33). Nevertheless, we understand the reviewer's comment on the nuance to be given to the intensity of this potential threat. We have therefore modified the text, replacing "major threat" by "significant threat" (lines 3 and 29).

We also would like to answer the reviewer’s comment on the potential environmental benefit associated with reduced N and P content in grains (mitigation of N emissions and P leaching). Indeed, if this reduced N and P content results from a lowered use efficiency of soil nutrients by plants, as suggested by several studies (Bloom 2010, Cassan 2023, Gojon 2023 and references therein), this may at the opposite favor N oxides emission and P leaching from the soil.

(2) Exaggerated Concerns:

The paper begins with the concern that carbon fertilization will lead to carbon dilution in our foods. While we indeed face numerous genuine threats in the coming decades, this particular issue is manageable. The increase in CO2 alone offers many opportunities for boosting yield. However, the heightened heat and increased evapotranspiration will pose massive challenges in many environments.

While there are indeed multiple threats that we are facing in the coming decades, we don't fully agree with this comment. At present, there's no evidence to say that the negative effect of CO2 on plant mineral content will be manageable. Furthermore, there is compelling evidence that altered mineral nutrition and mineral status of plants will be an important factor limiting the high CO2-induced increase in yield, as will be heat or increased evapotranspiration (see for instance Coskun et al (2016) Nutrient constraints on terrestrial carbon fixation: The role of Nitrogen. J. Plant Physiol. 203: 95-109; Jiang M (2020) Low phosphorus supply constrains plant responses to elevated CO2 : A meta-analysis. Glob Chang Biol 26: 5856-5873 ; Reich PB (2006) Nitrogen limitation constrains sustainability of ecosystem response to CO2. Nature 440: 922-925). Thus, although we do not negate the crucial importance of heat and water stress, we believe it is relevant to study the basic mechanisms responsible for the negative effect of CO2 on plant mineral composition.

Figure 4 in fact suggests that 43% of the REGMAP panel (cluster 3) is already pre-adapted to very high CO2 levels. This suggests annual species could adapt very rapidly.

We agree with the reviewer. However, this suggests that genetic variation exists in some ecotypes to support adaptation to elevated CO2. The purpose of this work is indeed to identify this genetic variation, in order to characterize the mechanisms behind.

(3) Assumptions on CO2 Levels:

The assumption of 900ppm seems to be based on a very extreme climate change scenario. Most people believe we will overshoot the 1.5°C scenario, however, it seems plausible that 2.5 to 3°C scenarios are more likely. This would correspond to around 500ppm of CO2. https://www.nature.com/articles/s41597-022-01196-7/tables/4

We agree with the reviewer that the CO2 concentration we used corresponds to a high value in the IPCC projections. That said, this value is currently considered very plausible: the following figure (from Smith and Myers (2018) Nature Climate Change) shows that current CO2 emissions align with the IPCC's most extreme model (RCP 8.5), which would result in a CO2 concentration of around 900 ppm in 2100. Furthermore, nothing allows to exclude the 4°C scenario in the 6th IPCC report.

Author response image 1.

(4) Focus on Real Challenges:

We have numerous real challenges, such as extreme heat and inconsistent rainfall, to address in the context of climate change. However, testing under extreme CO2 conditions and then asserting that carbon dilution will negatively impact nutrition is exaggerated.

While we fully agree that several threats linked to climate change exist, and all deserve to be studied, we find it questionable to consider that the potential effect of high CO2 on the mineral nutrition of plants is not a real challenge. The mineral nutrition of plants is already a current major environmental challenge. This perspective seems to reflect the reviewer's personal opinion rather than an analysis of our work.

In contrast, the FACE experiments are fundamental and are conducted at more realistic eCO2 levels. Understanding the interaction between a 20% increase in CO2 and new precipitation patterns is key for global carbon flux prediction.

Again, we do not fully understand this comment, as the aim of our study was not to perform a global carbon flux prediction, but to unravel genes and mechanisms underlying the negative effect of elevated CO2 on the nutrient content of Arabidopsis rosettes. However, we agree with the reviewer’s comment and with the fact that FACE are useful facilities to explore the CO2 response in more natural environments, and we highlight the fact that the decrease in mineral status of C3 plants has been widely documented in FACE studies. FACE experiments do not facilitate, however, to conduct fully controlled experiments (temperature, rainfall, wind and light intensities are not controllable in FACE), that allow to disentangle the mechanisms by which elevated CO2 regulates the signaling pathways associated with the plant mineral composition. In the longer term, studying the mechanisms we have identified in a more global context of climate change could be highly relevant.

As I look at the literature on commercial greenhouse tomato production, 1000ppm of eCO2 is common, but it also looks like the breeders and growers have already solved for flavor and nutrition under these conditions.

Indeed, tomato is often cultivated in CO2-enriched greenhouses at 1000 ppm. According to the literature, this results in a 20-25% reduction in vitamin C or lycopene, and requires a significantly higher nitrogen and water intake to reach expected sugar levels (Doddrell H (2023) Horticulture Research). In addition, the negative effect of elevated CO2 on tomato nutrient content seems to have significant repercussions on nutrition-health properties (Boufeldja (2023), Molecules).

Conclusion:

While the study provides valuable insights into the genetic underpinnings of Arabidopsis thaliana's response to elevated CO2 levels, it requires an entirely revised writeup, especially in its abstract, broader claims and implications. The manuscript would benefit from a more thorough introduction, a clearer definition of its scope, and a clear focus on the limits of this study.

We thank the reviewer for the comments made on our manuscript. In addition to the responses that we provide to these comments, we have modified the main text of the introduction, objectives and discussion to take these comments into consideration. We believe that this will significantly improve the manuscript.

Reviewer #2 (Public Review):

Strengths:

The authors have conducted a large, well-designed experiment to test the response to eCO2. Overall, the experimental design is sound and appropriate for the questions about how a change in CO2 affects the ionome of Arabidopsis. Most of the conclusions in this area are well supported by the data that the authors present.

We thank the reviewer for this positive appreciation.

Weakness:

While the authors have done good experiments, it is a big stretch from Arabidopsis grown in an arbitrary concentration of CO2 to relevance to human and animal nutrition in future climates. Arabidopsis is a great model plant, but its leaves are not generally eaten by humans or animals.

We agree with the reviewer’s comment. We recognized that implying a direct contribution of our work to human nutrition in the future climates is overstated, as mentioned by the reviewer 1 as well. This was not an intentional overstatement, as we have always been convinced that our work contributed to the understanding of the basic mechanisms involved in the negative regulation of plant mineral nutrition by high CO2. We have significantly modified the text to correct any misunderstanding of our work’s implication.

The authors don't justify their choice of a CO2 concentration. Given the importance of the parameter for the experiment, the rationale for selecting 900 ppm as elevated CO2 compared to any other concentration should be addressed. And CO2 is just one of the variables that plants will have to contend with in future climates, other variables will also affect elemental concentrations.

We agree with this comment. We added a justification of the high CO2 concentration used in this work in the Material and Methods section (lines 343-344). You can also read the explanation of this choice in the response to the reviewer 1’s point 3.

Given these concerns, I think the emphasis on biofortification for future climates is unwarranted for this study.

Anew, we agree with this comment and we have significantly modified the text to correct any misunderstanding of our work’s implication.

Additionally, I have trouble with these conclusions:

-Abstract "Finally, we demonstrate that manipulating the function of one of these genes can mitigate the negative effect of elevated CO2 on the plant mineral composition."

-Discussion "Consistent with these results, we show that manipulating TIP2;2 expressions with a knock-out mutant can modulate the Zn loss observed under high CO2."

The authors have not included the data to support this conclusion as stated. They have shown that this mutant increases the Zn content of the leaves when compared to WT but have not demonstrated that this response is different than in ambient CO2. This is an important distinction: one way to ameliorate the reduction of nutrients due to eCO2 is to try to identify genes that are involved in the mechanism of eCO2-induced reduction. Another way is to increase the concentration of nutrients so that the eCO2-induced reduction is not as important (i.e. a 10% reduction in Zn due to eCO2 is not as important if you have increased the baseline Zn concentration by 20%). The authors identified tip2 as a target from the GWAS on difference, but their validation experiment only looks at eCO2.

We thank the reviewer for this comment, and we agree with it. It is much more interesting, especially in the context of this paper, to analyze the function of a candidate gene not only in elevated CO2, but in both ambient and elevated CO2. Therefore, we added in Figure 7 data for the expression of TIP2;2 in contrasted haplotypes under ambient CO2, in comparison to those already presented under elevated CO2 (now Fig. 7C and 7D). This showed that TIP2;2 expression is lower in haplotype 0 also under ambient CO2. We also added in Figure 7 (Fig. 7E) the Zn level in WT and tip2;2-1 mutant under ambient CO2, in comparison to those already presented under elevated CO2. This showed that that the tip2;2-1 mutant line did not present any decrease in Zn shoot content in response to elevated CO2, in opposition to what is observed for the WT.

We have added comments associated to these new results in the Results and Discussion sections and in the discussion section (lines 233-242 in the results section, and lines 310-314 in the discussion section).

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

Reviewer Comments on the Article's Approach to Ionome Analysis

(1) Omission of Phosphorus from the Ionome:

It's surprising that phosphorus (P) was not measured in the ionome. After nitrogen (N), P is often the most limiting mineral for plant development and yield, making it a significant component of the ionome. Why did the authors omit this crucial element?

We agree with the reviewer that P is an important mineral for plant growth. The absence of data related to P content is due to feasibility constraints rather than oversight. The MP-AES instrument we used to analyze the ionome (except N and C, that we obtained from an Elementar Analyzer) would have required an extra-step and an extra-analysis to obtain data for macronutrient such as P or K. In the context of this large-scale experiment, we faced the necessity to compromise and proceed without these data.

(2) Relationship Between Leaf Ionome and Seed:

The manuscript lacks evidence demonstrating the relationship between the leaf ionome and the seed. This connection is vital to establish the study's aims as outlined in lines 20-24. If the central argument is that eCO2 threatens food security, it's essential for the authors to either:

• Provide evidence that eCO2 induces changes in the ionome profiles of seeds.

• Show that changes in the rosette leaf ionome lead to alterations in seed ionome profiles.

We agree with the reviewer. Although we know that seed ionome composition of Arabidopsis model accession such as Columbia is indeed negatively affected by eCO2, we do not provide the data that support some of the terms used in lines 20-24. The correspondence between leaf and seed ionome in natural population under eCO2 is certainly a next question that we will address. Therefore, to align our stated objectives with our data, we have modified the sentence in lines 20-24. We also added a comment on this point lines on the discussion section (lines 324-328).

(3) Analysis of Ionome in Rosette Leaves:

Why did the authors choose to analyze the ionome specifically in rosette leaves? Is there a known correlation between the ionome profile in rosette leaves and seeds?

See our answer to the above comment.

(4) Experimental Design Comments:

• The layout of the accession growouts, the methods of randomization, blocking, and controls/checks should be detailed.

• Were BLUEs (Best Linear Unbiased Estimators) or BLUPs (Best Linear Unbiased Predictors) employed to account for experimental design conditions? If not, it's recommended that they be used.

We thank the reviewer for this comment. A note on replicates has been added in the Method/Plant Material section. Concerning the BLUEs/BLUPs, although I am not familiar with their use, I do not think that these approaches are relevant in our experimental design. Indeed, we pooled 3 to 5 replicates for each accession to measure the ionome (as mentioned in the Method/Ionome analysis section – we realized this was perhaps not clear enough, and thus we reinforced this point in this section). Therefore, we do not have the variance data required to perform BLUEs/BLUPs.

(5) Carbon Dilution Effect:

The statement, "The first component of the PCA described a clear antagonistic trend between C content and the change of other mineral elements (Fig. 3B)..." suggests a well-understood carbon dilution effect. These results are anticipated and align with existing knowledge.

We thank the reviewer for this comment. However, this sentence does not relate to the biomass dilution hypothesis referred to by the reviewer. Indeed, the composition of each mineral (C and others) is expressed as a percentage of biomass, not as an absolute value. Therefore, this reflects more a probable effect of the increase in carbon compounds (notably soluble sugars), which could influence mineral composition.

(6) Heritability Estimates:

The authors should report both the broad-sense heritability and an estimate of heritability based on a GRM or Kinship matrix.

We thank the reviewer for this suggestion. We are skeptical of using a kinship matrix to estimate heritability in our study. Estimating narrow-sense heritability using a kinship matrix is conceptually based on the infinitesimal model of Fisher, thereby meaning that phenotypic variation is driven by hundreds to thousands of QTLs with small effects. If this is the case, GWAS conducted on several hundred (or even thousands) of genotypes will not be powerful enough to detect such QTLs. Accordingly, estimates of broad-sense heritability based on estimates of variance components can drastically differ from estimates of narrow-sense heritability based on the use of a kinship matrix, as illustrated in the study of Bergelson et al. (2019 Scientific Reports).

(7) Application of the Breeder's Equation:

It would be beneficial if the authors applied the breeder's equation to estimate the species' potential rate of response. Based on the allele frequency of the adapted cluster 3 (69 ecotypes or 43% frequency of Figure 3B), it seems plausible that the populations could adapt within 23 generations.

We thank the reviewer for this suggestion. Indeed, it would be really interesting to test whether sub-populations could adapt in comparison with others, and over what period of time. It is nevertheless not possible to do so using the Breeder’s equation in our case, as this requires fitness data under conditions of ambient or elevated CO2 (i.e. production of seeds) to be applied, and we do not have these data at the level of the whole population.

(8) Overall Quality:

In general, the authors have executed a high-quality ionome mapping experiment. However, the abstract, introduction, and discussion should be entirely rewritten and reframed.

We thank the reviewer for the positive evaluation of our experiment. As previously mentioned, we are for the most part in agreement with the comments made about the need to align our stated objectives with our experimental data and conclusions. To do so, we have rewritten part of the abstract, introduction and discussion. The details of these modifications are described in the responses made to each comment.

Here's a line-by-line list of suggestions on writing:

Line 30 would read better with a comma after thus (or by replacing thus with therefore and then a comma at the start of the sentence).

Line 33 nevertheless would read better in between commas.

Lines 45 - 48 sentence is too long, could probably divide it into two.

Lines 90 - 94 are hard to interpret, recommend rephrasing for clarity.

Line 130 - keep verbs in the past tense for consistency (ran instead of run).

Line 194 - what do the authors mean by crossed? I'm inferring they looked at the intersection of DEGs with the list of genes identified by GWA mapping, probably should use a more concise word.

There's a concurrent use of the adjective strong (Lines 80, 142, 144, 197, 245). I would advise using a more concise adjective or avoiding its use to let the reader form their own opinion on the data.

Lines 174-176 the cited reference (No. 15) is incorrect. The study by Katz et al. (2022) does not provide information on the role of ZIF1 in zinc sequestration mechanisms under elevated CO2 conditions.

We thank the reviewer for these detailed recommendations. We have corrected or rephrased the text according to these suggestions.

Reviewer #2 (Recommendations For The Authors):

Technical points:

900 ppm as elevated CO2: Given the importance of the parameter for the experiment, the rationale for selection 900 ppm as elevated CO2 compared to any other concentration should be addressed.

We acknowledge the reviewer's point and have previously addressed related aspects earlier in our response. In line with this, we have included a justification for this particular parameter in the Method section.

The authors do not mention what genotype was used for their root/shoot RNAseq experiment.

We thank the reviewer for this comment, and indeed, this information was not mentioned. This is now done, in the Method section.

Line 125: Spelling error "REGMPA".

This has been corrected.

Line 338: Removal of outlier observations - "Prior to GWAS and multivariate analyses such as PCA or clustering, mineral composition measures were pre-processed to remove technical outliers". The authors should mention the exact number of outliers that were removed and what the explicit criteria were for removal.

The number of outliers removed from each dataset is now indicated in Supplemental Table 7 (this is cited in the Method section). The explicit criteria used for this analysis is actually mentioned in the corresponding Method section: “the values positioned more than 5 median absolute deviations away from the median were removed from the dataset”.

Line 379: "Lowly expressed genes with an average value across conditions under 25 reads were excluded from the analysis". Providing information about the number of the lowly expressed genes that were removed from the analysis can help with the interpretation of the likelihood of the candidates selected being correct.

This is a standard procedure in RNAseq analysis. It avoids many false positives in the differential analysis of gene expression based on ratios (where a very small number in the denominator can lead to a very high variation in expression, of no real significance). For information, this step led to the removal of 11607 and 10121 genes for the shoot and root datasets.

Line 384: It's not clear how many biological replicates were used.

This has been corrected.

Additional comment: We have also become aware of a confusion concerning one of the candidate genes located close to GWA peaks: line 180 of the first version, we mentioned CAX1 (AT1G16380) for its role on nutrient deficiency response. There are actually two genes annotated as CAX1 in TAIR (both are cation exchangers), but the one involved in nutrient deficiency response is AT2G38170. We therefore removed the sentence mentioning AT1G16380/CAX1 as a potential candidate gene.

Author Response

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

Reviewer #1 (Public Review):

This paper performed a functional analysis of the poorly characterized pseudo-phosphatase Styxl2, one of the targets of the Jak/Stat pathway in muscle cells. The authors propose that Styxl2 is essential for de novo sarcomere assembly by regulating autophagic degradation of non-muscle myosin IIs (NM IIs). Although a previous study by Fero et al. (2014) has already reported that Styxl2 is essential for the integrity of sarcomeres, this study provides new mechanistic insights into the phenomenon. In vivo studies in this manuscript are compelling; however, I feel the contribution of autophagy in the degradation of NM IIs is still unclear.

Major concerns:

1) The contribution of autophagy in the degradation of Myh9 is still unclear to this reviewer.

It has been reported that autophagy is dispensable for sarcomere assembly in mice (Cell Metab, 2009, PMID; 1994508). In Fig. 7A, the authors showed that overexpressed Styxl2 downregulated the amount of ectopically expressed Myh9 in an ATG5-dependent manner in C2C12 cells; however, the experiment is far from a physiological condition. Therefore, the authors should test ATG5 knockdown and the genetic interaction between Styxl2 and ATG5 in vivo. That is, 1) loss of ATG5 on sarcomere assembly in zebrafish, and 2) the genetic interaction between Styxl2 and ATG5; co-injection of Styxl2 mRNA and ATG5-MO into the zebrafish embryos.

Our response: In fact, the reference cited by the reviewer (Cell Metab, 2009; PMID; 19945408) clearly indicated that autophagy is required for sarcomere assembly. Moreover, another paper using the fish extraocular muscle regeneration model (Autophagy, 2014, PMID: 27467399), also showed that the sarcomere structure was disrupted in the regenerated muscles when autophagy was inhibited by chloroquine. In addition, other references (Nature medicine, 2007, PMID: 17450150; Autophagy, 2010, PMID: 20431347) also showed that loss of Atg5 in mouse cardiac muscles led to disorganized sarcomere structure. We also performed the Atg5 knockdown experiments as suggested by the reviewer. However, the sarcomere structure defects were not so obvious as Styxl2 knockdown (see Author response image 1 below). In fact, it was reported that Atg5 knockdown may not be a desirable strategy to disrupt autophagy as it was found “--- only a small amount of Atg5 is needed for autophagy, knockdown of Atg5 to levels low enough to block autophagy might be difficult to achieve, --” (Nature medicine, 2007, PMID: 17450150). Due to the ineffectiveness of the Atg5 MO in our assays, we did not perform the second experiment suggested by the reviewer. Moreover, as Styxl2 is not a key component of the autophagy machinery, it is less likely that overexpression of Styxl2 alone can rescue the autophagy defects caused by Atg5.

Author response image 1.

The fish zygotes were injected with Atg5 or Ctrl MO. 48 hpf, the fish were stained with an anti-Actinin antibody. Some fast muscle fibers were disrupted when Atg5 was knocked down. The number in numerator at the bottom of each image represents fish embryos showing normal Actinin staining pattern, while that in denominator represents the total number of embryos examined. Scale bar, 10 µm.

2) As referenced, Yamamoto et al. reported that Myh9 is degraded by autophagy. Mechanistically, Nek9 acts as an autophagic adaptor that bridges Atg8 and Myh9 through interactions with both. Inconsistent with the model, the authors mentioned on page 12, lines 365-367, "A recent report showed that Myh9 could also undergo Nek9-mediated selective autophagy (Yamamoto et al., 2021), suggesting that Myh9 is ubiquitinated". I think it is not yet explored whether autophagic degradation of Myh9 requires its ubiquitination. Moreover, I cannot judge whether Myh9 is ubiquitinated in a Styxl2-dependent manner from the data in Fig. 7C. The author should test whether Nek9 is required for Myh9 degradation in muscles. If Nek plays a role in the Myh9 degradation, it would be better to remove Fig. 7C.

Our response: Indeed, as pointed out by the reviewer, it has not been explored whether Myh9 is ubiquitinated or not. However, it has been well-established that some proteins undergoing autophagic degradation are ubiquitinated, which are linked to Atg8/LC3 via p62 and NBR1 (Mol Cell, 2009, PMID: 19250911; J Biol Chem, 2007, PMID: 17580304). To improve the data quality, we repeated the Myh9 ubiquitination experiment in cells with or without Styxl2 by using a slightly different strategy: as shown in the revised Figure 7C, we first co-transfect HEK 293T cells with HA-Myh9, Myc-ubiquitin, and Flag-Styxl2. We then immunoprecipitated Myc-tagged Ubiquitin from the whole cell lysates, and then blot for HAMyh9. We detected an obvious increase in Ubiquitin-conjugated HA-Myh9 (revised Figure 7C). As suggested by the reviewer, we also tested whether knockdown of Nek9 affects the degradation of Myh9. We failed to detect an obvious effect (see Author response image 2 below) caused by Nek9 knockdown. One possible explanation for this negative result is that Nek9 itself is a negative regulator of selective autophagy (J Biol Chem, 2020, PMID: 31857374). By knocking it down, the functions of the autophagy machinery are expected to be enhanced instead of being impaired. This may explain why we failed to detect an effect on Myh9 degradation simply by knocking down Nek9. To further elucidate whether Nek9 is involved in Myh9 degradation in myoblasts, we may need to use a dominant-negative mutant of Nek9 missing the LCIII-binding motif as shown by Yamamoto (Nat Commun, 2021, PMID: 34078910). This will be addressed in our future study.

Author response image 2.

C2C12 cells were transfected with negative control siRNA (NC), siNek9#2 or siNek9#3. 18 h later, the cells were transfected with plasmids HA-Myh9 and Flag-Styxl2 or Flag-Stk24. After another 24 h, the cells were harvested for RT-qPCR (left panel) or western blot (right panel).

3) In Fig. 5F, the protein level of Styxl2 and Myh10 should be checked because the efficiency of Myh10-MO was not shown anywhere in this manuscript.

Our response: As suggested by the reviewer, a Western blot showing the protein levels of Myh10 was shown in Figure 5-figure supplement 1B.

Reviewer #2 (Public Review):

The authors investigated the role of the Jak1-Stat1 signaling pathway in myogenic differentiation by screening the transcriptional targets of Jak1-Stat1 and identified Styxl2, a pseudophosphatase, as one of them. Styxl2 expression was induced in differentiating muscles. The authors used a zebrafish knockdown model and conditional knockout mouse models to show that Styxl2 is required for de novo sarcomere assembly but is dispensable for the maintenance of existing sarcomeres. Styxl2 interacts with the non-muscle myosin IIs, Myh9 and Myh10, and promotes the replacement of these non-muscle myosin IIs by muscle myosin IIs through inducing autophagic degradation of Myh9 and Myh10. This function is independent of its phosphatase domain.

A previous study using zebrafish found that Styxl2 (previously known as DUSP27) is expressed during embryonic muscle development and is crucial for sarcomere assembly, but its mechanism remains unknown. This paper provides important information on how Styxl2 mediates the replacement of non-muscle myosin with muscle myosin during differentiation. This study may also explain why autophagy deficiency in muscles and the heart causes sarcomere assembly defects in previous mouse models.

Reviewer #3 (Public Review):

Wu and colleagues are characterising the function of Styxl2 during muscle development, a pseudo-phosphatase that was already described to have some function in sarcomere morphogenesis or maintenance (Fero et al. 2014). The authors verify a role for Styxl2 in sarcomere assembly/maintenance using zebrafish embryonic muscles by morpholino knockdown and by a conditional Styxl2 allele in mice (knocked-out in satellite cells with Pax7 Cre).

Experiments using a tamoxifen inducible Cre suggest that Styxl2 is dispensable for sarcomere maintenance and only needed for sarcomere assembly.

BioID experiments with Styxl2 in C2C 12 myoblasts suggest binding of nonmuscle myosins (NMs) to Styxl2. Interestingly, both NMs are downregulated when muscles differentiate after birth or during regeneration in mice. This down-regulation is reduced in the Styxl2 mutant mice, suggesting that Styxl2 is required for the degradation of these NMs.

Impressively, reducing one NM (zMyh10) by double morpholino injection in a Styxl2 morphant zebrafish, does improve zebrafish mobility and sarcomere structure. Degradation of Mhy9 is also stimulated in cell culture if Styxl2 is co-expressed. Surprisingly, the phosphatase domain is not needed for these degradation and sarcomere structure rescue effects. Inhibitor experiments suggest that Styxl2 does promote the degradation of NMs by promoting the selective autophagy pathway.

Strengths:

A major strength of the paper is the combination of various systems, mouse and fish muscles in vivo to test Styxl2 function, and cell culture including a C2C12 muscle cell line to assay protein binding or protein degradation as well as inhibitor studies that can suggest biochemical pathways.

Weakness:

The weakness of this manuscript is that the sarcomere phenotypes and also the western blots are not quantified. Hence, we rely on judging the results from a single image or blot. Also, Styxl2 role in sarcomere biology was not entirely novel.

Few high resolution sarcomere images are shown, myosins have not been stained for.

Reviewer #1 (Recommendations For The Authors):

Minor concerns:

4) The position of molecular weight markers should be shown in all Western blot data.

Our response: As suggested by the reviewer, the molecular weight markers have been added in the Western blot data.

5) Schematic models of Styxl2deltaN509 and N513 construct would be helpful for the readers.

Our response: A schematic has been added in Figure 6B (upper panel) to show Styxl2deltaN509 and Styxl2N513.

6) Several data were described but not shown (data not shown). I think the data need to be included in the main or supplemental figures.

Our response: As suggested by the reviewer, the raw data were now included in the Figure 6-figure supplement 1A and Figure 7-figure supplement 1.

Reviewer #2 (Recommendations For The Authors):

1) In Fig. 5E, the authors suggest that the needle touch response was improved by additional knockdown of Myh10. This is a bit confusing because the germline knockout of Myh10 is lethal (line 445). The authors should provide more explanation on this point. Additionally, it would be better to include Myh10-MO in Fig. 5E.

Our response:<br /> In line 445 of our original manuscript, we stated that germline knockout of mouse Myh10 gene is lethal based on a published report (Proc Natl Acad Sci USA, 1997, PMID: 9356462). Here, in zebrafish zygotes, we only knocked down zMyh10, thus, we do not expect to get a lethal phenotype. In addition, other groups who knocked down Myh10 in fish also did not get a lethal phenotype (Dev Biol, 2015, PMID: 25446029). As to the control involving Myh10MO in the experiment in Fig.5E, we did include it in our experiments. As we did not observe any obvious effects on either motility or sarcomere structures, we did not include the data set in the figure.

2) It was suggested that Myh9 and Myh10 form a complex (Rao et al. PLoS One 9, e114087, 2014). Thus, the IP experiments do not rule out the possibility that Styxl2 directly interacts with either Myh9 or Myh10 and indirectly with the other.

Our response: In known myosin-II complexes, different myosin molecules can associate with each other through their tail domains (Bioarchitecture, 2013, PMID: 24002531). Thus, if we use fulllength myosin molecules in our co-immunoprecipitation assays, it will be difficult to exclude the possibility raised by the reviewer. However, by using truncated myosin proteins, we showed that the head domain of either Myh9 or Myh10 could interact with Styxl2 in the absence of the tail domain (Figure 4E, F). This result strongly suggests that both Myh9 and Myh10 can independently interact with Styxl2.

Reviewer #3 (Recommendations For The Authors):

1) The western blot shown in Figure 3B supporting the induced deletion of Styxl2 should be quantified. Ideally, some other blots, e.g., in Figure 5, too. Please add the age of the mice in Figure 5B to the figure legend.

Our response:<br /> As suggested by the reviewer, we quantified the data in Figures.3B, 3F, 5B, 5D, and 7A and the data were included in the revised figures. In Fig.5B, we already indicated the age of the mice (i.e., P1) in the legend.

2) A quantification of the sarcomere phenotypes in the double knock-down of zMyh10 and Styxl2 compared to Styxl2 single would make the paper significantly stronger. Furthermore, a double morpholino control should be included to rule out any RNAi machinery 'dilution effect'.

Our response: As suggested by the reviewer, we quantified the sarcomere structures using the line scan analysis in ImageJ and the scan images were placed as inserts in the upper corner of the immunofluorescent images (revised Figures 5F, and 6C). To avoid potential “dilution effects”, in all the experiments involving the use of two different MOs, the total amount of MO was kept the same in all control samples by including a control MO (e.g., in samples treated with one specific MO, an equal amount of a control MO was also included, while in samples without any specific MO, twice as much control MO was used).

3) The sarcomere phenotypes in figure 6 should also be better quantified, for example using simple line scans of the alpha-actinin stains and assay periodicity or calculating the autocorrelation coefficients. How about myosin stains?

Our response: We quantified Figure 6C as suggested by the reviewer. We also performed myosin staining. The results were similar to that shown by the a-actinin antibody (see revised Figure 6-Fig supplement 1B).

4) Do the authors see periodic NMs patterns in developing mouse muscle fibers as indicated by the model in in in figure 7D? It is unclear if nonmuscle myosin is present in a PERIODIC pattern in early myofibrils. NM myosin periodic patterns that have been observed have a periodicity of only about 1 µm fitting the shorter length of the NM bipolar filaments (about 300 nm only, PMID 28114270).

Our response: The reviewer raised a good point here. Ideally, we should examine developing mouse muscle fibers to prove that NM shows periodic patterns. However, due to the difficulty in catching myocytes undergoing sarcomere assembly, the majority of the studies involving NM in sarcomeres use cultured cardiomyocytes. Using TA muscles from P1 new-born mice, we failed to detect the presence of NM in sarcomeres (see Author response image 3 below). Actually, nearly all the myofibers showed mature sarcomere pattern without the NM signal. More work is needed in the future to examine developing mouse fibers at different embryonic stages to look for the presence of NM in developing sarcomeres.

Author response image 3.

The TA muscles were collected from male and female P1 mice. The muscles were sectioned and co-stained for a-actinin (Actn) and Myh9. The majority of myofibrils is mature without the NM II signal. Scale bar, 10 µm.

5) Recent work suggested that mechanical tension is key to assemble the first long periodic myofibril containing immature sarcomeres. Tension is likely produced by a combination of NM and Mhc in the assembling sarcomeres themselves. This could be included in the introduction or discussion (PMIDs 24631244, 29316444, 29702642, 35920628).

Our response: We thank the reviewer for pointing to us additional relevant references. We have added them in the Introduction.

6) I suggest replacing "sarcomeric muscles" with "striated muscles".

Our response: We revised the term in the manuscript as suggested by the reviewer.

Author response:

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

Reviewer #1 (Public Review):

We appreciate the valuable and constructive comments of Reviewer #1 on our manuscript. We have addressed the comments from Reviewer #1 in the public review in the response to the recommendations for the authors, as the public review comments largely overlap with that of the recommendations for the authors.

Reviewer #1 (Recommendations For The Authors):

(1.1) Figure 1 did not use a mock-infected control for the development of R-loops but only a time before infection. I think it would have been a good control to have that after the same time of infection non-infected cells did not show increases in R-loops and this is not a product of the cell cycle.

We prepared our DRIPc-seq library using cell extracts harvested at 0, 3, 6, and 12 h post-infection (hpi), all at the same post-seeding time point. Each sample was infected with HIV-1 virus in a time-dependent manner. Therefore, it is unlikely that the host cellular R-loop induction observed in our DRIPc-seq results was due to R-loop formation during the cell cycle. In Lines 93–95 of the Results section of the revised manuscript, we have provided a more detailed description of our DRIPc-seq library experimental scheme. Thank you. 

(1.2) Figure 2 should have included a figure showing the proportion of DRIPc-seq peaks located in different genome features relative to one another instead of whether they were influenced by time post-infection. Figure 2C was performed in HeLa cells, but primary T cell data would have been more relevant as primary CD4+ T cells are more relevant to HIV infection.

We have included a new figure presenting the relative proportion of DRIPc-seq peaks mapped to different genomic features at each hpi (Fig. 2C of the revised manuscript). We found that the proportion of DRIPc-seq peaks mapped to various genomic compartments remained consistent over the hours following the HIV-1 infection. This further supports our original claim that HIV-1 infection does not induce R-loop enrichment at specific genomic features but that the accumulation of R-loops after HIV-1 infection is widely distributed.

We considered HeLa cells as the primary in vitro infection model, therefore, we conducted RNA-seq only on HeLa cells. However, we agree with the reviewer's opinion that data from primary CD4+ T cells may be more physiologically relevant. Nevertheless, as demonstrated in the new figure (Fig. 2C of the revised manuscript), HIV-1 infection did not significantly alter the proportion of R-loop peaks mapped to specific genomic compartments, such as gene body regions, in HeLa, primary CD4+ T, and Jurkat cells. Therefore, we anticipate no clear correlation between changes in gene expression levels and R-loop peak detection upon HIV-1 infection, even in primary T cells. Thank you.   

(1.3) Figure 5G is very hard to see when printed, is there a change in brightness or contrast that could be used? The arrows are helpful but they don't seem to be pointing to much.

We have highlighted the intensity of the PLA foci and magnified the images in Fig. 5G in the revised manuscript. While editing the images according to your suggestion, we found a misannotation regarding the multiplicity of infection in the number of PLA foci per nucleus quantification analysis graph in Fig. 5G of the original manuscript. We have corrected this issue and hope that it is now much clearer. 

(1.4) The introduction provided a good background for those who may not have a comprehensive understanding of DNA-RNA hybrids and R-loops, but the rationale that integration in non-expressed sequence implies that R-loops may be involved is very weak and was not addressed experimentally. A better rationale would have been to point out that, although integration in genes is strongly associated with gene expression, the association is not perfect, particularly in that some highly expressed genes are, nonetheless, poor integration targets.

In accordance with the reviewer's comment, we revised the Introduction. We have deleted the statement and reference in the introduction "... the most favored region of HIV-1 integration is an intergenic locus, ...”, which may overstate the relevance of the R-loop in HIV-1 integration events in non-expressed sequences. Instead, we introduced a more recent finding that high levels of gene expression do not always predict high levels of integration, together with the corresponding citation (Lines 46– 47 of the revised manuscript), according to the reviewer’s suggestion in the reviewer's public review 2)-(a).

(1.5) The discussion was seriously lacking in connecting their conclusions regarding R-loop targeting of integration to how integration works at the structural level, where it is very clear that concerted integration on the two DNA strands ca 5 bp apart is essential to correct, 2-ended integration. It is very difficult to visualize how this would be possible with the triple-stranded R-loop as a target. The manuscript would be greatly strengthened by an experiment showing concerted integration into a triplestranded structure in vitro using PICs or pure integrase.

We believe there has been a misunderstanding of our interpretation regarding the putative role of R-loop structures in the HIV-1 integration site mechanism because of some misleading statements in our original manuscript. Based primarily on our current data, we believe that R-loop structures are bound by HIV-1 integrase proteins and lead to HIV-1 viral genome integration into the vicinity regions of the host genomic R-loops. By carefully revising our manuscript, we found that the title, abstract, and discussion of our original manuscript includes phrases, such as “HIV-1 targets R-loops for integration,” which may overstate our finding on the role of R-loop in HIV-1 integration site selection. We replaced these phrases. For example, we used phrases, such as, “HIV-1 favors vicinity regions of R-loop for the viral genome integration,” in the revised manuscript. We apologize for the inconvenience caused by the unclear and nonspecific details of our findings.  

Using multiple biochemical experiments, we successfully demonstrated the interaction between the cellular R-loop and HIV-1 integrase proteins in cells and in vitro (Fig. 5 of the revised manuscript). However, we could not validate whether the center of the triple-stranded R-loops is the extraction site of HIV-1 integration, where the strand transfer reaction by integrase occurs. This is because an R-loop can be multi-kilobase in size (1, 2); therefore, we displayed a large-scale genomic region (30-kb windows) to present the integration sites surrounding the R-loop centers. Nevertheless, we believe that we validated R-loop-mediated HIV-1 integration in R-loop-forming regions using our pgR-poor and pgR-rich cell line models. When infected with HIV-1, pgR-rich cells, but not pgR-poor cells, showed higher infectivity upon R-loop induction in designated regions following DOX treatment (Fig. 3C and 3D of the revised manuscript). In addition, we quantified site-specific integration events in R-loop regions, and found that a greater number of integration events occurred in designated regions of the pgR-rich cellular genome upon R-loop induction by DOX treatment, but not in pgR-poor cells (Fig. 3E–G of the revised manuscript). 

We agree with the reviewer that an experiment showing the concerted integration of purified PICs into a triple-stranded structure in vitro would greatly strengthen our manuscript. We attempted the purification of viral DNA (vDNA)-bound PICs using either Sso7d-tagged HIV-1 integrase proteins or non-tagged HIV-1 integrase proteins (F185K/C280S) procured from the NIH HIV reagent program (HRP-20203), following the method described by Passos et al., Science, 2017; 355 (89-92) (3). Despite multiple attempts, we could not purify the nucleic acid-bound protein complexes for in vitro integration assays. However, we believe that pgR-poor and pgR-rich cell line models provide a strong advantage in specificity of our primer readouts. Compounded with our in cellulo observation, we believe that our work provides strong evidence for a causative relationship between R-loop formation/R-loop sites and HIV-1 integration.

Additionally, in the Discussion section of the revised manuscript, we have expanded our discussion on the role of genomic R-loops contributing in molding the host genomic environment for HIV-1 integration site selection, and the potential explanation on how R-loops are driving integration over long-range genomic regions. Thank you. 

(1.6) There are serious concerns with the quantitation of integration sites used here, which should be described in detail following line 503 but isn't. In Figure 3, E-G, they are apparently shown as reads per million, while in Figure 4B as "sites (%)" and in 4C as log10 integration frequency." Assuming the authors mean what they say, they are using the worst possible method for quantitation. Counting reads from restriction enzyme-digested, PCR-digested DNA can only mislead. At the numbers provided (MOI 0.6, 10 µg DNA assayed) there would be about 1 million proviruses in the samples assayed, so the probability of any specific site being used more than once is very low, and even less when one considers that a 10% assay efficiency is typical of integration site assays. Although the authors may obtain millions of reads per experiment, the number of reads per site is an irrelevant value, determined only by technical artefacts in the PCR reactions, most significantly the length of the amplicons, a function of the distance from the integration site to the nearest MstII site, further modified by differences in Tm. Better is to collapse identical reads to 1 per site, as may have been done in Figure 4B, however, the efficiency of integration site detection will still be inversely related to the length of the amplicon. Indeed, if the authors were to plot the read frequency against distance to the nearest MstII site, it is likely that they would get plots much like those in Figure 4B.

Detailed methods for integration site sequencing data processing are described in the Materials and Methods section of the revised manuscript (Line 621–631 of the revised manuscript). We primarily followed HIV-1 integration site sequencing data processing methods previously described by Li et al., mBio, 2020; 11(5) (4).  

While it may be correct that the HIV-1 integration event cannot occur more than once at a given site, our Fig. 3E, 4C, and 4D of the revised manuscript present the number of integration-site sequencing read counts expressed in reads-per-million (RPM) units or as log10-normalized values. Based on the number of mapped reads from the integration site sequencing results, we can infer that there was an integration event at this site, whether it was a single or multiple event.

