eLife Assessment

This study provides important insights into how Trypanosoma cruzi populations diversify surface protein expression, showing through single-cell RNA sequencing that trans-sialidase-like genes are expressed heterogeneously across individual parasites, a pattern with clear implications for immune evasion. The evidence is convincing, supported by robust single-cell transcriptomic analyses, consistent quantitative measures of expression heterogeneity, and integration with genomic organization that together argue against purely stochastic expression.

Reviewer #1 (Public review):

Summary:

The authors aimed to assess the variability in expression of surface protein multigene families between amastigote and trypomastigote Trypanosoma cruzi, as well as between individuals within each population. The analysis presented shows higher expression of multigene family transcripts in trypomastigotes compared to amastigotes and that there is variation in which copies are expressed between individual parasites. Notably, they find no clear subpopulations expressing previously characterised trans-sialidase groups and that no patterns of coexpressed TcS genes were evident within individual cells or subpopulations. They also note that TcS encoded in the core genome are more often expressed, compared to TcS genes encoded in other genome compartments.

Strengths:

Additionally, the authors successfully process methanol fixed parasites with the 10x Genomics platform. This approach is valuable for other studies where using live parasites for these methods is logistically challenging.

In this second submission the authors show the kallisto mapping approach used is as robust as possible, and that this approach outperforms STAR mapping.

Weaknesses:

The authors describe a single experiment, which lacks repeats, controls or complementation with other approaches and the investigation is limited to the trans-sialidase transcripts.

Comments on revised version:

Thank you to the authors for taking the time to thoroughly address the peer review. The main concerns have now been addressed, and the manuscript edited to make points of confusion clearer.

Reviewer #2 (Public review):

Summary:

This manuscript presents a valuable single-cell RNA-seq study on Trypanosoma cruzi, an important human parasite. It investigates the expression heterogeneity of surface proteins, particularly those from the trans-sialidase-like (TcS) superfamily, within amastigote and trypomastigote populations. The findings suggest a previously underappreciated level of diversity in TcS expression, which could have implications for understanding parasite-host interactions and immune evasion strategies. The use of single cell approaches to delve into population heterogeneity is strong. However, the study does have some limitations that need to be addressed.

The focus on single-cell transcriptional heterogeneity in surface proteins, especially the TcS family, in T. cruzi is novel. Given the important role of these proteins in parasite biology and host interaction, the findings have potential significance.

Strengths:

The key finding of heterogeneous TcS expression in trypomastigotes is well-supported. The analysis comparing multigene families, single-copy genes, and ribosomal proteins highlights the unusual nature of the variation in surface protein coding genes.

Weaknesses:

While the manuscript identifies TcS heterogeneity, the functional implications of the different expression profiles remain speculative. The authors state it may reflect differences in infectivity, but no direct experimental evidence supports this.

The manuscript lacks any functional validation of the single-cell findings. For instance, do the trypomastigote subpopulations identified based on TcS expression exhibit differences in infectivity, host cell tropism, or immune evasion? Such experiments would greatly strengthen the study.

The authors identify a subpopulation of TcS genes that are highly expressed in many cells. However, it is unclear if these correspond to previously characterized TcS members with specific functions.

The authors hypothesize that observed heterogeneity may relate to chromatin regulation. However, the study does not directly address these mechanisms. There are interesting connections to be made with what they identify as colocalization of genes within chromatin folding domains, but the authors do not fully explore this. It would be insightful to address these mechanisms in future work. [...]

Comments on revisions:

The novel version of the manuscript has improved and satisfied this reviewer.

Reviewer #3 (Public review):

The study aimed to address a fundamental question in T. cruzi and Chagas disease biology - how much variation is there in gene expression between individual parasites? This is particularly important with respect to the surface protein-encoding genes, which are mainly from massive repetitive gene families with 100s to 1000s of variant sequences in the genome. There is very little direct evidence for how expression of these genes is controlled. The authors conducted a single cell RNAseq experiment of in vitro cultured parasites with a mixture of amastigotes and trypomastigotes. Most of the analysis focused on the heterogeneity of gene expression patterns amongst trypomastigotes. They show that heterogeneity was very high for all gene classes, but surface-protein encoding genes were the most variable. Interestingly, in the case of the trans-sialidase genes, many sequence variants were detected in fewer than 5% of parasites while a subset of 31 others was detected in >40% if parasites, hinting at compartmentalised expression control within the gene family. The biology of the parasite (e.g. extensive post-transcriptional regulation) and potential technical caveats (e.g. high dropout rates across the genome) make it difficult to infer connections to actual protein expression on the parasite surface, but the results are a significant advance for the field.