We believe that the original annotation of y-axis, “Integration frequency,” may be misleading as it can be interpreted as a probability of any specific site being used for HIV-1 integration. Therefore, we corrected it as “number of mapped read” for clarity (Fig. 3E–G, 4C and 4D, and the corresponding figure legends of the revised manuscript). We apologize for any confusion. Thank you.

Other points:

(1.7) Overall: There are numerous grammatical and usage errors, especially in agreement of subject and verb, and missing articles, sometimes multiple times in the same sentence. These must be corrected prior to resubmission.

The revised manuscript was edited by a professional editing service. Thank you.

(1.8) Line 126-134: A striking result, but it needs more controls, as discussed above, including a dose-response analysis.

We determined the doses of NVP and RAL inhibitors in HeLa cells by optimizing the minimum dose of drug treatment that provided a sufficient inhibitory effect on HIV1 infection (Author response image 1). The primary objective of this experiment was to determine R-loop formation while reverse transcription or integration of the HIV-1 life cycle was blocked, therefore, we do not think that a dose-dependent analysis of inhibitors is required.

Author response image 1.

(A and B) Representative flow cytometry histograms of VSV-G-pseudotyped HIV-1-EGFP-infected HeLa cells at an MOI of 1, harvested at 48 hpi. The cells were treated with DMSO, the indicated doses of nevirapine (NVP) (A) or indicated doses of raltegravir (RAL) (B) for 24 h before infection. 

(1.9) Line 183: Please tell us what ECFP is and why it was chosen. Is there a reference for its failure to form R-loops?

Ibid: The human AIRN gene is a very poor target for HIV integration in PBMC.

A high GC skew value (> 0) is a predisposing factor for R-loop formation at the transcription site. This is because a high GC skew causes a newly synthesized RNA strand to hybridize to the template DNA strand, and the non-template DNA strand remains looped out in a single-stranded conformation (5) (Ref 36 in the revised manuscript). The ECFP sequence possessed a low GC skew value, as previously used for an R-loop-forming negative sequence (6) (Ref 17 of the revised manuscript). We have added this description and the corresponding references to Lines 188–192 of the revised manuscript.  

The human AIRN gene (RefSeq DNA sequence: NC_000006.12) sequence possesses a GC skew value of -0.04, in a window centered at base 2186, while the mouse AIRN (mAIRN) sequence is characterized by a GC skew value of 0.213. The ECFP sequence gave a GC skew value of -0.086 in our calculation. We anticipated that the human AIRN gene region does not form a stable R-loop, and in fact, it did not harbor R-loop enrichment upon HIV-1 infection in our DRIPc-seq data analysis of multiple cell types (Author response image 2)

Author response image 2.

Genome browser screenshot over the chromosomal regions in 20-kb windows centered on human AIRN showing results from DRIPc-seq in the indicated HIV-1-infected cells (blue, 0 hpi; yellow, 3 hpi; green, 6 hpi; red, 12 hpi)

(1.10) Line 190: You haven't shown dependence. Associated is a better word.

Thank you for the suggestion. We have changed “R-loop-dependent site-specific HIV-1 integration events...” to “R-loop-associated site-specific HIV-1 integration events...” (Line 198 of the revised manuscript) according to the reviewer’s suggestion in the revised manuscript. 

(1.11) Line 239: What happened to P1? What is the relationship of the P and N regions to genes?

We have added superimpositions of the P1 chromatin region on DRIPc-seq and the HIV-1 integration frequency to Figure 4C of the revised manuscript. We observed a relevant integration event within the P1 R-loop region, but to a lesser extent than in the P2 and P3 R-loop regions, perhaps because the P1 region has relatively less R-loop enrichment than the P2 and P3 regions, as examined by DRIP-qPCR in S3A Fig. of the revised manuscript.

Genome browser screenshots with annotations of accommodating genes in the P and N regions are shown in S2A–E Fig. of the revised manuscript, and RNA-seq analysis of the relative gene expression levels of the P1-3 and N1,2 R-loop regions are shown in S4 Table of the revised manuscript. Thank you.

(1.12) Line 261: But the binding affinity of integrase to the R-loop is somewhat weaker than to double-stranded DNA according to Figure 5A.

Nucleic acid substrates were loaded at the same molarity, and the percentage of the unbound fraction was calculated by dividing the intensity of the unbound fraction in each lane by the intensity of the unbound fraction in the lane with 0 nM integrase in the binding reaction. The calculated percentages of the unbound fraction from three independent replicate experiments are shown in Fig. 5A, right of the revised manuscript. In our analysis and measurements, the integrase proteins showed higher binding affinities to the R-loop and R-loop comprising nucleic acid structures than to dsDNA in vitro. We hope that this explanation clarifies this point. 

(1.13) Line 337: "accumulate". This is a not uncommon misinterpretation of the results of studies on the distribution of intact proviruses in elite controllers. The only possible correct interpretation of the finding is that proviruses form everywhere else but cells containing them are eliminated, most likely by the immune system.

Thank you for the suggestion. We have changed the Line 337 of the original manuscript to “... HIV-1 proviruses in heterochromatic regions are not eliminated but selected by immune system,” in Lines 361-363 of the revised manuscript. 

(1.14) Line 371 How many virus particles per cell does this inoculum amount to?

We determined the amount of GFP reporter viruses required to transduce ∼50% of WT Jurkat T cells, corresponding to an approximate MOI of 0.6. We repeatedly obtained 30–50% of VSV-G-pseudotyped HIV-1-EGFP positively infected cells for HIV1 integration site sequencing library construction for Jurkat T cells. 

(1.15) Line 503 and Figures 3 and 4: There must be a clear description of how integration events are quantitated.

Detailed methods for integration site sequencing data processing are described in the Materials and Methods section of the revised manuscript (Line 621–631 of the revised manuscript). We primarily followed HIV-1 integration site sequencing data processing methods previously described in Li et al., mBio, 2020; 11(5) (4).

Reviewer #2 (Public Review):

Retroviral integration in general, and HIV integration in particular, takes place in dsDNA, not in R-loops. Although HIV integration can occur in vitro on naked dsDNA, there is good evidence that, in an infected cell, integration occurs on DNA that is associated with nucleosomes. This review will be presented in two parts. First, a summary will be provided giving some of the reasons to be confident that integration occurs on dsDNA on nucleosomes. The second part will point out some of the obvious problems with the experimental data that are presented in the manuscript.

We appreciate your comments. We have carefully addressed the concerns expressed as follows (your comments are in italics):  

(2.1) 2017 Dos Passos Science paper describes the structure of the HIV intasome. The structure makes it clear that the target for integration is dsDNA, not an R-loop, and there are very good reasons to think that structure is physiologically relevant. For example, there is data from the Cherepanov, Engelman, and Lyumkis labs to show that the HIV intasome is quite similar in its overall structure and organization to the structures of the intasomes of other retroviruses. Importantly, these structures explain the way integration creates a small duplication of the host sequences at the integration site. How do the authors propose that an R-loop can replace the dsDNA that was seen in these intasome structures?

We do appreciate the current understanding of the HIV-1 integration site selection mechanism and the known structure of the dsDNA-bound intasome. Our study proposes an R-loop as another contributor to HIV-1 integration site selection. Recent studies providing new perspectives on HIV-1 integration site targeting motivated our current work. For instance, Ajoge et al., 2022 (7) indicated that a guanine-quadruplex (G4) structure formed in the non-template DNA strand of the R-loop influences HIV-1 integration site targeting. Additionally, I. K. Jozwik et al., 2022 (8) showed retroviral integrase protein structure bound to B-to-A transition in target DNA. R-loop structures are a prevalent class of alternative non-B DNA structures (9). We acknowledge the current understanding of HIV-1 integration site selection and explore how R-loop interactions may contribute to this knowledge in the Discussion section of our manuscript. 

Primarily based on our current data, we believe that R-loop structures are bound by HIV-1 integrase proteins and lead to HIV-1 viral genome integration into the vicinity regions of the host genomic R-loops, but we do not claim that R-loops completely replace dsDNA as the target for HIV-1 integration. An R-loop can be multi-kilobase in size and the R-loop peak length widely varies depending on the immunoprecipitation and library construction methods (1, 2), therefore, we could not validate whether the center of triple-stranded R-loops is the extraction site of HIV-1 integration where the strand transfer reaction by integrase occurs. Therefore, we replaced phrases such as, “HIV-1 targets R-loops for integration,” which may overstate our finding on the role of R-loop in HIV-1 integration site selection, with phrases, such as, “HIV-1 favors vicinity regions of R-loop for the viral genome integration,” in the revised manuscript. We apologize for the inconvenience caused by the unclear and non-specific details of our findings. Nevertheless, we believe that we validated R-loop-mediated HIV-1 integration in R-loop-forming regions using our pgR-poor and pgR-rich cell line models. We quantified site-specific integration events in the R-loop regions, and found that a greater number of integration events occurred in designated regions of the pgR-rich cellular genome upon R-loop induction by DOX treatment, but not in pgR-poor cells (Fig. 3E–G of the revised manuscript). 

dsDNA may have been the sole target of the intasome demonstrated in vitro possibly because dsDNA has only been considered as a substrate for in vitro intasome assembly. We hope that our work will initiate and advance future investigations on target-bound intasome structures by considering R-loops as potential new targets for integrated proteins and intasomes.  

(2.2) As noted above, concerted (two-ended) integration can occur in vitro on a naked dsDNA substrate. However, there is compelling evidence that, in cells, integration preferentially occurs on nucleosomes. Nucleosomes are not found in R loops. In an infected cell, the viral RNA genome of HIV is converted into DNA within the capsid/core which transits the nuclear pore before reverse transcription has been completed. Integration requires the uncoating of the capsid/core, which is linked to the completion of viral DNA synthesis in the nucleus. Two host factors are known to strongly influence integration site selection, CPSF6 and LEDGF. CPSF6 is involved in helping the capsid/core transit the nuclear pore and associate with nuclear speckles. LEDGF is involved in helping the preintegration complex (PIC) find an integration site after it has been released from the capsid/core, most commonly in the bodies of highly expressed genes. In the absence of an interaction of CPSF6 with the core, integration occurs primarily in the lamin-associated domains (LADs). Genes in LADs are usually not expressed or are expressed at low levels. Depending on the cell type, integration in the absence of CPSF6 can be less efficient than normal integration, but that could well be due to a lack of LEDGF (which is associated with expressed genes) in the LADs. In the absence of an interaction of IN with LEDGF (and in cells with low levels of HRP2) integration is less efficient and the obvious preference for integration in highly expressed genes is reduced. Importantly, LEDGF is known to bind histone marks, and will therefore be preferentially associated with nucleosomes, not R-loops. LEDGF fusions, in which the chromatin binding portion of the protein is replaced, can be used to redirect where HIV integrates, and that technique has been used to map the locations of proteins on chromatin. Importantly, LEDGF fusions in which the chromatin binding component of LEDGF is replaced with a module that recognizes specific histone marks direct integration to those marks, confirming integration occurs efficiently on nucleosomes in cells. It is worth noting that it is possible to redirect integration to portions of the host genome that are poorly expressed, which, when taken with the data on integration into LADs (integration in the absence of a CPSF6 interaction) shows that there are circumstances in which there is reasonably efficient integration of HIV DNA in portions of the genome in which there are few if any R-loops.

Although R-loops may not wrap around nucleosomes, long and stable R-loops likely cover stretches of DNA corresponding to multiple nucleosomes (10). For example, R-loops are associated with high levels of histone marks, such as H3K36me3, which LEDGF recognizes (2, 11). R-loops dynamically regulate the chromatin architecture. Possibly by altering nucleosome occupancy, positioning, or turnover, R-loop structures relieve superhelical stress and are often associated with open chromatin marks and active enhancers (2, 10). These features are also distributed over HIV-1 integration sites (12). In the Discussion section of the revised manuscript, we explored the R-loop molding mechanisms in the host genomic environment for HIV-1 integration site selection and its potential collaborative role with LEDGF/p75 and CPSF6 governing HIV-1 integration site selection. 

By carefully revising our original manuscript, with respect to the reviewer's comment, we recognized the need to tone down our statements. We found that the title, abstract, and discussion of our original manuscript includes phrases, such as, “HIV-1 targets Rloops for integration,” which may overstate our finding on the role of R-loop in HIV-1 integration site selection. We replaced these phrases. For example, we used phrases, such as “HIV-1 favors vicinity regions of R-loop for the viral genome integration,” in the revised manuscript. We apologize for the inconvenience caused by the unclear and non-specific details of our findings.

(2.3) Given that HIV DNA is known to preferentially integrate into expressed genes and that R-loops must necessarily involve expressed RNA, it is not surprising that there is a correlation between HIV integration and regions of the genome to which R loops have been mapped. However, it is important to remember that correlation does not necessarily imply causation.

We understand the reviewer's concern regarding the possibility of a coincidental correlation between the R-loop regions and HIV-1 integration sites, particularly when the interpretation of this correlation is primarily based on a global analysis. 

Therefore, we designed pgR-poor and pgR-rich cell lines, which we believe are suitable models for distinguishing between integration events driven by transcription and the presence of R-loops. Although the two cell lines showed comparable levels of transcription at the designated region upon DOX treatment via TRE promoter activation (Fig. 3B of the revised manuscript), only pgR-rich cells formed R-loops at the designated regions (Fig. 3C of the revised manuscript). When infected with HIV1, pgR-rich cells, but not pgR-poor cells, showed higher infectivity after DOX treatment (Fig. 3D of the revised manuscript). Moreover, we quantified site-specific integration events in the R-loop regions, and found that a greater number of integration events occurred in designated regions of the pgR-rich cellular genome upon R-loop induction by DOX treatment, but not in pgR-poor cells (Fig. 3E of the revised manuscript). Therefore, we concluded that transcriptional activation without an R-loop (in pgR-poor cells) may not be sufficient to drive HIV-1 integration. We believe that our work provides strong evidence for a causative relationship between R-loop formation/Rloop sites and HIV-1 integration. We hope that our explanation addresses your concerns. Thank you.

If we consider some of the problems in the experiments that are described in the manuscript:

(2.4) In an infected individual, cells are almost always infected by a single virion and the infecting virion is not accompanied by large numbers of damaged or defective virions. This is a key consideration: the claim that infection by HIV affects R-loop formation in cells was done with a VSVg vector in experiments in which there appears to have been about 6000 virions per cell. Although most of the virions prepared in vitro are defective in some way, that does not mean that a large fraction of the defective virions cannot fuse with cells. In normal in vivo infections, HIV has evolved in ways that avoid signaling infected the cell of its presence. To cite an example, carrying out reverse transcription in the capsid/core prevents the host cell from detecting (free) viral DNA in the cytoplasm. The fact that the large effect on R-loop formation which the authors report still occurs in infections done in the absence of reverse transcription strengthens the probability that the effects are due to the massive amounts of virions present, and perhaps to the presence of VSVg, which is quite toxic. To have physiological relevance, the infections would need to be carried out with virions that contain HIV even under circumstances in which there is at most one virion per cell.

Our virus production and in vitro and ex vivo HIV-1 infection experimental conditions, designed for infecting cell types, such as HeLa cells and primary CD4+ T cells with VSV-G pseudotyped HIV, were based on a comprehensive review of numerous references. At the very beginning of this study, we tested HIV-1-specific host genomic R-loop induction using empty virion particles (virus-like particles, VLP) or other types of viruses (non-retrovirus, SeV; retroviruses, FMLV and FIV), all produced with a VSV G protein donor. We could not include a control omitting the VSV G protein or using natural HIV-1 envelope protein to prevent viral spread in culture. We observed that despite all types of virus stocks being prepared using VSV-G, only cells infected with HIV-1 viruses showed R-loop signal enrichment (Author response image 3). Therefore, we omitted the control for the VSV G protein in subsequent analyses, such as DRIPcseq. We have also revised our manuscript to provide a clearer description of the experimental conditions. In particular, we now clearly stated that we used VSV-G pseudotyped HIV-1 in this study, throughout the abstract, results, and discussion sections of the revised manuscript. Thank you.

Author response image 3.

(A) Dot blot analysis of the R-loop in gDNA extracts from HIV-1 infected U2OS cells with MOI of 0.6 harvested at 6 hpi. The gDNA extracts were incubated with or without RNase H in vitro before membrane loading (anti-S9.6 signal). (B) Dot blot analysis of the R-loop in gDNA extracts from HeLa cells infected with 0.3 MOI of indicated viruses. The infected cells were harvested at 6 hpi. The gDNA extracts were incubated with or without RNase H in vitro before membrane loading (anti-S9.6 signal).

HIV-1 co-infection may also be expected in cell-free HIV-1 infections. However, it was previously suggested that the average number of infection events varies within 1.02 to 1.65 based on a mathematical model that estimates the frequency of multiple infections with the same virus (Figure 4c of Ito et al., Sci. Rep, 2017; 6559) (13). 

(2.5) Using the Sso7d version of HIV IN in the in vitro binding assays raises some questions, but that is not the real question/problem. The real problem is that the important question is not what/how HIV IN protein binds to, but where/how an intasome binds. An intasome is formed from a combination of IN bound to the ends of viral DNA. In the absence of viral DNA ends, IN does not have the same structure/organization as it has in an intasome. Moreover, HIV IN (even Sso7d, which was modified to improve its behavior) is notoriously sticky and hard to work with. If viral DNA had been included in the experiment, intasomes would need to be prepared and purified for a proper binding experiment. To make matters worse, there are multiple forms of multimeric HIV IN and it is not clear how many HIV INs are present in the PICs that actually carry out integration in an infected cell.

As the reviewer has noted, HIV IN, even with Sso7d tagging, is difficult. We attempted the purification of viral DNA (vDNA)-bound PICs using either Sso7d-tagged HIV-1 integrase proteins or non-tagged HIV-1 integrase proteins (F185K/C280S), procured from the NIH HIV reagent program (HRP-20203), following the method described by Passos et al., Science, 2017; 355 (89-92) (3). Despite multiple attempts, we were unable to purify the vDNA-bound IN protein complexes for in vitro assays. However, through multiple biochemical experiments, we believe that we have successfully demonstrated the interaction between cellular R-loops and HIV-1 integrase proteins both in cells and in vitro (Fig. 5A–F of the revised manuscript). We also observed a close association between integrase proteins and host cellular Rloops in HIV-1-infected cells, using a fluorescent recombinant virus (HIV-IN-EGFP) with intact IN-EGFP PICs (Fig. 5G of the revised manuscript). 

(2.6) As an extension of comment 2, the proper association of an HIV intasome/PIC with the host genome requires LEDGF and the appropriate nucleic acid targets need to be chromatinized.

The interaction between cellular R-loops and HIV-1 integrase proteins in HeLa cells endogenously expressing LEDGF/p75 was examined using reciprocal immunoprecipitation assays in Fig. 5C–F, S6B, and S6D Fig. of the revised manuscript. In addition, as discussed in more detail in our response to comment [28], we observed a close association between host cellular R-loops and HIV-1 integrase proteins by PLA assay, in HIV-1-infected HeLa cells. 

(2.7) Expressing any form of IN, by itself, in cells to look for what IN associates with is not a valid experiment. A major factor that helps to determine both where integration takes place and the sites chosen for integration is the transport of the viral DNA and IN into the nucleus in the capsid core. However, even if we ignore that important part of the problem, the IN that the authors expressed in HeLa cells won't be bound to the viral DNA ends (see comment 2), even if the fusion protein would be able to form an intasome. As such, the IN that is expressed free in cells will not form a proper intasome/PIC and cannot be expected to bind where/how an intasome/PIC would bind.

As discussed in more detail in our response to comment [2-8], we believe that our PLA experiment using the pVpr-IN-EGFP virus, which has previously been examined for virion integrity, as well as the IN-EGFP PICs (14), demonstrated a close association between host cellular R-loops and HIV-1 integrase proteins in HIV-1-infected cells. 

(2.8) As in comment 1, for the PLA experiments presented in Figure 5 to work, the number of virions used per cell (which differs from the MOI measured by the number of cells that express a viral marker) must have a high, which is likely to have affected the cells and the results of the experiment. However, there is the additional question of whether the IN-GFP fusion is functional. The fact that the functional intasome is a complex multimer suggests that this could be a problem. There is an additional problem, even if IN-GFP is fully functional. During a normal infection, the capsid core will have delivered copies of IN (and, in the experiments reported here, the IN-GFP fusion) into the nucleus that is not part of the intasome. These "free" copies of IN (here IN-GFP) are not likely to go to the same sites as an intasome, making this experiment problematic (comment 4).

The HIV-IN-EGFP virus stock was produced by polyethylenimine-mediated transfection of HEK293T cells with 6 µg of pVpr-IN-EGFP, 6 µg of HIV-1 NL4-3 noninfectious molecular clone (pD64E; NIH AIDS Reagent Program 10180), and 1 µg of pVSV-G as previously described in (14), and described in the Materials and Methods section of our manuscript. The pVpr-IN-EGFP vector used to produce HIV-1-IN-EGFP virus stock was provided by Anna Cereseto group (Albanese et al., PLOS ONE, 2008; 6(6); Ref 34 of the revised manuscript). It was previously reported that the HIV-1INEGFP virions produced by IN-EGFP trans-incorporation through Vpr are intact and infective viral particles (Figure 1 of Albanese et al., PLOS ONE, 2008; 6(6)). Therefore, we believe that the HIV-IN-EGFP used in our PLA experiments was functional. 

Additionally, Albanese et al. showed that the EGFP signal of HIV-IN-EGFP virions colocalizes with the viral protein matrix (p17MA) and capsid (P24CA) as well as with the newly synthesized cDNA produced by reverse transcriptase by labeling and visualizing the synthesized cDNA (14). In addition, the fluorescent recombinant virus (HIV-INEGFP) was structurally intact at the nuclear level (Figure 6 of Albanese et al., PLOS ONE, 2008; 6(6)). Therefore, we believe that our PLA experimental result is not likely misled as the reviewer concerns due to the integrity of the HIV-IN-EGFP virion as well as IN-EGFP PICs.

Furthermore, the in vitro HIV-1 infection setting of our PLA experiments was carefully determined based on multiple studies that performed image-based assays on HIV-1infected cells. For instance, Albanese et al. infected 4 × 104 cells with viral loads equivalent to 1.5 or 3 µg of HIV-1 p24 for their immunofluorescence analysis, in their previous report (14). We titrated the fluorescent HIV-1 virus stocks by examining both the multiplicity of infection (MOI) and quantifying the HIV-1 p24 antigen content (Author response image 4). In our calculation, we infected 5 × 104 HeLa cells with viral loads equivalent to 1.3 ug of HIV-1 p24, which is indicated as 2 MOI in Fig. 5G of our manuscript, for our PLA experiments. 

Image-Based Assays often require increased and enhanced signal for statistical robustness. For example, Achuthan et al. infected cells with VSV-G-pseudotyped HIV1 at the approximate MOI of 350 for vDNA and PIC visualization (15). Therefore, we believe our experimental condition for PLA experiments, which we carefully designed based on previous study that are frequently referred, are reasonable. We really hope that our discussion sufficiently addressed the reviewer’s concern. 

Author response image 4.

Gating strategy used to determine HIV-1-infectivity in HeLa cells at 48 hpi. Cells were infected with a known p24 antigen content in the stock of the VSV-G-pseudotyped HIV-1-EGFP-virus. The percentages of GFP-positive cell population are indicated.

(2.9) In the Introduction, the authors state that the site of integration affects the probability that the resulting provirus will be expressed. Although this idea is widely believed in the field, the actual data supporting it are, at best, weak. See, for example, the data from the Bushman lab showing that the distribution of integration sites is the same in cells in which the integrated proviruses are, and are not, expressed. However, given what the authors claim in the introduction, they should be more careful in interpreting enzyme expression levels (luciferase) as a measure of integration efficiency in experiments in which they claim proviruses are integrated in different places.

We thank the reviewer for the constructive comment. We have changed the statement in Lines 41–42 in the Introduction section of our original manuscript to “The chromosomal landscape of HIV-1 integration influences proviral gene expression, persistence of integrated proviruses, and prognosis of antiretroviral therapy.” (Lines 39-41 of the revised manuscript). We believe that this change can tone-down the relevance between the site of integration and the provirus expression level.

The piggyBac transposase randomly insert the “cargo (transposon)” into TTAA chromosomal sites of the target genome, generating efficient insertions at different genomic loci (16, 17). We believe that this random insertion of the pgR-poor/rich vector mediated by the piggyBac system allows us not to mislead the R-loop-mediated HIV1 integration site because of the genome locus bias of the vector insertion. Therefore, Figure 3 in our manuscript does not claim any relevance between the site of integration and the resulting provirus expression levels. Instead, as noted in Line 214 of the revised manuscript, using the luciferase reporter HIV-1 virus, we attempted to examine HIV-1 infection in cells with an "extra number of R-loops” in the host cellular genome. We observed that pgR-rich cells showed higher luciferase activity upon DOX treatment than pgR-poor cells (Fig. 3D of the revised manuscript). We believe that this is because a greater number of HIV-1 integration events may occur in pgR-rich cells, where DOX-inducible de novo R-loop regions are introduced. This has been further examined in Fig. 3E–G of the revised manuscript. We hope this explanation clarifies the Figure 3. Thank you. 

(2.10) Using restriction enzymes to create an integration site library introduces biases that derive from the uneven distribution of the recognition sites for the restriction enzymes.

As described in the Materials and Methods section, we adopted a sequencing library construction method using a previously established protocol (18, 19). Although we recognize the advantages of DNA fragmentation by sonication, in in vitro or ex vivo HIV-1 infection settings, where the multiplicity of infection is carefully determined based on multiple references, more copies of integrated viral sequences are expected compared to that in samples from infected patients (18). Therefore, in these settings, restriction enzyme-based DNA fragmentation and ligation-mediated PCR sequencing are well-established methods that provide significant data sources for HIV-1 integration site sequencing (15, 20-22). Furthermore, our data showing the proportion of integration sites over R-loop regions (Fig. 4B of the revised manuscript) are presented alongside the respective random controls (i.e., proportion of integration sites within the 30-kb windows centered on randomized DRIPc-seq peaks, gray dotted lines; control comparisons between randomized integration sites with DRIPc-seq peaks, black dotted lines; and randomized integration sites with randomized DRIPcseq peaks, gray solid lines), which do not show such a correlation between the HIV-1 integration sites and nearby areas of the R-loop regions. Therefore, we believe that our results from the integration site sequencing data analysis are unlikely to be biased. 

Reviewer #3 (Public Review):

In this manuscript, Park and colleagues describe a series of experiments that investigate the role of R-loops in HIV-1 genome integration. The authors show that during HIV-1 infection, R-loops levels on the host genome accumulate. Using a synthetic R-loop prone gene construct, they show that HIV-1 integration sites target sites with high R-loop levels. They further show that integration sites on the endogenous host genome are correlated with sites prone to R-loops. Using biochemical approaches, as well as in vivo co-IP and proximity ligation experiments, the authors show that HIV-1 integrase physically interacts with R-loop structures.

My primary concern with the paper is with the interpretations the authors make about their genome-wide analyses. I think that including some additional analyses of the genome-wide data, as well as some textual changes can help make these interpretations more congruent with what the data demonstrate. Here are a few specific comments and questions:

We are grateful for the time and effort we spent on our behalf and the reviewer’s appreciation for the novelty of our work, in particular, R-loop induction by HIV-1 infection and the correlation between host R-loops and the genomic site of HIV-1 integration. In the following sections, we provide our responses to your comments and suggestions. Your comments are in italics. We have carefully addressed the following issues.

(3.1) I think Figure 1 makes a good case for the conclusion that R-loops are more easily detected HIV-1 infected cells by multiple approaches (all using the S9.6 antibody). The authors show that their signals are RNase H sensitive, which is a critical control. For the DRIPc-Seq, I think including an analysis of biological replicates would greatly strengthen the manuscript. The authors state in the methods that the DRIPc pulldown experiments were done in biological replicates for each condition. Are the increases in DRIPc peaks similar across biological replicates? Are genomic locations of HIV-1-dependent peaks similar across biological replicates? Measuring and reporting the biological variation between replicate experiments is crucial for making conclusions about increases in R-loop peak frequency. This is partially alleviated by the locus-specific data in Figure S3A. However, a better understanding of how the genome-wide data varies across biological replicates will greatly enhance the quality of Figure 1.

DRIPc-seq experiments were conducted with two biological replicates. To define consensus DRIPc-seq peaks using these two replicates, we used two methods applicable to ChIP-seq analysis: the irreproducible discovery rate (IDR) method and sequencing data pooling. We found that the sequencing data pooling method yielded significantly more DRIPc-seq peaks than consensus peak identification through IDR, and we decided to utilize R-loop peaks from pooled sequencing data for our downstream analyses, as described in the figure legends and Materials and Methods of the revised manuscript. 

As noted by the reviewer, it is important to verify whether the increasing trend in the number of R-loop peaks and genomic locations of HIV-1 dependent R-loops were consistently observed across the two biological replicates. Therefore, we independently performed R-loop calling on each replicate of the sequencing data of primary CD4+ T cells from two individual donors to verify that the increase in R-loop numbers was consistent (Author response image 5). Additionally, the overlap of the R-loop peaks between the two replicates was statistically significant across the genome (Author response table 1). Thank you.

Author response image 5.

Bar graph indicating DRIPc-seq peak counts for HIV-1-infected primary CD4+ T cells harvested at the indicated hours post infection (hpi). Pre-immunoprecipitated samples were untreated (−) or treated (+) with RNase H, as indicated. Each dot corresponds to an individual data set from two biologically independent experiments.

Author response table 1.

DRIPc-seq peak length and Chi-square p-value in CD4+ T cells from individual donor 1 and 2 

(3.2) I think that the conclusion that R-loops "accumulate" in infected cells is acceptable, given the data presented. However, in line 134 the authors state that "HIV1 infection induced host genomic R-loop formation". I suggest being very specific about the observation. Accumulation can happen by (a) inducing a higher frequency of the occurrence of individual R-loops and/or (b) stabilizing existing R-loops. I'm not convinced the authors present enough evidence to claim one over the other. It is altogether possible that HIV-1 infection stabilizes R-loops such that they are more persistent (perhaps by interactions with integrase?), and therefore more easily detected. I think rephrasing the conclusions to include this possibility would alleviate my concerns.

We thank the reviewer for the considerable discussion on our manuscript. We have now changed Line 134 to, “HIV-1 infection induces host genomic R-loop enrichment” (Lines 132-133 of the revised manuscript), and added a new conclusion sentence implicating the possible explanation for the R-loop signal enrichment upon HIV-1 infection (Lines 133–135 of the revised manuscript), according to the reviewer's suggestion.    

(3.3) A technical problem with using the S9.6 antibody for the detection of R-loops via microscopy is that it cross-reacts with double-stranded RNA. This has been addressed by the work of Chedin and colleagues (as well as others). It is absolutely essential to treat these samples with an RNA:RNA hybrid-specific RNase, which the authors did not include, as far as their methods section states. Therefore, it is difficult to interpret all of the immunofluorescence experiments that depend on S9.6 binding.

We understand the reviewer's concern regarding the cross-reactivity of the S9.6 antibody with more abundant dsRNA, particularly in imaging applications. We carefully designed the experimental and analytical methods for R-loop detection using microscopy. For example, we pre-extracted the cytoplasmic fraction before staining with the S9.6 antibody and quantified the R-loop signal by subtracting the nucleolar signal. Both of these steps were taken to eliminate the possibility of misdetecting Rloops via microscopy because of the prominent cytoplasmic and nucleolar S9.6 signals, which primarily originate from ribosomal RNA. In addition, we included R-loop negative control samples in our microscopy analysis that were subjected to intensive RNase H treatment (60U/mL RNase H for 36 h) and observed a significant reduction in the S9.6 signal (Figure 1E of the revised manuscript). RNase H-treated samples served as essential and widely accepted negative controls for R-loop detection. 

We would like to point out that recent studies have reported strong intrinsic specificity of S9.6 anybody for DNA:RNA hybrid duplex over dsDNA and dsRNA, along with the structural elucidations of S9.6 antibody recognition of hybrids (23, 24). Therefore, our interpretation of host cellular R-loop enrichment after HIV-1 infection using S9.6 antibodies in multiple biochemical approaches is well supported. Nevertheless, we agree with the reviewer's opinion that additional negative controls for the detection of R-loops via microscopy, such as RNase T1-and RNase III-treated samples, could improve the robustness and accuracy of R-loop imaging data (25).  

(3.4) Given that there is no clear correlation between expression levels and R-loop peak detection, combined with the data that show increased detection of R-loop frequency in non-genic regions, I think it will be important to show that the R-loop forming regions are indeed transcribed above background levels. This will help alleviate possible concerns that there are technical errors in R-loop peak detection.

Figures S5D and S5E in the revised manuscript show the relative gene expression levels of the R-loop-forming positive regions (P1-3) and the referenced Rloop-positive loci (RPL13A and CALM3). The gene expression levels of these R-loopforming regions were significantly higher than those of the ECFP or mAIRN genes without DOX treatment, which can be considered background levels of transcription in cells. Thank you. 

(3.5) In Figures 4C and D the hashed lines are not defined. It is also interesting that the integration sites do not line up with R-loop peaks. This does not necessarily directly refute the conclusions (especially given the scale of the genomic region displayed), but should be addressed in the manuscript. Additionally, it would greatly improve Figure 4 to have some idea about the biological variation across replicates of the data presented 4A.

We thank the reviewer for the considerable comment on our study. First of all, we added an annotation for the dashed lines in the figure legends of Figures 4C and 4D in the revised manuscript.

We agree with the reviewer's interpretation of the relationship between the integration sites and R-loop peaks. Primarily based on our current data, we believe R-loop structures are bound by HIV-1 integrase proteins and lead HIV-1 viral genome integration into the “vicinity” regions of the host genomic R-loops. We displayed a large-scale genomic region (30-kb windows) to present integration sites surrounding R-loop centers because an R-loop can be multi-kilobase in size (1, 2). Depending on the immunoprecipitation and library construction methods, the R-loop peaks varied in size, and the peak length showed a wide distribution (Figure 3B of Malig et al., 2020, Figure 1B of Sanz et al., 2016, and Figure 2A of the revised manuscript). Therefore, presenting integration site events within a wide window of R-loop peaks could be more informative and better reflect the current understanding of R-loop biology.

R-loop formation recruits diverse chromatin-binding protein factors, such as H3K4me1, p300, CTCF, RAD21, and ZNF143 (Figure 6A and 6B of Sanz et al., 2016) (26), which allow R-loops to exhibit enhancer and insulator chromatin states, which can act as distal regulatory elements (26, 27). We have demonstrated physical interactions between host cellular R-loops and HIV-1 integrase proteins (Figure 5 of the revised manuscript), therefore, we believe that this ‘distal regulatory element-like feature’ of the R-loop can be a potential explanation for how R-loops drive integration over longrange genomic regions.

According to your suggestion, we added this explanation to the relevant literature in the Discussion section of the revised manuscript.

Author response image 6 which represents the biological variation across replicates of the data shown in Figure 4A. The integration site sequencing data for Jurkat cells were adopted from SRR12322252 (4), which consists of the integration site sequencing data of HIV-1-infected wild type Jurkat cells with one biological replicate. We hope that our explanations and discussion have successfully addressed your concerns. Thank you. 

Author response image 6.

Bar graphs showing the quantified number of HIV-1 integration sites per Mb pair in total regions of 30-kb windows centered on DRIPc-seq peaks from HIV-1 infected HeLa cells and primary CD4+ T cells (magenta) or non-R-loop region in the cellular genome (gray). Each dot corresponds to an individual data set from two biologically independent experiments.

(3.6) The authors do not adequately describe the Integrase mutant that they use in their biochemical experiments in Figure 5A. Could this impact the activity of the protein in such a way that interferes with the interpretation of the experiment? The mutant is not used in subsequent experiments for Figure 5 and so even though the data are consistent with each other (and the conclusion that Integrase interacts with R-loops) a more thorough explanation of why that mutant was used and how it impacts the biochemical activity of the protein will help the interpretation of the data presented in Figure 5.

We appreciate the reviewer’s suggestions. In our EMSA analysis, we purified and used Sso7d-tagged HIV-1 integrase proteins with an active-site amino acid substitution, E152Q. First, we used the Sso7d-tagged HIV-1 integrase protein, as it has been suggested in previous studies that the fusion of small domains, such as Sso7d (DNA binding domain) can significantly improve the solubility of HIV integrase proteins without affecting their ability to assemble with substrate nucleic acids and their enzymatic activity (Figure 1B of Li et al., PLOS ONE, 2014;9 (8) (28, 29). We used an integrase protein with an active site amino acid substitution, E152Q, in our mobility shift assay, because the primary goal of this experiment was to examine the ability of the protein to bind or form a complex with different nucleic acid substrates. We thought that abolishing the enzymatic activity of the integrase protein, such as 3'-processing that cleaves DNA substrates, would be more appropriate for our experimental objective. This Sso7d tagged- HIV-1 integrase with the E152Q mutation has also been used to elucidate the structural model of the integrase complex with a nucleic acid substrate by cryo-EM (3) and has been shown to not disturb substrate binding.   Based on the reviewer’s comments, we have added a description of the E152Q mutant integrase protein in Lines 268–270 of the revised manuscript. Thank you.

Reviewer #3 (Recommendations For The Authors):

The paper suffers from many grammatical errors, which sometimes interfere with the interpretations of the experiments. In the view of this reviewer, the manuscript must be carefully revised prior to publication. For example, lines 247-248 "Intasomes consist of HIV-1 viral cDNA and HIV-1 coding protein, integrases." It is unclear from this sentence whether there are multiple integrases or multiple proteins that interact with the viral genome to facilitate integration. This makes the subsequent experiments in Figure 5 difficult to interpret. There are many other examples, too numerous to point out individually.

We thoughtfully revised the original manuscript, making the best efforts to provide clearer details of our findings. We believe that we have made substantial changes to the manuscript, including Lines 247–248 of the original manuscript that the reviewer noted. Furthermore, the revised manuscript was edited by a professional editing service. Thank you.     (1) M. Malig, S. R. Hartono, J. M. Giafaglione, L. A. Sanz, F. Chedin, Ultra-deep Coverage Singlemolecule R-loop Footprinting Reveals Principles of R-loop Formation. J Mol Biol 432, 22712288 (2020).

(2) L. A. Sanz et al., Prevalent, Dynamic, and Conserved R-Loop Structures Associate with Specific Epigenomic Signatures in Mammals. Mol Cell 63, 167-178 (2016).

(3) D. O. Passos et al., Cryo-EM structures and atomic model of the HIV-1 strand transfer complex intasome. Science 355, 89-92 (2017).

(4) W. Li et al., CPSF6-Dependent Targeting of Speckle-Associated Domains Distinguishes Primate from Nonprimate Lentiviral Integration. mBio 11,  (2020).

(5) P. A. Ginno, Y. W. Lim, P. L. Lott, I. Korf, F. Chedin, GC skew at the 5' and 3' ends of human genes links R-loop formation to epigenetic regulation and transcription termination. Genome Res 23, 1590-1600 (2013).

(6) S. Hamperl, M. J. Bocek, J. C. Saldivar, T. Swigut, K. A. Cimprich, Transcription-Replication Conflict Orientation Modulates R-Loop Levels and Activates Distinct DNA Damage Responses. Cell 170, 774-786 e719 (2017).

(7) H. O. Ajoge et al., G-Quadruplex DNA and Other Non-Canonical B-Form DNA Motifs Influence Productive and Latent HIV-1 Integration and Reactivation Potential. Viruses 14,  (2022).

(8) I. K. Jozwik et al., B-to-A transition in target DNA during retroviral integration. Nucleic Acids Res 50, 8898-8918 (2022).

(9) F. Chedin, C. J. Benham, Emerging roles for R-loop structures in the management of topological stress. J Biol Chem 295, 4684-4695 (2020).

(10) F. Chedin, Nascent Connections: R-Loops and Chromatin Patterning. Trends Genet 32, 828838 (2016).

(11) P. B. Chen, H. V. Chen, D. Acharya, O. J. Rando, T. G. Fazzio, R loops regulate promoterproximal chromatin architecture and cellular differentiation. Nat Struct Mol Biol 22, 9991007 (2015).

(12) A. R. Schroder et al., HIV-1 integration in the human genome favors active genes and local hotspots. Cell 110, 521-529 (2002).

(13) Y. Ito et al., Number of infection events per cell during HIV-1 cell-free infection. Sci Rep 7, 6559 (2017).

(14) A. Albanese, D. Arosio, M. Terreni, A. Cereseto, HIV-1 pre-integration complexes selectively target decondensed chromatin in the nuclear periphery. PLoS One 3, e2413 (2008).

(15) V. Achuthan et al., Capsid-CPSF6 Interaction Licenses Nuclear HIV-1 Trafficking to Sites of Viral DNA Integration. Cell Host Microbe 24, 392-404 e398 (2018).

(16) X. Li et al., piggyBac transposase tools for genome engineering. Proc Natl Acad Sci U S A 110, E2279-2287 (2013).

(17) Y. Cao et al., Identification of piggyBac-mediated insertions in Plasmodium berghei by next generation sequencing. Malar J 12, 287 (2013).

(18) E. Serrao, P. Cherepanov, A. N. Engelman, Amplification, Next-generation Sequencing, and Genomic DNA Mapping of Retroviral Integration Sites. J Vis Exp,  (2016).

(19) K. A. Matreyek et al., Host and viral determinants for MxB restriction of HIV-1 infection. Retrovirology 11, 90 (2014).

(20) G. A. Sowd et al., A critical role for alternative polyadenylation factor CPSF6 in targeting HIV-1 integration to transcriptionally active chromatin. Proc Natl Acad Sci U S A 113, E10541063 (2016).

(21) B. Lucic et al., Spatially clustered loci with multiple enhancers are frequent targets of HIV-1 integration. Nat Commun 10, 4059 (2019).

(22) P. K. Singh, G. J. Bedwell, A. N. Engelman, Spatial and Genomic Correlates of HIV-1 Integration Site Targeting. Cells 11,  (2022).

(23) C. Bou-Nader, A. Bothra, D. N. Garboczi, S. H. Leppla, J. Zhang, Structural basis of R-loop recognition by the S9.6 monoclonal antibody. Nat Commun 13, 1641 (2022).

(24) Q. Li et al., Cryo-EM structure of R-loop monoclonal antibody S9.6 in recognizing RNA:DNA hybrids. J Genet Genomics 49, 677-680 (2022).

(25) J. A. Smolka, L. A. Sanz, S. R. Hartono, F. Chedin, Recognition of RNA by the S9.6 antibody creates pervasive artifacts when imaging RNA:DNA hybrids. J Cell Biol 220,  (2021).

(26) L. A. Sanz, F. Chedin, High-resolution, strand-specific R-loop mapping via S9.6-based DNARNA immunoprecipitation and high-throughput sequencing. Nat Protoc 14, 1734-1755 (2019).

(27) M. Merkenschlager, D. T. Odom, CTCF and cohesin: linking gene regulatory elements with their targets. Cell 152, 1285-1297 (2013).

(28) M. Li, K. A. Jurado, S. Lin, A. Engelman, R. Craigie, Engineered hyperactive integrase for concerted HIV-1 DNA integration. PLoS One 9, e105078 (2014).

(29) M. Li et al., A Peptide Derived from Lens Epithelium-Derived Growth Factor Stimulates HIV1 DNA Integration and Facilitates Intasome Structural Studies. J Mol Biol 432, 2055-2066 (2020).

Author Response

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

General remarks for the Editor and the Reviewers

We would like to thank the Editor and the Reviewers for their feedback. Below we address their comments and present our point-by-point responses as well as the related changes in the manuscript.

In addition to these changes, in a few cases we have found it necessary to move some texts and provide some additional explanations within the manuscript. We emphasize that these amendments have been made for only technical reasons, and do not alter the results and conclusions of the paper, but may help to render the text more coherent and understandable to readers with little knowledge of the subject.

These minor corrections are:

• We extended the Introduction section by a sentence (lines 40-42) that is intended to fit the proposed template directed, non-enzymatic replication mechanism into a more general prebiotic evolutionary context, thus emphasizing its biological relevance. This sentence includes an additional reference (Rosenberger et al., 2021).

• Two very methodologically oriented and repeated descriptions of random sequence generation have been moved to the Methods section (lines 178-185) from the Results section (lines 336-339 and lines 351-354).

• We complemented the Data availability statement with licensing information (lines 684-685).

• Further minor changes (also indicated by red texts) have been implemented to remedy logical and grammatical glitches.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

Szathmary and colleagues explore the parabolic growth regime of replicator evolution. Parabolic growth occurs when nucleic acid strain separation is the rate-limiting step of the replication process which would have been the case for non-enzymatic replication of short oligonucleotide that could precede the emergence of ribozyme polymerases and helicases. The key result is that parabolic replication is conducive to the maintenance of genetic diversity, that is, the coexistence of numerous master sequences (the Gause principle does not apply). Another important finding is that there is no error threshold for parabolic replication except for the extreme case of zero fidelity.

Strengths:

I find both the analytic and the numerical results to be quite convincing and well-described. The results of this work are potentially important because they reveal aspects of a realistic evolutionary scenario for the origin of replicators.

Weaknesses:

There are no obvious technical weaknesses. It can be argued that the results represent an incremental advance because many aspects of parabolic replication have been explored previously (the relevant publications are properly cited). Obviously, the work is purely theoretical, experimental study of parabolic replication is due. In the opinion of this reviewer, though, these are understandable limitations that do not actually detract from the value of this work.

We are grateful that this Reviewer appreciates our work. We completely agree that the ultimate validation must come from experiments. It is important to stress that in this field theory often preceded experimental work by decades, and the former often guided the latter. We hope that for the topic of the present paper experiments will follow considerably faster.

Reviewer #2 (Public Review):

Summary:

A dominant hypothesis concerning the origin of life is that, before the appearance of the first enzymes, RNA replicated non-enzymatically by templating. However, this replication was probably not very efficient, due to the propensity of single strands to bind to each other, thus inhibiting template replication. This phenomenon, known as product inhibition, has been shown to lead to parabolic growth instead of exponential growth. Previous works have shown that this situation limits competition between alternative replicators and therefore promotes RNA population diversity. The present work examines this scenario in a model of RNA replication, taking into account finite population size, mutations, and differences in GC content. The main results are (1) confirmation that parabolic growth promotes diversity, but that when the population size is small enough, sequences least efficient at replicating may nevertheless go extinct; (2) the observation that fitness is not only controlled by the replicability of sequences, but also by their GC content; (3) the observation that parabolic growth attenuates the impact of mutations and, in particular, that the error threshold to which exponentially growing sequences are subject can be exceeded, enabling sequence identity to be maintained at higher mutation rates.

Strengths:

The analyses are sound and the observations are intriguing. Indeed, it has been noted previously that parabolic growth promotes coexistence, its role in mitigating the error threshold catastrophe - which is often presented as a major obstacle to our understanding of the origin of life - had not been examined before.

Weaknesses:

Although all the conclusions are interesting, most are not very surprising for people familiar with the literature. As the authors point out, parabolic growth is well known to promote diversity (SzathmaryGladkih 89) and it has also been noted previously that a form of Darwinian selection can be found at small population sizes (Davis 2000).

Given that under parabolic growth, no sequence is ever excluded for infinite populations, it is also not surprising to find that mutations have a less dramatic exclusionary impact.

In the two articles cited (Szathmary-Gladkih 1989 and Davis 2000) the subexponentiality of the system was implemented in a mechanistic way, by introducing the exponent 0 < 𝑝 < 1. Although the behaviour of these models is more or less consistent with experimental findings (von Kiedrowski, 1986; Zielinski and Orgel, 1987), the divergence of per capita growth rates (𝑥̇/𝑥) at very low concentrations–which guarantees the ability to maintain unlimited diversity in the case of infinite population sizes–makes this formal approach partly unrealistic.

To avoid the possible artefacts of this mechanistic approach, and as there are no previous studies analysing the diversity maintaining ability of finite populations of parabolic replicators in an individual-based model context, we implemented a simplified template replication mechanism leading to parabolic growth and analysed the dynamics in an individual-based stochastic model context. The key point of our investigation is that considerable diversity can be maintained in the system even when the population size is quite small.

Regarding the Reviewer’s comment on selection: Darwinian selection can only occur in a simple subexponential dynamics if the ratio of replicabilities diverges, cf. Eq. (8) and the preceding paragraph in Davis, 2000.

Our results also show (Figs. 4B and 4C) that high mutation rates and the error threshold problem can still be considered as a major limiting factor for parabolically replicating systems in terms of their diversity-maintaining ability. In the light of the above, potential mechanisms to relax the error threshold in such systems, one of which is demonstrated in the present study, seem to be important steps to account for the sequence diversification and increase in molecular complexity during the early evolution of RNA replicators.

A general weakness is the presentation of models and parameters, whose choices often appear arbitrary. Modeling choices that would deserve to be further discussed include the association of the monomers with the strands and the ensuing polymerization, which are combined into a single association/polymerization reaction (see also below), or the choice to restrict to oligomers of length L = 10. Other models, similar to the one employed here, have been proposed that do not make these assumptions, e.g. Rosenberger et al. Self-Assembly of Informational Polymers by Templated Ligation, PRX 2021. To understand how such assumptions affect the results, it would be helpful to present the model from the perspective of existing models.

The assumption of one-step polymerization reactions that we used here is a common technique for modelling template replication of sequence-represented replicators [see, e.g., Fontana and Schuster, 1998 (10.1126/science.280.5368.1451), Könnyű et al., 2008 (10.1186/1471-2148-8267), Vig-Milkovics et al, 2019 (10.1016/j.jtbi.2018.11.020) or Szilágyi et al., 2020 (10.1371/journal.pgen.1009155)]. This is because assuming base-to-base polymerisation of the copy would lead to a very large number of different types of intermediates, which a Gillespietype stochastic simulation algorithm could not handle in reasonable computation times, even if the sequences were relatively short. For comparison, in our model, where polymerization is one-step, the characteristic time of a simulation for 𝐿 = 10, 𝑁 = 105 and 𝛿 = 0.01 was 552 hours.

Note that in Rosenberg et al. (PRX 2021), in contrast to a pioneering work [Fernando et al, 2007 (10.1007/s00239-006-0218-4)], sequences of replicators are not represented, which makes this approach completely inapplicable to our case, in which sequence defines the fitness. In sum, we suggest that this valid criticism points to possible future work.

The values of the (many) parameters, often very specific, also very often lack justifications. For example, why is the "predefined error factor" ε = 0.2 and not lower or higher? How would that affect the results?

A general remark. For the more important parameters , several values were used to test the behaviour of the model (see Table 1), but due to the considerable number of parameters, it is impossible to examine all possible combinations. 𝑐+ = 1 fixes the timescale, 𝐿 is set to 10 to obtain reasonable running times (see above).

𝜀 characterizes how replicability decreases as the number of mutations increases. In the manuscript we used the following default vector: 𝜀 = (0.05, 0.2, 1) in which the third element corresponds to the mutation-free sequence, so it must to be 1. The first element determines the baseline replicability (see Methods), which we preferred not to change because it would fundamentally alter the ratio of replication propensities to association and dissociation propensities (as the substantial amount of complementary sequences of the master sequences are of baseline replicability) and thus would alter the reaction kinetics to an extent that it is not comparable with the original results. Therefore, only the second element can be adjusted. Accordingly, we have analysed the behaviour of the model in the cases of a steeper and a more gradual loss of replicability using the following two vectors, respectively: 𝜀, = (0.05, 𝟎. 𝟎𝟓, 1) and 𝜀,, = (0.05, 𝟎. 𝟓, 1). The choice of 𝜀, is chemically more plausible, since for very short oligomers the loss of chemical activity and replicability as a function of the number of mutations can be very sharp. We performed a series of simulations with all possible combinations of 𝛿 = 0.001, 0.005, 0.1 and 𝑁 = 103, 104, 105 for 𝜀′ and 𝜀,,in the constant population and chemostat model context (36 different runs). For other parameters, we took the default values, see Table 1. These values also correspond to the parameters we used in Figures 2 and 6. The results show that the steeper loss of replicability (𝜀,) slightly increases the diversity maintaining ability of the system, whereas the more gradual loss of replicability (𝜀,,) moderately decreases the diversity-maintaining ability of the system, and that these shifts are more pronounced in the constant population size model (Author response image 1) than in the chemostat model (Author response image 2). Altogether, these results confirm that the qualitative outcome of the model is robust in a wide range of loss of replicability (𝜀 vector) values.

Author response image 1.

Replicator coexistence in the constant population model with different loss of replicability (𝜀 vector) values. Within a given combination of 𝛿 and 𝑁 parameter values, the upper panel corresponds to the steeper loss of replicability (𝜀!), the middle panel to the default 𝜀 vector (Figure 2A), and the bottom panel to the more gradual loss of replicability vector (𝜀!!). Within each 𝛿; 𝑁 parameter combination, the same master sequence set was used with the three different 𝜀 vectors for comparability.

Author response image 2.

Replicator coexistence in the chemostat model with different loss of replicability (𝜀 vector) values. Within a given combination of 𝛿 and 𝑁 parameter values, the upper panel corresponds to the steeper loss of replicability (𝜀!), the middle panel to the default 𝜀 vector (Figure 6A), and the bottom panel to the more gradual loss of replicability vector (𝜀!!). Within each 𝛿; 𝑁 parameter combination, the same master sequence set was used with the three different 𝜀 vectors for comparability.

Similarly, in equation (11), where does the factor 0.8 come from?

This factor scales the decay rate of duplex sequences (𝑐"!") as the function of the binding energy

(𝐸b). The value of 0.8 is an arbitrary choice, the value should be in the interval (0,1) and is only relevant in the chemostat model. It is expected to have a similar effect on the dynamics as the duplex decay factor parameter 𝑓, which we have investigated in a wide range of different values (cf. Table 1, Fig. 6), although 𝑓 is independent of the binding energy (𝐸/): increasing/decreasing the 0.8 factor is expected to decrease/increase the average total population size. We have investigated the diversity maintaining ability of the system at smaller (0.6) and larger (0.9) parameter values at different population sizes (𝑁 ≈ 103, 104 and 105) and at different replicability distances (δ = 0.001, 0.005 and 0.01) as shown in Fig. 6. We have found that the number of coexisting master types changes very little in response to changes in this factor. Only two shifts could be detected (underlined): factor 0.9 combined with 𝑁 ≈ 104 and 𝛿 = 0.001 caused the number of surviving master types to decrease by one, while factor 0.9 combined with 𝑁 ≈ 103 and 𝛿 = 0.01 caused the number of surviving master types to increase by one (Author response table 1). Factor 0.6 produced the same number of surviving types as the default (Author response table 1). In summary, the model shows marked robustness to changes in the values of this parameter.

Author response table 1.

Number of coexisting master types in the chemostat model with different binding energy dependent duplex decay rates. Within each 𝛿; 𝑁 parameter combination, the same master sequence set was used with the three different factor values: 0.6, 0.8 (the original) and 0.9 for comparability.

Why is the kinetic constant for duplex decay reaction 1.15e10−8?

Note that this value is the minimum of the duplex decay rate, Table 1 correctly shows the interval of this kinetic constant as: [1.15 ⋅ 10-8, 6.4 ⋅ 10-5]. Both values are derived from the basic parameters of the system and can be computed according to Eq. (11). The minimum: as the parameter set corresponding to this value is: . The maximum: with .

Are those values related to experiments, or are they chosen because specific behaviors can happen only then?

See above.

The choice of the model and parameters potentially impact the two main results, the attenuation of the error threshold and the role of GC content:

Regarding the error threshold, it is also noted (lines 379-385) that it disappears when back mutations are taken into account. This suggests that overcoming the error threshold might not be as difficult as suggested, and can be achieved in several ways, which calls into question the importance of the particular role of parabolic growth. Besides, when the concentration of replicators is low, product inhibition may be negligible, such that a "parabolic replicator" is effectively growing exponentially and an error catastrophe may occur. Do the authors think that this consideration could affect their conclusion? Can simulations be performed?

The assumption of back mutation only provides a theoretical solution to the error threshold problem: back mutation guarantees a positive (non-zero) concentration of a master type, but, since the probability of back mutation is generally very low, this equilibrium concentration may be extremely low, or negligible for typical system sizes. Consequently, back mutation alone does not solve the problem of the error catastrophe: in our system back mutation is present (the probability that a sequence with 𝑘 errors mutates back to a master sequence is 𝜇k(1−𝜇)L-k), and the diversity-maintaining ability is limited. The effect of back mutation decreases exponentially with increasing sequence length.

Regarding the role of the GC content, GC-rich oligomers are found to perform the worst but no rationale is provided.

For GC-rich oligonucleotides the dissociation probability of a template-copy complex is relatively low (cf. Eqs. (9, 10)), thus they have a relatively low number of offspring, cf. lines 557-561: “a relatively high dissociation probability and the consequential higher propensity of being in a simple stranded form provides an advantage for sequences with relatively low GC content in terms of their replication affinity, that is, the expected number of offspring in case of such variants will be relatively high.”. Note that the simulation results shown in Fig. 3A, demonstrate the realization of this effect with prepared sequences (along a GC content gradient).

One may assume that it happens because GC-rich sequences are comparatively longer to release the product. However, it is also conceivable that higher GC content may help in the polymerization of the monomers as the monomers attach longer on the template (as described in Eq. (9)). This is an instance where the choice to pull into a single step the association and polymerization reactions are pulled into a single step independent of GC content may be critical.

It would be important to show that the result arises from the actual physics and not from this modeling choice.

Some more specific points that would deserve to be addressed:

• Line 53: it is said that p "reflects how easily the template-reaction product complex dissociates". This statement is not correct. A reaction order p<1 reflects product inhibition, the propensity of templates to bind to each other, not slow product release. Product release can be limiting, yet a reaction order of 1 can be achieved if substrate concentrations are sufficiently high relative to oligomer concentrations (von Kiedrowski et al., 1991).

We think the key reference is Von Kiedrowski (1993) in this case. Other things being equal, his Table 1 on p. 134 shows that a sufficient increase in 𝐾4, i.e., the stability of the duplex (template and copy) (association rate divided by dissociation rate) throws the system into the parabolic regime. This is what we had in mind. In order to clarify this, we modified the quoted sentence thus: “In this kinetics, the growth order is equal or close to 0.5 (i.e., the dynamics is sub-exponential) because increased stability of the template-copy complex (rate of association divided by dissociation) promotes parabolic growth (von Kiedrowski et al., 1991; von Kiedrowski & Szathmáry, 2001).”

• Population size is a key parameter, and a comparison is made between small (10^3) and large (10^5) populations, but without explaining what determines the scale (small/large relative to what?).

The “small” value (103) corresponds to the smallest meaningful population size, significantly smaller population sizes (e.g. 102) cannot maintain the 10 master types (or any subset of them) and are chemically unrealistic. The “large value” (105) is the largest population size for which simulation times are still acceptable, in the case of 106 the runtimes are in the order of months.

• In the same vein, we might expect size not to be the only important parameter, but also concentration.

With constant volume population size and concentration are strictly coupled.

• Lines 543-546: if understanding correctly, the quantitative result is that the error threshold rises from 0.1 in the exponential case to 0.196 in the parabolic. Are the authors suggesting that a factor of 2 is a significant difference?

In this paragraph we compared the empirical error threshold of our system (which is close to 𝑝"#$ = 0.15) with the error threshold of the well-known single peak fitness landscape (which can be approximated by ) as a reference case. To make the message even clearer we have extended the last sentence (lines 596-597) as follows: “but note that applying this approach to our system is a serious oversimplification”. The 0.196 is simply the probability of error-free replication of a sequence when , but we have removed this sentence (“corresponding to the replication accuracy of a master sequence”) from the manuscript as it seems to be confusing.

• Figure 3C: this figure shows no statistically significant effect?

Thank you for pointing out this. We statistically tested the hypothesis that the GC content between the survived and the extinct master subsets are different. This analysis revealed that the differences between these two groups are statistically significant, which we now included in the manuscript at lines 380-390: “A direct investigation of whether the sequence composition of the master types is associated with their survival outcome was conducted using the data from the constant population model simulation results (Figure 2). In these data, the average GC content was measured to be lower in the surviving master subpopulations than in the extinct subpopulations (Figure 3C). To determine whether this difference was statistically significant, nonparametric, two-sample Wilcoxon rank-sum tests (Hollander & Wolfe, 1999) were performed on the GC content of the extinct-surviving master subsets. The GC content was significantly different between these two groups in all nine investigated parameter combinations of population size (N) and replicability distance (δ) at p<0.05 level, indicating a selective advantage for a lower GC content in the constant population model context. The exact p values obtained from this analysis are shown in Figure 3C.”

• line 542: "phase transition-like species extension (Figure 4B)": such a clear threshold is not apparent.

Thank you for pointing out the incorrect phrasing. As there is no clear threshold in the number of coexisting types as a function of the mutation rate, we removed the “phase transition-like” expression: “However, when finite population sizes and stochastic effects are taken into account, at the largest investigated per-base mutation rate (𝑝mut = 0.15), the summed relative steady-state master frequencies approach zero (Figure 4C) with accelerating species extinction (Figure 4B), indicating that this value is close to the system׳s empirical error threshold.” (lines 589-594).

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

On the whole, the work is well done and presented, there are no major recommendations. It seems a good idea to cite and briefly discuss this recent paper: https://pubmed.ncbi.nlm.nih.gov/36996101/ which develops a symbiotic scenario of the coevolution of primordial replicators and reproducers that appears to be fully compatible with the results of the current work.

Thank you for bringing this article to our attention. We have inserted the following sentence at lines 621-624: “The demonstrated diversity-maintaining mechanism of finite parabolic populations can be used as a plug-in model to investigate the coevolution of naked and encapsulated molecular replicators (e.g., Babajanyan et al., 2023).”

The manuscript is well written, but there are some minor glitches that merit attention. For example:

l. 5 "carriers presents a problem, because product formation and mutual hybridization" - "mutual" is superfluous here, delete

l. 13 "amplification. In addition, sequence effects (GC content) and the strength of resource" - hardly "effects" - should be 'features' or 'properties'

l. 41 "If enzyme-free replication of oligomer modules with a high degree of sequence" - "modules" here is only confusing - simply, "oligomers"

l. 44 "under ecological competition conditions with which distinct replicator types with different" - delete "with" etc, there are many such minor glitches that are best corrected.

Thank you for pointing out, we have corrected! Other drafting errors, glitches, superfluous sentences have also been corrected.

Reviewer #2 (Recommendations For The Authors):

None

Editor (Recommendations For The Authors):

In the manuscript, it appears that coexistence is assessed at a given point in time, while figures seem to show that it remains time-dependent. It would be great if the authors could clarify this and/or discuss this.

We appreciate you bringing this to our attention, as we have indeed missed to elaborate on this important point. The steady state characteristic of the coexistence is assessed in our model in the following way: the relative frequency of each master sequence is tested for the condition of ≥ 100- (cut-off relative frequency for survival) in every 2,000th replication step in the interval between 10,000 replication steps before termination and actual termination (10= replication steps). If the above condition is true more than once, we consider the master type in question as survived (we have included this explanation in the Methods section: lines 258-268). Although this relatively narrow time interval can still be regarded as a snapshot of the state of the system, according to our numerical experiences, the resulting measure is a reliable quantitative indicator of the apparent stability of species coexistence in the parabolic dynamics.

Author Response

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

Responses to reviewers’ comments

(1) The rationale of selecting tNOX/ENOX2 as a potential target of 4-dmH, but not heliomycin, is unclear by taking a biased approach. Thus, there is high possibility that 4-dmH binds to other proteins involved in apoptosis inhibition. An unbiased screen to identify 4-dmH-binding proteins would be a better approach unless there is a clear and logical rationale.

We apologize for this oversight. In response to this comment, we rewrote the abstract, reorganized the results, and added more references to better introduce tNOX/ENOX2.

A) Under the “4-dmH, but not heliomycin, targets intracellular tNOX, an upstream regulator of SIRT1” result section:

We next addressed the molecular mechanisms underlying SIRT1 inhibition and concurrent cell death by these two compounds in oral cancer cells. Being an NAD+-dependent protein deacetylase, SIRT1 activity is primarily governed by NAD+/NADH ratio, thus, there exists a positive correlation between these two [1-9]. We then questioned whether these two compounds inhibit SIRT1 by affecting the intracellular NAD+/NADH levels, and were surprised to find that 4-dmH, but not heliomycin, caused a prominent inhibition of intracellular NAD+/NADH ratio (revised Fig. 7a). The discrepancy in their ability to reduce NAD+ generation led us to explore the role of a tumor-associated NADH oxidase (tNOX, ENOX2) in 4-dmH-suppressed SIRT1 and apoptosis induction. We have previously reported that tNOX inhibition reduced the intracellular NAD+/NADH ratio and SIRT1 deacetylase activity, increasing p53 acetylation and apoptosis [10-13]. In the light of this information, we assessed the effect of the compounds on tNOX expression and found that 4-dmH, but not heliomycin, considerably diminished tNOX protein expression in a concentration-dependent manner (Fig. 7b).

B) To demonstrate that our results from ligand-binding assays (CETSA) were specific to tNOX, we conducted more CETSA experiments to exclude PARP or NOX4 targets of 4-dmH. PARP acts as a DNA damage sensor and also a NAD+-consuming enzyme, affecting many cellular functions [14]. NOX4 belongs to the NOX family of NADPH oxidases that mediate electron transport through intracellular membranes and is also shown to be involved in tumorigenesis [15, 16]. We show that 4-dmH treatments did not seem to increase the melting temperature of PARP or NOX4, excluding those two proteins as potential targets of 4-dmH (revised Fig. 8c).

Author response image 1.

(2) The authors should show whether heliomycin indeed does not induce apoptosis, while 4-dmH cannot induce autophagy.

We have reorganized and revised our manuscript and figures (Fig. 5 and Fig. 6) to better demonstrate the different cell death pathways associated with heliomycin and 4-dmH. Using flow cytometry, we show that heliomycin, but not 4-dmH, induced autophagy in two lines of oral cancer cells (Fig. 5a). In the revision, we moved up the analysis of apoptosis by JC-1 staining to Figure 5 (revised Fig. 5b). We also reorganized the protein analysis to demonstrate better the downregulation of pro-apoptotic Bak and Puma and a lack of caspase 3-directed PARP cleavage, indicating the ineffective apoptosis by heliomycin (revised Fig. 5c). Similarly, we found that the absence of upregulation of ULK1, Atg 5, Atg7, and cleaved LC3-II provided evidence for the inadequate autophagy by 4-dmH (revised Fig. 5d). Attached please see the revised Figure 5.

Author response image 2.

(3) They should demonstrate whether genetic knockdown of tNOX, SirT1, or both tNOX and SirT1 induces apoptosis or autophagy and also reduces malignant properties of oral cancer cells.

A) In the revision, we conducted more experiments utilizing the RNAi-knockdown to understand the role of tNOX on the regulation of apoptosis or autophagy. Our results indicate that the tNOX-depletion effectively provoked spontaneous apoptosis and autophagy in SAS cells (revised Fig. 7e). However, given that SIRT1 per se is not the focus of this present study and SIRT1-knockdown has been shown to increase apoptotic population by other groups [17] [18], we decided not to pursue it further.

Author response image 3.

B) In our earlier studies, we have adequately demonstrated that tNOX confers a survival advantage for cancer cells. For example, tNOX-deficiency by RNA interference in cancer cells abolishes cancer phenotypes, reducing NAD+ production, proliferation, and migration/invasion while increasing apoptosis [19-22]. On the other hand, tNOX-overexpressing in non-cancerous cells stimulates the growth of cells, decreases doubling time, and enhances cell migration [23-26].

(4) The authors should examine whether overexpression of SirT1 or tNOX in cells treated with heliomycin or 4-dmH could nullify heliomycin-induced autophagy and 4-dmH-induced apoptosis. Also, instead of overexpressing tNOX, they can supplement NAD into cells treated with 4-dmH.

A) The utilization of tNOX overexpression has been previously reported in several studies, demonstrating that tNOX-overexpressing in non-cancerous cells stimulates the growth of cells, decreases doubling time, and enhances cell migration [23-26]. However, in our experiences, the effect of tNOX overexpression in cancer cells is much less apparent than that in non-cancerous cells. Thus, we decided not to study it further, given that our results from tNOX knockdown have evidently signified the role of tNOX in the regulation of apoptosis and autophagy.

B) Since SIRT1 is not the major focus of this present study and SIRT1-overexpression has been shown to reduce stress-mediated apoptosis by other groups [27, 28], we decided not to pursue it further.

C) The systemic deterioration in NAD+ level has been correlated with many diseases and aging [29-31]. In this regard, NAD+ administration was reported to attenuate doxorubicin-induced apoptosis in the liver of mice, suggesting a protective effect [32]. The administration of nicotinamide riboside (NR), a precursor of NAD+, was also demonstrated to prevent ROS generation and apoptosis in the mouse sepsis models [33]. With data from these animal studies already demonstrating the benefits of NAD+ supplements, we decided not to conduct similar experiments in a cell-based setting.

(5) Related to Fig. 5C and 6a, the authors should examine the effects of heliomycin and 4-dmH on the cell cycle profiles, Annexin V positivity, and colony formation.

We added the results from colony-forming assays and revealed that both compounds exhibited high growth-suppressive ability against oral cancer cells (revised Fig. 6c). Nevertheless, we showed that the diminution in growth by the compounds was least likely to arise from cell cycle arrest mediated by these two compounds (revised Fig. 6d). Due to the possible interference of the fluorescence wavelength of heliomycin/derivative, we examined JC-1 staining rather than Annexin V positivity. The apoptotic effect of the compounds was demonstrated in revised Fig. 5b in the revision.

Author response image 4.

(6) They should also examine whether either or both heliomycin and 4-dmH induce reactive oxygen species (ROS).

In our previous report, we examined the effects of heliomycin and 4-dmH on oxidative stress utilizing H2DCFDA [34]. The dye fluoresces in the presence of intracellularly generated reactive oxygen species (ROS). We showed that 4-dmH significantly induced the generation of ROS generation. However, no marked ROS generation was observed in cells exposed to heliomycin.

(7) Related to Fig. 9d, they should mutate amino acid residue(s) in tNOX that are crucial for the 4-dmH-tNOX binding, including Ile 90, Lys98, Pro111, Pro113, Leu115, Pro117, and Pro118, to examine whether these mutants lose the binding to 4-dmH and fail to rescue 4-dmH-induced apoptosis, unlike wild-type tNOX.

For further evaluation of the importance of the consistent interaction residues in the three docked compound-tNOX complexes, the seven interaction residues on tNOX were substituted with alanine or glycine amino acids and then simulated the protein structures. The simulated protein structures appear slightly different from the original tNOX structure. Overall, the root mean square difference between the original tNOX structure and the structures with residues substituted by alanine or glycine amino acids was estimated at 3.339 or 4.024 angstroms (Å), respectively (Fig. S1a). The simulated protein structures were also employed to conduct the docking analysis for 4-dmH. The results of further docking analysis revealed that 4-dmH could bind within the same pocket of different types of tNOX structures but with varying orientations (Fig. S1b). This observation also suggests that the replacement of both key residues with alanine or glycine could result in a reduction of the binding affinity of 4-dmH to tNOX, with values of -8.2 and -7.6 kcal/mol, respectively. Moreover, the substitution of both key residues with alanine or glycine also reduces the number of the original interacting residues and interaction forces in core moieties in the 4-dmH-tNOX complexes (Fig. S1c and S1d). Together, our experimental results and molecular docking simulations are consistent with the notion that 4-dmH possesses a better affinity ability for tNOX than for SIRT1.

Author response image 5.

The simulated tNOX structures (a, b) and the binding modes of 4-dmH after docking study (c, d). (a) Superimposition of three types of tNOX structures, including the original tNOX structure (orange) and the critical residues in tNOX protein substituted with alanine (magenta) or glycine (cyan). The substituted residues were shown as sticks. (b) Superimposition of the docked 4-dmH (blue). (c) Schematic presentations of possible interactions between 4-dmH and the interacted residues in tNOX protein substituted with alanine. (d) Schematic presentations of possible interactions between 4-dmH and the interacted residues in tNOX protein substituted with glycine. The key residues were identified based on the best docking pose of 4-dmH. The red circles and ellipses indicate the identical residues that interacted with different types of tNOX structures.

(8) Related to Fig. 10a, heliomycin appears to also reduce tNOX levels (although the extent is not as robust as 4-dmH), which is not expected since heliomycin does not bind to tNOX. They should compare the effects of heliomycin and 4-dmH on reducing the protein levels of tNOX. If heliomycin does not change the tNOX protein levels, then they need to discuss why heliomycin reduces tNOX levels in vivo.

In our previous studies, we have shown that tNOX knockdown partially attenuates SIRT1 expression and represses growth in various cancer cell types, such as lung [22], bladder [20], and stomach [13]. We also observed that tNOX is acetylated/ubiquitinated under certain stresses and SIRT1 depletion affects tNOX expression (data not shown). It is speculated that SIRT1 deacetylates tNOX and modulates its protein stability. Thus, there is a reciprocal regulation between tNOX and SIRT1. Although heliomycin does not bind to tNOX, its inhibition of SIRT1 activity/expression might also have an impact on tNOX expression.

(9) Related to Fig. 10F, if tNOX is an upstream regulator of SirT1 and both heliomycin and 4-dmH ultimately target SirT1, it is unclear why heliomycin and 4-dmH cause different biological outcomes. One explanation is that tNOX has apoptosis-inhibiting function other than supporting (or independent of) SirT1 and hence 4-dmH-mediated tNOX inhibition causes apoptosis rather than autophagy. They should explain and discuss more about whether tNOX-inhibiting/binding function of 4-dmH is sufficient to explain the different biological outcomes from heliomycin.

Thank you for this valuable suggestion. Indeed, in our earlier studies, we have adequately demonstrated that tNOX-deficiency by RNA interference in cancer cells abolishes cancer phenotypes, reducing NAD+ production, proliferation, and migration/invasion while increasing apoptosis; thus, tNOX confers a survival advantage for cancer cells [19-22]. On the other hand, tNOX-overexpressing in non-cancerous cells stimulates the growth of cells, decreases doubling time, and enhances cell migration [23-26]. With these lines of evidence, we believe that tNOX not only supports but also exerts functions independent of SIRT1. The tNOX- and SIRT1-inhibiting function of 4-dmH, thus, results in the different biological outcomes from the SIRT1-binding heliomycin.

(10) They should examine the effects of heliomycin and 4-dmH on cell viability of non-tumor cells to examine their toxicities.

Using cell impedance measurements, we also examined the effects of heliomycin and 4-dmH on the proliferation of human non-cancerous BEAS-2B cells. Our results demonstrated that heliomycin did not exhibit cytotoxicity toward human non-cancerous BEAS-2B cells (revised Fig. 6a). Furthermore, the water-soluble 4-dmH effectively diminished cell proliferation in a dose-dependent manner in oral cancer cells, but much less apparent in that of BEAS-2B cells (revised Fig. 6b). Similar results were reported in our previous study, indicating that 4-dmH displayed much higher IC50 values against non-cancerous human dermal microvascular endothelium HMEC-1 cells compared to those of tumor cells [34].

Author response image 6.

(11) They should consistently use either tNOX or ENOX2 to avoid confusion.

Thank you for the suggestion. We have now consistently used tNOX throughout the manuscript. However, for the revised Figure 7d, the commercially available antibody to ENOX2 (from Proteintech, Rosemont, IL, USA) is different from the one to tNOX (produced in our laboratory) and this is the only place we have used ENOX2 rather than tNOX.

References

1) Mouchiroud L, Houtkooper RH, Moullan N, Katsyuba E, Ryu D, Canto C, Mottis A, Jo YS, Viswanathan M, Schoonjans K et al: The NAD(+)/Sirtuin Pathway Modulates Longevity through Activation of Mitochondrial UPR and FOXO Signaling. Cell 2013, 154(2):430-441.

2) He S, Gao Q, Wu X, Shi J, Zhang Y, Yang J, Li X, Du S, Zhang Y, Yu J: NAD(+) ameliorates endotoxin-induced acute kidney injury in a sirtuin1-dependent manner via GSK-3beta/Nrf2 signalling pathway. J Cell Mol Med 2022, 26(7):1979-1993.

3) Donmez G: The neurobiology of sirtuins and their role in neurodegeneration. Trends Pharmacol Sci 2012, 33(9):494-501.

4) Teertam SK, Phanithi PB: Up-regulation of Sirtuin-1/autophagy signaling in human cerebral ischemia: possible role in caspase-3 mediated apoptosis. Heliyon 2022, 8(12):e12278.

5) Li BY, Peng WQ, Liu Y, Guo L, Tang QQ: HIGD1A links SIRT1 activity to adipose browning by inhibiting the ROS/DNA damage pathway. Cell reports 2023, 42(7):112731.

6) Bai P, Canto C, Oudart H, Brunyanszki A, Cen Y, Thomas C, Yamamoto H, Huber A, Kiss B, Houtkooper RH et al: PARP-1 inhibition increases mitochondrial metabolism through SIRT1 activation. Cell Metab 2011, 13(4):461-468.

7) Ma Y, Nie H, Chen H, Li J, Hong Y, Wang B, Wang C, Zhang J, Cao W, Zhang M et al: NAD(+)/NADH metabolism and NAD(+)-dependent enzymes in cell death and ischemic brain injury: current advances and therapeutic implications. Curr Med Chem 2015, 22(10):1239-1247.

8) Fulco M, Schiltz RL, Iezzi S, King MT, Zhao P, Kashiwaya Y, Hoffman E, Veech RL, Sartorelli V: Sir2 regulates skeletal muscle differentiation as a potential sensor of the redox state. Mol Cell 2003, 12(1):51-62.

9) Yang Y, Liu Y, Wang Y, Chao Y, Zhang J, Jia Y, Tie J, Hu D: Regulation of SIRT1 and Its Roles in Inflammation. Front Immunol 2022, 13:831168.

10) Tikhomirov AS, Shchekotikhin AE, Lee YH, Chen YA, Yeh CA, Tatarskiy VV, Jr., Dezhenkova LG, Glazunova VA, Balzarini J, Shtil AA et al: Synthesis and Characterization of 4,11-Diaminoanthra[2,3-b]furan-5,10-diones: Tumor Cell Apoptosis through tNOX-Modulated NAD(+)/NADH Ratio and SIRT1. Journal of medicinal chemistry 2015, 58(24):9522-9534.

11) Chang CF, Islam A, Liu PF, Zhan JH, Chueh PJ: Capsaicin acts through tNOX (ENOX2) to induce autophagic apoptosis in p53-mutated HSC-3 cells but autophagy in p53-functional SAS oral cancer cells. Am J Cancer Res 2020, 10(10):3230-3247.

12) Lin CY, Islam A, Su CJ, Tikhomirov AS, Shchekotikhin AE, Chuang SM, Chueh PJ, Chen YL: Engagement with tNOX (ENOX2) to Inhibit SIRT1 and Activate p53-Dependent and -Independent Apoptotic Pathways by Novel 4,11-Diaminoanthra[2,3-b]furan-5,10-diones in Hepatocellular Carcinoma Cells. Cancers (Basel) 2019, 11(3).

13) Chen HY, Cheng HL, Lee YH, Yuan TM, Chen SW, Lin YY, Chueh PJ: Tumor-associated NADH oxidase (tNOX)-NAD+-sirtuin 1 axis contributes to oxaliplatin-induced apoptosis of gastric cancer cells. Oncotarget 2017, 8(9):15338-15348.

14) Xu Q, Liu X, Mohseni G, Hao X, Ren Y, Xu Y, Gao H, Wang Q, Wang Y: Mechanism research and treatment progress of NAD pathway related molecules in tumor immune microenvironment. Cancer Cell Int 2022, 22(1):242.

15) Brandes RP, Weissmann N, Schroder K: Nox family NADPH oxidases: Molecular mechanisms of activation. Free Radic Biol Med 2014, 76:208-226.