(1) Limit of detection and gene dropouts.

An average of ~1100 genes are detected per parasite which indicates a dropout rate of over 90%. It appears that RNA for the "average" single copy 'core' gene is only detected in around 3% of the parasites sampled (Figure 2c: ~100 / 3192). While comparable with some other trypanosome scRNAseq studies, this remains a caveat to the interpretation that high cell-to-cell variability in gene expression is explained by biological factors. The argument would be more convincing if the dropout rates and expression heterogeneity were minimal for highly expressed housekeeping genes. The authors are appropriately cautious in their interpretation and acknowledge the need for further validation.

(2) Heterogeneity across the board.

The authors focus on the relative heterogeneity in RNA abundance for surface proteins from the multicopy gene families vs core genes. While multicopy gene sequences do show significantly more cell-to-cell variability, there is still surprisingly high inequality of expression amongst genes in other classes including single copy housekeeping and ribosomal genes. Again the biological relevance of the comparison is uncertain and the authors acknowledge the need for further investigation.

This study provides some tantalising evidence that the expression of surface genes may vary substantially between individual parasites in a single clonal population. The study is also amongst the very first to apply scRNAseq to T. cruzi, so the broader data set will be an important resource for researchers in the field.

Comment on revised version:

The manuscript is significantly improved. The revised explanations and figures make several aspects of the data analysis and interpretation much clearer to me now. Thanks to the authors.

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

The authors aimed to assess the variability in the expression of surface protein multigene families between amastigote and trypomastigote Trypanosoma cruzi, as well as between individuals within each population. The analysis presented shows higher expression of multigene family transcripts in trypomastigotes compared to amastigotes and that there is variation in which copies are expressed between individual parasites. Notably, they find no clear subpopulations expressing previously characterised trans-sialidase groups. The mapping accuracy to these multicopy genes requires demonstration to confirm this, and the analysis could be extended further to probe the features of the top expressed genes and the other multigene families also identified as variable.

Strengths:

The authors successfully process methanol-fixed parasites with the 10x Genomics platform. This approach is valuable for other studies where using live parasites for these methods is logistically challenging.

Weaknesses:

The authors describe a single experiment, which lacks controls or complementation with other approaches and the investigation is limited to the trans-sialidase transcripts.

It would be more convincing to show either bioinformatically or by carrying out a controlled experiment, that the sequencing generated has been mapped accurately to different members of multigene families to distinguish their expression. If mapping to the multigene families is inaccurate, this will impact the transcript counts and downstream analysis.

We thank the reviewer for raising these important points.

We agree that the analysis of multigene families at the single-cell level is an important question, particularly given the heterogeneity observed across several of them. However, the aim of this short report is not to provide a comprehensive analysis of the entire experiment, but rather to focus on what we consider an important biological phenomenon observed in TcTS genes.

Regarding the mapping accuracy of the reads, we acknowledge that this can limit the disambiguation of highly similar multicopy transcripts. This is, in fact, a common challenge when analyzing transcriptomic data from T. cruzi.

To address this issue, we analyzed the sequence identity of the 3′ ends of TcS transcripts (defined as the 3′UTR plus 20% of the CDS region). As shown in Author response image 1, these regions display a median sequence identity of approximately 25%, indicating that sufficient sequence divergence exists for mapping algorithms to use during read assignment.

In addition, it is important to note that kallisto, the software used in our analysis, was specifically designed to address multimapping reads through pseudoalignment combined with an expectation-maximization algorithm that probabilistically assigns reads across compatible transcripts.

To directly assess performance, we simulated reads from the T. cruzi transcriptome used in this study (3′UTRs plus 20% of the CDS regions) and compared two mapping/counting strategies: (a) transcriptome pseudoalignment using kallisto, and (b) genome alignment followed by counting using STAR + featureCounts. The latter approximates the strategy implemented in CellRanger, the standard pipeline for quantifying expression levels from 10X Genomics single cell RNA-seq data. We found that kallisto recovered the simulated “true” counts with substantially higher accuracy than STAR + featureCounts (Pearson correlation: all genes, 0.991 vs 0.595; surface protein genes, 0.9996 vs 0.827; trans-sialidase (TcS) genes, 0.9998 vs 0.773). These results indicate that pseudoalignment is currently the optimal strategy for recovering the relative expression of highly similar gene family members (Author response image 1 C).

Author response image 1

(A) Distribution of pairwise sequence identity values calculated among the 3′-end regions of all transcripts (defined as the 3′UTR plus 20% of the coding sequence). (B) Distribution of read mapping coordinates over all multigene family transcripts normalized as percentage of the gene length (C) Scatter plots showing the correlation between estimated transcript counts obtained using kallisto (red) and STAR + featureCounts (grey) versus the corresponding simulated ground-truth values.

Reviewer #2 (Public review):

Summary:

This manuscript presents a valuable single-cell RNA-seq study on Trypanosoma cruzi, an important human parasite. It investigates the expression heterogeneity of surface proteins, particularly those from the trans-sialidase-like (TcS) superfamily, within amastigote and trypomastigote populations. The findings suggest a previously underappreciated level of diversity in TcS expression, which could have implications for understanding parasite-host interactions and immune evasion strategies. The use of single-cell approaches to delve into population heterogeneity is strong. However, the study does have some limitations that need to be addressed.

The focus on single-cell transcriptional heterogeneity in surface proteins, especially the TcS family, in T. cruzi is novel. Given the important role of these proteins in parasite biology and host interaction, the findings have potential significance.

Strengths:

The key finding of heterogeneous TcS expression in trypomastigotes is well-supported. The analysis comparing multigene families, single-copy genes, and ribosomal proteins highlights the unusual nature of the variation in surface protein-coding genes.

Weaknesses:

While the manuscript identifies TcS heterogeneity, the functional implications of the different expression profiles remain speculative. The authors state it may reflect differences in infectivity, but no direct experimental evidence supports this.

The manuscript lacks any functional validation of the single-cell findings. For instance, do the trypomastigote subpopulations identified based on TcS expression exhibit differences in infectivity, host cell tropism, or immune evasion? Such experiments would greatly strengthen the study.

We thank the reviewer for their careful reading of the manuscript. We agree that obtaining experimental evidence on the influence of multiple multigene families would represent a significant advancement in the field. However, we would like to emphasize that this study is presented as a short communication centered on a specific and biologically relevant observation within a single multigene family. The aim of the manuscript is to highlight what we consider an important biological phenomenon that raises hypotheses to be tested in future work.

The influence of phenotypic heterogeneity and its possible advantages under environmental pressures has been previously proposed for Trypanosoma cruzi, related trypanosomatids, and other biological systems, ranging from bacteria to tumors (Seco-Hidalgo 2015, doi: 10.1098/rsob.150190 and Luzak 2021, doi: 10.1146/annurev-micro-040821-012953, for a comprehensive review on this topic). While the reviewer is correct in noting that our model does not demonstrate a functional role for TcTS heterogeneity, the experimental approaches required to address this question in a large multigene family are highly complex. This is particularly challenging in T. cruzi, where the study of multigene families is limited by the restricted set of available molecular biology tools (such as RNAi). Therefore, further experimental validation of these observations falls outside the scope of this short report.

In this revised version, we have included additional validation and clarification of the results, as well as a more explicit discussion of their limitations. In addition, we present a preliminary analysis exploring potential mechanisms that could coordinate the observed expression patterns of the TcTS family.

The authors identify a subpopulation of TcS genes that are highly expressed in many cells. However, it is unclear if these correspond to previously characterized TcS members with specific functions.

The TcS subgroup with a high frequency of detection comprises 31 genes, none of which belong to the catalytically active Group I trans-sialidases. Instead, this subgroup includes members of Groups II, III, IV, V, VI, and VIII. This information has been added to Supplementary Table 3 and is now stated in the revised manuscript.