16) Gong S, Wang S, Shao M: NADPH Oxidase 4: A Potential Therapeutic Target of Malignancy. Front Cell Dev Biol 2022, 10:884412.

17) Wang Y, Sui Y, Niu Y, Liu D, Xu Q, Liu F, Zuo K, Liu M, Sun W, Wang Z et al: PBX1-SIRT1 Positive Feedback Loop Attenuates ROS-Mediated HF-MSC Senescence and Apoptosis. Stem Cell Rev Rep 2023, 19(2):443-454.

18) Wang X, Lu Y, Tuo Z, Zhou H, Zhang Y, Cao Z, Peng L, Yu D, Bi L: Role of SIRT1/AMPK signaling in the proliferation, migration and invasion of renal cell carcinoma cells. Oncol Rep 2021, 45(6).

19) Liu SC, Yang JJ, Shao KN, Chueh PJ: RNA interference targeting tNOX attenuates cell migration via a mechanism that involves membrane association of Rac. Biochem Biophys Res Commun 2008, 365(4):672-677.

20) Lin MH, Lee YH, Cheng HL, Chen HY, Jhuang FH, Chueh PJ: Capsaicin Inhibits Multiple Bladder Cancer Cell Phenotypes by Inhibiting Tumor-Associated NADH Oxidase (tNOX) and Sirtuin1 (SIRT1). Molecules 2016, 21(7).

21) Cheng HL, Lee YH, Yuan TM, Chen SW, Chueh PJ: Update on a tumor-associated NADH oxidase in gastric cancer cell growth. World J Gastroenterol 2016, 22(10):2900-2905.

22) Lee YH, Chen HY, Su LJ, Chueh PJ: Sirtuin 1 (SIRT1) Deacetylase Activity and NAD(+)/NADH Ratio Are Imperative for Capsaicin-Mediated Programmed Cell Death. J Agric Food Chem 2015, 63(33):7361-7370.

23) Islam A, Su AJ, Zeng ZM, Chueh PJ, Lin MH: Capsaicin Targets tNOX (ENOX2) to Inhibit G1 Cyclin/CDK Complex, as Assessed by the Cellular Thermal Shift Assay (CETSA). Cells 2019, 8(10).

24) Su YC, Lin YH, Zeng ZM, Shao KN, Chueh PJ: Chemotherapeutic agents enhance cell migration and epithelial-to-mesenchymal transition through transient up-regulation of tNOX (ENOX2) protein. Biochim Biophys Acta 2012, 1820(11):1744-1752.

25) Zeng ZM, Chuang SM, Chang TC, Hong CW, Chou JC, Yang JJ, Chueh PJ: Phosphorylation of serine-504 of tNOX (ENOX2) modulates cell proliferation and migration in cancer cells. Experimental cell research 2012, 318(14):1759-1766.

26) Chueh PJ, Wu LY, Morre DM, Morre DJ: tNOX is both necessary and sufficient as a cellular target for the anticancer actions of capsaicin and the green tea catechin (-)-epigallocatechin-3-gallate. Biofactors 2004, 20(4):235-249.

27) Ran D, Zhou D, Liu G, Ma Y, Ali W, Yu R, Wang Q, Zhao H, Zhu J, Zou H et al: Reactive Oxygen Species Control Osteoblast Apoptosis through SIRT1/PGC-1alpha/P53(Lys382) Signaling, Mediating the Onset of Cd-Induced Osteoporosis. J Agric Food Chem 2023.

28) Zhang Z, Chen X, Liu S: Role of Sirtuin-1 in Neonatal Hypoxic-Ischemic Encephalopathy and Its Underlying Mechanism. Med Sci Monit 2020, 26:e924544.

29) McReynolds MR, Chellappa K, Baur JA: Age-related NAD(+) decline. Exp Gerontol 2020, 134:110888.

30) Xie N, Zhang L, Gao W, Huang C, Huber PE, Zhou X, Li C, Shen G, Zou B: NAD(+) metabolism: pathophysiologic mechanisms and therapeutic potential. Signal Transduct Target Ther 2020, 5(1):227.

31) Zapata-Perez R, Wanders RJA, van Karnebeek CDM, Houtkooper RH: NAD(+) homeostasis in human health and disease. EMBO Mol Med 2021, 13(7):e13943.

32) Wang B, Ma Y, Kong X, Ding X, Gu H, Chu T, Ying W: NAD(+) administration decreases doxorubicin-induced liver damage of mice by enhancing antioxidation capacity and decreasing DNA damage. Chem Biol Interact 2014, 212:65-71.

33) Hong G, Zheng D, Zhang L, Ni R, Wang G, Fan GC, Lu Z, Peng T: Administration of nicotinamide riboside prevents oxidative stress and organ injury in sepsis. Free Radic Biol Med 2018, 123:125-137.

34) Nadysev GY, Tikhomirov AS, Lin MH, Yang YT, Dezhenkova LG, Chen HY, Kaluzhny DN, Schols D, Shtil AA, Shchekotikhin AE et al: Aminomethylation of heliomycin: Preparation and anticancer characterization of the first series of semi-synthetic derivatives. European journal of medicinal chemistry 2018, 143:1553-1562.

Author response:

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

Public Reviews:

Reviewer 1:

(1) General comment: The evidence for these highly novel, potentially interesting roles (of the exocyst) would need to be more compelling to support direct involvement.

We wish to thank the reviewer for his/her comments, and for considering that the proposed functions are highly novel and potentially interesting. To strengthen the evidence supporting the new roles of the exocyst, we have performed a number of additional experiments that are depicted in novel figures or figure panels of the new version of the manuscript. Particularly, we aimed at providing further support of the direct involvement of the exocyst in different steps of the regulated secretory pathway. Please see the details below.

(2) For instance, the localization of exocyst to Golgi or to granule-granule contact sites does not seem substantial.

We have performed quantitative colocalization studies, as suggested by the reviewer to further substantiate our initial findings. We have carefully analysed GFP-Sec15 distribution in relation to the Golgi complex and secretory Glue granules at relevant time points of salivary gland development. Overall, we found that GFP-Sec15 distribution is dynamic during salivary gland development. Before Glue synthesis (72 h AEL), Sec15 was observed in close association (defined as a distance equal to, or less than 0.6 µm) with the Golgi complex (please see below Author response image 1). This association was lost once Glue granules have begun to form (96 h AEL). Importantly, we do not see relevant association between GFP-Sec15 and the ER (please see Author response image 2). These observations support our conclusion that the exocyst plays a role at the Golgi complex. New images supporting these conclusions, as well as quantitative data, have been included in Figure 5 of the new version of the manuscript. In addition, real time imaging, as well as 3D reconstruction analyses, confirming the close association between Sec15 and Golgi cisternae are now included in the manuscript. Please see Supplementary Videos 1-3. These new data are described in the text lines 200-210 of the Results section and text lines 359368 of the Discussion section.

Interestingly, at the time when Sec15-Golgi association is lost (96 h AEL), Sec15 foci associate instead with newly formed secretory granules (< 1µm diameter). This association persists during secretory granule maturation (100-116 h AEL), when Sec15 foci localize specifically in between neighbouring, immature secretory granules. When maturation has ended and Glue granule exocytosis begins (116-120 h AEL), this localization between granules is lost. These observations are consistent with a role of the exocyst in homotypic fusion during SG maturation. We have included new images showing that association between Sec15 and secretory granules is dynamic and depends on the developmental stage. We have quantified this association both during maturation and at a stage when SGs are already mature. We have in addition performed a 3D reconstruction analysis of these images to confirm the close association between Sec15 and immature SGs. These new data are now depicted in Figure 7BC, Supplementary Videos 4-5, and described in text lines 216-221 of the Results section. In addition, a lower magnification image is provided below in this letter (Author response image 3), quantifying the proportion of Sec15 foci localized in between SGs (yellow arrows) relative to the total number of Sec15 foci (yellow arrows + green arrowheads).

Author response image 1.

Criteria utilized to define Sec15 focithat were“associated” or“not associated” withthe trans-Golgi network in the experiments of Figure 5C-E of the manuscript.When the distance between maximal intensities of GFP-Sec15 and Golgi-RFP signals was equal or less than 0.6 m, the signals were considered “associated” (upper panels). When the distance was more than 0.6 m, the signals were considered “not associated” (lower panels).

Author response image 2.

Criteria utilized to define Sec15 focithat were“associated” or“not associated” withthe ERin the experiments of Figure 5A-Bof the manuscript.When the distance between maximal intensities of GFP-Sec15 and KDEL-RFP signals was equal or less than 0.6 m, the signals were considered “associated”. When the distance was more than 0.6 m, the signals were considered “not associated”.

Author response image 3.

(A) GFP-Sec15 foci (cyan) and SGs (red) are shown in cells bearing Immature SGs or (B) with mature SGs. Yellow arrows indicate GFP-Sec15 foci localized in between SGs; green arrowheads indicate GFP-Sec15 foci that arenot in between SGs. (C) Quantification of the percentage (%) of Sec15 foci localized in between SGs respect to the total number of Sec15 foci in cells filled with immature SGs (ISG)vs cells with mature SGs (MSG).

It is interesting to mention that previous evidence from mammalian cultured cells (Yeaman et al,  2001) show that the exocyst localizes both at the trans-Golgi network and at the plasma membrane, weighing in favour of our claim that the exocyst is required at various steps of the exocytic pathway. Thus, the exocyst may play multiple roles in the secretion pathway in other biological models as well. This concept has now been included at the Discussion section of the revised version of the manuscript (lines 359-368).

To make the conclusions of our work clearer, in the revised version of the manuscript, we have now included a graphical abstract, summarizing the dynamic localization of the exocyst in relation to the processes of SG biogenesis, maturation and exocytosis reported in our work. 

(3) Instead, it is possible that defects in Golgi traffic and granule homotypic fusion are not due to direct involvement of the exocyst in these processes, but secondary to a defect in canonical exocyst roles at the plasma membrane. A block in the last step of glue exocytosis could perhaps propagate backward in the secretory pathway to disrupt Golgi complexes or cause poor cellular health due to loss of cell polarity or autophagy.

We thank the reviewer for these thoughtful comments. We have performed a number of additional experiments to assess “cellular health” or to identify possible defects in cell polarity after knock-down of exocyst subunits. These new data have been included in new supplementary figures 5 and 6 of the revised version of the manuscript (please see below). 

In our view, the precise localization of GFP-Sec15 at the Golgi complex (Figure 5C-E), as well as in between immature secretory granules (Figure 7B-D), argues in favour of a direct involvement of the exocyst in SG biogenesis and homofusion respectively. 

We truly appreciate the comment of the reviewer raising the possibility that the defects that we observe at early steps of the pathway (SG biogenesis and SG maturation) may actually stem from a backward effect of the role of the exocyst in SG-plasma membrane tethering. We wish to respectfully point out that the processes of biogenesis, maturation and plasma membrane tethering/fusion of SGs do not occur simultaneously in the Drosophila larval salivary gland in vivo, as they do in other secretory model systems (i.e. cell culture). In this regard, the experimental model is unique in terms of synchronization. In each cell of the salivary gland, the three processes (biogenesis, maturation and exocytosis) occur sequentially, and controlled by developmental cues. At the developmental stage when SGs fuse with the plasma membrane, SG biogenesis has already ceased many hours earlier: SG biogenesis occurs at 96-100 hours after egg lay (AEL), SG maturation takes place at 100-112 hours AEL, and SG-plasma membrane fusion happens only when all SGs have undergone maturation and are ready to fuse with the plasma membrane at 116-120 h AEL. Thus, in our view it is not conceivable that a defect in SG-plasma membrane tethering/fusion (116-120 h AEL) may affect backwards the processes of SG biogenesis or SG maturation, which have occurred earlier in development (96-112 h AEL).

As suggested by the reviewer, we have analysed several markers of cellular health and cell polarity, comparing conditions of exocyst subunit silencing (exo70RNAi, sec3RNAi or exo84RNAi) with wild type controls (whiteRNAi). These new data are depicted in Supplementary Figures 5 and 6, and described in lines 172-179 of the Results section of the revised version of the manuscript. Noteworthy, for these experiments we have applied silencing conditions that block secretory granule maturation, bringing about mostly immature SGs. Our analyses included: 1) Subcellular distribution of PI(4,5)P2, 2) subcellular distribution of the tetraspanin CD63, 3) of Rab11, 4) of filamentous actin, and 5) of CD8. We have also compared 6) nuclear size and nuclear general morphology, 7) the number and distribution of mitochondria, 8) morphology and subcellular distribution of the cis- and 9) trans-Golgi networks. Finally, 10) we have compared basal autophagy in salivary cells with or without knocking down exocyst subunits. The markers that we have analysed behaved similarly to those of control salivary glands, suggesting that the observed defects in regulated exocytosis indeed reflect different roles of the exocyst in the secretory pathway, rather than poor cellular health or impaired cell polarity.  

Our conclusions are in line with previous studies in which apico-basal polarity, Golgi complex morphology and distribution, as well as apical membrane trafficking were also evaluated in exocyst mutant backgrounds, finding no anomalies (Jafar-Nejad et al, 2005). 

Conversely, in studies in which apical polarity was disturbed by interfering with Crumbs levels, SG biogenesis, maturation and exocytosis were not affected (Lattner et al, 2019), indicating that these processes not necessarily interfere with one another.  

(4) Final recommendation: In the absence of stronger evidence for these other exocyst roles, I would suggest focusing the study on the canonical role (interesting, as it was previously reported that Drosophila exocyst had no function in the salivary gland and limited function elsewhere [DOI: 10.1034/j.1600-0854.2002.31206.x]), and leave the alternative roles for discussion and deeper study in the future.  

We appreciate the reviewer´s recommendation. However, we believe that the major strength of our work is the discovery of non-canonical roles of the exocyst complex, unrelated to its function as a tethering complex for vesicle-plasma membrane fusion. We believe that in the new version of our manuscript, we provide stronger evidence supporting the two novel roles of the exocyst:

a) Its participation in maintaining the normal structure of the Golgi complex, and b) Its function in secretory granule maturation.

Reviewer 2:

(5) General comment: A key strength is the breadth of the assays and study of all 8 exocyst subunits in a powerful model system (fly larvae). Many of the assays are quantitated and roles of the exocyst in early phases of granule biogenesis have not been ascribed. 

We are grateful that the reviewer appreciates the novelty of our contribution.

(6) However there are several weaknesses, both in terms of experimental controls, concrete statements about the granules (better resolution), and making a clear conceptual framework. Namely, why do KD of different exocysts have different effects on presumed granule formation

The reviewer has raised a point that is central to the interpretation of all our data throughout the manuscript. The short answer is that the extent of RNAi-dependent silencing of exocyst subunits determines the phenotype: 

1) Maximum silencing affects Golgi complex morphology and prevents SG biogenesis. 2) Intermediate silencing blocks SG maturation, without affecting Golgi complex morphology and SG biogenesis. 3) Weak silencing blocks SG tethering and fusion with the plasma membrane, without affecting Golgi complex morphology, SG biogenesis or SG maturation. 

In other words, 1) Low levels of exocyst subunits are sufficient for normal Golgi complex morphology and SG biogenesis. 2) Intermediate levels of exocyst subunits are sufficient for SG maturation (and also sufficient for SG biogenesis). 3) High levels of exocyst subunits are required for SG tethering and subsequent fusion with the plasma membrane. 

Based on the above notion, we have exploited the fact that temperature can fine-tune the level of Gal4/UAS-dependent transcription, thereby achieving different levels of silencing, as shown by Norbert Perrimon et al in their seminal paper “the level of RNAi knockdown can also be altered by using Gal4 lines of various strengths, rearing flies at different temperatures, or via coexpression of UAS-Dicer2” (Perkins et al, 2015). 

We found in our system that indeed, by applying appropriate silencing conditions (RNAi line and temperature) to any of the eight subunits of the exocyst, we have been able to obtain one of the three alternative phenotypes: Impaired SG biogenesis, or impaired SG maturation, or impaired SG tethering/fusion with the plasma membrane.

These concepts are summarized below in Author response image 4. Please see also at point 26, the general comment of Reviewer #3. 

We have conducted qRT-PCR assays to provide experimental support to the notions summarized above in Author response image 4. We measured the remaining levels of mRNAs of some of the exocyst subunits, after inducing RNAi-mediated silencing at different temperatures, or with different RNAi transgenic lines. The remaining RNA levels after silencing correlate well with the observed phenotypes, following the predictions of Author response image 4 and summarized in Author response image 5. These new data are now shown in Supplementary Figure 2 of the revised version of the manuscript, and described in lines 153-159 at the Results section.

(7) Why does just overexpression of a single subunit (Sec15) induce granule fusion?

The reviewer raises a very important point. Based on available data from the literature, Sec15 behaves as a seed for assembly of the holocomplex and it also mediates the recruitment of the holocomplex to SGs through its interaction with Rab11 (Escrevente et al, 2021; Bhuin and Roy, 2019; Wu et al, 2005; Zhang et al, 2004; Guo et al, 1999). Thus, overexpression of Sec15 is expected to enhance exocyst assembly, thereby potentiating the activities carried out by the complex in the cell, including SG homofusion. In the revised version of the manuscript we have also performed the overexpression of Sec8, finding that, unlike Sec15, Sec8 fails to induce homotypic fusion. These results were expected, as they confirm that Sec8 does not behave as a seed for mounting the whole complex. These new data have been included in Figure 7E-H, and are described in text lines 221-229 of the Results section. 

Author response image 4.

Conceptual model of RNAi expression at different temperatures , remaining levels of mRNA/protein levels and phenotypes obtained at each temperature.

Author response image 5.

qRT-PCR assays presented in Supplementary Figure 2 are shown in combination with the phenotypes observed at each of the conditions analyzed. Note the correlation between phenotypes and the extent of mRNA downregulation.

(8) While the paper is fascinating, the major comments need to be addressed to really be able to make better sense of this work, which at present is hard to disentangle direct vs. secondary effects, especially as much of the TGN seems to be altered in the KDs.  

We hope that our response to point 6) has helped to clarify this important point raised by the Reviewer. After applying silencing conditions where normal structure of the trans-Golgi network is impaired, SG biogenesis does not occur. Thus, since SGs do not form, it is not conceivable to detect defects in SG maturation or SG fusion with the plasma membrane in the same cell.

(9) The authors conveniently ascribe many of the results to the holocomplex, but their own data (Fig. 4 and Fig. 6) are at odds with this.

This is another central point of our work, so we thank the reviewer for his/her comment. In Figures 4A, 7A and 9A of the revised version of the manuscript, we show that, by inducing appropriate levels of silencing of any of the 8 subunits of the exocyst, each of the three alternative phenotypic manifestations can occur. In our opinion, this argues in favour of a function for the whole exocyst complex in each of the three specific activities proposed in our study: 1) SG biogenesis, 2) SG maturation, and 3) SG tethering/fusion with the plasma membrane. In detailed characterizations of these three phenotypes performed throughout the study, we decided to induce silencing of just two or three of the subunits of the exocyst, assuming that the whole complex accounts the mechanisms involved.

Major comments

(10) Resolution not sufficient. Identification of "mature secretory granules" (MSG) in Fig. 3 is based on low-resolution images in which the MSG are not clearly seen (see control in Fig. 3A) and rather appear as a diffuse haze, and not as clear granules. There may be granules here, but as shown it is not clear. Thus it would be helpful to acquire images at higher resolution (at the diffraction limit, or higher) to see and count the MSG.

We thank the reviewer for raising this point, as it may not be straightforward to the reader to identify the SGs throughout the figures of our study. To make it clearer, in Figure 3A (magnified insets on the right), we have delimitated individual SGs with a green dotted line, and included diagrams (far right), which we hope will help the identification of SGs. In Figure 3B, we show that after silencing Sec84, a mosaic phenotype was observed: In some cells SGs fail to undergo maturation, and remain smaller than normal. In other cells of this mosaic phenotype, biogenesis of SGs was impaired and the fluorescent cargo remained trapped in a mesh-like structure (that we later show that corresponds to the ER). The dotted line marks individual SGs, and the diagrams included on the right intend to help the interpretation of the phenotype. The mesh-like structures where Sgs3-GFP was retained are also marked with dotted line, and schematized on the right. These new schemes are described in the Figure 3 caption of the revised version of the manuscript.

We wish to mention that all the confocal images depicted in this figure and throughout the manuscript  have been captured at high resolution, with a theoretical resolution limit of 168177nm (d = γ/2NA). Given that secretory granules range from 0.8-7µm in diameter, the resolution is more than sufficient to clearly resolve these structures. 

(11) Note: the authors are not clear on which objective was used. Maybe the air objective as the resolution appears poor).  

In this particular figure, we have utilized a Plan-Apochromat 63X/1.4NA oil objective of the inverted Carl Zeiss LSM 880 confocal microscope (mentioned in materials and methods).

(12) They need to prove that the diffuse Sgs3-GFP haze is indeed due to MSG.  

If we interpret correctly the concern of the reviewer, what he/she calls “diffuse haze” is actually the distribution of Sgs3-GFP within individual SGs, which, as previously reported by other authors, is not homogeneous at this stage (Syed et al. 2022). We hope that the diagrams that we have included in Figure 3 A, B (point 10) will help the readers interpreting the images.   

(13) Related it is unclear what are the granule structures that correspond to Immature secretory granules (ISG) and cells with mesh-like structures (MLS)?

We are confident that the diagrams now included in Figure 3A and B will help the interpretation, and particularly to identify immature granules and the mesh-like structure generated after silencing of exocyst subunits.

(14) Similarly, Sgs3 images of KD of 8 exocyst subunits were interpreted to be identical, in Fig. 4, but the resolution is poor.

We hope that the issue related to resolution of our images has been properly addressed in the response to point 10) of this letter. In Figure 4A, we show that after silencing of any of the 8 subunits (with the appropriate conditions), in all cases SG biogenesis was impaired, and Sgs3GFP was instead retained in a mesh-like structure. Images obtained after silencing different exocyst subunits are of course not identical, but in all cases, a mesh-like structure has replaced the formation of SGs (Figure 4A). Hopefully, the diagrams now included in Figure 3A and B help the correct interpretation of the phenotypes throughout the study.

To demonstrate that the structure in which Sgs3-GFP was retained upon exocyst complex knockdown corresponds to the ER, we performed a colocalization analysis between Sgs3-GFP and the ER markers GFP-KDEL or Bip-sfGFP-HDEL, after which we calculated the Pearsons Coefficient, which indicated substantial colocalization (Figure 4B-G and Supplementary Figures 7 and 8). These new data are described in lines 196-199 of the revised version of the manuscript. To facilitate the visualization of the results, in the revised version of the manuscript we have included magnified cropped areas of the images shown in Figure 4A.

(15) What is remarkable is a highly variable effect of different subunit KD on the percentage of cells with MLS (Fig. 4C). Controls = 100 %, Exo70=~75% (at 19 deg), Sec3 = ~30%, Sec10 = 0%, Exo84 = 100% ... This is interesting for the functional exocyst is an octameric holocomples, thus why the huge subunit variability in the phenotypes? The trivial explanation is either: i) variable exocyst subunit KD (not shown) or ii) variability between experiments (no error bars are shown). Both should be addressed by quantification of the KD of different proteins and secondly by replicating the experiments.

We agree with the reviewer statement. We believe that both, variability of KD efficiency (i) and variability between experiments (ii) contribute to the variable effect observed after knocking down the different subunits. As detailed in the response to point 6), we have performed qRT-PCR determinations to confirm that the severity of the phenotype depends on the efficiency of RNAimediated silencing. We chose to analyse in detail the effect on the subunits exo70 and sec3, which were those with the highest phenotypic differences between the three silencing temperatures utilized. We found that as expected, the levels of silencing were temperaturedependent, being higher at 29°C and lower at 19°C. These data were included in Supplementary Figure 2, and described lines 153-159 of the Results section and also summarized in Author response images 4 and 5 of this rebuttal letter.

We thank the reviewer for his/her comment on the replication of experiments and statistics. We failed to include detailed numerical information in the original submission, such as the number of replicas and standard deviations of the data depicted in Figure 3C and Supplementary Figure 1, so we apologize for this omission. In the revised version of the manuscript, we have included a table (Supplementary Table 3) in which all the raw data of Figure 3C and Supplementary Figure 1, including standard deviations, are now depicted.

(16) If their data holds up then the underlying mechanism here needs to be considered.

(Note: there is some precedent from the autophagy field of differential exocyst effects)

Our proposed mechanism is essentially that the holocomplex is required for multiple processes along the secretory pathway. Each of these actions (Golgi structure maintenance, SG maturation and SG tethering/fusion with the plasma membrane) requires different amounts of holocomplex activity, being this the reason why each phenotype manifests at different levels of RNAi-mediated silencing (Author response image 4 of this letter). The model predicts that Golgi structure maintenance requires minimal levels of complex activity, and that is why strong knock-down of exocyst subunits is required to obtain this phenotype. In line with our results, it has been reported that other tethering complexes of the CATCHR family are also required for maintaining Golgi cisternae stuck together (D'Souza et al, 2020; Khakurel and Lupashin, 2023; Liu et al, 2019). One possibility is that the exocyst may play a redundant role in the maintenance of the normal structure of the Golgi complex, along with other CATCHR complexes. This potential redundancy could explain why severe exocyst knock-down is required to observe structural anomalies at this organelle. On the other end of the spectrum, we propose that tethering/fusion with the plasma membrane is very susceptible to even slight reduction of complex activity, so that mild RNAi-mediated silencing is sufficient to provoke defects in this process. This proposed model is depicted in Author response image 4 and discussed in lines 395-405 of the Discussion section. 

(17) In the salivary glands the authors state that the exocyst is needed for Sgs3-GFP exit from the ER. First, Pearson's coefficient should be shown so as to quantitate the degree of ER localizations of all KDs.

We thank the reviewer for this comment that helped us to strengthen the observation that when SG biogenesis is impaired, Sgs3-GFP remains trapped in the ER. In the revised version of the manuscript, we have calculated Pearson´s coefficient to assess colocalization between ER markers (GFP-KDEL or Bip-sfGFP-HDEL) and Sgs3-GFP in salivary gland cells that express sec15RNAi. The Pearson’s coefficient was around 0.6 for both ER markers, indicating that colocalization with Sgs3-GFP was substantial (Supplementary Figure 8, text lines 196-199 of the Results section).

(18) Second, there should be some rescue performed (if possible) to support specificity. 

As suggested by the reviewer, we have performed a rescue experiment of the phenotype provoked by the expression of sec15 RNAi, which consisted on the retention of Sgs3-GFP in the endoplasmic reticulum: Expression of Sec15-GFP reverted substantially the ER retention phenotype, rescuing SG biogenesis and also SG maturation in most cells (over 60% of the cells). These new data are now shown in Supplementary Figure 4, and described in lines 168-171 of the Results section.

(19) Third, importantly other proteins that should traffic to the PM need to be shown to traffic normally so as to rule out a non-specific effect.

We have addressed this issue (also mentioned by Reviewer #1), by analyzing the localization of a number of polarization markers, finding that the overall polarization of the cell was not affected by loss of function of exocyst subunits. Please, see our response to the point 3) raised by Reviewer #1. The new data showing cell polarization markers are shown in Supplementary Figure 6 of the revised version of the manuscript, and described on text lines 172-179 of the Results section.

(20) It is unclear from their model (Fig. 5) why after exocyst KD of Sec15 the cis-Golgi is more preserved than the TGN, which appears as large vacuoles. This is not quantitated and not shown for the 8 subunits.

We thank the reviewer for this relevant comment. We agree that the phenotype of either, sec15 or sec3 loss-of-function cells manifests differently with cis-Golgi and trans-Golgi markers. While the cis-Golgi marker looked fragmented and aggregated, the trans-Golgi marker adopted a swollen appearance. However, in our view, the different appearance of the two markers does not necessarily imply that one compartment is more preserved than the other. In the revised version of the manuscript, we have quantified the penetrance of the phenotypes provoked by sec15 or sec3 silencing, using both cis-Golgi and trans-Golgi markers. In both cases, the penetrance was high, although even higher with the trans-Golgi marker. These new data are now depicted in Supplementary Figure 9 of the revised version of the manuscript. 

It is interesting to mention that in HeLa cells, as well as in the retinal epithelial cell line hTERT, Golgi phenotypes similar to those we have described here have been reported after loss-offunction of other tethering complexes, which were shown to maintain the Golgi cisternae stuck together, including the GOC and GARP complexes (D'Souza et al, 2020, Khakurel and Lupashin, 2023; Shijie Liu et al, 2019). As we did throughout our work, not every aspect of the analysis included the silencing of all eight subunits. In this case, we chose to silence Sec3 and Sec15. Please note that we have modified the model depicted in Figure 6E-F, to highlight the cis- and transGolgi phenotypes upon exocyst knock-down, as well as the localization of the exocyst in cisternae of the Golgi complex.

(21) Acute/Chronic control: It would be nice to acutely block the exocyst so as to better distinguish if the effects observed are primary or secondary effects (e.g. on a recycling pathway).

We thank the reviewer for raising this important issue. To address this point, and to be able to induce silencing of exocyst subunits at specific time intervals of larval development, we utilized a strategy based on a thermosensitive variant of the Gal4 inhibitor Gal80 (Gal80ts)(Lee and Luo, 1999). We blocked Gal4 activity (and therefore RNAi expression) by maintaining the larvae at 18 °C during the 1st and 2nd instars (until 120 hours after egg lay), and then induced the activity of Gal4 specifically at the 3rd larval instar by raising the temperature to 29 ºC, a condition in which Gal80ts becomes inactive. After silencing the expression of sec3 or sec15 at the 3rd larval instar only, the phenotype was very similar to that observed after chronic silencing of exocyst subunits (larvae maintained at 29 ºC all throughout development, where Gal4 was never inhibited). These observations suggest that the defects observed in the secretory pathway after knock down of exocyst subunits reflect genuine functions of the exocyst in this pathway, rather than a secondary effect derived from impaired development of the salivary glands at early larval stages. These new results are now shown in Supplementary Figure 3, and described in manuscript lines 160-171 of the Results section.   

(22) Granule homotypic fusion. Strangely over-expression of just one subunit, Sec15-GFP, made giant secretory granules (SG) that were over 8 microns big! Why is that, especially if normally the exocyst is normally a holocomplex. Was this an effect that was specific to Sec15 or all exocyst subunits? Is the Sec15 level rate limiting in these cells? It may be that a subcomplex of Sec15/10 plays earlier roles, but in any case this needs to be addressed across all (or many) of the exocyst subcomplex members.

Please, see our response to point 7) of this letter. Sec15 is believed to act as a seed for the formation of the whole complex.

(23) In summary, there are clearly striking effects on secretory granule biogenesis by dysfunction of the exocyst, however right now it is hard to disentangle effects on ERGolgi traffic, loss of the TGN, and a problem in maturation or fusion of granules. 

As discussed in detail in our response to the point 3 raised by Reviewer #1, the secretory pathway is highly synchronized in each of the cells of the Drosophila salivary gland. SG biogenesis, SG maturation and SG fusion with the plasma membrane never occur simultaneously in the same cell. Thus, in a cell in which ER-Golgi traffic is impaired (and SG biogenesis does not occur), SGs do not exist, and therefore, they cannot exhibit defects in the process of maturation or fusion with the plasma membrane. In summary, we believe that our work has shown that in Drosophila larval salivary glands the exocyst holocomplex is required for (at least) three functions along the secretory pathway: 1) To maintain the appropriate Golgi complex architecture, thus enabling ERGolgi transport; 2) For secretory granule maturation: both, homotypic fusion and acquisition of maturation factors; 3) For secretory granule exocytosis: secretory granule tethering to enable subsequent fusion with the plasma membrane. As mentioned above (point 6 of this letter), these three functions require different amounts of the holocomplex, and therefore can be revealed by inducing different levels of silencing.  

(24) It is also confusing if the entire exocyst holocomplex or subcomplex plays a key role 

The fact that, by silencing any of the subunits (with the appropriate conditions) it is possible obtain any of the 3 phenotypes (impaired SG biogenesis, impaired SG maturation or impaired SG fusion with the plasma membrane) argues in favour of a function of the complex as a whole in each of these three functions.

Reviewer 3:

(25) General comment: Freire and co-authors examine the role of the exocyst complex during the formation and secretion of mucins from secretory granules in the larval salivary gland of Drosophila melanogaster. Using transgenic lines with a tagged Sgs3 mucin the authors KD expression of exocyst subunit members and observe a defect in secretory granules with a heterogeneity of phenotypes. By carefully controlling RNAi expression using a Gal4-based system the authors can KD exocyst subunit expression to varying degrees. The authors find that the stronger the inhibition of expression of exocyst the earlier in the secretory pathway the defect. The manuscript is well written, the model system is physiological, and the techniques are innovative.

We appreciate the reviewer´s assessment of our work. 

(26) My major concern is that the evidence underlying the fundamental claim of the manuscript that "the exocyst complex participates" in multiple secretory processes lacks direct evidence.

We thank the reviewer for raising this important issue. We believe that the analysis of Sec15 subcellular localization during salivary gland development (Figures 5, 7B-D and 9E-F), in combination with the detailed analysis of the phenotypes provoked by loss-of-function of each of the exocyst subunits, provide evidence supporting multiple functions of the exocyst in the secretory pathway. We have also included 3D reconstructions and videos of GFP-Sec15 colocalization with Golgi and SG markers to support exocyst localization associated to these structures (Supplementary Videos 1-7), text lines 200-210; 216-221 and 303-305.

(27) It is clear from multiple lines of evidence, which are discussed by the authors, that exocyst is essential for an array of exocytic events. The fundamental concern is that loss of homeostasis on the plasma membrane proteome and lipidome might have severe pleiotropic effects on the cell.

We agree with the reviewer that this is an important point that needed to be addressed. As discussed in detail above at the response to point 3 raised by Reviewer #1, we have analysed several plasma membrane markers (including a PI(4,5)P2 lipid reporter), and found that overall, plasma membrane integrity and polarity were not substantially affected (Supplementary Figure 6). In addition, we have analyzed several markers of general cellular “health” that indicate that salivary gland cells do not seem to be distressed by the reduction of exocyst complex activity (Supplementary Figure 5). These new data are described in lines 172-179 of the Results section.

(28) Perhaps the authors have more evidence that exocyst is important for homeotypic fusion of the SGs, as supported by the localisation of Sec15 on the fusion sites.

We believe that the fact that, by silencing any of the exocyst subunits (with the appropriate conditions), immature smaller-than-normal granules were observed, argus in favour that the exocyst as a whole participates in SG homofusion (Figure 7A). In addition, we have included more images, quantifications, 3D reconstructions and videos of GFP-Sec15 localized just at the contact sites between immature SGs. We have quantified and compared GFP-Sec15 localization at immature SG vs its localization at mature SGs, finding that localizes preferentially at immature SGs, supporting a role of the exocyst as a tethering complex during homotypic fusion (shown Figure 7B-C and Supplementary Videos 4-6, and described in lines 216-221 of the Results section). Please see also our response to the point 2 raised by reviewer 1 in this rebuttal letter, and to Author response image 3 above in this letter.

(29) The second question that I think is important to address is, what exactly do the varying RNAi levels correspond to in terms of experiments, and have these been validated? Due to the fundamental claim being that the severity of the phenotype being correlated with the level of KD, I think validation of this model is absolutely essential.  

We thank the Reviewer for raising this important point, and agree it was lacking in the original version of our manuscript. As discussed in our response to the point 6) raised by Reviewer #2, we have performed qRT-PCR determinations for exo70 and sec3 mRNA levels after inducing silencing of these subunits at different temperatures, or with different RNAi transgenic lines. The remnant mRNA levels correlate well with the observed phenotypes. Please see Supplementary Figure 2 of the revised manuscript, and Author response image 5 of this rebuttal letter; described in lines 155-159 of the Results section. 

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

-  The authors assert in the discussion that exocyst involvement in constitutive secretion is well documented. This is based on a very recent study in mammalian culture cells. Therefore, I would not dismiss the issue as completely settled. Furthermore, a previous study of Drosophila sec10 reported no roles outside the ring gland (DOI: 10.1034/j.1600-0854.2002.31206.x).

We have included these observations in the Discussion section. Lines 326-329.

-  A salivary gland screening by Julie Brill's lab reported exocyst components as hits (DOI: 10.1083/jcb.201808017).

We have referred to this paper in the Discussion section. Lines 326-329.

-  It should be explained in more detail what is measured in graphs 7C, F, and others quantifying fluorescence around secretory granules. Looking at the images, the decrease in Rab1 and Rab11 seems less convincing.

We have made a clearer description of how fluorescence intensity was measured in the Methods section lines 558-561. Also, we have uploaded a source data file in which the raw data of each experiment used for quantifications are disclosed. 

Please note that the data indicates that Rab11 levels are higher in sec5 (Figure 8J-L) and sec3 (supplementary Figure 11M-R).

Reviewer #2 (Recommendations For The Authors):

No major issues.

Writing - The authors should better frame their interpretations of other studies of the exocyst that include the role in autophagy, Palade body trafficking, and differential roles of the subunits.

We have discussed these specific points in the Discussion section, lines 348-355 and 409-410.

Minor - Fig. 6A: Why are variable temperatures (19-29 deg C used for the 8 KD experiments)?

Please show it all at the same temperature (control too).

The need for the usage of specific temperatures to obtain specific phenotypes with each of the RNAi lines used was explained in point 6 of this letter.

Reviewer #3 (Recommendations For The Authors):

In the abstract, the authors refer to the exocytic process and go on to describe secretory granule biogenesis and exocytosis. However, there are many exocytic processes aside from secretory granule biogenesis, and I think the authors should clarify this.

Corrected in the Abstract. Lines 19-21

Page 17 Thomas, 2021 reference, there is a glitch with the reference.

Thanks for noticing. Fixed.

References

Bhuin T, Roy JK. Developmental expression, co-localization and genetic interaction of exocyst component Sec15 with Rab11 during Drosophila development. Exp Cell Res. 2019 Aug 1;381(1):94-104. doi: 10.1016/j.yexcr.2019.04.038. Epub 2019 May 7. PMID: 31071318.

D'Souza Z, Taher FS, Lupashin VV. Golgi inCOGnito: From vesicle tethering to human disease. Biochim Biophys Acta Gen Subj. 2020 Nov;1864(11):129694. doi: 10.1016/j.bbagen.2020.129694. Epub 2020 Jul 27. PMID: 32730773; PMCID: PMC7384418.

Escrevente C, Bento-Lopes L, Ramalho JS, Barral DC. Rab11 is required for lysosome exocytosis through the interaction with Rab3a, Sec15 and GRAB. J Cell Sci. 2021 Jun 1;134(11):jcs246694. doi: 10.1242/jcs.246694. Epub 2021 Jun 8. PMID: 34100549; PMCID: PMC8214760.

Guo W, Roth D, Walch-Solimena C, Novick P. The exocyst is an effector for Sec4p, targeting secretory vesicles to sites of exocytosis. EMBO J. 1999 Feb 15;18(4):1071-80. doi: 10.1093/emboj/18.4.1071. PMID: 10022848; PMCID: PMC1171198.

Jafar-Nejad H, Andrews HK, Acar M, Bayat V, Wirtz-Peitz F, Mehta SQ, Knoblich JA, Bellen HJ. Sec15, a component of the exocyst, promotes notch signaling during the asymmetric division of Drosophila sensory organ precursors. Dev Cell. 2005 Sep;9(3):351-63. doi: 10.1016/j.devcel.2005.06.010. PMID: 16137928.

Khakurel A, Lupashin VV. Role of GARP Vesicle Tethering Complex in Golgi Physiology. Int J Mol Sci. 2023 Mar 23;24(7):6069. doi: 10.3390/ijms24076069. PMID: 37047041; PMCID: PMC10094427.

Lattner J, Leng W, Knust E, Brankatschk M, Flores-Benitez D. Crumbs organizes the transport machinery by regulating apical levels of PI(4,5)P2 in Drosophila. Elife. 2019 Nov 7;8:e50900. doi: 10.7554/eLife.50900. PMID: 31697234; PMCID: PMC6881148.