The authors hypothesize that observed heterogeneity may relate to chromatin regulation. However, the study does not directly address these mechanisms. There are interesting connections to be made with what they identify as the colocalization of genes within chromatin folding domains, but the authors do not fully explore this. It would be insightful to address these mechanisms in future work.

In response to the reviewer’s and editorial team’s request for additional mechanistic insight into the regulatory processes that may be involved in the observed patterns, we have expanded the revised manuscript to discuss how the genomic context of TcS loci could contribute to the observed heterogeneity in TcS expression. As noted in the original version of the manuscript, TcS genes and other surface-protein gene families are largely partitioned into discrete genomic compartments, whose expression has been reported to be regulated by epigenetic control of chromatin-folding domains (doi.org/10.1038/s41564-023-01483-y). However, we previously showed that TcS genes detected in a high proportion of cells are, in most cases, dispersed throughout the genome, arguing against a model in which their preferential expression results from colocalization within a small number of ubiquitously activated chromatin domains. In response to the reviewer’s suggestion, we performed a more detailed analysis of the genomic locations of these TcS genes. We found that many of them are localized within the core compartment (new Figure 5). Because the core compartment is enriched for conserved, housekeeping genes that typically display more constitutive expression (doi.org/10.1038/s41564-023-01483-y), whereas the disruptive compartment is enriched for lineage-specific multigene families associated with variable, stage-specific, and recently reported stochastic expression (doi.org/10.1038/s41467-025-64900-2), our results are consistent with a model in which compartment-specific regulatory mechanisms (in addition to post-transcriptional regulation) influence the differential cellular expression of core- versus disruptive-located TcS genes. We have incorporated these results and discussion in the revised manuscript.

The merging of technical replicates needs further justification and explanation as they were not processed through separate experimental conditions. While barcodes were retained, it would be informative to know how well each technical replicate corresponds with the other. If both datasets were sequenced on the same lane, the inclusion of technical replicates adds noise to the analysis.

Regarding technical details, we now include the total number of mapped reads and average number of reads mapped per cell (new paragraph in the Methods section.

The technical replicates consist of a single Illumina library that was sequenced in two separate runs. As this approach is expected to be highly reproducible, we merged both runs into a single count table. To support this decision, we assessed the concordance between the two sequencing runs and observed an almost perfect correlation between them (Author response image 2).

Author response image 2.

Correlation analysis of number of reads assigned to cells between technical replicate 1 and technical replicate 2.

While the number of cells sequenced (3192) seems reasonable, it's not clear how much the conclusions are affected by the depth of sequencing. A more detailed description of the sequencing depth and its impact on gene detection would be valuable.

We detected a mean of 1088 genes per cell. Based on the 15,319 annotated protein-coding genes in the reference genome, this represents 7.1% of the T. cruzi protein-coding gene complement detected in each cell.

Across the entire dataset, a total of 14,321 genes were detected in at least one cell, representing 93.5% of all annotated protein-coding genes. This suggests that our experiment captured a broad representation of the parasite's transcriptome.

This per-cell detection rate is characteristic of droplet-based scRNA-seq and is consistent with other trypanosomatid studies. For example, the T. brucei single-cell atlas (Hutchinson et al., 2021) reported a median detection of 1052 genes per cell. In the case of T. cruzi, the recently published pre-print of the T. cruzi single cell atlas from Laidlaw & García-Sánchez et al. reported a mean between 298 and 928 genes detected per cell (depending on the sample).

This information is now included in Methods.

While most of the methods are clear, the way in which the subsampled gene lists were generated could be more thoroughly described, as some details are not clear for the subsampling of single-copy genes.

The subsampling method was originally described in the Figure 2 legend; to better highlight this approach, we have now moved its description to the Methods section.

Some of the figures are difficult to interpret. For example, the color scaling in the heatmap of Supplementary Figure 3B is not self-explanatory and it is hard to extract meaningful conclusions from the graph.

We agree with the reviewer in this assessment. We have now modified the figures to be more self-explanatory and better reflect the conclusions.