Lee T, Luo L. Mosaic analysis with a repressible cell marker for studies of gene function in neuronal morphogenesis. Neuron. 1999 Mar;22(3):451-61. doi: 10.1016/s08966273(00)80701-1. PMID: 10197526.

Liu S, Majeed W, Grigaitis P, Betts MJ, Climer LK, Starkuviene V, Storrie B. Epistatic Analysis of the Contribution of Rabs and Kifs to CATCHR Family Dependent Golgi Organization. Front Cell Dev Biol. 2019 Aug 2;7:126. doi: 10.3389/fcell.2019.00126. PMID: 31428608; PMCID: PMC6687757.

Perkins LA, Holderbaum L, Tao R, Hu Y, Sopko R, McCall K, Yang-Zhou D, Flockhart I, Binari R, Shim HS, Miller A, Housden A, Foos M, Randkelv S, Kelley C, Namgyal P, Villalta C, Liu LP, Jiang X, Huan-Huan Q, Wang X, Fujiyama A, Toyoda A, Ayers K, Blum A, Czech B, Neumuller R, Yan D, Cavallaro A, Hibbard K, Hall D, Cooley L, Hannon GJ, Lehmann R, Parks A, Mohr SE, Ueda R, Kondo S, Ni JQ, Perrimon N. The Transgenic RNAi Project at Harvard Medical School: Resources and Validation. Genetics. 2015 Nov;201(3):843-52. doi: 10.1534/genetics.115.180208. Epub 2015 Aug 28. PMID: 26320097; PMCID: PMC4649654.

Wu S, Mehta SQ, Pichaud F, Bellen HJ, Quiocho FA. Sec15 interacts with Rab11 via a novel domain and affects Rab11 localization in vivo. Nat Struct Mol Biol. 2005 Oct;12(10):879-85. doi: 10.1038/nsmb987. Epub 2005 Sep 11. PMID: 16155582.

Yeaman C, Grindstaff KK, Wright JR, Nelson WJ. Sec6/8 complexes on trans-Golgi network and plasma membrane regulate late stages of exocytosis in mammalian cells. J Cell Biol. 2001 Nov 12;155(4):593-604. doi: 10.1083/jcb.200107088. Epub 2001 Nov 5. PMID: 11696560; PMCID: PMC2198873.

Zhang XM, Ellis S, Sriratana A, Mitchell CA, Rowe T. Sec15 is an effector for the Rab11 GTPase in mammalian cells. J Biol Chem. 2004 Oct 8;279(41):43027-34. doi: 10.1074/jbc.M402264200. Epub 2004 Jul 29. PMID: 15292201.

Author Response

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

eLife assessment

The authors present a potentially useful model involving Ca2+ signaling in inflammasome activation. As it stands, it was felt that the data were not sufficient to support the model and the claims of the study are inadequately presented.

Public Reviews:

Reviewer #1 (Public Review):

This manuscript proposes a complex unclear model involving Ca2+ signaling in inflammasome activation. The experimental approaches used to study the calcium dynamics are problematic and the results shown are of inadequate quality. The major claims of this manuscript are not adequately substantiated.

Major concerns:

(1) The analysis of lysosomal Ca2+release is being carried out after many hours of treatment. Such evidence is not meaningful to claim that PA activates Ca2+ efflux from lysosome and even if this phenomenon was robust, it is not doubtful that such kinetics are meaningful for the regulation of inflammasome activation. Furthermore, the evidence for lysosomal Ca2+ release is indirect and relies on a convoluted process that doesn't make any conceptual sense to me. In addition to these major shortcomings, the indirect evidence of perilysosomal Ca2+ elevation is also of very poor quality and from the standpoint of my expertise in calcium signaling, the data are incredulous. The use of GCaMP3-ML1, transiently transfected into BMDMs is highly problematic. The efficiency of transfection in BMDMs is always extremely low and overexpression of the sensor in a few rare cells can lead to erroneous observations. The overexpression also results in gross mislocalization of such membrane-bound sensors. The accumulation of GCaMP3-ML1 in the ER of these cells would prevent any credible measurements of perilysosomal Ca2+ signals. A meaningful investigation of this process in primary macrophages requires the generation of a mouse line wherein the sensor is expressed at low levels in myeloid cells, and shown to be localized almost exclusively in the lysosomal membrane. The mechanistic framework built around these major conceptual and technical flaws is not especially meaningful and since these are foundational results, I cannot take the main claims of this study seriously.

Ans) We agree with the reviewer’s concern that transfection efficiency could be low in BMDMs together with possible mislocalization of GCAMP3-ML1. However, in our experiment, transfection of BMDM with test plasmids resulted in good expression of test proteins. Below, we present our data showing good transfection efficiency of BMDM cells, while a different plasmid was employed.

Author response image 1.

(2) The cytosolic Ca2+ imaging shown in Figure 1C doesn't make any sense. It looks like a snapshot of basal Ca2+ many hours after PA treatment - calcium elevations are highly dynamic. Snapshot measurements are not helpful and analyses of Calcium dynamics requires a recording over a certain timespan. Unfortunately, this technical approach has been used throughout the manuscript. Also, BAPTA-AM abrogates IL-1b secretion because IL-1b transcription is Ca2+ dependent - the result shown in figure 1D does not shed light on anything to do with inflammasome activation and it is misleading to suggest that.

Ans) We agree with the reviewer’s concern that snapshot could lead to false conclusion. We have not traced cytosolic Ca2+ content after treatment with LPS + PA. However, we have traced lysosomal Ca2+ and ER Ca2+ for more than 15 min, which was presented in Figure 4B. We also agree with the comment that BAPTA-AM might affect transcription of pro-IL-1β. We have conducted immunoblot analysis after treatment with LPS+PA in the presence of BAPTA-AM. Protein band of pro-IL-1β was not affected by BAPTA-AM treatment suggesting no effect of BAPTA-AM on transcription or translation of pro-IL-1β, which was added to Figure 1D, as suggested.

(3) Trpm2-/- macrophages are known to be hyporesponsive to inflammatory stimuli - the reduced secretion of IL-1b by these macrophages is not novel. From a mechanistic perspective, this study does not add much to that observation and the proposed role of TRPM2 as a lysosomal Ca2+ release channel is not substantiated by good quality Ca2+ imaging data (see point 3 above). Furthermore, the study assumes that TRPM2 is a lysosomal ion channel. One paper reported TRPM2 in the lysosomes but this is a controversial claim, with no replication or further development in the last 14 years. This core assumption can be highly misleading to readers unfamiliar with TRPM2 biology and it is necessary to present credible evidence that TRPM2 is functional in the lysosomal membrane of macrophages. Ideally, this line of investigation should rest on robust demonstration of TRPM2 currents in patch-clamp electrophysiology of lysosomes. If this is not technically feasible for the authors, they should at least investigate TRPM2 localization on lysosomal membranes of macrophages.

Ans) We agree with the reviewer’s comment that TRPM2. However, we have shown that TRPM2 current was not activated in the plasma membrane of BMDMs after treatment with LPS+PA. We also agree with the reviewer’s comment that inflammatory cytokine release from TRPM2 KO cells or inflammasome response of TRPM2 KO macrophages to ROS or nanoparticles has been reported to be reduced; however, the role of TRPM2 in metabolic inflammation or inflammasome activation in response to lipid stimulators has not been shown, as discussed in the new lines 9-10 from the bottom of page 18. Regarding the role of lysosomal TRPM2 in inflammation, we have shown that bafilomycin A1 treatment abrogated increase of cytosolic Ca2+ by LPS+PA (Figure 3-figure supplement 1D), supporting the role of lysosome and lysosomal Ca2+ in inflammasome activation by LPS+PA.

We agree with the reviewer’s comment that TRPM2 expression on lysosome needs to be tested. We conducted confocal microscopy after immunofluorescence staining using anti-TRMP2 and -LAMP2 antibodies, which showed a certain portion of TRPM2 was colocalized with LAMP-2. This result substantiating TRPM2 expression on lysosome of macrophages was incorporated as Figure 2-figure supplement 1A.

(4) Apigenin and Quercetin are highly non-specific and their effects cannot be attributed to CD38 inhibition alone. Such conclusions need strong loss of function studies using genetic knockouts of CD38 - or at least siRNA knockdown. Importantly, if indeed TRPM2 is being activated downstream of CD38, this should be easily evident in whole cell patch clamp electrophysiology. TRPM2 currents can be resolved using this technique and authors have Trpm2-/- cells for proper controls. Authors attempted these experiments but the results are of very poor quality. If the TRPM2 current is being activated through ADPR generated by CD38 (in response to PA stimulation), then it is very odd that authors need to include 200 uM cADPR to see TRPM2 current (Fig. 3A). Oddly, even these data cast great doubt on the technical quality of the electrophysiology experiments. Even with such high concentrations of cADPr, the TRPM2 current is tiny and Trpm2-/- controls are missing. The current-voltage relationship is not shown, and I feel that the results are merely reporting leak currents seen in measurements with substandard seals. Also 20 uM ACA is not a selective inhibitor of TRPM2 - relying on ACA as the conclusive diagnostic is problematic.

Ans) We agree with the reviewer’s comment that effects of apigenin and quercetin could be due to mechanisms other than inhibition of CD38-mediated inflammasome activation. Indeed, that is the reason we have used TRPM2 KO mice and cells. Small TRPM2 current after treatment with high concentrations of cADPr might suggest the minor role of plasma membrane of TRPM2 in macrophage. Regarding concern about ACA, we added data showing inhibition of IL-1β release in response to LPS+PA by ACA as a new Figure 3-figure supplement 1A.

(5) TRPM2 is expressed in many different cell lines. The broad metabolic differences observed by the authors in the Trpm2-/- mice cannot be attributed to macrophage-mediated inflammation. Such a conclusion requires the study of mice wherein Trpm2 is deleted selectively in macrophages or at least in the cells of the myeloid lineage.

Ans) We agree with the reviewer’s comment that TRPM2 in cells other than macrophage might have affected the results. Thus, we have conducted in vitro stimulation of TRPM2-KO primary peritoneal macrophages with LPS+PA. We have observed that IL-1β release of TRPM2-KO macrophages in response in vitro treatment with LPS+PA was significantly lower than that from wild-type macrophages (Figure 2C & D), showing the role of TRPM2 in macrophages in inflammasome activation by LPS+PA, which could be independent of TRPM2 in tissues or cells other than macrophages.

(6) The ER-Lysosome Ca2+ refilling experiments rely on transient transfection of organelle-targeted sensors into BMDMs. See point #1 to understand why I find this approach to be highly problematic. Furthermore, the data procured are also not convincing and lack critical controls (localization of sensors has not been demonstrated and their response to acute mobilization of Ca2+ has not been shown to inspire any confidence in these results).

Ans) We agree with the reviewer’s comment that transfection or ER-targeted Ca2+ sensor could have artifactual effects. However, we have studied ER-Lysosome Ca2+ experiment using not only GEM-CEPIAer but also using D1ER, a FRET-based ER Ca2+ sensor which has an advantage of short distance of molecular interaction. Thus, we believe that changes of ER Ca2+ after treatment with LPS+PA is not due to an artifactual effect. Multiple contact between VAPA and ORP1L (Figure 4E) also supports ER-lysosome contact, likely facilitating ER-lysosome Ca2+ flux.

(7) Authors claim that SCOE is coupled to K+ efflux. But there is no credible evidence that SOCE is activated in PA stimulated macrophages. The data shown in Fig 4 supp 1 do not investigate SOCE in a reliable manner - the conclusion is again based on snapshot measurements and crude non-selective inhibitors. The correct way to evaluate SOCE is to record cytosolic Ca2+ elevations over a period of time in absence and presence of extracellular Ca2+. However, even such recordings can be unreliable since the phenomenon is being investigated hours after PA stimulation. So, the only definitive way to demonstrate that Orai channels are indeed active during this process is through patch clamp electrophysiology of PA stimulated cells.

Ans) We agree with the reviewer’s comment that the final proof of SOCE activation is activation of Orai channel evidenced by electrophysiology. However, we have shown STIM1 aggregation colocalized with Ora1, which is another strong evidence of SOCE channel activation (Vaca L. Cell Calcium 47:199, 2010). Such a paper showing the role of SOCE aggregation in SOCE activation was incorporated in the text (line 4 from the bottom of page 10) and References.

Reviewer #2 (Public Review):

In this manuscript by Kang et. al., the authors investigated the mechanisms of K+-efflux-coupled SOCE in NLRP3 inflammasome activation by LP（LPS+PA, and identified an essential role of TRPM2-mediated lysosomal Ca2+ release and subsequent IP3Rs-mediated ER Ca2+ release and store depletion in the process. K+ efflux is shown to be mediated by a Ca2+-activated K+ channel (KCa3.1). LP-induced cytosolic Ca2+ elevation also induced a delayed activation of ASK1 and JNK, leading to ASC oligomerization and NLRP3 inflammasome activation. Overall, this is an interesting and comprehensive study that has identified several novel molecular players in metabolic inflammation. The manuscript can benefit if the following concerns could be addressed:

(1) The expression of TRPM2 in the lysosomes of macrophages needs to be more definitively established. For instance, the cADPR-induced TRPM2 currents should be abolished in the TRPM2 KO macrophages. Can you show the lysosomal expression of TRPM2, either with an antibody if available or with a fluorescently-tagged TRPM2 overexpression construct?

Ans) We agree with the reviewer’s comment that TRPM2 expression on lysosome needs to be tested. We conducted confocal microscopy after immunofluorescent staining using anti-TRMP2 and -LAMP2 antibodies, which showed a certain portion of TRPM2 was colocalized with LAMP2. This result was incorporated as Figure 2-figure supplement 1A.

(2) Can you use your TRPM2 inhibitor ACA to pharmacologically phenocopy some results, e.g., about [Ca2+]ER, [Ca2+]LY, and [Ca2+]i from the TRPM2 knockout? Ans) We agree with the reviewer’s comment that the effect of ACA on other experimental results needs to be shown. We did not study the effect of ACA on Ca2+ flux; however, we have observed that ACA inhibited IL-1β release in response to LPS+PA. This data was incorporated as Figure 3-figure supplement 1A.

Author response image 2.

(3) In Fig. S4A, bathing the cells in zero Ca2+ for three hours might not be ideal. Can you use a SOCE inhibitor, e.g, YM-58483, to make the point?

Ans) We agree with the reviewer’s comment that SOCE inhibitor experiment would be necessary in addition to the experiment employing zero Ca2+. In fact, we have already used two SOCE inhibitors (2-APB and BTP2) (Figure 4-fig. supplement 1 B-D. Particularly, BTP2 experiment could eliminate possible role of ER Ca2+ inhibition that might occur when 2-APB was employed.

(4) In Fig. 1A, you need a positive control, e.g., ionomycin, to show that the GPN response was selectively reduced upon LP treatment.

Ans) We did not employ ionomycin as a control in this study. In our previous study using other agents inducing lysosomal Ca2+ efflux, we have observed lysosomal Ca2+ efflux with intact subsequent ionomycin response. While we did not include ionomycin in the current paper, we are positive that ionomycin response would be preserved.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

See Public Review.

Reviewer #2 (Recommendations For The Authors):

(5) In Fig. 4B, the red label should read "BAPTA-1 Dextran", but not "GAPTA-1 Dextran".

(6) Writing should be improved in many sections.

Author Response

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

Response to the Referee Comments We would like to express our appreciation to the editor and the reviewers for their thoughtful comments and constructive suggestions on the manuscript. We agree with most of the comments and have carefully revised the manuscript accordingly. The revisions are highlighted in red font in the revised manuscript. Below are point-by-point responses to the referee’s comments.

Public Reviews:

Reviewer #1 (Public Review):

Microglia are increasingly recognized as playing an important role in shaping the synaptic circuit and regulating neural dynamics in response to changes in their surrounding environment and in brain states. While numerous studies have suggested that microglia contribute to sleep regulation and are modulated by sleep, there has been little direct evidence that the morphological dynamics of microglia are modulated by the sleep/wake cycle. In this work, Gu et al. applied a recently developed miniature two-photon microscope in conjunction with EEG and EMG recording to monitor microglia surveillance in freely-moving mice over extended period of time. They found that microglia surveillance depends on the brain state in the sleep/wake cycle (wake, non-REM, or REM sleep). Furthermore, they subjected the mouse to acute sleep deprivation, and found that microglia gradually assume an active state in response. Finally, they showed that the state-dependent morphological changes depend on norepinephrine (NE), as chemically ablating noradrenergic inputs from locus coeruleus abolished such changes; this is in agreement with previous publications. The authors also showed that the effect of NE is partially mediated by β2-adrenergic receptors, as shown with β2-adrenergic receptor knock-out mice. Overall, this study is a technical tour de force, and its data add valuable direct evidence to the ongoing investigations of microglial morphological dynamics and its relationship with sleep. However, there are a number of details that need to be clarified, and some conclusions need to be corroborated by more control experiments or more rigorous statistical analysis. Specifically:

1. The number of branch points per microglia shown here (e.g., Fig. 2g) is much lower than the values of branch points in the literature, e.g., Liu T et al., Neurobiol. Stress 15: 100342, 2021 (mouse dmPFC, IHC); Liu YU et al., Nat. Neurosci. 22: 1771-81, 2019 (mouse S1, in vivo 2P imaging). The authors need to discuss the possible source of such discrepancy.

Thank you for raising this important point. Two reasons may account for this difference. Firstly, the difference in the definition of branch points in the software. Liu YU et al. used the Sholl analysis of image J software to analyze the number of branch points of microglia. Sholl analysis defines the number of branch points as the number of crossings between branches and concentric circles of increasing radii. We reconstructed microglia morphology using Imaris, a software that defines branching points based on the number of bifurcation points. The number of bifurcations calculated represents the number of microglia branch points. Secondly, this and previous studies found that more branching points present in the state of anesthesia. The morphological characteristics of microglia in head-fixed mice under anesthesia was reported by Liu T et al. and the microglia reconstruction results presented by the authors are indeed more complex than ours. In short, this is an aspect that we have been paying attention to, and the main reasons for this difference may lie in the definition of branch points, analysis methods and related choice of thresholds. True differences in brain states and the heterogeneity of microglia in different brain regions may also contribute to the apparent discrepancy.

1. Microglia process end-point speed (Fig. 2h, o): here the authors show that the speed is highest in the wake state and lowest in NREM, which agrees with the measurement on microglia motility during wakefulness vs NREM in a recent publication (Hristovska I et al., Nat. Commun. 13: 6273, 2022). However, Hristovska et al. also reported lower microglia complexity in NREM vs wake state, which seems to be the opposite of the finding in this paper. The authors need to discuss the possible source of such differences.

This is also an important point. Hristovska et al. reported the morphodynamic characteristics of microglia during wakefulness and NREM sleep. It is worth noting that the sleep state of the mice in their experiments was unnatural due to the head fixation and body limitations, the duration of NREM sleep (sleep stability) being quite different from the NREM sleep analyzed under natural sleep. The limitations of this approach are also discussed by Hristovska et al. “Even though sleep episodes were, as anticipated, shorter than those observed in freely moving animals, changes in neuronal activity characteristic of NREM sleep were monitored by EEG recordings, and changes in morphodynamics were observed during single episodes. Several episodes of REM sleep were detected, but they were too short and rare to be analyzed reliably.” The unnatural sleep state would lead to an increase in the microarousal state, and ultimately a change in the structure of the sleep state, which may be the main reason for the difference in microglia behavior from our natural sleep. We have discussed this in the revised manuscript. Please see line 292298.

1. Fig. 3: the authors used single-plane images to analyze the morphological changes over 3 or 6 hours of SD, which raises the concern that the processes imaged at the baseline may drift out of focus, leading to the dramatic reduction in process lengths, surveillance area, and number of branch points. In fact, a previous study (Bellesi M et al., J. Neurosci. 37(21): 5263-73, 2017) shows that after 8 h SD, the number of microglia process endpoints per cell and the summed process length per cell do not change significantly (although there is a trend to decline). The authors may confirm their findings by either 3D imaging in vivo, or 3D imaging in fixed tissue.

Three lines of evidence indicate that microglia morphology changes in Fig 3 are due to SD, rather than variations in the focal plane. First, our single-plane images were quite stable over 3 or 6 hours of SD, though occasional reversible drifts might happen due to sudden motions. Second, per your suggestion, further experiments and analysis of 3D imaging were performed to monitor microglia dynamics during sleep deprivation. The new result is shown in revised Fig. S3 C-D: the length of microglia branches and the number of branching points were significantly reduced after SD, in agreement with the results of single-plane imaging. Furthermore, we detected no significant difference in microglia branching characteristics during 6h sleep deprivation in 2AR KO mice (Fig.S4), and this indirectly affirmed that singleplane imaging is stable enough for detecting true changes in branching during SD.

1. Fig. 4b: the EEG and EMG signals look significantly different from the example given in Fig. 2a. In particular, the EMG signal appears completely flat except for the first segment of wake state; the EEG power spectrum for REM appears dark; and the wake state corresponds to stronger low frequency components (below ~ 4 Hz) compared to NREM, which is the opposite of Fig. 2a. This raises the concern whether the classification of sleep stage is correct here.

Thank you for insightful comments. We carefully examined the behavioral video of Figure 4b, there were occasionally microarousal events indicated by slow head rotation during NREM sleep, while the companion EMG signals were completely flat, which is atypical during sleep wake cycle. The microarousal events were not excluded from sleep, which makes this set of data unrepresentative and contrary to Fig.4b. In our revised manuscript, we replaced it with more representative data that can clearly and consistently distinguish between different brain states in mice on EMG and EEG. Please see revised Fig.2a, page 34; revised Fig.4b, page 37.

1. Fig. 4 NE dynamics. • How long is a single continuous imaging session for NE? • When monitoring microglia surveillance, the authors were able to identify wake or NREM states longer than 15 min, and REM states longer than 5 min. Here the authors selected wake/NREM states longer than 1 min and REM states longer than 30 s. What makes such a big difference in the time duration selected for analysis? • Also, the definition of F0 is a bit unclear. Is the same F0 used throughout the entire imaging session, or is it defined with a moving window?

A single continuous session of NE imaging usually took about 1 hour. Subsequent analysis was performed on imaging data from each recording that included wake, NREM sleep, and REM sleep. Because of the different time scales of microglia morphological dynamic (relatively slow) and NE signals (fast), we used different time windows in the previous analysis in the previous version of the manuscript.

Per your suggestion, we have now set the same time window selection criteria for both microglia morphological and NE dynamic analysis: for wake and NREM sleep durations longer than 1 minute, and REM sleep durations longer than 30 seconds. We updated the Methods and all statistics in related figures, please see line 151-154, 481-485, 490-492; Fig. 2e-g and 2l-n, page 34. F0 definition is now explained in the Methods section. Please see line 521-522.

1. Fig. 5b: how does the microglia morphology in LC axon ablation mice compare with wild type mice under the wake state? The text mentioned "more contracted" morphology but didn't give any quantification. Also, the morphology of microglia in the wake state (Fig. 5b) appears very different from that shown in Fig. S3C1 (baseline). What is the reason?

The morphology of microglia is indeed heterogeneous and variable, affected by factors including brain state, brain region, microenvironmental changes, along with animal-to-animal difference. We didn’t perform the microglia morphology comparison between the LC axon ablation mice and wild type mice and, in view of this, we removed the description of “more contracted morphology” from the main text. It should also be noted that, as we primarily focused on changes of a microglia in different states over time by selfcomparison, we minimized possible effects of heterogeneity in microglia morphology on our conclusions.

1. The relationship between NE level and microglia dynamics. Fig. 4C shows that the extracellular NE level is the highest in the wake state and the lowest in REM. Previous studies (Liu YU et al., Nat. Neurosci. 22(11):1771-1781, 2019; Stowell RD et al., Nat. Neurosci. 22(11): 1782-1792, 2019) suggest that high NE tone corresponds to reduced microglia complexity and surveillance. Hence, it would be expected that microglia process length, branch point number, and area/volume are higher in REM than in NREM. However, Fig. 2l-n show the opposite. How should we understand this ?

Your point is well-taken. On the one hand, our data clearly showed that NE is critically involved in the brain state-dependent microglia dynamic surveillance, with evidence from the ablation of the LC-NE projection and from the β2AR knockout animal model.

On the other hand, we also understand that NE is not the sole determinant, so the relationship between the NE level and the complexity and surveillance may not be unique.

In this regard, other potential modulators also present dynamic during sleepwake cycle and may partake in the regulation of microglia dynamic surveillance. previous studies (Liu YU et al., 2019; Stowell RD et al., 2019) have shown that microglia can be jointly affected by surrounding neuronal activity and NE level during wake. It has been reported that LC firing stops (Aston-Jones et al., 1981; Rasmussen et al., 1986), while inhibitory neurons, such as PV neurons and VIP neurons, become relatively active during REM sleep (Brécier et al., 2022). ATP level in basal forebrain is shown to be higher in REM than NREM (Peng et al., 2023). In addition, our own preliminary result (Author response image 1) also showed a higher adenosine level in REM than NREM in somatosensory cortex. Last but not the least, we found that β2AR knockout failed to abolish microglial responses to sleep state switch and SD stress altogether.

In brief, microglia are highly sensitive to varied changes in the surrounding environment, and many a modulator may participate in the microglia dynamic during sleep state. This may underlie the microglia complexity difference between REM and NREM. Future investigations are warranted to delineate the signal-integrative role of microglia in physiology and under stress. We have discussed the pertinent points in the revised manuscript. Please see line 343-354.

Author response image 1.

Extracellular adenosine levels in somatosensory cortex in different brain states. AAV2/9-hSyn-GRABAdo1.0 (Peng W. et al., Science. 2020) was injected into the somatosensory cortex (A/P, -1 mm; M/L, +2 mm; D/V, -0.3 mm). Data from the same recording are connected by lines. n = 9 from 3 mice.

Reviewer #2 (Public Review):

The manuscript describes an approach to monitor microglial structural dynamics and correlate it to ongoing changes in brain state during sleep-wake cycles. The main novelty here is the use of miniaturized 2p microscopy, which allows tracking microglia surveillance over long periods of hours, while the mice are allowed to freely behave. Accordingly, this experimental setup would permit to explore long-lasting changes in microglia in a more naturalistic environment, which were previously not possible to identify otherwise. The findings could provide key advances to the research of microglia during natural sleep and wakefulness, as opposed to anesthesia. The main findings of the paper are that microglia increase their process motility and surveillance during REM and NREM sleep as compared to the awake state. The authors further show that sleep deprivation induces opposite changes in microglia dynamics- limiting their surveillance and size. The authors then demonstrate potential causal role for norepinephrine secretion from the locus coeruleus (LC) which is driven by beta 2 adrenergic receptors (b2AR) on microglia. However, there are several methodological and experimental concerns which should be addressed.

The major comments are summarized below:

1. The main technological advantage of the 2p miniaturized microscope is the ability to track single cells over sleep cycles. A main question that is unclear from the analysis and the way the data is presented is: are the structural changes in microglia reversible? Meaning, could the authors provide evidence that the same cell can dynamically change in sleep state and then return to similar size in wakefulness? The same question arises again with the data which is presented for anesthesia, is this change reversible?

As revealed by long-term free behavioral mTPM imaging, the brain-statedependent morphological changes in microglia were reproducible and reversible. Author response image 2 shows that microglia displayed reversible dynamic changes during multiple rounds of sleep-wake transition. Author response image 3 shows that microglia dynamics induced by anesthesia also exhibited reversibility.

Author response image 2.

Long-term tracking of microglia process area in different brain states. Data analysis used 8 cells. Data total of 31 time points were selected from in vivo imaging data and were used to characterize the morphological changes of microglia over a continuous 7-hour period.

Author response image 3.

Reversible changes of microglial process length, area, number of branch points under anesthesia. Wake group: 30 minute-accommodation to new environment; Isoflurane group: 1.5% in air applied at a flow rate of 0.4 L/min for 30 minutes; Recovery group: 30 minutes after recovery from anesthesia. n = 9 cells from 3 mice for each group.

1. The binary comparison between brain states is misleading, shouldn't the changes in structural dynamics compared to the baseline of the state onset? The authors method describes analysis of the last 5 minutes in each sleep/wake state. However, these transitions are directional- for instance, REM usually follows NREM, so the description of a decrease in length during REM sleep could be inaccurate.

As you know, the time scale of microglia morphological dynamic is relatively slow, so we analyzed the microglia morphological dynamic of the last part (30s in the revised manuscript) of each state instead of the state onset, allowing time for stabilization of the microglia response to inter-state transition.

Further, we compared microglia dynamic between two NREM groups transiting to different subsequent states: group1 (NREM to REM) vs group2 (NREM to Wake). This precaution was to exclude the directional effect of state transitions. Our results showed that there was no difference in microglial length, area, number of branching points between the two NREM groups (Author response image 4), indicating that the last 30s of each NREM was not affected by its following state and that it’s reasonable to perform binary comparison.

Author response image 4.

Microglial morphological length, area change, and number of branch points of the last 30s of NREM sleep followed by REM or Wake. n = 9 cells from 3 mice for each group.

1. Sleep deprivation- again, it is unclear whether these structural changes are reversible. This point is straightforward to address using this methodology by measuring sleep following SD. In addition, the authors chose a method to induce sleep deprivation that is rather harsh. It is unclear if the effect shown is the result of stress or perhaps an excess of motor activity.

We adopted the method of forced exercise as it has been commonly used for sleep deprivation (Pandi-Perumal et al., 2007; Nollet M et al., 2020), though it does have the potential limitation of excess of motor activity.

In light of your comments and suggestion, we presented new data demonstrating that sleep duration of the mice, mostly NREM sleep, increased compensatively (ZT9-10) after the 6-hour sleep deprivation (ZT2-8) (revised Fig. S3B). This result shows that sleep deprivation indeed increase sleep pressure in the mice. As the sleep pressure was eased during recovery sleep, morphological changes of microglia were reversed over a timescale of several hours (revised Fig. S3 E-J).

1. The authors perform measurements of norepinephrine with a recently developed GRAB sensor. These experiments are performed to causally link microglia surveillance during sleep to norepinephrine secretion. They perform 2p imaging and collect data points which are single neurons, and it is unclear why the normalization and analysis is performed for bulk fluorescence similar to data obtained with photometry.

We did not perform single-neuron analysis for two reasons. First, our experimental conditions, e.g., the expression of the NE indicator and the control of imaging laser intensity, did not yield sufficient signal-to-noise to clearly discriminate individual neurons with two-photon imaging. Second, NE signal may play a modulatory role, and fluorescence changes appeared to be global, rather than local or cell-specific. Therefore, we analyzed fluorescence changes in different brain states over the whole field-of-view in Fig. 4, rather than at the subregional or single-cell level.

1. The experiments involving b2AR KO mice are difficult to interpret and do not provide substantial mechanistic insight. Since b2AR are expressed throughout numerous cell types in the brain and in the periphery, it is entirely not clear whether the effects on microglia dynamics are direct. The conclusion and the statement regarding the expression of b2AR in microglia is not supported by the references the authors present, which simply demonstrate the existence and function of b2AR in microglia. In addition, these mice show significant changes in sleep pattern and increased REM sleep. This could account for reasons for the changes in microglia structure rather than the interpretation that these are direct effects.

To summarize, the main conclusions of the paper require further support with analysis of existing data and experimental validation.

Previous studies have revealed that norepinephrine (NE) has a modulating effect on microglial dynamics through β2AR pathway (Stowell RD et al., 2019; Liu YU et al., 2019). Stowell et al. and Liu et al. use in vivo two-photon imaging to demonstrate that microglia dynamics differ between awake and anesthetized mice and to highlight the roles of NE and β2AR in these states (Gyoneva S et al., 2013; Stowell RD et al., 2019; Liu YU et al., 2019). To evaluate the direct effect of β2AR on microglial dynamics, Stowell et al. administered the β2AR agonist clenbuterol to anesthetized mice and found that this decreased the motility, arbor complexity, and process coverage of microglia in the parenchyma (Stowell RD et al., 2019). Inhibition of β2AR by antagonist ICI-118,551 in awake mice recapitulated the effects of anesthesia by enhancing microglial arborization and surveillance (Stowell RD et al., 2019). In addition, it has been shown microglia expressed higher numbers of β2ARs than any other cells in the brain (Zhang et al., 2014).

To this end, our current work provided new evidence to support the involvement of the LC-NE-β2AR axis in modulating microglia dynamics both during natural sleep-wake cycle and under SD stress. While we were aware the limitation of using pan-tissue β2AR knockout model that precluded us from pinpointing role of microglial β2AR, it is safe to state that β2-adrenergic receptor signaling plays a significant role in the sleep-state dependent microglia dynamic surveillance, based on the present and previous data.

We have discussed this in the revised manuscript. Please see line 324-354. As you suggested, we added references to support the statement regarding the expression of β2AR in microglia (please see line 333).

Recommendations for the authors: please note that you control which, if any, revisions, to undertake

Reviewer #1 (Recommendations For The Authors):

Some technical details need to be clarified. Also, please double-check for typos.

1. In vivo imaging preparation: how long is the recovery time between window/EEG implantation surgery and imaging/recording?

Imaging data were collected one month after the surgery. We have added descriptions to the methods section of the revised manuscript. Please see line 419.

1. Statistical analysis: the authors used t-test or ANOVA without first checking whether the data pass the normality test. If the data does not follow a normal distribution, nonparametric tests would be more appropriate.

Per your suggestion, we performed the test of statistical significance using parametric (ANOVA) if past the normality test, or the non-parametric (Friedman) tests for non-normal data. Please see line 533-535.

1. Fig. 1b needs a minor change. In the figure, the EMG electrodes appear to be connected to the brain as well.

We have corrected this oversight. Thank you.

1. Fig. 1c: it would be helpful to give examples of raw EEG and EMG traces for REM and NREM separately.

Raw traces are now shown as suggested. Please see Fig. 1c, page 32.

1. Fig. 1h: is each data point one microglia or one end-point?

In Fig. 1h, each data represents the average speed of all branches of one microglia, not one end-point.

1. Sleep deprivation starts at 9 am. What time corresponds to Zeitgeber Time 0 (ZT0, the beginning of the light phase)?

We now clarified that 9 am corresponds to Zeitgeber time 2. Please see line 196.

1. Line 61: the authors referred to Ramon y Cajal's original suggestion that microglia dynamics are coupled to the sleep-wake cycle. However, the cited paper only indicates that Cajal suggested a role of astrocytes in the sleep-wake cycle, not microglia. In addition, there is a typo in the line: there should be a space between "Ramon" and "y" in Cajal's name.

We have updated the statement and reference literature to point out the microglia’s involvement in the sleep-wake cycle. The typo was corrected. Please see line 64-65.

1. Fig. S3B: As each group has only 3 mice, it is unclear how t-test can yield p < 0.01 or even 0.001.

We checked the original data again and it was correct. This small p-values may be due to the small intra-group difference of control group.

1. Line 251-253, "Figure 4h-n" should be "Figure 5h-n"?

We have revised it. Please see line 265-266.

1. Fig. 5h: the receptor should be "adrenergic receptor", not "adrenal receptor".

We changed the term to “adrenergic receptor”. Please see Fig 5h.

1. Fig. 5g, n: the number of data points is apparently less than the sample size given in the figure legend. Perhaps some data points have exactly the same value so they overlap? The authors may consider plotting identical values with a slight shift so that the number of data points shown matches the actual sample size, to avoid confusion.

Yes, we have added small jitters so different data points can be seen to avoid confusion. Please see Fig. 5n.

1. There are some typos (e.g., Line 217, "he" should be "the") and some incomplete references (e.g., [13], [22], [34], [35] lack volume and page number, [15] and [39] lack publisher information). Some references have inconsistent formats (e.g., "Journal of Neuroscience" is sometimes abbreviated and sometimes not). Please correct these.

We have corrected these oversights. Please see references, page 27.

Reviewer #2 (Recommendations For The Authors):

Major issues:

1. Re-analyze the data in a manner that allows to follow and compare the same cells over different state transitions. This is necessary to evaluate the reversibility of microglia structure. In addition, consider analysis of the change from the beginning to the end of each state.

As shown in response figure 2, microglia dynamics were reversible during multiple rounds of sleep-wake transition.

1. It would be nice to see the raw data obtained over time, at least for Figure 1, before offline correction of movement to evaluate the imaging quality and level of drift during imaging.

We agree to your good suggestion. Please see the supporting material video.

1. It would be helpful to add an analysis of the percent time spent in each state for the 10 hour recordings.

Advice has been adopted. Please see revised Fig. S4C.

1. In Figure 2 the results are from 15 cells from several animals. How much do the results vary between mice? It will be helpful to show if this varies between different mice by labeling cells from each mouse differently.

In Author response image 5, in which we have labeled the distribution of data points from seven mice, there was mixed distribution of data from different animals at each brain state, but no clear animal-to-animal difference.

Author response image 5.

Quantitative analysis of microglial length based on multi-plane microglial imaging. n = 17 cells from 7 mice for each group. In right panel, each color codes data from the same animal.

1. SD- please add some quantification for sleep and EEG to show that the manipulation really caused sleep deprivation. To address the confound of forced movement and stress, it might be helpful to add quantification of movement compared to an undisturbed wakefulness.

We have added related data (revised Fig. S3B), as suggested. Please see line 196-197.

1. The DSP4 application should be also performed with NE measurements to verify the specific of the NE signal measured as well as the DSP4 toxin.

Following your suggestion, we have added DSP4 data in revised Fig. S4B.

1. Some suggested refined experiments for the b2AR KO are: a-A conditional b2AR KO in microglia, as cited in the work. b- Local application of a b2 blocker during SD. c- Imaging of NE dynamics in the b2 animals. If NE dynamics during natural sleep cycle are perturbed, then this suggests upstream mechanisms rather than direct microglia effects as suggested by the authors.

We agree that the current study cannot pinpoint a direct effect of microglia harbored β2AR. We have discussed this limitation in the revised manuscript.

Please see line 324-354.

Minor:

1. Typo on page 4 (microcopy instead of microscopy).

It was corrected. Please see line 87.

1. Typo page 11- 'and he largest changes in NE' - supposed to be 'the'.

We have corrected these mistakes. Please see line 228.

1. Fig. 4- there are several units missing in the figure in panel b: the top is Hz, but what does the color bar indicate exactly? 2 what? both for theta/delta and for NE. We have modified this figure and legend for clarity. Please see Fig. 4, page 37.

2. Bottom of page 12- referring to figure 4 but talking about figure 5.

The typo was corrected. Please see line 265-266.

Reference

1. Aston-Jones G, Bloom FE. Activity of norepinephrine-containing locus coeruleus neurons in behaving rats anticipates fluctuations in the sleep-waking cycle. J Neurosci. 1, 876–886 (1981).

2. Bellesi M, de Vivo L, Chini M, Gilli F, Tononi G, Cirelli C. Sleep loss promotes astrocytic phagocytosis and microglial activation in mouse cerebral cortex. J Neurosci. 37, 5263–5273 (2017).

3. Brécier A, Borel M, Urbain N, Gentet LJ. Vigilance and behavioral state-dependent modulation of cortical neuronal activity throughout the sleep/wake cycle. J Neurosci. 42, 4852–66 (2022).

4. Dworak M, McCarley RW, Kim T, Kalinchuk AV, Basheer R. Sleep and brain energy levels: ATP changes during sleep. J Neurosci. 30, 9007-16 (2010).

5. Gyoneva S., Traynelis SF. Norepinephrine modulates the motility of resting and activated microglia via different adrenergic receptors. J Biol Chem. 288, 15291302 (2013).

6. Kjaerby C, Andersen M, Hauglund N, Untiet V, Dall C, Sigurdsson B, Ding F, Feng J, Li Y, Weikop P, Hirase H, Nedergaard M. Memory-enhancing properties of sleep depend on the oscillatory amplitude of norepinephrine. Nat Neurosci. 25, 1059–1070 (2022).

7. Liu T, Lu J, Lukasiewicz K, Pan B, Zuo Y. Stress induces microglia-associated synaptic circuit alterations in the dorsomedial prefrontal cortex. Neurobiology of Stress. 15, 100342 (2021).

8. Liu YU, Ying Y, Li Y, Eyo UB, Chen T, Zheng J, Umpierre AD, Zhu J, Bosco DB, Dong H, Wu LJ. Neuronal network activity controls microglial process surveillance in awake mice via norepinephrine signaling. Nat Neurosci. 22, 1771–1781 (2019).

9. Nollet M, Wisden W, Franks NP. Sleep deprivation and stress: a reciprocal relationship. Interface Focus. 10, 20190092 (2020).

10. Pandi-Perumal SR, Cardinali DP, Chrousos GP. 2007. Neuroimmunology of sleep. New York, NY: Springer.

11. Peng W, Liu X, Ma G, Wu Z, Wang Z, Fei X, Qin M, Wang L, Li Y, Zhang S, Xu M. Adenosine-independent regulation of the sleep-wake cycle by astrocyte activity. Cell Discov. 9, 16 (2023).

12. Peng W, Wu Z, Song K, Zhang S, Li Y, Xu M. Regulation of sleep homeostasis mediator adenosine by basal forebrain glutamatergic neurons. Science. 369, 6508 (2020).

13. Rasmussen K, Morilak DA, Jacobs BL. Single unit activity of locus coeruleus neurons in the freely moving cat: I. During naturalistic behaviors and in response to simple and complex stimuli. Brain Research. 371, 324–334 (1986).

14. Stowell RD, Sipe GO, Dawes RP, Batchelor HN, Lordy KA, Whitelaw BS, Stoessel MB, Bidlack JM, Brown E, Sur M, Majewska AK. Noradrenergic signaling in the wakeful state inhibits microglial surveillance and synaptic plasticity in the mouse visual cortex. Nat Neurosci. 22, 1782-1792 (2019).

15. Umpierre AD, Bystrom LL, Ying Y, Liu YU, Worrell G, Wu LJ. Microglial calcium signaling is attuned to neuronal activity in awake mice. Elife. 27, e56502 (2020).

16. Wang Z, Fei X, Liu X, Wang Y, Hu Y, Peng W, Wang YW, Zhang S, Xu M. REM sleep is associated with distinct global cortical dynamics and controlled by occipital cortex. Nat Commun. 13, 6896 (2022).

17. Zhang Y, Chen K, Sloan SA, Bennett ML, Scholze AR, O’Keeffe S, Phatnani HP, Guarnieri P, Caneda C, Ruderisch N, Deng S, Liddelow SA, Zhang C, Daneman R, Maniatis T, Barres BA, Wu JQ. An RNA-sequencing transcriptome and splicing database of glia, neurons, and vascular cells of the cerebral cortex. J Neurosci. 34, 11929–11947 (2014).

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

The paper proposes an interesting perspective on the spatio-temporal relationship between FC in fMRI and electrophysiology. The study found that while similar network configurations are found in both modalities, there is a tendency for the networks to spatially converge more commonly at synchronous than asynchronous time points. However, my confidence in the findings and their interpretation is undermined by an apparent lack of justification for the expected outcomes for each of the proposed scenarios, and in the analysis pipeline itself.

Main Concerns

(1) Figure 1 makes sense to me conceptually, including the schematics of the trajectories, i.e.

Scenario 1: Temporally convergent, same trajectories through connectome state space

Scenario 2: Temporally divergent, different trajectories through connectome state space

However, based on my understanding I am concerned that these scenarios do not necessarily translate into the schematic CRP plots shown in Figure 2C, or the statements in the main text:

For Scenario 1: "epochs of cross-modal spatial similarity should occur more frequently at on-diagonal (synchronous) than off-diagonal (asynchronous) entries, resulting in an on-/off-diagonal ratio larger than unity"

For Scenario 2: "epochs of spatial similarity could occur equally likely at on-diagonal and off-diagonal entries (ratio≈1)"

Where do the authors get these statements and the schematics in Figure 2C from? Are they based on previous literature, theory, or simulations?

I am not convinced based on the evidence currently in the paper, that the ratio of off- to on-diagonal entries (and under what assumptions) is a definitive way to discriminate between scenarios 1 and 2.

For example, what about the case where the same network configuration reoccurs in both modalities at multiple time points? It seems to me that one would get a CRP with entries occurring equally on the on-diagonal as on the off-diagonal, regardless of whether the dynamics are matched between the two modalities or not (i.e. regardless of scenario 1 or 2 being true).

This thought experiment example might have a flaw in it, and the authors might ultimately be correct, but nonetheless, a systematic justification needs to be provided for using the ratio of off- to on-diagonal entries to discriminate between scenarios 1 and 2 (and under what assumptions it is valid).

In the absence of theory, a couple of ways I can think of to gain insight into this key aspect are:

(1) Use surrogate data for scenarios 1 and 2:

a. For scenario 1: Run the CRP using a single modality. E.g. feed in the EEG into the analysis as both modality 1 AND modality 2. This should provide at least one example of CRP under scenario 1 (although it does not ensure that all CRPs under this scenario will look like this, it is at least a useful sanity check)

b. For scenario 2: Run the CRP using a single modality plus a shuffled version. E.g. feed in the EEG into the analysis as both modality 1 AND a temporally shuffled version of the EEG as modality 2. The temporal shuffling of the EEG could be done by simply splitting the data into blocks of say ~10s and then shuffling them into a new order. This should provide a version of the CRP under scenario 2 (although it does not ensure that all CRPs under this scenario will look like this, it is at least a useful sanity check).

(2) Do simulations, with clearly specified assumptions, for scenarios 1 and 2. One way of doing this is to use a simplified (state-space) setup and randomly simulate N spatially fixed networks that are independently switching on and off over time (i.e. "activation" is 0 or 1). Note that this would result in a N-dimensional connectome state space.

The authors would only need to worry about simulating the network activation time courses, i.e. they would not need to bother with specifying the spatial configuration of each network, instead, they would make the implied assumption that each of these networks has the same spatial configuration in modality 1 and modality 2.

With that assumption, the CRP calculation should simply correspond to calculating, at each time i in modality 1 and time j in modality 2, the number of networks that are activating in both modality 1 and modality 2, by using their activation time courses. Using this, one can simulate and compute the CRPs for the two scenarios:

a. Scenario 1: where the simulated activation timecourses are set to be the same between both modalities

b. Scenario 2: where the simulated activation timecourses are simulated separately for each of the modalities

We thank the reviewer for raising this important matter as it directly relates to our study hypothesis. To address this point, we chose to focus on the first of the two alternative suggestions of the reviewer, as it provides evidence based on empirical data. In line with the reviewer’s suggestion 1, recurrence plots have indeed been previously applied to connectome dynamics data from the same modality [Hansen et al., NeuroImage 2015; Fig. 2B]. As shown in the referenced study, where the recurrence plot has been estimated within fMRI connectome dynamics, the on-diagonal entries have noticeably larger correlation values in comparison to off-diagonal entries. As the authors state, this contrast emphasizes the autocorrelation of connectome dynamics in their single modality recurrence plot. Extending these findings to our cross-modal recurrence plots, more synchronicity of connectome dynamics across fMRI and EEG will -by theory- translate into stronger correlation values along the diagonal axis as it represents neighboring timepoints in the data. On the other hand, less cross-modal synchronicity translates to a lack of such correlation prevalence along the diagonal axis.

Complementing these statements with empirical data, Author response image 1 shows the fMRI-to-iEEG and fMRI-to-fMRI CRPs side by side as suggested by the reviewer. For simplicity, we thresholded each CRP at the top 5% of entries and calculated their corresponding on-/off-diagonal ratios. The on/off-diagonal ratio for fMRI-to-fMRI CRP was 4.32 ± 6.26 across -5 to +5 TR lags (with a maximum of 16.56 at a lag of one TR), while this value was 1.00 ± 0.31 for fMRI-to-iEEG CRP. Thus, it becomes apparent that synchronicity of connectome dynamics directly translates to the on-/off-diagonal ratio in CRP.

Author response image 1.

Sample CRP shown for a subject for comparing two cases: fMRI-to-iEEG (left) and fMRI-to-fMRI (right). The comparison shows that in the presence of genuine synchronous connectome dynamics, as expected for the within-molality case (right panel), the on-/off-diagonal ratio is expected to show noticeably higher values. This figure establishes a strong link between our proposed metric of on-/off-diagonal ratio and the extent of synchronicity of connectome dynamics.

Author response image 2.

On-/off-diagonal ratio in the fMRI-to-fMRI recurrence plot is considerably higher than the cross-modal fMRI-to-iEEG case. Horizontal axis shows the lag where the metric was calculated in the CRP. The bars reflect the group average metric while the whickers show standard deviation. Note that for the within-modality case, ratio is not defined at lag zero because of identical connectome frames.

(2) Choices in the analysis pipeline leading up to the computation of FC in fMRI or EEG will affect the quality of information available in the FC. For example, but not only, the choice of parcellation (in the study, the number of parcels is very high given the number of EEG sensors). I think it is important that we see the impact of the chosen pipeline on the time-averaged connectomes, an output that the field has some idea about what is sensible. This would give confidence that the information being used in the main analyses in the paper is based on a sensible footing and relates to what the field is used to thinking about in terms of FC. This should be trivial to compute, as it is just a case of averaging the time-varying FCs being used for the CRP over all time points. Admittedly, this approach is less useful for the intracranial EEG.

We agree with the reviewer on ensuring that the time-averaged FC aligns with expectations of the field and prior work. For this reason, our supplementary analysis already included an analysis that replicates the well-established (albeit modest) spatial similarity between fMRI static connectome and EEG/iEEG static connectomes:

“In scalp EEG-fMRI data, cross-modal spatial (2D) Pearson correlation of group-level time-averaged connectomes between fMRI and EEG-FCAmp or fMRI and EEG-FCPhase were calculated across all frequency bands. The average spatial correlation value across frequency bands r = 0.28 and r = 0.28 for EEG-FCAmp and EEG-FCPhase, respectively. The spatial correlation values across all frequency bands and connectivity measures were significantly higher than the corresponding null distributions generated by phase-permuted group-level fMRI-FC spatial organization (p<0.005; 200 repetitions; FDR-corrected at q<0.05 for the number of frequency bands). …. Of note, the small effect sizes are strongly in line with prior literature (Hipp and Siegel, 2015; Wirsich et al., 2017; Betzel et al., 2019) and may point to possible divergence in the dynamic domain as investigated in the main manuscript.”

This replication directly confirms the validity of our selected atlas for further investigations into the connectome dynamics. We acknowledge that with 64 EEG channels, one can only estimate a relatively coarse connectome. Among the well-known coarse atlases, we chose the Desikan-Killiany atlas as it is based on anatomical features, eliminating possible biases towards a particular functional data modality. Moreover, this atlas has been commonly used for multimodal functional connectivity studies, facilitating the confirmation of prior findings in the time-averaged domain [Deligianni et al. Front. Neurosci 2104, Wirsich et al. NeuroImage, 2020, Wirsich et al., NeuroImage 2021].

(3) Leakage correction. The paper states: "To mitigate this issue, we provide results from source-localized data both with and without leakage correction (supplementary and main text, respectively)." Given that FC in EEG is dominated by spatial leakage (see Hipp paper), then I cannot see how it can be justified to look at non-spatial leakage correction results at all, let alone put them up front as the main results. All main results/figures for the scalp EEG should be done using spatial leakage-corrected EEG data.

We agree that relying on leakage-uncorrected scalp EEG alone would be problematic. It is for this reason that the intracranial data constructs the core of our results, emphasizing that the observed multiplex architecture of connectomes is indeed present in the absence of source leakage. Only when this finding is established in the intracranial EEG, do we provide the scalp EEG data as a generalization to whole-cortex coverage connectomes of healthy subjects. Moreover, it is known that existing source-leakage correction algorithms may inadvertently remove some of the genuine zero-lag connectivity. For instance, Finger and colleagues have shown that the similarity of functional connectivity to structural connectivity diminishes after correction for source-leakage (Finger et. al, PLOS Comp. Biol. 2016). Therefore, we have deliberately chosen to include our generalization findings before source-leakage correction (main text) as well as after source-leakage correction reflecting a more stringent approach (supplementary analysis). Importantly, our conclusions hold true for both before and after source-leakage correction.

Reviewer #2 (Public Review):

Summary:

The study investigates the brain's functional connectivity (FC) dynamics across different timescales using simultaneous recordings of intracranial EEG/source-localized EEG and fMRI. The primary research goal was to determine which of three convergence/divergence scenarios is the most likely to occur.

The results indicate that despite similar FC patterns found in different data modalities, the time points were not aligned, indicating spatial convergence but temporal divergence.

The researchers also found that FC patterns in different frequencies do not overlap significantly, emphasizing the multi-frequency nature of brain connectivity. Such asynchronous activity across frequency bands supports the idea of multiple connectivity states that operate independently and are organized into a multiplex system.

Strengths:

The data supporting the authors' claims are convincing and come from simultaneous recordings of fMRI and iEEG/EEG, which has been recently developed and adapted.

The analysis methods are solid and involve a novel approach to analyzing the co-occurrence of FC patterns across modalities (cross-modal recurrence plot, CRP) and robust statistics, including replication of the main results using multiple operationalizations of the functional connectome (e.g., amplitude, orthogonalized, and phase-based coupling).

In addition, the authors provided a detailed interpretation of the results, placing them in the context of recent advances and understanding of the relationships between functional connectivity and cognitive states.

Weaknesses:

Despite the impressive work, the paper still lacks some analyses to make it complete.

Firstly, the effect of the window size is unclear, especially in the case of different frequencies where the number of cycles that fall in a window will vary drastically. A typical oscillation lasts just a few cycles (see Myrov et al., 2024), and brain states are usually short-lived because of meta-stability (see Roberts et al., 2019).

We now replicate our results with an additional window size. Please see section “Recommendations for the authors”.

Secondly, the authors didn't examine frequencies lower than 1Hz despite similarities between fMRI and infra-slow oscillations found in prior literature (see Palva et al., 2014; Zhang et al., 2023).

We address this issue below. Please see section “Recommendations for the authors”.

On a minor note, the phase-locking value (PLV) is positively biased for EEG data (see Palva et al., 2018) and a different metric for phase coupling could be a more appropriate choice (e.g., iPLV/wPLI, see Vinck et al., 2011).

While iPLV and wPLI are not positively biased, they may reduce genuine zero-phase connectivity as they were initially designed to address spurious zero-phase connectivity from source leakage in scalp EEG. Indeed, PLV connectivity is shown to be more strongly correlated with structural connectivity than wPLI and other phase coupling methods [Finger et al., PLOS Comp. Biol. 2016], emphasizing that it contains genuine connectivity that may be lacking when zero-phase connectivity is removed. We chose PLV because it is a widely used functional connectivity metric, particularly in intracranial data where source leakage is not a critical concern. Thus, using PLV facilitates cross-study comparisons including to our prior work [e.g. Mostame et al. NeuroImage 2020, Mostame et al. J Neurosci 2021].

The repository with the code is also unavailable.

Thank you for bringing this to our attention. We have now made our repository publicly accessible at: https://github.com/connectlab/Mostame2024_Multiplex_iEEG_fMRI.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

The window widths used to compute FC as a function of time are an important aspect, so I feel that this should be briefly described up-front in the main Results text.

Methods. "Finally, to compensate for the time lag between hemodynamic and neural responses of the brain (Logothetis et al., 2001), we shifted the fMRI-FC time course 6 seconds backwards in time." What about the effects of temporal blurring from the HRF? Do we need to care about that?

We agree with the importance to investigate the effect if temporal blurring of the HRF. The main text already included a replication of findings from CRPs generated using fMRI data and EEG amplitude signals convolved with the canonical HRF. This method serves as an alternative to the 6-second shifting. Both approaches produced similar results.

Methods. In fMRI connectome computation it is common to look at partial correlation rather than full correlation. Partial correlation focuses more on direct connections. It would be good if the paper acknowledged and justified why it is OK to use full correlation.

We have now added a brief explanation in this regard in the main text (Methods section) as follows:

“In fMRI connectome computation, some prior work has used partial correlation instead of full correlation. Partial correlation emphasizes direct connections by calculating correlation between any pair of bran regions after regressing out the timeseries of all other regions. However, we have opted to use full correlation because this permits interpretation of our outcomes in the context of the vast existing literature that uses full correlations in fMRI including the majority of bimodal (EEG-fMRI) connectome studies (e.g. Tagliazucchi et al., 2012; Deligianni et al., 2014; Wirsich et al., 2017b, 2020, 2021; Allen et al., 2018).”

The paper should relate the results to findings showing clear links between simultaneously recorded EEG and fMRI beyond FC. E.g. Mantini (PNAS) 2007 and Van De Ville (PNAS) 2010 to name two.

In line with this important point, we have extended the existing discussion section that compares our outcomes to EEG-fMRI beyond functional connectivity:

“Prior multi-modal studies of neural dynamics have predominantly aimed at methodologically cross-validating hemodynamic and electrophysiological observations, thus focusing on their convergence. These important foundational studies include e.g., the cross-modal comparison of region-wise (Mukamel et al., 2005; Nir et al., 2007) or ICN-wise (Mantini et al., 2007) activity fluctuations, instantaneous activity maps (Hunyadi et al., 2019; Zhang et al., 2020) or EEG microstates (Van de Ville 2010), infraslow connectome states (Abreu et al., 2020), or connection-wise FC including studies in the iEEG-fMRI and scalp EEG-fMRI data used in the current study (Ridley et al., 2017; and Wirsich et al., 2020, respectively). In contrast to this prior work, the current study investigated the highly time-resolved cross-modal temporal relationship at the level of FC patterns distributed over all available pairwise connections, and found a connectome-level temporal divergence. The discrepancy between temporal divergence in our study and convergence in prior studies implies that infraslow fluctuations of activity in individual regions or of FC in individual region-pairs observable in both modalities (prior studies) are neurally distinct from connectome-wide FC dynamics observable separately in each modality (current study). Indeed, we confirmed the existence of infraslow electrophysiological FC dynamics driving cross-modal temporal associations at the level of individual connections (Fig. S3) …”

Reviewer #2 (Recommendations For The Authors):

(1) Check different window sizes and stability of the FC patterns as a function of it.

We thank the reviewer for the helpful feedback. We agree that the window size could possibly affect the estimation of individual connectome frames, particularly given that neural processes unfold at hundreds of milliseconds rather than seconds. However, we expect that the asynchronous nature of cross-modal convergence observed in our data would remain intact regardless of the specific window length used for FC calculations. To confirm this, we replicated some of our main analyses in the iEEG-fMRI data with a window length of 500ms (as opposed to 3s, equivalent to one TR) as follows:

First, we showed that changing the window length does not substantially impact the overall architecture of the connectomes (Author response image 3). Particularly, the time-averaged connectome patterns across different frequency bands were all strongly correlated between the two analyses (500ms and 3s window lengths).

Author response image 3.

Time-averaged connectome patterns are highly replicable when calculated using 3s or 500ms window lengths. Horizontal axis represents frequency bands, while each dot represents a subject. Vertical axis shows 2D Pearson correlation of the two connectomes. The group average within each frequency band is marked by a horizontal line.

Second, we replicated our major findings of CRP and its on-/off-diagonal ratio in the iEEG-fMRI dataset using a window length of 500ms for FC calculations. Indeed, the data does not show a substantial difference in the on-/off-diagonal ratios of the CRP entries between the 3s and 500ms window lengths. Specifically, the ratio was equal to 1.02 ± 0.07 for 500ms window length, emphasizing absence of significant temporal convergence of the connectome dynamics (see Author response image 4). A paired t-test between group-averaged ratios across different lags confirms a lack of significant difference between the two analyses (p= 0.50). This finding further emphasizes the genuine asynchronous nature of connectome dynamics across the neural timescales measured in fMRI and electrophysiology. We have added this analysis to the supplementary data.

Author response image 4.

On-/off-diagonal ratio is shown across lags for both analyses: 3s window length (blue) and 500ms window length (red). Each bar shows the mean across subjects, while the whiskers show the corresponding standard deviations.

(2) Try to decrease the lowest frequency of the analysis below 1Hz or just compute it for multiple log-spaced frequencies from infra-slow delta to high-gamma band.

Thank you for pointing out this matter. We do not expect considerable signal in the frequency range below the current lower bound of delta (1Hz) because as in most other EEG recordings, EEG was not recorded in DC setting and has a hardware high-pass filter of 0.1Hz. Nonetheless, we investigated the power spectral density of our iEEG-fMRI data and found that there is indeed little signal power left in the available infraslow range [0.5 – 1 Hz] after the preprocessing steps (Author response image 5).

Author response image 5.

Power spectral density of all subjects in the fMRI-iEEG dataset shows lack of sufficient power in the infraslow range. Infraslow range signals are almost always filtered out during recording unless the recording setup includes a DC amplifier. The infraslow signal of EEG that is often considered correlated with the fMRI signals in the literature most commonly are extracted from the slow-changing envelope of the bandlimited signals, like envelope of gamma oscillations.

Accordingly, when the iEEG signals are filtered within the range of [0.5, 1], there is little signal variation observed in the signal timeseries, contrasting the adjacent delta band signal (Author response image 6). Importantly, the power envelope of the delta band (and all other canonical bands not shown here) comprise major fluctuations in the infraslow range, as expected. We would like to emphasize that the existing studies addressing infraslow EEG signal dynamics typically consider the infraslow envelope fluctuations of band-limited signals in traditional frequency bands [e.g. Nir et. al, Nat Neurosci 2008] rather than direct recordings in the infraslow frequency range. Investigating HRF-convolved EEG signals similarly captures the infraslow characteristics of the timeseries [e.g. Mantini et al. PNAS 2007, Sadaghiani et al., J Neurosci 2010] (and note that HRF-convolved analyses are included as supplementary investigation in the current study). To the best of our knowledge, very few studies have investigated direct infraslow EEG signals using DC EEG, and we are aware of only two DC-EEG studies with concurrent fMRI [Hiltunen et al., J Neurosci 2014, Grooms et al., Brain Connectivity 2017]. The infraslow correlates of fMRI in electrophysiological signals reported in prior work therefore reflect the slow changes in faster activity or connectivity of traditional frequency bands, which is indeed already included in the current study.

Author response image 6.

Sample timeseries of the iEEG signal of the nine subjects (nine rows) for a 400 second interval. Blue signals show the bandlimited delta with its envelope shown as darker blue. The red signal represents the infraslow signal component left in the data, which is much lower in power.

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

Ritvo and colleagues present an impressive suite of simulations that can account for three findings of differentiation in the literature. This is important because differentiation-in which items that have some features in common, or share a common associate are less similar to one another than are unrelated items-is difficult to explain with classic supervised learning models, as these predict the opposite (i.e., an increase in similarity). A few of their key findings are that differentiation requires a high learning rate and low inhibitory oscillations, and is virtually always asymmetric in nature.

This paper was very clear and thoughtful-an absolute joy to read. The model is simple and elegant, and powerful enough to re-create many aspects of existing differentiation findings. The interrogation of the model and presentation of the findings were both extremely thorough. The potential for this model to be used to drive future work is huge. I have only a few comments for the authors, all of which are relatively minor.

(1) I was struck by the fact that the "zone" of repulsion is quite narrow, compared with the zone of attraction. This was most notable in the modeling of Chanales et al. (i.e., just one of the six similarity levels yielded differentiation). Do the authors think this is a generalizable property of the model or phenomenon, or something idiosyncratic to do with the current investigation? It seems curious that differentiation findings (e.g., in hippocampus) are so robustly observed in the literature despite the mechanism seemingly requiring a very particular set of circumstances. I wonder if the authors could speculate on this point a bit-for example, might the differentiation zone be wider when competitor "pop up" is low (i.e., low inhibitory oscillations), which could help explain why it's often observed in hippocampus? This seems related a bit to the question about what makes something "moderately" active, or how could one ensure "moderate" activation if they were, say, designing an experiment looking at differentiation.

We thank the reviewer for this comment. In the previous version of the manuscript, in the section entitled “Differentiation Requires a High Learning Rate and Is Sensitive to Activation Dynamics”, we discussed some reasons why differentiation may be more likely to be found in the hippocampus – namely, the high learning rate of the hippocampus and the sparsity of hippocampal activation patterns (pp. 27-28):

“These results have implications for where to look for differentiation in the brain. Our finding that differentiation requires a high learning rate suggests that differentiation will be more evident in the hippocampus than in neocortex, insofar as hippocampus is thought to have a higher learning rate than neocortex (McClelland et al., 1995). In keeping with this prediction, numerous studies have found differentiation effects in hippocampus but not in neocortical regions involved in sensory processing (e.g., Chanales et al., 2017; Favila et al., 2016; Zeithamova et al., 2018). At the same time, some studies have found differentiation effects in neocortex (e.g., Schlichting et al., 2015; Wammes et al., 2022). One possible explanation of these neocortical differentiation effects is that they are being ``propped up’’ by top-down feedback from differentiated representations in the hippocampus. This explanation implies that disruptions of hippocampal processing (e.g., lesions, stimulation) will eliminate these neocortical differentiation effects; we plan to test this prediction in future work.

Additionally, the simulations where we adjusted the oscillation amount (using our model of Schlichting et al., 2015) imply that differentiation will be most evident in brain regions where it is relatively hard to activate competitors. Given the U shape of the NMPH learning rule, limiting competitor activity makes it less likely that plasticity will ``cross over'' from weakening (and differentiation) to strengthening (and integration). Thus, within the hippocampus, subregions with sparser activity (e.g., dentate gyrus, and to a lesser extent, CA3; Barnes et al., 1990, GoodSmith et al., 2017; West et al., 1991) will be more prone to differentiation. There is strong empirical support for this prediction. For example, Wammes et al. (2022) manipulated the similarity of stimuli in a statistical learning experiment and found that moderate levels of visual similarity were associated with significant differentiation in the dentate gyrus but not other subregions. Also, numerous studies have found greater differentiation in dentate gyrus / CA3 than in CA1 (e.g., Dimsdale-Zucker et al., 2018; Wanjia et al., 2021; Molitor et al., 2021; Kim et al., 2017; but see Zheng et al., 2021).”

In the revised draft we have supplemented this discussion with a new section entitled “Reconciling the Prevalence of Differentiation in the Model and in the Data” (pp. 30-31):

“A key lesson from our model is that, from a computational perspective, it is challenging to obtain differentiation effects: The region of parameter space that gives rise to differentiation is much smaller than the one that gives rise to integration (for further discussion of this issue, see the section in Methods on Practical Advice for Getting the Model to Show Differentiation). However, the fact that integration is more prevalent in our simulations across parameter configurations does not mean that integration will be more prevalent than differentiation in real-life circumstances. What really matters in predicting the prevalence of differentiation in real life is how the parameters of the brain map on to parameters of the model: If the parameters of the brain align with regions of model parameter space that give rise to differentiation (even if these regions are small), this would explain why differentiation has been so robustly observed in extant studies. Indeed, this is exactly the case that we sought to make above about the hippocampus – i.e., that its use of especially sparse coding and a high learning rate will give rise to the kinds of neural dynamics that cause differentiation (as opposed to integration). As another example, while it is true that half of the overlap conditions in our simulation of Chanales et al. (2021) give rise to integration, this does not imply that integration will occur half of the time in the Chanales et al. (2021) study; it may be that the levels of overlap that are actually observed in the brain in Chanales et al. (2021) are more in line with the levels of overlap that give rise to differentiation in our model.”

(2) With real fMRI data we know that the actual correlation value doesn't matter all that much, and anti-correlations can be induced by things like preprocessing decisions. I am wondering if the important criterion in the model is that the correlations (e.g., as shown in Figure 6) go down from pre to post, versus that they are negative in sign during the post learning period. I would think that here, similar to in neural data, a decrease in correlation would be sufficient to conclude differentiation, but would love the authors' thoughts on that.

We thank the reviewer for bringing this up. In the paper, we define differentiation as the moving apart of representations – so we agree with the reviewer that it would be appropriate to conclude that differentiation is taking place when correlations go down from pre to post.

In addition to the definitional question (“what counts as differentiation”), one can also ask the mechanistic question of what is happening in the model at the (simulated) neuronal level in conditions where differentiation (i.e., an average decrease in similarity from pre to post) occurs. Here, the model’s answer is clear: When the similarity of two pairmates decreases, it is because the pairmates have acquired anticorrelated representations at the (simulated) neuronal level. When similarity decreases on average from pre to post, but the average “post” similarity value is not negative, this is because there is a mix of outcomes across runs of the model (due to variance in the initial, random model weights and also variance in the order in which items are presented across training epochs) – some runs lead to differentiation (manifested as anticorrelated pairmate representations) whereas others lead to no change or integration. The average pre-to-post change depends on the relative frequencies with which these different outcomes occur.

We have made several edits to the paper to clarify this point.

We added a new section under “Results” in our simulation of Chanales et al. (2021) entitled, “Pairs of Items that Differentiate Show Anticorrelated Representations” (p. 15):

“Figure 6B also highlights that, for learning rates where robust differentiation effects occur in aggregate (i.e., there is a reduction in mean pattern similarity, averaging across model runs), these aggregate effects involve a bimodal distribution across model runs: For some model runs, learning processes give rise to anticorrelated representations, and for other model runs the model shows integration; this variance across model runs is attributable to random differences in the initial weight configuration of the model. The aggregate differentiation effect is therefore a function of the proportion of model runs showing differentiation (here, anticorrelation) and the proportion of model runs showing integration. The fact that differentiation shows up as anticorrelation in the model's hidden layer relates to the learning effects discussed earlier:

Unique competitor units are sheared away from (formerly) shared units, so the competitor ends up not having any overlap with the target representation (i.e., the level of overlap is less than you would expect due to chance, which mathematically translates into anticorrelation). We return to this point and discuss how to test for anticorrelation in the Discussion section.”

We added new text to the “Take-Home Lessons” section in the Chanales et al. (2021) simulation (p. 17):

“In particular, the simulations expose some important boundary conditions for when representational change can occur according to the NMPH (e.g., that differentiation depends on a large learning rate, but integration does not), and the simulations provide a more nuanced account of exactly how representations change (e.g., that differentiation driven by the NMPH is always asymmetric, whereas integration is sometimes asymmetric and sometimes symmetric; and that, when differentiation occurs on a particular model run, it tends to give rise to anticorrelated representations in the model's hidden layer).”

We added new text to the “Nature of Representational Change” section in the Favila et al. (2016) simulation (p. 21):

“Figure 8 - Supplement 1 also indicates that, as in our simulation of Chanales et al. (2021), individual model runs where differentiation occurs show anticorrelation between the pairmate representations, and gradations in the aggregate level of differentiation that is observed across conditions reflect differences in the proportion of trials showing this anticorrelation effect.”

We added new text to the “Take-Home Lessons” section in the Favila et al. (2016) simulation (p.21):

“As in our simulation of \cite{chanales2021adaptive}, we found that the NMPH-mediated differentiation was asymmetric, manifested as anticorrelation between pairmate representations on individual model runs, and required a high learning rate, leading to abrupt representational change.”

We added new text to the “Nature of Representational Change” section in the Schlichting et al. (2015) simulation (p. 26):

“Also, as in our other simulations, when differentiation occurs on a particular model run it tends to give rise to anticorrelated representations (results not shown).”

We added new text to the “Take-Home Lessons” section in the Schlichting et al. (2015) simulation (pp. 26-27):

“As in the other versions of our model, differentiation requires a high learning rate, and – on model runs when it occurs – it is asymmetric and gives rise to anticorrelated representations.”

We added new text at the start of the Discussion (p. 27):

“In addition to qualitatively replicating the results from the studies we simulated, our model gives rise to several novel predictions – most notably, that differentiation driven by the NMPH requires a rapid learning rate and, when it occurs for a particular pair of items, it is asymmetric and gives rise to anticorrelated representations.”

We also added a new section in the Discussion entitled “Testing the Model's Prediction about Anticorrelation”, which (among other things) highlights the reviewer’s point that fMRI pattern similarity values can be affected by preprocessing choices (p. 30):

“Even though we operationally define differentiation as a reduction in similarity with learning, the way that it actually shows up on individual model runs is as anticorrelation between pairmates; in the model, the size of the aggregate differentiation effect is determined by the proportion of model runs that show this anticorrelation effect (vs. no change or integration). This implies that, if we could get a clean measurement of the similarity of pairmates in an experiment, we might see a multimodal distribution, with some pairmates showing anticorrelation, and others showing increased correlation (integration) or no change in similarity. This kind of clean readout of the similarity of individual pairs might be difficult to obtain with fMRI; it is more feasible that this could be obtained with electrophysiology. Another challenge with using fMRI to test this prediction is that anticorrelation at the individual-neuron level might not scale up to yield anticorrelation at the level of the BOLD response; also, fMRI pattern similarity values can be strongly affected by preprocessing choices – so a negative pattern similarity value does not necessarily reflect anticorrelation at the individual-neuron level. A final caveat is that, while we predict that differentiation will show up as anticorrelation in the brain region that gives rise to the differentiation effect, this might not translate into anticorrelation in areas that are downstream of this region (e.g., if the hippocampus is the source of the differentiation effect, we would expect anticorrelation there, but not necessarily in neocortical regions that receive input from the hippocampus; we revisit this point later in the discussion, when we address limitations and open questions).”

We added new text in the Discussion, under “Limitations and Open Questions” (p. 31):

“Importantly, while hippocampus can boost the representation of unique features in neocortex, we expect that neocortex will continue to represent shared perceptual features (e.g., in Favila et al., 2016, the fact that both pairmates are photos of barns). For this reason, in paradigms like the one used by Favila et al. (2016), the predicted effect of hippocampal differentiation on neocortical representations will be a reduction in pattern similarity (due to upregulation in the representation of unique pairmate features) but neocortex should not cross over into anticorrelation in these paradigms (due to its continued representation of shared perceptual features). Indeed, this is exactly the pattern that Wanjia et al. (2021) observed in their study, which used similar stimuli to those used in Favila et al. (2016).”

Lastly, we updated the Abstract (p. 1)

“What determines when neural representations of memories move together (integrate) or apart (differentiate)? Classic supervised learning models posit that, when two stimuli predict similar outcomes, their representations should integrate. However, these models have recently been challenged by studies showing that pairing two stimuli with a shared associate can sometimes cause differentiation, depending on the parameters of the study and the brain region being examined. Here, we provide a purely unsupervised neural network model that can explain these and other related findings. The model can exhibit integration or differentiation depending on the amount of activity allowed to spread to competitors – inactive memories are not modified, connections to moderately active competitors are weakened (leading to differentiation), and connections to highly active competitors are strengthened (leading to integration). The model also makes several novel predictions – most importantly, that when differentiation occurs as a result of this unsupervised learning mechanism, it will be rapid and asymmetric, and it will give rise to anticorrelated representations in the region of the brain that is the source of the differentiation. Overall, these modeling results provide a computational explanation for a diverse set of seemingly contradictory empirical findings in the memory literature, as well as new insights into the dynamics at play during learning.”

(3) For the modeling of the Favila et al. study, the authors state that a high learning rate is required for differentiation of the same-face pairs. This made me wonder what happens in the low learning rate simulations. Does integration occur?

For the same-face condition of the Favila simulation, lowering learning rate does not result in an overall integration effect:

Author response image 1.

In other cases, we do see integration emerge at lower learning rates – e.g., in the Schlichting interleaved condition we see a small integration effect emerge for a learning rate value of 0.3:

Author response image 2.

Our view is that, while integration can emerge at low learning rates, it is not a reliable property of the model – in some cases, there is a “window” of learning rates where there is enough learning to drive integration but not enough to drive differentiation, and in other cases there is not. Given this lack of reliability across simulations, we would prefer not to discuss this in the paper.

This paradigm has a lot of overlap with acquired equivalence, and so I am thinking about whether these are the sorts of small differences (e.g., same-category scenes and perhaps a high learning rate) that bias the system to differentiate instead of integrate.

We agree that it would be very interesting to use the model to explore acquired equivalence and related phenomena, but we think it is out of scope of the current paper. We have added some text to the Discussion under “Limitations and Open Questions” (p. 32):

“Another important future direction is to apply the model to a wider range of learning phenomena involving representational change – for example, acquired equivalence, which (like some of the studies modeled here) involves linking distinct stimuli to a shared associate (see, e.g., Honey and Hall, 1989; Shohamy and Wagner, 2008; Myers et al., 2003; Meeter et al., 2009; de Araujo Sanchez and Zeithamova, 2023). It is possible that some of these phenomena might be better explained by supervised learning, or a mixture of unsupervised and supervised learning, than by unsupervised learning alone.”

(4) For the simulations of the Schlichting et al. study, the A and B appear to have overlap in the hidden layer based on Figure 9, despite there being no similarity between the A and B items in the study (in contrast to Favila et al., in which they were similar kinds of scenes, and Chanales et al., in which they were similar colors). Why was this decision made? Do the effects depend on some overlap within the hidden layer? (This doesn't seem to be explained in the paper that I saw though, so maybe just it's a visualization error?)

Overlap in the pretrained hidden representations of A and B is not strictly necessary for these effects – it would be possible to reconfigure other parameters to get high levels of competition even if there were no overlap (e.g., by upregulating the strengths of connections from shared input features). Having said that, it is definitely true that overlap between the pretrained hidden representations boosts competition, and we think it is justified to posit this in the Schlichting simulation. We have now added an explanation for this in the paper (p. 23):

“New text in Schlichting, “Knowledge Built into the Network”

Matching the previous two simulations, we pretrained the weights so the hidden representations of the stimuli initially had 2/6 units in common. Even though the A and B stimuli used in the actual experiment did not have obvious feature overlap (they were randomly selected novel objects), it is important to note that the hidden layer is not simply a representation of the sensory features of the A and B stimuli; the hidden layer also receives input from the output layer, which represents the shared associate of A and B (X). We think that the presence of this shared associate justifies our use of initially-overlapping hidden representations.”

(5) It seems as though there were no conditions under which the simulations produced differentiation in both the blocked and intermixed conditions, which Schlichting et al. observed in many regions (as the present authors note). Is there any way to reconcile this difference?

We thank the reviewer for bringing this up. If we set the connection strength between X (in the output layer) and A (in the hidden layer) in the blocked condition to .9 instead of .999 (keeping this connection strength at .8 for the interleaved condition) and we set Osc to .0615, we observe differentiation in both conditions.

Rather than replacing the original results in the paper, which would entail re-making the associated videos, etc., we have added a supplementary figure (Figure 10 - Supplement 1), which is included on p. 46.

We also added the following to the Results section of the Schlichting simulation in the main text (p. 26):

“Figure 10 - Supplement 1 shows results from an alternative parameterization where, in the low-oscillation-amplitude condition, differentiation is observed in both the blocked and interleaved conditions (mirroring results from Schlichting et al., 2015, who found differentiation in both conditions in several regions of interest, including parts of the hippocampus and medial prefrontal cortex).”

(6) A general question about differentiation/repulsion and how it affects the hidden layer representation in the model: Is it the case that the representation is actually "shifted" or repelled over so it is no longer overlapping? Or do the shared connections just get pruned, such that the item that has more "movement" in representational space is represented by fewer units on the hidden layer (i.e., is reduced in size)? I think, if I understand correctly, that whether it gets shifted vs. reduce would depend on the strength of connections along the hidden layer, which would in turn depend on whether it represents some meaningful continuous dimension (like color) or not. But, if the connections within the hidden layer are relatively weak and it is the case that representations become reduced in size, would there be any anticipated consequences of this (e.g., cognitively/behaviorally)?

The representations are shifted – this is discussed in the Chanales results section:

“Because the activity ``set point'' for the hidden layer (determined by the kWTA algorithm) involves having 6 units active, and the unique parts of the competitor only take up 4 of these 6 units, this leaves room for activity to spread to additional units. Given the topographic projections in the output layer, the model is biased to ``pick up'' units that are adjacent in color space to the currently active units; because activity cannot flow easily from the competitor back to the target (as a result of the aforementioned severing of connections), it flows instead {\em away} from the target, activating two additional units, which are then incorporated into the competitor representation. This sequence of events (first a severing of the shared units, then a shift away from the target) completes the process of neural differentiation, and is what leads to the behavioral repulsion effect in color recall (because the center-of-mass of the color representation has now shifted away from the target).”