Reviewer #3 (Public review):

The study aimed to address a fundamental question in T. cruzi and Chagas disease biology - how much variation is there in gene expression between individual parasites? This is particularly important with respect to the surface protein-encoding genes, which are mainly from massive repetitive gene families with 100s to 1000s of variant sequences in the genome. There is very little direct evidence for how the expression of these genes is controlled. The authors conducted a single-cell RNAseq experiment of in vitro cultured parasites with a mixture of amastigotes and trypomastigotes. Most of the analysis focused on the heterogeneity of gene expression patterns amongst trypomastigotes. They show that heterogeneity was very high for all gene classes, but surface-protein encoding genes were the most variable. In the case of the trans-sialidase gene family, many sequence variants were only detected in a small minority of parasites. The biology of the parasite (e.g. extensive post-transcriptional regulation) and potential technical caveats (e.g. high dropout rates across the genome) make it difficult to infer what this might mean for actual protein expression on the parasite surface.

We thank the reviewer for this important comment, highlighting a central challenge when studying trypanosomatid biology. We acknowledge that in most eukaryotes and particularly in T. cruzi, where there is a predominant role of post-transcriptional regulation, mRNA levels are not always directly correlated with protein abundance, as previously reported by us and others (10.1186/s12864-015-1563-8, 10.1128/msphere.00366-21, 10.1590/S0074-02762011000300002, 10.1042/bse0510031). Nevertheless, steady-state transcript levels obtained by RNA-seq remain informative for assessing differential gene expression, and this approach has been widely used as a proxy for the study of gene expression profiles in T. cruzi (10.7717/peerj.3017, 10.1371/journal.ppat.1005511, 10.1016/j.jbc.2023.104623, 10.3389/fcimb.2023.1138456, 10.1186/s13071-023-05775-4).

It's also interesting to note that recent proteomic analyses (10.1038/s41467-025-64900-2) have revealed substantial heterogeneity in the expression of surface proteins, including trans-sialidases, supporting the idea that the transcriptional heterogeneity we observe reflects a genuine biological feature that propagates to the protein level.

We have now added a sentence to the discussion acknowledging this limitation and discussed the results from Cruz-Saavedra, et al. in the revised manuscript.

(1) Limit of detection and gene dropouts

An average of ~1100 genes are detected per parasite which indicates a dropout rate of over 90%. It appears that RNA for the "average" single copy 'core' gene is only detected in around 3% of the parasites sampled (Figure 2c: ~100 / 3192). This may be comparable with some other trypanosome scRNAseq studies, but this still seems to be a major caveat to the interpretation that high cell-to-cell variability in gene expression is explained by biological rather than technical factors. The argument would be more convincing if the dropout rates and expression heterogeneity were minimal for well-known highly expressed genes e.g. tubulin, GAPDH, and ribosomal RNAs. Admittedly, in their Final Remarks, the authors are very cautious in their interpretation, but it would be good to see a more thorough discussion of technical factors that might explain the low detection rates and how these could be tested or overcome in future work.

(2) Heterogeneity across the board

The authors focus on the relative heterogeneity in RNA abundance for surface proteins from the multicopy gene families vs core genes. While multicopy gene sequences do show more cell-to-cell variability, the differences (Figure 2D) are roughly average Gini values of 0.99 vs 0.97 (single copy) or 0.95 (ribosomal). Other studies that have applied similar approaches in other systems describe Gini values of < 0.2-0.25 for evenly expressed "housekeeping" genes (PMIDs 29428416, 31784565). Values observed here of >0.9 indicate that the distribution for all gene classes is extremely skewed and so the biological relevance of the comparison is uncertain.

We recognize the limitations imposed by gene dropout in our data, as highlighted by the reviewer. Unfortunately, gene dropout is an inherent limitation of 10x genomics data. Trypanosomatids are not an exception in this regard, and the general metrics of the single-cell RNA-seq data in other reports are equivalent to those obtained in our experiment.

Despite this important limitation, we believe that our comparative analyses (the contrast between TcS and ribosomal protein expression) provide valuable insights into a biological phenomenon with potential functional relevance for the parasite. Furthermore, we are actively working on generating single-cell RNA-seq data using alternative methodologies that improve gene dropout rates. We anticipate that these future studies will help clarify the extent of the phenomenon described in this work.