Reviewer #2 (Public Review):

This paper addresses an important computational problem in learning and memory. Why do related memory representations sometimes become more similar to each other (integration) and sometimes more distinct (differentiation)? Classic supervised learning models predict that shared associations should cause memories to integrate, but these models have recently been challenged by empirical data showing that shared associations can sometimes cause differentiation. The authors have previously proposed that unsupervised learning may account for these unintuitive data. Here, they follow up on this idea by actually implementing an unsupervised neural network model that updates the connections between memories based on the amount of coactivity between them. The goal of the authors' paper is to assess whether such a model can account for recent empirical data at odds with supervised learning accounts. For each empirical finding they wish to explain, the authors built a neural network model with a very simple architecture (two inputs layers, one hidden layer, and one output layer) and with prewired stimulus representations and associations. On each trial, a stimulus is presented to the model, and inhibitory oscillations allow competing memories to pop up. Pre-specified u-shaped learning rules are used to update the weights in the model, such that low coactivity leaves model connections unchanged, moderate coactivity weakens connections, and high coactivity strengthens connections. In each of the three models, the authors manipulate stimulus similarity (following Chanales et al), shared vs distinct associations (following Favila et al), or learning strength (a stand in for blocked versus interleaved learning schedule; following Schlichting et al) and evaluate how the model representations evolve over trials.

As a proof of principle, the authors succeed in demonstrating that unsupervised learning with a

simple u-shaped rule can produce qualitative results in line with the empirical reports. For instance, they show that pairing two stimuli with a common associate (as in Favila et al) can lead to *differentiation* of the model representations. Demonstrating these effects isn't trivial and a formal modeling framework for doing so is a valuable contribution. Overall, the authors do a good job of both formally describing their model and giving readers a high level sense of how their critical model components work, though there are some places where the robustness of the model to different parameter choices is unclear. In some cases, the authors are very clear about this (e.g. the fast learning rate required to observe differentiation). However, in other instances, the paper would be strengthened by a clearer reporting of the critical parameter ranges.

We thank the reviewer for raising this point. The interdependence of parameters in our model makes it infeasible to identify critical parameter ranges. We have added a paragraph to the “Approach to Parameterization and Data Fitting” section in the Methods to address this point (p. 33):

“The overall goal of this modeling work is to account for key empirical regularities regarding differentiation and integration and to establish boundary conditions on these regularities. As such, the modeling work described below focuses more on qualitative fits to general properties of the data space than on quantitative fits to results from specific studies. Automatic parameter optimization is not feasible for this kind of model, given the large number of model parameters and the highly interactive, nonlinear nature of competitive dynamics in the model; consequently, model fitting was done by hand.

These complex interactions between parameters also make it infeasible to list “critical parameter ranges” for generating particular model outcomes. Our experience in working with the model has been that activation dynamics are what matter most for learning, and that disparate parameter sets can give rise to the same activation dynamics and -- through this -- the same learning effects; likewise, similar parameter sets can give rise to different activation dynamics and different learning outcomes. Consequently, in this paper we have focused on characterizing the dynamics that give rise to different learning effects (and how they can be affected by local parameter perturbations, e.g., relating to learning rate and oscillation size), rather than the – impossible, we believe – task of enumerating the full set of parameter configurations that give rise to a particular result.”

For instance, it's clear from the manipulation of oscillation strength in the model of Schlichting et al that this parameter can dramatically change the direction of the results. The authors do report the oscillation strength parameter values that they used in the other two models, but it is not clear how sensitive these models are to small changes in this value.

In some cases, the effects of oscillation strength are relatively smooth. For example, in the Favila simulation, increasing the oscillation amplitude Osc effectively recapitulates the U-shaped curve (i.e., higher levels of Osc lead to more competitor activation, which initially leads to weakening / differentiation but then gives way to strengthening / integration), as is shown for the Favila Different Face condition in this plot:

Author response image 3.

In the Chanales 2/6 overlap condition, the effects of varying Osc are more nonlinear:

Author response image 4.

We think this is attributable to the increased “all-or-none” recurrent dynamics in this simulation (due to the recurrent projections within the output layer), which make it more difficult to evoke moderate (vs. high) levels of activation. This difficulty in reliably obtaining graded activation dynamics is likely a consequence of the small-scale (“toy”) nature of the model and the simple inhibitory mechanisms employed here, as opposed to being a generalizable property of the brain – presumably, the actual brain employs more nuanced and effective means of controlling activation. Furthermore, we don’t think that the high prevalence of integration in the model’s parameter space necessarily translates into a prediction that integration should be more prevalent overall – see the new “Reconciling the Prevalence of Differentiation in the Model and in the Data” section described in response to one of the reviewer’s other points below. Due to the paper already being quite long, we have opted not to include the above plots / discussion in the paper.

Similarly, it's not clear whether the 2/6 hidden layer overlap (only explicitly manipulated in the model of Chanales et al) is required for the other two models to work.

When we were parameterizing the model, we opted to keep the 2/6 level of overlap for all of the simulations and we adjusted other parameters to fit the data; in part, this was because overlap can only be adjusted in discrete jumps, whereas other influential parameters in the model can be adjusted in a more graded, real-valued way. Our use of 2/6 overlap (as opposed to, say, 1/6 or 3/6 overlap) for the Favila and Schlichting models was done out of convenience, and should not be interpreted as a strong statement that this particular level of overlap is necessary for obtaining differentiation; we could easily get the model to show differentiation given other overlap levels by adjusting other parameters.

Finally, though the u-shaped learning rule is essential to this framework, the paper does little formal investigation of this learning rule. It seems obvious that allowing the u-shape to collapse too much toward a horizontal line would reduce the model's ability to account for empirical results, but there may be other more interesting features of the learning rule parameterization that are essential for the model to function properly.

Given that the paper is already quite long, we have opted not to include further exploration of the parameters of the U-shaped learning rule in the paper. However, for the reviewer’s information, we report the effects of a few illustrative manipulations of these parameters below. As a general principle, the effects of these manipulations make sense in light of the theoretical framework described in the paper.

For example, the parameter “DRevMag” controls the size of the negative “dip” in the U-shaped curve (more negative values = a larger dip). Given that this negative dip is essential for severing weights to competitors and causing differentiation, shifting DRevMag upwards towards zero should shift the balance of the model away from differentiation and towards integration. This is indeed what we observe, as shown in this parameter sweep from the Chanales simulation:

Author response image 5.

As another example: The “DRev” parameter controls where the U-shaped curve transitions from negative weight change to positive weight change. Lower values of DRev mean that the region of coactivity values leading to negative weight change will be smaller, and the region of coactivity values leading to positive weight change will be larger. As such, we would expect that lower values of DRev would bias the model toward integration. That is indeed the case, as shown in this parameter sweep from the Schlichting Blocked simulation:

Author response image 6.

There are a few other points that may limit the model's ability to clearly map onto or make predictions about empirical data. The model(s) seems very keen to integrate and do so more completely than the available empirical data suggest. For instance, there is a complete collapse of representations in half of the simulations in the Chanales et al model and the blocked simulation in the Schlichting et al model also seems to produce nearly complete integration Even if the Chanales et al paper had observed some modest behavioral attraction effects, this model would seem to over-predict integration. The author's somewhat implicitly acknowledge this when they discuss the difficulty of producing differentiation ("Practical Advice for Getting the Model to Show Differentiation") and not of producing integration, but don't address it head on.

We thank the reviewer for this comment – R1 had a similar comment. We have added a new section to the Discussion to address this point (p. 30):

“Reconciling the Prevalence of Differentiation in the Model and in the Data.

A key lesson from our model is that, from a computational perspective, it is challenging to obtain differentiation effects: The region of parameter space that gives rise to differentiation is much smaller than the one that gives rise to integration (for further discussion of this issue, see the section in Methods on Practical Advice for Getting the Model to Show Differentiation). However, the fact that integration is more prevalent in our simulations across parameter configurations does not mean that integration will be more prevalent than differentiation in real-life circumstances. What really matters in predicting the prevalence of differentiation in real life is how the parameters of the brain map on to parameters of the model: If the parameters of the brain align with regions of model parameter space that give rise to differentiation (even if these regions are small), this would explain why differentiation has been so robustly observed in extant studies. Indeed, this is exactly the case that we sought to make above about the hippocampus – i.e., that its use of especially sparse coding and a high learning rate will give rise to the kinds of neural dynamics that cause differentiation (as opposed to integration). As another example, while it is true that half of the overlap conditions in our simulation of Chanales et al. (2021) give rise to integration, this does not imply that integration will occur half of the time in the Chanales et al. (2021) study; it may be that the levels of overlap that are actually observed in the brain in Chanales et al. (2021) are more in line with the levels of overlap that give rise to differentiation in our model.”

Second, the authors choice of strongly prewiring associations in the Chanales and Favila models makes it difficult to think about how their model maps onto experimental contexts where competition is presumably occurring while associations are only weakly learned. In the Chanales et al paper, for example, the object-face associations are not well learned in initial rounds of the color memory test. While the authors do justify their modeling choice and their reasons have merit, the manipulation of AX association strength in the Schlichting et al model also makes it clear that the association strength has a substantial effect on the model output. Given the effect of this manipulation, more clarity around this assumption for the other two models is needed.

We thank the reviewer for bringing this up. We have edited the section entitled “A Note on Prewiring Representations” in the Methods to further justify our choice to prewire associations in the Chanales and Favila models (p. 37):

“In our model, our practice of ``prewiring'' memory representations for the A and B pairmates serves two functions. In some cases, it is meant to stand in for actual training (as in the blocked / interleaved manipulation; the connections supporting the AX association are prewired to be stronger in the blocked condition than in the interleaved condition). However, the other, more fundamental role of prewiring is to ensure that the A and B input patterns evoke sparse distributed representations in the hidden layer (i.e., where some units are strongly active but most other units are inactive). In the real brain, this happens automatically because the weight landscape has been extensively sculpted by both experience and evolution. For example, in the real hippocampus, when the second pairmate is presented for the first time, it will evoke a sparse distributed representation in the CA3 subfield (potentially overlapping with the first pairmate’s CA3 representation) even before any learning of the second pairmate has occurred, due to the strong, sparse mossy fiber projections that connect the dentate gyrus to CA3 (McNaughton & Morris, 1987). As discussed above, we hypothesize that this initial, partial overlap between the second pairmate’s representation and the first pairmate’s representation can lead to pop-up of the unique features of the first pairmate’s representation, triggering learning that leads to differentiation or integration. In our small-scale model, we are effectively starting with a ``blank brain''; in the absence of prewiring, the A and B inputs would activate overly diffuse representations that do not support these kinds of competitive dynamics. As such, prewiring in our model is necessary for proper functioning. The presence of prewired A and B representations should therefore not be interpreted as reflecting a particular training history (except in the blocked / interleaved case above); rather, these prewired representations constitute the minimum step we would take to ensure well-defined competitive dynamics in our small-scale model.

The fact that connection strengths serve this dual function – sometimes reflecting effects of training (as in our simulation of Schlichting et al., 2015) and in other cases reflecting necessary prewiring – complicates the interpretation of these strength values in the model. Our view is that this is a necessary limitation of our simplified modeling approach – one that can eventually be surmounted through the use of more biologically-detailed architectures (see Limitations and Open Questions in the Discussion).”

Overall, this is strong and clearly described work that is likely to have a positive impact on computational and empirical work in learning and memory. While the authors have written about some of the ideas discussed in this paper previously, a fully implemented and openly available model is a clear advance that will benefit the field. It is not easy to translate a high-level description of a learning rule into a model that actually runs and behaves as expected. The fact that the authors have made all their code available makes it likely that other researchers will extend the model in numerous interesting ways, many of which the authors have discussed and highlighted in their paper.

Reviewer #3 (Public Review):

This paper proposes a computational account for the phenomenon of pattern differentiation (i.e., items having distinct neural representations when they are similar). The computational model relies on a learning mechanism of the nonmonotonic plasticity hypothesis, fast learning rate and inhibitory oscillations. The relatively simple architecture of the model makes its dynamics accessible to the human mind. Furthermore, using similar model parameters, this model produces simulated data consistent with empirical data of pattern differentiation. The authors also provide insightful discussion on the factors contributing to differentiation as opposed to integration. The authors may consider the following to further strengthen this paper:

The model compares different levels of overlap at the hidden layer and reveals that partial overlap seems necessary to lead to differentiation. While I understand this approach from the perspective of modeling, I have concerns about whether this is how the human brain achieves differentiation. Specifically, if we view the hidden layer activation as a conjunctive representation of a pair that is the outcome of encoding, differentiation should precede the formation of the hidden layer activation pattern of the second pairmate. Instead, the model assumes such pattern already exists before differentiation. Maybe the authors indeed argue that mechanistically differentiation follows initial encoding that does not consider similarity with other memory traces?

Related to the point above, because the simulation setup is different from how differentiation actually occurs, I wonder how valid the prediction of asymmetric reconfiguration of hidden layer connectivity pattern is.

We thank the reviewer for this comment. In the revised manuscript, we have edited the “Note on Prewiring Representations” in the Methods to clarify how our assumptions about prewiring relate to what we really think is happening in the brain (p. 37):

“In our model, our practice of ``prewiring'' memory representations for the A and B pairmates serves two functions. In some cases, it is meant to stand in for actual training (as in the blocked / interleaved manipulation; the connections supporting the AX association are prewired to be stronger in the blocked condition than in the interleaved condition). However, the other, more fundamental role of prewiring is to ensure that the A and B input patterns evoke sparse distributed representations in the hidden layer (i.e., where some units are strongly active but most other units are inactive). In the real brain, this happens automatically because the weight landscape has been extensively sculpted by both experience and evolution. For example, in the real hippocampus, when the second pairmate is presented for the first time, it will evoke a sparse distributed representation in the CA3 subfield (potentially overlapping with the first pairmate’s CA3 representation) even before any learning of the second pairmate has occurred, due to the strong, sparse mossy fiber projections that connect the dentate gyrus to CA3 (McNaughton & Morris, 1987). As discussed above, we hypothesize that this initial, partial overlap between the second pairmate’s representation and the first pairmate’s representation can lead to pop-up of the unique features of the first pairmate’s representation, triggering learning that leads to differentiation or integration. In our small-scale model, we are effectively starting with a ``blank brain''; in the absence of prewiring, the A and B inputs would activate overly diffuse representations that do not support these kinds of competitive dynamics. As such, prewiring in our model is necessary for proper functioning. The presence of prewired A and B representations should therefore not be interpreted as reflecting a particular training history (except in the blocked / interleaved case above); rather, these prewired representations constitute the minimum step we would take to ensure well-defined competitive dynamics in our small-scale model.

The fact that connection strengths serve this dual function – sometimes reflecting effects of training (as in our simulation of Schlichting et al., 2015) and in other cases reflecting necessary prewiring – complicates the interpretation of these strength values in the model. Our view is that this is a necessary limitation of our simplified modeling approach – one that can eventually be surmounted through the use of more biologically-detailed architectures (see Limitations and Open Questions in the Discussion).”

Although as the authors mentioned, there haven't been formal empirical tests of the relationship between learning speed and differentiation/integration, I am also wondering to what degree the prediction of fast learning being necessary for differentiation is consistent with current data. According to Figure 6, the learning rates lead to differentiation in the 2/6 condition achieved differentiation after just one-shot most of the time. On the other hand, For example, Guo et al (2021) showed that humans may need a few blocks of training and test to start showing differentiation.

We thank the reviewer for mentioning this. We have added a paragraph to the “Differentiation Requires a High Learning Rate and Is Sensitive to Activity Dynamics” section of the Discussion that addresses this point (pp. 28-29):

“Although the results from Wanjia et al. (2021) provide strong support for the model's prediction that differentiation will be abrupt, they raise another question: What explains variance across items in when this abrupt change takes place? The answer to this question remains to be seen, but one possibility is encoding variability: If we assume that participants stochastically sample (i.e., attend to) the features of the scene pairmates, it is possible that participants might initially fail to sample the features that distinguish the scene pairmates, which can be quite subtle – and if the distinguishing features of the pairmates are not represented in high-level visual regions (i.e., the pairmates are represented in these regions as having the same features), this could delay the onset of differentiation until the point at which the distinguishing features happen (by chance) to be sampled.”

Related to the point above, the high learning rate prediction also seems to be at odds with the finding that the cortex, which has slow learning (according to the theory of complementary learning systems), also shows differentiation in Wammes et al (2022).

We now address this point in the section of the Discussion entitled “Differentiation Requires a High Learning Rate and Is Sensitive to Activity Dynamics” (p. 27):

“Our finding that differentiation requires a high learning rate suggests that differentiation will be more evident in the hippocampus than in neocortex, insofar as hippocampus is thought to have a higher learning rate than neocortex (McClelland et al., 1995). In keeping with this prediction, numerous studies have found differentiation effects in hippocampus but not in neocortical regions involved in sensory processing (e.g., Chanales et al., 2017; Favila et al., 2016; Zeithamova et al., 2018). At the same time, some studies have found differentiation effects in neocortex (e.g., Schlichting et al., 2015; Wammes et al., 2022). One possible explanation of these neocortical differentiation effects is that they are being ``propped up’’ by top-down feedback from differentiated representations in the hippocampus.”

More details about the learning dynamics would be helpful. For example, equation(s) showing how activation, learning rate and the NMPH function work together to change the weight of connections may be added. Without the information, it is unclear how each connection changes its value after each time point.

We thank the reviewer for this comment. We have made two major changes to address this concern. First, we have edited the “Learning” section within “Basic Network Properties” in the main text (pp. 6-7):

“Connection strengths in the model between pairs of connected units x and y were adjusted at the end of each trial (i.e., after each stimulus presentation) as a U-shaped function of the coactivity of x and y, defined as the product of their activations on that trial. The parameters of the U-shaped learning function relating coactivity to change in connection strength (i.e., weakening / strengthening) were specified differently for each projection where learning occurs (bidirectionally between the input and hidden layers, the hidden layer to itself, and the hidden to output layer). Once the U-shaped learning function for each projection in each version of the model was specified, we did not change it for any of the various conditions. Details of how we computed coactivity and how we specified the U-shaped function can be found in the Methods section.”

Second, we have added the requested equations to the “Learning” part of the Methods (pp. 37-38):

The right side of the function, strong activation leads to strengthening of the connectivity, which I assume will lead to stronger activation on the next time point. The model has an upper limit of connection strength to prevent connection from strengthening too much. The same idea can be applied to the left side of the function: instead of having two turning points, it can be a linear function such that low activation keeps weakening connection until the lower limit is reached. This way the NMPH function can take a simpler form (e.g., two line-segments if you think the weakening and strengthening take different rates) and may still simulate the data.

We thank the reviewer for mentioning this. We have added a new paragraph in the “Learning” section of the Methods to justify the particular shape of the learning curve (pp. 38-39):

“Evidence for the U-shaped plasticity function used here (where low activation leads to no change, moderate activation leads to weakening, and higher levels of activation lead to strengthening) was previously reviewed in Ritvo et al. (2019). In brief, there are three lines of work that support the U shape: First, multiple neurophysiological studies have found that moderate postsynaptic depolarization leads to synaptic weakening and higher levels of depolarization lead to synaptic strengthening (e.g., Artola et al., 1990; Hansel et al., 1996). Second, human neuroscience studies have used pattern classifiers, applied to fMRI and EEG data, to measure memory activation, and have related this measure to subsequent memory accessibility; several studies using this approach have found that low levels of activation lead to no change in memory strength, moderate levels of activation lead to impaired subsequent memory, and higher levels of activation lead to increased subsequent memory (e.g., Newman and Norman, 2010; Detre et al., 2013; Kim et al., 2014; for related findings, see Lewis-Peacock and Norman, 2014; Wang et al., 2019). Third, a recent human fMRI study by Wammes et al. (2022) manipulated memory activation by varying the visual similarity of pairmates and observed a U-shaped function relating visual similarity to representational change in the hippocampus, whereby low levels of pairmate similarity were associated with no change, moderate levels of similarity were associated with differentiation, and the differentiation effect went away at higher levels of similarity.

We have also included a pointer to this new paragraph in the “Nonmonotonic Plasticity Hypothesis” section of Introduction (p. 2):

(for further discussion of the empirical justification for the NMPH, see the Learning subsection in the Methods)”

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

A few additional minor things about data presentation and the like:

(1) Figure 1 legend - a more general description of how to interpret the figure might be helpful for more naive readers (e.g., explaining how one can visualize in the schematic that there is overlap in the hidden layer between A and B). Also, from the Figure 1 depiction, it's not clear what is different about the setup from the initial left hand side panels in A, B, C, to make it such that activity spreads strongly to A in panel A, weakly in panel B, and not at all in panel C since the weights are the same. Is there a way to incorporate this into the graphic, or describe it in words?

To address this point, we have added the following text to the Figure 1 caption (p. 3):

“Note that the figure illustrates the consequences of differences in competitor activation for learning, without explaining why these differences would arise. For discussion of circumstances that could lead to varying levels of competitor activation, see the simulations described in the text.”

(2) I believe not all of the papers cited on lines 193-195 actually have similarity manipulations in them. I'd recommend double checking this list and removing those less relevant to the statement.

Thank you for pointing this out; we have removed the Ballard reference and we have clarified what we mean by similarity reversal (p. 7):

“The study was inspired by recent neuroimaging studies showing ``similarity reversals'', wherein stimuli that have more features in common (or share a common associate) show less hippocampal pattern similarity (Favila et al., 2016; Schlichting et al., 2015; Molitor et al., 2021; Chanales et al., 2017; Dimsdale-Zucker et al., 2018; Wanjia et al., 2021; Zeithamova et al., 2018; Jiang et al., 2020; Wammes et al., 2022).”

(3) I wanted a bit more detail about how the parameters were set in the main paper, not just in the methods. Even something as brief as noting that model fitting was done by hand by tweaking parameters to re-create the empirical patterns (if I'm understanding correctly) would have been helpful for me.

To address this point, we have added the following text under “Basic Network Properties” (p. 4):

“Our goal was to qualitatively fit key patterns of results from each of the aforementioned studies. We fit the parameters of the model by hand as they are highly interdependent (see the Methods section for more details).”

(4) In Figure 4E, it would be helpful to describe the x and y axes of the MDS plots in the legend.

To address this point, we have added the following new text to the Figure 4 caption that clarifies how the MDS plots were generated (p. 11):

“MDS plots were rotated, shifted, and scaled such that pairmate 1before is located at (0,0), pairmate 2before is located directly to the right of pairmate 1before, and the distance between pairmate 1before and pairmate 2before is proportional to the baseline distance between the pairmates.”

(5) Figure 6 - at first I thought the thicker line was some sort of baseline, but I think it is just many traces on top of one another. If other readers may be similarly confused, perhaps this could be stated.

Thanks for this comment. We have updated Figure 6 (p. 16).

We have also updated the caption.

I am having a lot of difficulty understanding the terms "competitor-to-competitor,"

"competitor-to-target/shared," and "target/shared-to-target/shared," and therefore I don't fully get Figure 5. I think it might be helpful to expand the description of these terms where they are first introduced in the paper (p. 13?). I think I am missing something crucial here, and I am not quite sure what that is-which I know is not very helpful! But, to narrate my confusion a bit, I thought that these terms would somehow relate to connections between different connections of the network. For example is competitor-to-competitor within the hidden layer? Or is this somehow combining across relevant connections that might span different pairs of layers in the model? And, I really have no idea why it is "target/shared."

Thank you for these comments. We have updated Figure 5 and we have also made several changes to the main text and the figure caption to address these points.

Changes to the main text (p. 13):

“Whether symmetric or asymmetric integration occurs depends on the relative strengths of connections between pairs of unique competitor units (competitor-competitor connections) compared to connections between unique competitor units and shared units (competitor-shared connections) after the first trial (Figure 5; note that the figure focuses on connections between hidden units, but the principle also applies to connections that span across layers). Generally, coactivity between unique competitor units (competitor-competitor coactivity) is less than coactivity between unique competitor units and shared units (competitor-shared coactivity), which is less than coactivity between unique target units and shared units (target-shared coactivity).”

(7) Relatedly in Figure 13, I understand how some competitor-to-target/shared connections could be spared in the bottom instance given panel B. However, I'm struggling to understand how that relates to the values in the corresponding chart in panel A. What about panel A, bottom (vs. the top) means lower coactivities between some competitor-to-target/shared? Is it because if the noise level is higher, the "true" activation of competitor-to-target/shared connections is weaker? I think again, I'm missing something critical here! and wonder if other readers may be in the same situation. (I know the authors described this also on p. 36, but I'm still confused!)

We have updated Figure 13 to clarify these points.

(8)  In Figure 9, I believe there is no caption for panel D. Also, it looks as though the item unit active for A and B is the same. I wonder if this is an error?

Thank you for catching these errors! They have both been fixed.

Reviewer #2 (Recommendations For The Authors):

-Perhaps I missed it, but I think defining coactivity (how it is computed) in the main text would be useful for readers, as this is critical for understanding the model. I did find it in the methods.

We thank the reviewer for this suggestion. We have updated the “Learning” section within “Basic Network Properties” in the main text to address this point (pp. 6-7):

“Connection strengths in the model between pairs of connected units x and y were adjusted at the end of each trial (i.e., after each stimulus presentation) as a U-shaped function of the coactivity of x and y, defined as the product of their activations on that trial. The parameters of the U-shaped learning function relating coactivity to change in connection strength (i.e., weakening / strengthening) were specified differently for each projection where learning occurs (bidirectionally between the input and hidden layers, the hidden layer to itself, and the hidden to output layer). Once the U-shaped learning function for each projection in each version of the model was specified, we did not change it for any of the various conditions. Details of how we computed coactivity and how we specified the U-shaped function can be found in the Methods section.”

-The modeling results in the different face condition are at odds with the data for the Favila et al model (they observe some differentiation in the paper and the model predicts no change). This could be due to a number of unmodeled factors, but it is perhaps worth noting.

Thank you for pointing this out. It is possible to better capture the pattern of results observed by Favila et al. in their paper (with some differentiation in the different-face condition and even more differentiation in the same-face condition) by slightly adjusting the model parameters (specifically, by setting the oscillation amplitude Osc for the hidden layer to .1 instead of .067).

Rather than replacing the old (Osc \= .067) results in the paper, which would entail re-making the associated videos, etc., we have added a supplementary figure (Figure 8 - Supplement 1; see p.45):

We also added new text to the Favila Results, under “Differentiation and Integration” (p. 20):

“Note also that the exact levels of differentiation that are observed in the different-face and same-face conditions are parameter dependent; for an alternative set of results showing some differentiation in the different-face condition (but still less than is observed in the same-face condition), see Figure 8 - Supplement 1.”

-Related to my comment in the public review about pre-wiring associations, in the caption for Figure 9 (Schlichting model), the authors report "In both conditions, the pre-wired connection linking the "item B" hidden units to the "item X" output unit is set to .7. In the interleaved condition, the connection linking the "item A" hidden units to the "item X" output unit is set to .8, to reflect some amount of initial AX learning. In the blocked condition, the connection linking the "item A" hidden units to the "item X" output unit is set a higher value (.999), to reflect extra AX learning." What are the equivalent values for the other models, especially the Favila model since the structure is the same as Schlichting? I understood all the "strong" connections to be .99 unless otherwise stated. If that's the case, I don't understand why the blocked Schlichting model and the Favila model produce opposite effects. More clarity would be useful here.

We have added a new paragraph to the results section for the Schlicting model (under “Differentiation and Integration”) to clarify why the blocked Schlichting model and the Favila model show different results (p. 24):

“Note that the key feature driving integration in the blocked condition of this simulation is not the high strength of the connection from X to A on its own – rather, it is the asymmetry in the pretrained connection strengths from X to A (.999) and from X to B (.7). This asymmetry, which is meant to reflect the extensive training on A-X that occurred before the initial presentation of B-X, results in the A-X hidden representation decisively winning the competition during B-X presentation, which then leads to the B input also being linked to this representation (i.e., integration). It is instructive to compare this to the same-face condition from our simulation of Favila et al. (2016): In that simulation, the two pairmates are also linked strongly (.99 initial connection strength) to a shared associate, but in that case the connections are equally strong, so there is more balanced competition -- in this case, the competitor representation only comes to mind moderately (instead of displacing the target representation), so the result is differentiation instead of integration.”

-The meaning of the different colored dots in Figure 5 is bit hard to keep track of, even given the legend labels. The figure might benefit from a model sketch highlighting each of the different coactivity types. The left side of Fig 13 was useful but again somehow mapping on the colors would help further. Another note on these figures: what does having two dots of each color mean? Is it just an illustration of the variance? There would be more dots if there was one dot per coactivity value.

We have updated Figure 5 and Figure 13 to clarify these points (including a clarification that the dots only represent a subset of the possible pairings between units).

-While I appreciate the goal of the paper is to account for these three studies, readers who aren't familiar with or specifically interested in these studies may appreciate a small amount of intuition on why formalizing unsupervised learning models may be broadly important for computational investigations of learning/memory/cognition.

We have added the following text under “Basic Network Properties” in the Introduction to address this point (p. 4):

“Achieving a better understanding of unsupervised learning is an important goal for computational neuroscience, given that learning agents have vastly more opportunities to learn in an unsupervised fashion than from direct supervision (for additional discussion of this point, see, e.g., Zhuang et al., 2021).”

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

Summary:

This paper presents a compelling and comprehensive study of decision-making under uncertainty. It addresses a fundamental distinction between belief-based (cognitive neuroscience) formulations of choice behaviour with reward-based (behavioural psychology) accounts. Specifically, it asks whether active inference provides a better account of planning and decision-making, relative to reinforcement learning. To do this, the authors use a simple but elegant paradigm that includes choices about whether to seek both information and rewards. They then assess the evidence for active inference and reinforcement learning models of choice behaviour, respectively. After demonstrating that active inference provides a better explanation of behavioural responses, the neuronal correlates of epistemic and instrumental value (under an optimised active inference model) are characterised using EEG. Significant neuronal correlates of both kinds of value were found in sensor and source space. The source space correlates are then discussed sensibly, in relation to the existing literature on the functional anatomy of perceptual and instrumental decision-making under uncertainty.

Strengths:

The strengths of this work rest upon the theoretical underpinnings and careful deconstruction of the various determinants of choice behaviour using active inference. A particular strength here is that the experimental paradigm is designed carefully to elicit both information-seeking and reward-seeking behaviour; where the information-seeking is itself separated into resolving uncertainty about the context (i.e., latent states) and the contingencies (i.e., latent parameters), under which choices are made. In other words, the paradigm - and its subsequent modelling - addresses both inference and learning as necessary belief and knowledge-updating processes that underwrite decisions.

The authors were then able to model belief updating using active inference and then look for the neuronal correlates of the implicit planning or policy selection. This speaks to a further strength of this study; it provides some construct validity for the modelling of belief updating and decision-making; in terms of the functional anatomy as revealed by EEG. Empirically, the source space analysis of the neuronal correlates licences some discussion of functional specialisation and integration at various stages in the choices and decision-making.

In short, the strengths of this work rest upon a (first) principles account of decision-making under uncertainty in terms of belief updating that allows them to model or fit choice behaviour in terms of Bayesian belief updating - and then use relatively state-of-the-art source reconstruction to examine the neuronal correlates of the implicit cognitive processing.

Response: We are deeply grateful for your careful review of our work and for the thoughtful feedback you have provided. Your dedication to ensuring the quality and clarity of the work is truly admirable. Your comments have been invaluable in guiding us towards improving the paper, and We appreciate your time and effort in not just offering suggestions but also providing specific revisions that I can implement. Your insights have helped us identify areas where I can strengthen the arguments and clarify the methodology.

Comment 1:

The main weaknesses of this report lies in the communication of the ideas and procedures. Although the language is generally excellent, there are some grammatical lapses that make the text difficult to read. More importantly, the authors are not consistent in their use of some terms; for example, uncertainty and information gain are sometimes conflated in a way that might confuse readers. Furthermore, the descriptions of the modelling and data analysis are incomplete. These shortcomings could be addressed in the following way.

First, it would be useful to unpack the various interpretations of information and goal-seeking offered in the (active inference) framework examined in this study. For example, it will be good to include the following paragraph:

"In contrast to behaviourist approaches to planning and decision-making, active inference formulates the requisite cognitive processing in terms of belief updating in which choices are made based upon their expected free energy. Expected free energy can be regarded as a universal objective function, specifying the relative likelihood of alternative choices. In brief, expected free energy can be regarded as the surprise expected following some action, where the expected surprise comes in two flavours. First, the expected surprise is uncertainty, which means that policies with a low expected free energy resolve uncertainty and promote information seeking. However, one can also minimise expected surprise by avoiding surprising, aversive outcomes. This leads to goal-seeking behaviour, where the goals can be regarded as prior preferences or rewarding outcomes.

Technically, expected free energy can be expressed in terms of risk plus ambiguity - or rearranged to be expressed in terms of expected information gain plus expected value, where value corresponds to (log) prior preferences. We will refer to both decompositions in what follows; noting that both decompositions accommodate information and goal-seeking imperatives. That is, resolving ambiguity and maximising information gain have epistemic value, while minimising risk or maximising expected value have pragmatic or instrumental value. These two kinds of values are sometimes referred to in terms of intrinsic and extrinsic value, respectively [1-4]."

Response 1: We deeply thank you for your comments and corresponding suggestions about our interpretations of active inference. In response to your identified weaknesses and suggestions, we have added corresponding paragraphs in the Methods section (The free energy principle and active inference, line 95-106):

“Active inference formulates the necessary cognitive processing as a process of belief updating, where choices depend on agents' expected free energy. Expected free energy serves as a universal objective function, guiding both perception and action. In brief, expected free energy can be seen as the expected surprise following some policies. The expected surprise can be reduced by resolving uncertainty, and one can select policies with lower expected free energy which can encourage information-seeking and resolve uncertainty. Additionally, one can minimize expected surprise by avoiding surprising or aversive outcomes (oudeyer et al., 2007; Schmidhuber et al., 2010). This leads to goal-seeking behavior, where goals can be viewed as prior preferences or rewarding outcomes.

Technically, expected free energy can also be expressed as expected information gain plus expected value, where the value corresponds to (log) prior preferences. We will refer to both formulations in what follows. Resolving ambiguity, minimizing risk, and maximizing information gain has epistemic value while maximizing expected value have pragmatic or instrumental value. These two types of values can be referred to in terms of intrinsic and extrinsic value, respectively (Barto et al., 2013; Schwartenbeck et al., 2019).”

Oudeyer, P. Y., & Kaplan, F. (2007). What is intrinsic motivation? A typology of computational approaches. Frontiers in neurorobotics, 1, 108.

Schmidhuber, J. (2010). Formal theory of creativity, fun, and intrinsic motivation (1990–2010). IEEE transactions on autonomous mental development, 2(3), 230-247.

Barto, A., Mirolli, M., & Baldassarre, G. (2013). Novelty or surprise?. Frontiers in psychology, 4, 61898.

Schwartenbeck, P., Passecker, J., Hauser, T. U., FitzGerald, T. H., Kronbichler, M., & Friston, K. J. (2019). Computational mechanisms of curiosity and goal-directed exploration. elife, 8, e41703.

Comment 2:

The description of the modelling of choice behaviour needs to be unpacked and motivated more carefully. Perhaps along the following lines:

"To assess the evidence for active inference over reinforcement learning, we fit active inference and reinforcement learning models to the choice behaviour of each subject. Effectively, this involved optimising the free parameters of active inference and reinforcement learning models to maximise the likelihood of empirical choices. The resulting (marginal) likelihood was then used as the evidence for each model. The free parameters for the active inference model scaled the contribution of the three terms that constitute the expected free energy (in Equation 6). These coefficients can be regarded as precisions that characterise each subjects' prior beliefs about contingencies and rewards. For example, increasing the precision or the epistemic value associated with model parameters means the subject would update her beliefs about reward contingencies more quickly than a subject who has precise prior beliefs about reward distributions. Similarly, subjects with a high precision over prior preferences or extrinsic value can be read as having more precise beliefs that she will be rewarded. The free parameters for the reinforcement learning model included..."

Response 2: We deeply thank you for your comments and corresponding suggestions about our description of the behavioral modelling. In response to your identified weaknesses and suggestions, we have added corresponding content in the Results section (Behavioral results, line 279-293):

“To assess the evidence for active inference over reinforcement learning, we fit active inference (Eq.9), model-free reinforcement learning, and model-based reinforcement learning models to the behavioral data of each participant. This involved optimizing the free parameters of active inference and reinforcement learning models. The resulting likelihood was used to calculate the Bayesian Information Criterion (BIC) (Vrieze 2012) as the evidence for each model. The free parameters for the active inference model (AL, AI, EX, prior, and α) scaled the contribution of the three terms that constitute the expected free energy in Eq.9. These coefficients can be regarded as precisions that characterize each participant's prior beliefs about contingencies and rewards. For example, increasing α means participants would update their beliefs about reward contingencies more quickly, increasing AL means participants would like to reduce ambiguity more, and increasing AI means participants would like to learn the hidden state of the environment and avoid risk more. The free parameters for the model-free reinforcement learning model are the learning rate α and the temperature parameter γ and the free parameters for the model-based are the learning rate α, the temperature parameter γ and prior (the details for the model-free reinforcement learning model can be seen in Eq.S1-11 and the details for the model-based reinforcement learning model can be seen Eq.S12-23 in the Supplementary Method). The parameter fitting for these three models was conducted using the `BayesianOptimization' package in Python (Frazire 2018), first randomly sampling 1000 times and then iterating for an additional 1000 times.”

Vrieze, S. I. (2012). Model selection and psychological theory: a discussion of the differences between the Akaike information criterion (AIC) and the Bayesian information criterion (BIC). Psychological methods, 17(2), 228.

Frazier, P. I. (2018). A tutorial on Bayesian optimization. arXiv preprint arXiv:1807.02811.

Comment 3:

In terms of the time-dependent correlations with expected free energy - and its constituent terms - I think the report would benefit from overviewing these analyses with something like the following:

"In the final analysis of the neuronal correlates of belief updating - as quantified by the epistemic and intrinsic values of expected free energy - we present a series of analyses in source space. These analyses tested for correlations between constituent terms in expected free energy and neuronal responses in source space. These correlations were over trials (and subjects). Because we were dealing with two-second timeseries, we were able to identify the periods of time during decision-making when the correlates were expressed.

In these analyses, we focused on the induced power of neuronal activity at each point in time, at each brain source. To illustrate the functional specialisation of these neuronal correlates, we present whole-brain maps of correlation coefficients and pick out the most significant correlation for reporting fluctuations in selected correlations over two-second periods. These analyses are presented in a descriptive fashion to highlight the nature and variety of the neuronal correlates, which we unpack in relation to the existing EEG literature in the discussion. Note that we did not attempt to correct for multiple comparisons; largely, because the correlations observed were sustained over considerable time periods, which would be almost impossible under the null hypothesis of no correlations."

Response 3: We deeply thank you for your comments and corresponding suggestions about our description of the regression analysis in the source space. In response to your suggestions, we have added corresponding content in the Results section (EEG results at source level, line 331-347):

“In the final analysis of the neural correlates of the decision-making process, as quantified by the epistemic and intrinsic values of expected free energy, we presented a series of linear regressions in source space. These analyses tested for correlations over trials between constituent terms in expected free energy (the value of avoiding risk, the value of reducing ambiguity, extrinsic value, and expected free energy itself) and neural responses in source space. Additionally, we also investigated the neural correlate of (the degree of) risk, (the degree of) ambiguity, and prediction error. Because we were dealing with a two-second time series, we were able to identify the periods of time during decision-making when the correlates were expressed. The linear regression was run by the "mne.stats.linear regression" function in the MNE package (Activity ~ Regressor + Intercept). Activity is the activity amplitude of the EEG signal in the source space and regressor is one of the regressors that we mentioned (e.g., expected free energy, the value of reducing ambiguity, etc.).

In these analyses, we focused on the induced power of neural activity at each time point, in the brain source space. To illustrate the functional specialization of these neural correlates, we presented whole-brain maps of correlation coefficients and picked out the brain region with the most significant correlation for reporting fluctuations in selected correlations over two-second periods. These analyses were presented in a descriptive fashion to highlight the nature and variety of the neural correlates, which we unpacked in relation to the existing EEG literature in the discussion. Note that we did not attempt to correct for multiple comparisons; largely, because the correlations observed were sustained over considerable time periods, which would be almost impossible under the null hypothesis of no correlations.”

Comment 4:

There was a slight misdirection in the discussion of priors in the active inference framework. The notion that active inference requires a pre-specification of priors is a common misconception. Furthermore, it misses the point that the utility of Bayesian modelling is to identify the priors that each subject brings to the table. This could be easily addressed with something like the following in the discussion:

"It is a common misconception that Bayesian approaches to choice behaviour (including active inference) are limited by a particular choice of priors. As illustrated in our fitting of choice behaviour above, priors are a strength of Bayesian approaches in the following sense: under the complete class theorem [5, 6], any pair of choice behaviours and reward functions can be described in terms of ideal Bayesian decision-making with particular priors. In other words, there always exists a description of choice behaviour in terms of some priors. This means that one can, in principle, characterise any given behaviour in terms of the priors that explain that behaviour. In our example, these were effectively priors over the precision of various preferences or beliefs about contingencies that underwrite expected free energy."

Response 4: We deeply thank you for your comments and corresponding suggestions about the prior of Bayesian methods. In response to your suggestions, we have added corresponding content in the Discussion section (The strength of the active inference framework in decision-making, line 447-453):

“However, it may be the opposite. As illustrated in our fitting results, priors can be a strength of Bayesian approaches. Under the complete class theorem (Wald 1947; Brown 1981), any pair of behavioral data and reward functions can be described in terms of ideal Bayesian decision-making with particular priors. In other words, there always exists a description of behavioral data in terms of some priors. This means that one can, in principle, characterize any given behavioral data in terms of the priors that explain that behavior. In our example, these were effectively priors over the precision of various preferences or beliefs about contingencies that underwrite expected free energy.”

Wald, A. (1947). An essentially complete class of admissible decision functions. The Annals of Mathematical Statistics, 549-555.

Brown, L. D. (1981). A complete class theorem for statistical problems with finite sample spaces. The Annals of Statistics, 1289-1300.

Reviewer #2 (Public Review):

Summary:

Zhang and colleagues use a combination of behavioral, neural, and computational analyses to test an active inference model of exploration in a novel reinforcement learning task.

Strengths:

The paper addresses an important question (validation of active inference models of exploration). The combination of behavior, neuroimaging, and modeling is potentially powerful for answering this question.

Response: We want to express our sincere gratitude for your thorough review of our work and for the valuable comments you have provided. Your attention to detail and dedication to improving the quality of the work are truly commendable. Your feedback has been invaluable in guiding us towards revisions that will strengthen the work. We have made targeted modifications based on most of the comments. However, due to factors such as time and energy constraints, we have not added corresponding analyses for several comments.

Comment 1:

The paper does not discuss relevant work on contextual bandits by Schulz, Collins, and others. It also does not mention the neuroimaging study of Tomov et al. (2020) using a risky/safe bandit task.

Response 1:

We deeply thank you for your suggestions about the relevant work. We now discussion and cite these representative papers in the Introduction section (line 42-55):

“The decision-making process frequently involves grappling with varying forms of uncertainty, such as ambiguity - the kind of uncertainty that can be reduced through sampling, and risk - the inherent uncertainty (variance) presented by a stable environment. Studies have investigated these different forms of uncertainty in decision-making, focusing on their neural correlates (Daw et al., 2006; Badre et al., 2012; Cavanagh et al., 2012).

These studies utilized different forms of multi-armed bandit tasks, e.g the restless multi-armed bandit tasks (Daw et al., 2006; Guha et al., 2010), risky/safe bandit tasks (Tomov et al., 2020; Fan et al., 2022; Payzan et al., 2013), contextual multi-armed bandit tasks (Schulz et al., 2015; Schulz et al., 2015; Molinaro et al., 2023). However, these tasks either separate risk from ambiguity in uncertainty, or separate action from state (perception). In our work, we develop a contextual multi-armed bandit task to enable participants to actively reduce ambiguity, avoid risk, and maximize rewards using various policies (see Section 2.2) and Figure 4(a)). Our task makes it possible to study whether the brain represents these different types of uncertainty distinctly (Levy et al., 2010) and whether the brain represents both the value of reducing uncertainty and the degree of uncertainty. The active inference framework presents a theoretical approach to investigate these questions. Within this framework, uncertainties can be reduced to ambiguity and risk. Ambiguity is represented by the uncertainty about model parameters associated with choosing a particular action, while risk is signified by the variance of the environment's hidden states. The value of reducing ambiguity, the value of avoiding risk, and extrinsic value together constitute expected free energy (see Section 2.1).”

Daw, N. D., O'doherty, J. P., Dayan, P., Seymour, B., & Dolan, R. J. (2006). Cortical substrates for exploratory decisions in humans. Nature, 441(7095), 876-879.

Badre, D., Doll, B. B., Long, N. M., & Frank, M. J. (2012). Rostrolateral prefrontal cortex and individual differences in uncertainty-driven exploration. Neuron, 73(3), 595-607.

Cavanagh, J. F., Figueroa, C. M., Cohen, M. X., & Frank, M. J. (2012). Frontal theta reflects uncertainty and unexpectedness during exploration and exploitation. Cerebral cortex, 22(11), 2575-2586.

Guha, S., Munagala, K., & Shi, P. (2010). Approximation algorithms for restless bandit problems. Journal of the ACM (JACM), 58(1), 1-50.

Tomov, M. S., Truong, V. Q., Hundia, R. A., & Gershman, S. J. (2020). Dissociable neural correlates of uncertainty underlie different exploration strategies. Nature communications, 11(1), 2371.

Fan, H., Gershman, S. J., & Phelps, E. A. (2023). Trait somatic anxiety is associated with reduced directed exploration and underestimation of uncertainty. Nature Human Behaviour, 7(1), 102-113.

Payzan-LeNestour, E., Dunne, S., Bossaerts, P., & O’Doherty, J. P. (2013). The neural representation of unexpected uncertainty during value-based decision making. Neuron, 79(1), 191-201.

Schulz, E., Konstantinidis, E., & Speekenbrink, M. (2015, April). Exploration-exploitation in a contextual multi-armed bandit task. In International conference on cognitive modeling (pp. 118-123).

Schulz, E., Konstantinidis, E., & Speekenbrink, M. (2015, November). Learning and decisions in contextual multi-armed bandit tasks. In CogSci.

Molinaro, G., & Collins, A. G. (2023). Intrinsic rewards explain context-sensitive valuation in reinforcement learning. PLoS Biology, 21(7), e3002201.

Levy, I., Snell, J., Nelson, A. J., Rustichini, A., & Glimcher, P. W. (2010). Neural representation of subjective value under risk and ambiguity. Journal of neurophysiology, 103(2), 1036-1047.

Comment 2:

The statistical reporting is inadequate. In most cases, only p-values are reported, not the relevant statistics, degrees of freedom, etc. It was also not clear if any corrections for multiple comparisons were applied. Many of the EEG results are described as "strong" or "robust" with significance levels of p<0.05; I am skeptical in the absence of more details, particularly given the fact that the corresponding plots do not seem particularly strong to me.

Response 2: We deeply thank you for your comments about our statistical reporting. We have optimized the fitting model and rerun all the statistical analyses. As can be seen (Figure 6, 7, 8, S3, S4, S5), the new regression results are significantly improved compared to the previous ones. Due to the limitation of space, we place the other relevant statistical results, including t-values, std err, etc., on our GitHub (https://github.com/andlab-um/FreeEnergyEEG). Currently, we have not conducted multiple comparison corrections based on Reviewer 1’s comments (Comments 3) “Note that we did not attempt to correct for multiple comparisons; largely, because the correlations observed were sustained over considerable time periods, which would be almost impossible under the null hypothesis of no correlations”.

Author response image 1.

Comment 3:

The authors compare their active inference model to a "model-free RL" model. This model is not described anywhere, as far as I can tell. Thus, I have no idea how it was fit, how many parameters it has, etc. The active inference model fitting is also not described anywhere. Moreover, you cannot compare models based on log-likelihood, unless you are talking about held-out data. You need to penalize for model complexity. Finally, even if active inference outperforms a model-free RL model (doubtful given the error bars in Fig. 4c), I don't see how this is strong evidence for active inference per se. I would want to see a much more extensive model comparison, including model-based RL algorithms which are not based on active inference, as well as model recovery analyses confirming that the models can actually be distinguished on the basis of the experimental data.

Response 3: We deeply thank you for your comments about the model comparison details. We previously omitted some information about the comparison model, as classical reinforcement learning is not the focus of our work, so we put the specific details in the supplementary materials. Now we have placed relevant information in the main text (see the part we have highlighted in yellow). We have now added the relevant information regarding the model comparison in the Results section (Behavioral results, line 279-293):

“To assess the evidence for active inference over reinforcement learning, we fit active inference (Eq.9), model-free reinforcement learning, and model-based reinforcement learning models to the behavioral data of each participant. This involved optimizing the free parameters of active inference and reinforcement learning models. The resulting likelihood was used to calculate the Bayesian Information Criterion (BIC) as the evidence for each model. The free parameters for the active inference model (AL, AI, EX, prior, and α) scaled the contribution of the three terms that constitute the expected free energy in Eq.9. These coefficients can be regarded as precisions that characterize each participant's prior beliefs about contingencies and rewards. For example, increasing α means participants would update their beliefs about reward contingencies more quickly, increasing AL means participants would like to reduce ambiguity more, and increasing AI means participants would like to learn the hidden state of the environment and avoid risk more. The free parameters for the model-free reinforcement learning model are the learning rate α and the temperature parameter γ and the free parameters for the model-based are the learning rate α, the temperature parameter γ and prior (the details for the model-free reinforcement learning model can be found in Eq.S1-11 and the details for the model-based reinforcement learning model can be found in Eq.S12-23 in the Supplementary Method). The parameter fitting for these three models was conducted using the `BayesianOptimization' package in Python, first randomly sampling 1000 times and then iterating for an additional 1000 times.”

We have now incorporated model-based reinforcement learning into our comparison models and placed the descriptions of both model-free and model-based reinforcement learning algorithms in the supplementary materials. We have also changed the criterion for model comparison to Bayesian Information Criterion. As indicated by the results, the performance of the active inference model significantly outperforms both comparison models.

Sorry, we didn't do model recovery before, but now we have placed the relevant results in the supplementary materials. From the result figures, we can see that each model fits its own generated simulated data well:

“To demonstrate how reliable our models are (the active inference model, model-free reinforcement learning model, and model-based reinforcement learning model), we run some simulation experiments for model recovery. We use these three models, with their own fitting parameters, to generate some simulated data. Then we will fit all three sets of data using these three models.

The model recovery results are shown in Fig.S6. This is the confusion matrix of models: the percentage of all subjects simulated based on a certain model that is fitted best by a certain model. The goodness-of-fit was compared using the Bayesian Information Criterion. We can see that the result of model recovery is very good, and the simulated data generated by a model can be best explained by this model.”

Author response image 2.

Comment 4:

Another aspect of the behavioral modeling that's missing is a direct descriptive comparison between model and human behavior, beyond just plotting log-likelihoods (which are a very impoverished measure of what's going on).

Response 4: We deeply thank you for your comments about the comparison between the model and human behavior. Due to the slight differences between our simulation experiments and real behavioral experiments (the "you can ask" stage), we cannot directly compare the model and participants' behaviors. However, we can observe that in the main text's simulation experiment (Figure 3), the active inference agent's behavior is highly consistent with humans (Figure 4), exhibiting an effective exploration strategy and a desire to reduce uncertainty. Moreover, we have included two additional simulation experiments in the supplementary materials, which demonstrate that active inference may potentially fit a wide range of participants' behavioral strategies.

Author response image 3.

(An active inference agent with AL=AI=EX=0. It can accomplish tasks efficiently like a human being, reducing the uncertainty of the environment and maximizing the reward.)

Author response image 4.

(An active inference agent with AL=AI=0, EX=10. It will only pursue immediate rewards (not choosing the "Cue" option due to additional costs), but it can also gradually optimize its strategy due to random effects.)

Author response image 5.

(An active inference agent with EX=0, AI=AL=10. It will only pursue environmental information to reduce the uncertainty of the environment. Even in "Context 2" where immediate rewards are scarce, it will continue to explore.) (a) shows the decision-making of active inference agents in the Stay-Cue choice. Blue corresponds to agents choosing the "Cue" option and acquiring "Context 1"; orange corresponds to agents choosing the "Cue" option and acquiring "Context 2"; purple corresponds to agents choosing the "Stay" option and not knowing the information about the hidden state of the environment. The shaded areas below correspond to the probability of the agents making the respective choices. (b) shows the decision-making of active inference agents in the Stay-Cue choice. The shaded areas below correspond to the probability of the agents making the respective choices. (c) shows the rewards obtained by active inference agents. (d) shows the reward prediction errors of active inference agents. (e) shows the reward predictions of active inference agents for the "Risky" path in "Context 1" and "Context 2".

Comment 5:

The EEG results are intriguing, but it wasn't clear that these provide strong evidence specifically for the active inference model. No alternative models of the EEG data are evaluated.

Overall, the central claim in the Discussion ("we demonstrated that the active inference model framework effectively describes real-world decision-making") remains unvalidated in my opinion.

Response 5: We deeply thank you for your comments. We applied the active inference model to analyze EEG results because it best fit the participants' behavioral data among our models, including the new added results. Further, our EEG results serve only to verify that the active inference model can be used to analyze the neural mechanisms of decision-making in uncertain environments (if possible, we could certainly design a more excellent reinforcement learning model with a similar exploration strategy). We aim to emphasize the consistency between active inference and human decision-making in uncertain environments, as we have discussed in the article. Active inference emphasizes both perception and action, which is also what we wish to highlight: during the decision-making process, participants not only passively receive information, but also actively adopt different strategies to reduce uncertainty and maximize rewards.

Reviewer #3 (Public Review):

Summary:

This paper aims to investigate how the human brain represents different forms of value and uncertainty that participate in active inference within a free-energy framework, in a two-stage decision task involving contextual information sampling, and choices between safe and risky rewards, which promotes a shift from exploration to exploitation. They examine neural correlates by recording EEG and comparing activity in the first vs second half of trials and between trials in which subjects did and did not sample contextual information, and perform a regression with free-energy-related regressors against data "mapped to source space." Their results show effects in various regions, which they take to indicate that the brain does perform this task through the theorised active inference scheme.

Strengths:

This is an interesting two-stage paradigm that incorporates several interesting processes of learning, exploration/exploitation, and information sampling. Although scalp/brain regions showing sensitivity to the active-inference-related quantities do not necessarily suggest what role they play, it can be illuminating and useful to search for such effects as candidates for further investigation. The aims are ambitious, and methodologically it is impressive to include extensive free-energy theory, behavioural modelling, and EEG source-level analysis in one paper.

Response: We would like to express our heartfelt thanks to you for carefully reviewing our work and offering insightful feedback. Your attention to detail and commitment to enhancing the overall quality of our work are deeply admirable. Your input has been extremely helpful in guiding us through the necessary revisions to enhance the work. We have implemented focused changes based on a majority of your comments. Nevertheless, owing to limitations such as time and resources, we have not included corresponding analyses for a few comments.

Comment 1:

Though I could surmise the above general aims, I could not follow the important details of what quantities were being distinguished and sought in the EEG and why. Some of this is down to theoretical complexity - the dizzying array of constructs and terms with complex interrelationships, which may simply be part and parcel of free-energy-based theories of active inference - but much of it is down to missing or ambiguous details.

Response 1: We deeply thank you for your comments about our work’s readability. We have significantly revised the descriptions of active inference, models, research questions, etc. Focusing on active inference and the free energy principle, we have added relevant basic descriptions and unified the terminology. We have added information related to model comparison in the main text and supplementary materials. We presented our regression results in clearer language. Our research focused on the brain's representation of decision-making in uncertain environments, including expected free energy, the value of reducing ambiguity, the value of avoiding risk, extrinsic value, ambiguity, and risk.

Comment 2:

In general, an insufficient effort has been made to make the paper accessible to readers not steeped in the free energy principle and active inference. There are critical inconsistencies in key terminology; for example, the introduction states that aim 1 is to distinguish the EEG correlates of three different types of uncertainty: ambiguity, risk, and unexpected uncertainty. But the abstract instead highlights distinctions in EEG correlates between "uncertainty... and... risk" and between "expected free energy .. and ... uncertainty." There are also inconsistencies in mathematical labelling (e.g. in one place 'p(s|o)' and 'q(s)' swap their meanings from one sentence to the very next).

Response 2: We deeply thank you for your comments about the problem of inconsistent terminology. First, we have unified the symbols and letters (P, Q, s, o, etc.) that appeared in the article and described their respective meanings more clearly. We have also revised the relevant expressions of "uncertainty" throughout the text. In our work, uncertainty refers to ambiguity and risk. Ambiguity can be reduced through continuous sampling and is referred to as uncertainty about model parameters in our work. Risk, on the other hand, is the inherent variance of the environment and cannot be reduced through sampling, which is referred to as uncertainty about hidden states in our work. In the analysis of the results, we focused on how the brain encodes the value of reducing ambiguity (Figure 8), the value of avoiding risk (Figure 6), and (the degree of) ambiguity (Figure S5) during action selection. We also analyzed how the brain encodes reducing ambiguity and avoiding risk during belief update (Figure 7).

Comment 3:

Some basic but important task information is missing, and makes a huge difference to how decision quantities can be decoded from EEG. For example:

- How do the subjects press the left/right buttons - with different hands or different fingers on the same hand?

Response 3: We deeply thank you for your comments about the missing task information. We have added the relevant content in the Methods section (Contextual two-armed bandit task and Data collection, line 251-253):

“Each stage was separated by a jitter ranging from 0.6 to 1.0 seconds. The entire experiment consists of a single block with a total of 120 trials. The participants are required to use any two fingers of one hand to press the buttons (left arrow and right arrow on the keyboard).”

Comment 4:

- Was the presentation of the Stay/cue and safe/risky options on the left/right sides counterbalanced? If not, decisions can be formed well in advance especially once a policy is in place.

Response 4: The presentation of the Stay/cue and safe/risky options on the left/right sides was not counterbalanced. It is true that participants may have made decisions ahead of time. However, to better study the state of participants during decision-making, our choice stages consist of two parts. In the first two seconds, we ask participants to consider which option they would choose, and after these two seconds, participants are allowed to make their choice (by pressing the button).

We also updated the figure of the experiment procedure as below (We circled the time that the participants spent on making decisions).

Author response image 6.

Comment 5:

- What were the actual reward distributions ("magnitude X with probability p, magnitude y with probability 1-p") in the risky option?

Response 5: We deeply thank you for your comments about the missing task information. We have placed the relevant content in the Methods section (Contextual two-armed bandit task and Data collection, line 188-191):

“The actual reward distribution of the risky path in "Context 1" was [+12 (55%), +9 (25%), +6 (10%), +3 (5%), +0 (5%)] and the actual reward distribution of the risky path in "Context 2" was [+12 (5%), +9 (5%), +6 (10%), +3 (25%), +0 (55%)].”

Comment 6:

The EEG analysis is not sufficiently detailed and motivated.

For example,

- why the high lower-filter cutoff of 1 Hz, and shouldn't it be acknowledged that this removes from the EEG any sustained, iteratively updated representation that evolves with learning across trials?

Response 6: We deeply thank you for your comments about our EEG analysis. The 1Hz high-pass filter may indeed filter out some useful information. We chose a 1Hz high-pass filter to filter out most of the noise and prevent the noise from affecting our results analysis. Additionally, there are also many decision-related works that have applied 1Hz high-pass filtering in EEG data preprocessing (Yau et al., 2021; Cortes et al., 2021; Wischnewski et al., 2022; Schutte et al., 2017; Mennella et al., 2020; Giustiniani et al., 2020).

Yau, Y., Hinault, T., Taylor, M., Cisek, P., Fellows, L. K., & Dagher, A. (2021). Evidence and urgency related EEG signals during dynamic decision-making in humans. Journal of Neuroscience, 41(26), 5711-5722.

Cortes, P. M., García-Hernández, J. P., Iribe-Burgos, F. A., Hernández-González, M., Sotelo-Tapia, C., & Guevara, M. A. (2021). Temporal division of the decision-making process: An EEG study. Brain Research, 1769, 147592.

Wischnewski, M., & Compen, B. (2022). Effects of theta transcranial alternating current stimulation (tACS) on exploration and exploitation during uncertain decision-making. Behavioural Brain Research, 426, 113840.

Schutte, I., Kenemans, J. L., & Schutter, D. J. (2017). Resting-state theta/beta EEG ratio is associated with reward-and punishment-related reversal learning. Cognitive, Affective, & Behavioral Neuroscience, 17, 754-763.

Mennella, R., Vilarem, E., & Grèzes, J. (2020). Rapid approach-avoidance responses to emotional displays reflect value-based decisions: Neural evidence from an EEG study. NeuroImage, 222, 117253.

Giustiniani, J., Nicolier, M., Teti Mayer, J., Chabin, T., Masse, C., Galmès, N., ... & Gabriel, D. (2020). Behavioral and neural arguments of motivational influence on decision making during uncertainty. Frontiers in Neuroscience, 14, 583.

Comment 7:

- Since the EEG analysis was done using an array of free-energy-related variables in a regression, was multicollinearity checked between these variables?

Response 7: We deeply thank you for your comments about our regression. Indeed, we didn't specify our regression formula in the main text. We conducted regression on one variable each time, so there was no need for a multicollinearity check. We have now added the relevant content in the Results section (“EEG results at source level” section, line 337-340):

“The linear regression was run by the "mne.stats.linear regression" function in the MNE package (Activity ~ Regressor + Intercept). Activity is the activity amplitude of the EEG signal in the source space and regressor is one of the regressors that we mentioned (e.g., expected free energy, the value of reducing ambiguity, etc.).”

Comment 8:

- In the initial comparison of the first/second half, why just 5 clusters of electrodes, and why these particular clusters?

Response 8: We deeply thank you for your comments about our sensor-level analysis. These five clusters are relatively common scalp EEG regions to analyze (left frontal, right frontal, central, left parietal, and right parietal), and we referred previous work analyzed these five clusters of electrodes (Laufs et al., 2006; Ray et al., 1985; Cole et al., 1985). In addition, our work pays more attention to the analysis in source space, exploring the corresponding functions of specific brain regions based on active inference models.

Laufs, H., Holt, J. L., Elfont, R., Krams, M., Paul, J. S., Krakow, K., & Kleinschmidt, A. (2006). Where the BOLD signal goes when alpha EEG leaves. Neuroimage, 31(4), 1408-1418.

Ray, W. J., & Cole, H. W. (1985). EEG activity during cognitive processing: influence of attentional factors. International Journal of Psychophysiology, 3(1), 43-48.

Cole, H. W., & Ray, W. J. (1985). EEG correlates of emotional tasks related to attentional demands. International Journal of Psychophysiology, 3(1), 33-41.

Comment 9:

How many different variables are systematically different in the first vs second half, and how do you rule out less interesting time-on-task effects such as engagement or alertness? In what time windows are these amplitudes being measured?

Response 9 (and the Response for Weaknesses 11): There were no systematic differences between the first half and the second half of the trials, with the only difference being the participants' experience. In the second half, participants had a better understanding of the reward distribution of the task (less ambiguity). The simulation results can well describe these.

Author response image 7.

As shown in Figure (a), agents can only learn about the hidden state of the environment ("Context 1" (green) or "Context 2" (orange)) by choosing the "Cue" option. If agents choose the "Stay" option, they will not be able to know the hidden state of the environment (purple). The risk of agents is only related to wh

ether they choose the "Cue" option, not the number of rounds. Figure (b) shows the Safe-Risky choices of agents, and Figure (e) is the reward prediction of agents for the "Risky" path in "Context 1" and "Context 2". We can see that agents update the expected reward and reduce ambiguity by sampling the "Risky" path. The ambiguity of agents is not related to the "Cue" option, but to the number of times they sample the "Risky" path (rounds).

In our choosing stages, participants were required to think about their choices for the first two seconds (during which they could not press buttons). Then, they were asked to make their choices (press buttons) within the next two seconds. This setup effectively kept participants' attention focused on the task. And the two second during the “Second choice” stage when participants decide which option to choose (they cannot press buttons) are measured for the analysis of the sensor-level results.

Comment 10:

In the comparison of asked and not-asked trials, what trial stage and time window is being measured?

Response 10: We have added relevant descriptions in the main text. The two second during the “Second choice” stage when participants decide which option to choose (they cannot press buttons) are measured for the analysis of the sensor-level results.

Author response image 8.

Comment 11:

Again, how many different variables, of the many estimated per trial in the active inference model, are different in the asked and not-asked trials, and how can you know which of these differences is the one reflected in the EEG effects?

Response 11: The difference between asked trials and not-asked trials lies only in whether participants know the specific context of the risky path (the level of risk for the participants). A simple comparison indeed cannot tell us which of these differences is reflected in the EEG effects. Therefore, we subsequently conducted model-based regression analysis in the source space.

Comment 12:

The authors choose to interpret that on not-asked trials the subjects are more uncertain because the cue doesn't give them the context, but you could equally argue that they don't ask because they are more certain of the possible hidden states.

Response 12: Our task design involves randomly varying the context of the risky path. Only by choosing to inquire can participants learn about the context. Participants can only become increasingly certain about the reward distribution of different contexts of the risky path, but cannot determine which specific context it is. Here are the instructions for the task that we will tell the participants (line 226-231).

"You are on a quest for apples in a forest, beginning with 5 apples. You encounter two paths: 1) The left path offers a fixed yield of 6 apples per excursion. 2) The right path offers a probabilistic reward of 0/3/6/9/12 apples, and it has two distinct contexts, labeled "Context 1" and "Context 2," each with a different reward distribution. Note that the context associated with the right path will randomly change in each trial. Before selecting a path, a ranger will provide information about the context of the right path ("Context 1" or "Context 2") in exchange for an apple. The more apples you collect, the greater your monetary reward will be."

Comment 13:

- The EEG regressors are not fully explained. For example, an "active learning" regressor is listed as one of the 4 at the beginning of section 3.3, but it is the first mention of this term in the paper and the term does not arise once in the methods.

Response 13: We have accordingly revised the relevant content in the main text (as in Eq.8). Our regressors now include expected free energy, the value of reducing ambiguity, the value of avoiding risk, extrinsic value, prediction error, (the degree of) ambiguity, reducing ambiguity, and avoiding risk.

Comment 14:

- In general, it is not clear how one can know that the EEG results reflect that the brain is purposefully encoding these very parameters while implementing this very mechanism, and not other, possibly simpler, factors that correlate with them since there is no engagement with such potential confounds or alternative models. For example, a model-free reinforcement learning model is fit to behaviour for comparison. Why not the EEG?

Response 14: We deeply thank you for your comments. Due to factors such as time and effort, and because the active inference model best fits the behavioral data of the participants, we did not use other models to analyze the EEG data. At both the sensor and source level, we observed the EEG signal and brain regions that can encode different levels of uncertainties (risk and ambiguity). The brain's uncertainty driven exploration mechanism cannot be explained solely by a simple model-free reinforcement learning approach.

Recommendations for the authors:

Response: We have made point-to-point revisions according to the reviewer's recommendations, and as these revisions are relatively minor, we have only responded to the longer recommendations here.

Reviewer #1 (Recommendations For The Authors)

I enjoyed reading this sophisticated study of decision-making. I thought your implementation of active inference and the subsequent fitting to choice behaviour - and study of the neuronal (EEG) correlates - was impressive. As noted in my comments on strengths and weaknesses, some parts of your manuscript with difficult to read because of slight collapses in grammar and an inconsistent use of terms when referring to the mathematical quantities. In addition to the paragraphs I have suggested, I would recommend the following minor revisions to your text. In addition, you will have to fill in some of the details that were missing from the current version of the manuscript. For example:

Recommendation 1:

Which RL model did you use to fit the behavioural data? What were its free parameters?

Response 1: We have now added information related to the comparison models in the behavioral results and supplementary materials. We applied both simple model-free reinforcement learning and model-based reinforcement learning. The free parameters for the model-free reinforcement learning model are the learning rate α and the temperature parameter γ, while the free parameters for the model-based approach are the learning rate α, the temperature parameter γ, and the prior.

Recommendation 2:

When you talk about neuronal activity in the final analyses (of time-dependent correlations) what was used to measure the neuronal activity? Was this global power over frequencies? Was it at a particular frequency band? Was it the maximum amplitude within some small window et cetera? In other words, you need to provide the details of your analysis that would enable somebody to reproduce your study at a certain level of detail.

Response 2: In the final analyses, we used the activity amplitude at each point in the source space for our analysis. Previously, we had planned to make our data and models available on GitHub to facilitate easier replication of our work.

Reviewer #3 (Recommendations For The Authors)

Recommendation 1:

It might help to explain the complex concepts up front, to use the concrete example of the task itself - presumably, it was designed so that the crucial elements of the active inference framework come to the fore. One could use hypothetical choice patterns in this task to exemplify different factors such as expected free energy and unexpected uncertainty at work. It would also be illuminating to explain why behaviour on this task is fit better by the active inference model than a model-free reinforcement learning model.

Response 1: Thank you for your suggestions. We have given clearer explanations to the three terms in the active inference formula: the value of reducing ambiguity, the value of avoiding risk, and the extrinsic value (Eq.8), which makes it easier for readers to understand active inference.

In addition, we can simply view active inference as a computational model similar to model-based reinforcement learning, where the expected free energy represents a subjective value, without needing to understand its underlying computational principles or neurobiological background. In our discussion, we have argued why the active inference model fits the participants' behavior better than our reinforcement learning model, as the active inference model has an inherent exploration mechanism that is consistent with humans, who instinctively want to reduce environmental uncertainty (line 435-442).

“Active inference offers a superior exploration mechanism compared with basic model-free reinforcement learning  (Figure 4 (c)). Since traditional reinforcement learning models determine their policies solely on the state, this setting leads to difficulty in extracting temporal information (Laskin et al., 2020) and increases the likelihood of entrapment within local minima. In contrast, the policies in active inference are determined by both time and state. This dependence on time (Wang et al., 2016) enables policies to adapt efficiently, such as emphasizing exploration in the initial stages and exploitation later on. Moreover, this mechanism prompts more exploratory behavior in instances of state ambiguity. A further advantage of active inference lies in its adaptability to different task environments (Friston et al., 2017). It can configure different generative models to address distinct tasks, and compute varied forms of free energy and expected free energy.”

Laskin, M., Lee, K., Stooke, A., Pinto, L., Abbeel, P., & Srinivas, A. (2020). Reinforcement learning with augmented data. Advances in neural information processing systems, 33, 19884-19895.

Wang, J. X., Kurth-Nelson, Z., Tirumala, D., Soyer, H., Leibo, J. Z., Munos, R., ... & Botvinick, M. (2016). Learning to reinforcement learn. arXiv preprint arXiv:1611.05763.

Friston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., & Pezzulo, G. (2017). Active inference: a process theory. Neural computation, 29(1), 1-49.

Recommendation 2:

Figure 1A provides a key example of the lack of effort to help the reader understand. It suggests the possibility of a concrete example but falls short of providing one. From the caption and text, applied to the figure, I gather that by choosing either to run or to raise one's arms, one can control whether it is daytime or nighttime. This is clearly wrong but it is what I am led to think by the paper.

Response 2: Thank you for your suggestion, which we had not considered before. In this figure, we aim to illustrate that "the agent receives observations and optimizes his cognitive model by minimizing variational free energy → the agent makes the optimal action by minimizing expected free energy → the action changes the environment → the environment generates new observations for the agent." We have now modified the image to be simpler to prevent any possible confusion for readers. Correspondingly, we removed the figure of a person raising their hand and the shadowed house in Figure a.

Author response image 9.

Recommendation 3:

I recommend an overhaul in the labelling and methodological explanations for consistency and full reporting. For example, line 73 says sensory input is 's' and the cognitive model is 'q(s),' and the cause of the sensory input is 'p(s|o)' but on the very next line, the cognitive model is 'p(s|o)' and the causes of sensory input are 'q(s).' How this sensory input s relates to 'observations' or 'o' is unclear, and meanwhile, capital S is the set of environmental states. P seems to refer to the generative distribution, but it also means probability.

Response 3: Thank you for your advice. Now we have revised the corresponding labeling and methodological explanations in our work to make them consistent. However, we are not sure how to make a good modification to P here. In many works, P can refer to a certain probability distribution or some specific probabilities.

Recommendation 4:

Even the conception of a "policy" is unclear (Figure 2B). They list 4 possible policies, which are simply the 4 possible sequences of steps, stay-safe, cue-risky, etc, but with no contingencies in them. Surely a complete policy that lists 'cue' as the first step would entail a specification of how they would choose the safe or risky option BASED on the information in that cue

Response 4: Thank you for your suggestion. In active inference, a policy actually corresponds to a sequence of actions. The policy of "first choosing 'Cue' and then making the next decision based on specific information" differs from the meaning of policy in active inference.

Recommendation 5:

I assume that the heavy high pass filtering of the EEG (1 Hz) is to avoid having to baseline-correct the epochs (of which there is no mention), but the authors should directly acknowledge that this eradicates any component of decision formation that may evolve in any way gradually within or across the stages of the trial. To take an extreme example, as Figure 3E shows, the expected rewards for the risky path evolve slowly over the course of 60 trials. The filter would eliminate this.

Response 5: Thank you for your suggestion. The heavy high pass filtering of the EEG (1 Hz) is to minimize the noise in the EEG data as much as possible.

Recommendation 6:

There is no mention of the regression itself in the Methods section - the section is incomplete.

Response 6: Thank you for your suggestion. We have now added the relevant content in the Results section (EEG results at source level, line 337-340):

“The linear regression was run by the "mne.stats.linear regression" function in the MNE package (Activity ∼ Regressor + Intercept, Activity is the activity amplitude of the EEG signal in the source space and regressor is one of the regressors that we mentioned).”

Recommendation 7:

On Lines 260-270 the same results are given twice.

Response 7: Thank you for your suggestion. We have now deleted redundant content.

Recommendation 8:

Frequency bands are displayed in Figure 5 but there is no mention of those in the Methods. In Figure 5b Theta in the 2nd half is compared to Delta in the 1st half- is this an error?

Response 8: Thank you for your suggestion. It indeed was an error (they should all be Theta) and now we have corrected it.

Author response image 10.

Author Response

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

Reviewer 1

Major points:

R1C1: I appreciate that the data are aligned, in some points, with related studies of this niche. However, it would help the reader to have this alignment explored more extensively in the Discussion as well.

Answer: We acknowledge that the discussion would benefit from additional comparisons to the available datasets. We thus add the following comment after the first paragraph of the discussion: “Previous studies of the different sub-populations of SVZ progenitors were carried out using transcriptomic approaches based on the expression of various more or less specific markers. These approaches have made it possible to identify quiescent and activated neural stem cells as well as mature neuroblasts, but have been faced with the strong influence of the cell cycle on cell clustering. Indeed, neural progenitors in these studies cycling have been gathered in either “mitotic” clusters (Llorens et al. 2015, Zywitza et al. 2018, Cebrian et al. 2021) or “neural progenitor cells” clusters (Dulken et al. 2017) that had no clear biological significance and hindering identification of subtypes of SVZ cycling progenitors. Our study, combining, for the first time, characterization of Facs-isolated cells and an irradiation-based model of sequential regeneration, allowed to clearly distinguish the molecular profiles of TAP and iNB among cycling progenitors reflecting differences in their in vitro and in vivo respective potentials”.

R1C2: The data on multilineage differentiation, both in culture and upon engraftment, would be greatly strengthened by quantification. What is the relative yield of TUJ1/DCX-positive cells versus the other marker combinations? Specifically regarding the multilineage differentiation in vitro - because different media conditions are used to generate each lineage, it may be difficult to determine relative yield. Could a differentiation system that allows production of all 3 lineages be used instead?

If the fraction of non-DCX/TUJ1-labeled progeny is low, particularly in vivo, this might suggest that while multilineage differentiation is possible, it is a much less likely cellular state outcome than production of mature neuroblasts. Some suggested references with examples of the culture conditions, experimental conditions, and discussions highlighted in the public review: Culture conditions that allow simultaneous trilineage differentiation. PMID: 17615304 Influence of culture conditions on potency: similar to issues covered in PMID: 21549325.

Answer: We agree with the reviewer that quantification of a multilineage differentiation in vitro would improve the characterization of the relative potencies of the different SVZ progenitor.

According to PMID: 17615304 and PMID: 21549325, and in agreement with our own experience, the only culture condition that allows neurosphere-derived neural progenitors to differentiate in vitro into the three lineages is the removal of mitogens from the culture medium. However, this does not work on freshly isolated SVZ cells, which remain in an undifferentiated state in this condition.

This is why we chose to use specific differentiation media for each of the 3 lineages as in Figure 1C. It is also for this reason that we performed as many experiments as possible in vivo rather than in vitro as in Figure S2. In the new version, we have added a quantitative analysis of stainings by antibodies against GFAP, CNPase or DCX of GFP-positive cells persisting at IS, where high number of grafted cells were found in Figure S2B. This was performed by using the NIS software measuring eGFP-, GFAP-, CNPase- and DCX-positive areas. The intersection between each marker and eGFP areas was then determined as a percentage of staining (Figure S2C). The results showed that approximately one third of GFP+ cells expressed GFAP or DCX. The quantitative analysis of CNPase expression was complicated by CNPase-positive host cells, but the stronger CNPase staining in eGFP-positive areas clearly revealed the expression of CNPase by a significant proportion of eGFP-positive cells.

R1C3: Additionally, for claims similar to what is currently made in the text, it would be extremely valuable to confirm the purity of the sort for each population - for example by fixing and staining the sorted fraction with additional antibodies that confirm cell identity.

Answer: We have previously shown in Daynac et al. 2013 that s-iNB expressed the neuroblast markers CD24 and DCX, but also markers of neural progenitors such as Mash1, a basic helix-loop-helix transcription factor. As suggested by the reviewer, we have further investigated the expression of other markers of neural progenitors by sorted cells. The results showed that the proportion of DLX2+ cells a marker of proliferating progenitors (Doetsch et al. 2002) was very high in aNSC/TAP (98%) and progressively decreased in iNB (82%) and mNB (25%). Similarly, the expression of the transcription factor SOX2 that plays an essential role in the maintenance of neural progenitors (PMID: 25126380) accounted for 78% of aNSC/TAP, 70% of iNB and 17% of mNB.

Altogether, these new data confirmed the identity of the different cell populations and particularly that of iNB. They are commented at the beginning of the Results and shown in Figure S1.

R1C4: Line 125: GFAP alone doesn't necessarily indicate a "conversion to NSCs" - this conclusion could be greatly strengthened by inclusion of more markers, particularly at the protein level, or cyto-architectural studies.

Answer: We agree with the reviewer that GFAP expression alone is not sufficient to evidence the presence of NSC in the SVZ. We have thus modified the text accordingly: “Importantly, eGFP+ cells were present in the SVZ of all the animals transplanted with eGFP+s-iNB and eGFP+s-NSC/TAP (Fig. 1Db, Fig. 1Dc), some of them expressing GFAP indicating the generation of astrocytes, and therefore possibly NSC”.

R1C5: Could these cellular states be reflective of preferential translation of DCX? It would be very helpful to see the flow cytometry sort data for iNBs / mNBs used in Figure 6, particularly if these cells were also fixed and stained directly for DCX protein.

Answer: As suggested by the reviewer, freshly FAC-sorted iNB and mNB were fixed and labelled with an anti-DCX monoclonal antibody after permeabilization. As shown in the figure below, we found a higher level of DCX expression in mNB than in iNB. Therefore, this result tends to indicate that the proliferation capacity is somehow related to the level of DCX expression. However, because of the relatively low importance of this result, we decided not to include them in the manuscript.