Our results reveal a small subset of TcS genes that are frequently detected across cells, a pattern that is not compatible with random detection unless these genes were highly expressed and preferentially captured by random sampling. However, as shown in Figure 4b, many genes expressed at comparable levels are not detected at high frequencies. In line with this, Figure 4c shows that within individual cells, the detected TcS genes exhibit similar expression levels. Finally, we confirmed that this frequently detected subset shows high read counts at the bulk RNA-seq level (Figure 4 - Figure Supplement 1), consistent with the fact that these TcS are frequent in the population even when they are not specially highly expressed within each cell. Taken together, these findings argue against a purely random sampling of TcS genes and support the interpretation that this pattern reflects an underlying biological feature. We agree that further validation will be required. Accordingly, since the initial submission, we have been careful to frame our conclusions conservatively, explicitly noting that dropout remains a limitation of these data that could influence the observed patterns. In the revised version, we have strengthened this point by including a specific statement in the final remarks. Our interpretation is presented as a working hypothesis that is fully compatible with the observations reported here and may be informative for the field. To better reflect this reasoning, we have revised Figure 4b, expanded the discussion, and explicitly included this limitation in the final remarks of the revised manuscript.

Nevertheless, this study does provide some tantalising evidence that the expression of surface genes may vary substantially between individual parasites in a single clonal population. The study is also amongst the very first to apply scRNAseq to T. cruzi, so the broader data set will be an important resource for researchers in the field.

We thank the reviewer for highlighting the relevance of our study and for their positive assessment of the potential significance of these observations. We also agree that the dataset generated here may represent a useful resource for the community.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

(1) In Figures 1c and 1d, it would be useful to include the genes as the plot titles.

We agree with the reviewer that including gene names in the plot makes the panels more self-explanatory. We have added gene names to the updated version of Figure 1.

(2) Can you include the read lengths of the sequencing and whether this is sufficient to map accurately to very similar genes of the same multigene family? As stated in the public summary, this would make the data far more convincing as standard 10x chromium cannot distinguish similar gene copies unless a longer read 2 is used. Given that only the 3' end is targeted, is this enough to distinguish the TcS and other mutligene family transcripts?

We thank the reviewer for raising this important point. We agree that short 3′ biased reads can limit the disambiguation of highly similar multicopy transcripts. This is, in fact, a common challenge when analyzing transcriptomic data from T. cruzi.

To address this issue, we analyzed the sequence identity of the 3′ ends of TcS transcripts (defined as the 3′UTR plus 20% of the CDS region). As shown in Author response image 1, these regions display a median sequence identity of approximately 25%, indicating that sufficient sequence divergence exists for mapping algorithms to use during read assignment.

In addition, it is important to note that kallisto, the software used in our analysis, was specifically designed to address multimapping reads through pseudoalignment combined with an expectation-maximization algorithm that probabilistically assigns reads across compatible transcripts.

To directly assess performance, we simulated reads from the T. cruzi transcriptome used in this study (3′UTRs plus 20% of the CDS regions) and compared two mapping/counting strategies: (a) transcriptome pseudoalignment using kallisto, and (b) genome alignment followed by counting using STAR + featureCounts. The latter approximates the strategy implemented in CellRanger, the standard pipeline for quantifying expression levels from 10X Genomics single cell RNA-seq data. We found that kallisto recovered the simulated “true” counts with substantially higher accuracy than STAR + featureCounts (Pearson correlation: all genes, 0.991 vs 0.595; surface protein genes, 0.9996 vs 0.827; trans-sialidase (TcS) genes, 0.9998 vs 0.773). These results indicate that pseudoalignment is currently the optimal strategy for recovering the relative expression of highly similar gene family members (Author response image 1C).

The length of the R2 read (91bp) was included in Methods (line 411).

(3) It is stated that 'single copy' genes also include 'low copy number genes". What does this include exactly? Is it more actuate to say non-surface protein genes?

The distinction we aim to make is between multigene families and the rest of the genome. Most multigene families encode surface proteins, but not all surface protein genes belong to multigene families. To clarify this point we included a sentence in methods to reflect that when we describe “surface proteins” we are referring to surface proteins coded by multigene families (line 453). In addition, long-read genomic DNA sequencing and assembly have revealed that many genes previously believed to be single-copy are actually duplicated at low copy numbers (doi.org/10.1099/mgen.0.000177). For this reason, we extend the concept of “single-copy” genes to include those that have only a few duplicates.

(4) It is stated in line 127 that TcS have particular high heterogeneity - it does not look that way by eye compared to the other multigene families. Can statistic be used to prove this, or simply state the decision was made to focus on the TcS?

As noticed by the reviewer, all multigene families show significantly higher heterogeneity compared to single-copy genes, as stated in the text and shown in figure legends from Figure 2, Supplementary Figure 1 and the new Supplementary Table 2.

That said, it was not the statistical results that guided our decision to focus on TcS, but rather their well-established biological relevance in T. cruzi. As suggested, we have now emphasized this rationale more clearly in the revised text (lines 160-167).

Besides, recent work has shown that TcS genes exhibit a bimodal distribution of expression levels using bulk RNA-seq data, in contrast to core genes and other multigene families (doi.org/10.1038/s41467-025-64900-2, doi.org/10.1038/s41564-023-01483-y). This distinct regulatory behavior further justifies our decision to examine TcS separately.

(5) Expression of different TcS has been investigated between the different life cycle stages for a few individual genes previously (Freitas et al). Can the authors not extend this investigation to all the genes detect by scRNA-seq here to demonstrate those with higher/lower expression in amastigotes vs trypomastigotes building on Figure 2A? Are particular groups linked to either stage?

We performed this analysis and did not observe any correlation between TcS groups and life cycle stage. In all cases TcS were more frequently detected in trypomastigotes. This difference was statistically significant for all groups except group VII, likely due to the low number of genes analyzed in this group (Author response image 3).

Author response image 3.

Per-gene number of expressing cells by TcS group and life-stage. Boxplots show, for each TcS group (I–VIII), the distribution across genes of the number of cells in which the gene is detected. Each point represents a single TcS; Amastigote cells: green points/boxes, Trypomastigote cells: salmon points/boxes. The y-axis is on log10 scale. Asterisks indicate statistically significant differences from the comparison between Amastigote and Trypomastigote within each TcS group, assessed using a paired two-sided Wilcoxon signed-rank test: * p < 0.05, ** p < 0.01, *** p < 0.001.

(6) What exactly is the Z-score shown in Figure 2B?

In this analysis num_multigene represents the number of multigene family genes detected in each individual cell. For every cell, we counted how many genes from our predefined multigene family gene list has detectable expression (more than zero UMI counts); in the UMAP plot, this value is reflected by the size of each point. On the other hand, z_multigene captures the relative expression level of multigene family genes within each cell. This metric is calculated by summing the UMI counts of all multigene family genes per cell and then standardizing this value across the dataset using a z-score transformation, such that positive values reflect above-average multigene family expression and negative values reflect below-average levels. In the UMAP plot, this metric determines the color scale of each point. Taking together num_multigene and z_multigene allow us to distinguish cells that express multigene family genes broadly (high gene counts), strongly (high relative expression), both, or neither, and to relate these patterns to identified cell populations.

We included a short description in legend of the new version of Figure 2 (lines 176-180).

(7) For the reclustering of trypomastigotes based on TcS genes alone, please show the UMAP and discuss why the resolution giving two clusters is chosen? I assume increasing the resolution does not reveal clusters of cells express one of the 8 groups of TcS for example?

We appreciate the reviewer’s suggestion. In this analysis, our goal was to test whether the phenotypic heterogeneity previously reported in trypomastigotes could be recapitulated using TcS genes alone, as prior studies described two major transcriptomic phenotypes within this stage.

Increasing the clustering resolution did not reveal subclusters corresponding to the eight TcS sequence groups. This might reflect the fact that these groups are defined based on sequence similarity rather than on expression patterns, as noted by Freitas et al. (doi:10.1371/journal.pone.0025914).

(8) In Figure 4B, there may be an upward trend in the level of expression and the number of cells a transcript is detected in? It would be worth showing this is or is not the case with statistics if possible.

The number of genes detected in a high proportion of cells is low, which limits the statistical power of this analysis. Also, substantial dispersion is observed within the 0-5% interval. Nevertheless, this figure is presented primarily to highlight that a considerable number of highly expressed genes are detected in only a small fraction of cells. If expression level were the main determinant of detection frequency across cells, one would expect very few highly expressed genes to fall within the 0-5% interval. Contrary to this expectation, among the 50 highest expressed TcS genes, 62% are detected in fewer than 5% of cells, and even among the top 10 most highly expressed TcS genes, 40% fall within this lowest detection group. To facilitate this interpretation, we modified the figure (new Figure 4b) to explicitly highlight the top 50 most expressed TcS genes and incorporated this discussion into the main text of the revised manuscript (lines 244-251), making the conclusion clearer to the reader.

(9) Do the cells group instead by expression of any of the other multigene families not investigated in detail?

It is possible that additional transcriptional substructure among trypomastigotes is driven by the expression of other multigene families beyond TcS. In this short report (with limited number of figures, words, etc.), we focused specifically on the trans-sialidase family as discussed earlier. A more comprehensive analysis including other large surface gene families (MASPs, mucins, GP63) is planned as part of ongoing work and will be presented in future reports.

Reviewer #2 (Recommendations for the authors):

This reviewer suggests the conduction of functional experiments in follow-up studies to establish links between TcS expression profiles and parasite behavior and into potential regulatory mechanisms responsible for the observed TcS heterogeneity, particularly focusing on epigenetic modifications. It would be interesting to correlate the highly expressed TcS members identified here with previously characterized TcS isoforms and provide more description regarding which particular groups and TcS members are driving the findings. It would benefit from further clarification regarding sequencing depth, technical replication merging, subsampling, and specific parameters for alignment methods and more information regarding the specific statistical tests and their applicability to the data.

This is a promising single-cell study with potentially high significance. The manuscript is well-written, and the analyses are reasonably well-executed. However, the current manuscript is limited by a lack of functional validation and mechanistic insights. The addition of further analyses and experiments, as suggested, will strengthen the conclusions and increase the impact of the work.

We thank the reviewer for their careful reading of the manuscript. As suggested, we have performed additional validation and clarification of the results, as well as a more explicit discussion of their limitations. In addition, we have included a preliminary analysis exploring potential mechanisms that could be coordinating the observed expression patterns of the TcS family (see below). Even though we consider relevant and interesting to experimentally validate these results, given the inherent difficulties in studying multigene families in T. cruzi, an organism with a very limited set of molecular biology tools (such as RNAi), further experimental validation of these observations is outside of the scope of this short report.

Regarding the reviewer’s question, we studied if any TcS subgroup could be driving our observations. However, we did not find any correlations indicating that a particular group was associated with any of our findings. We now include TcS group information to Supplementary Table 3.

Regarding technical details, we now included the total number of mapped reads (line 422) and average number of reads mapped per cell (new paragraph in the Methods section, line 432-436).  

The technical replicates consist of a single Illumina library that was sequenced in two separate runs. As this approach is expected to be highly reproducible, we merged both runs into a single count table, as stated in line 424. To support this decision, we assessed the concordance between the two sequencing runs and observed an almost perfect correlation between them (Author response image 2).

The subsampling method was originally described in the Figure 2 legend; to better highlight this approach, we have now moved its description to the Methods section (line 456).

The specific kallisto parameters used are stated in Methods (line 418-419). We now included that default options were used unless otherwise specified (line 419-420).

In response to the reviewer’s and editorial team’s request for additional mechanistic insight into the regulatory processes that may be involved in the observed patterns, we have expanded the revised manuscript to discuss how the genomic context of TcS loci could contribute to the observed heterogeneity in TcS expression. As noted in the original version of the manuscript, TcS genes and other surface-protein gene families are largely partitioned into discrete genomic compartments, whose expression has been reported to be regulated by epigenetic control of chromatin-folding domains (doi.org/10.1038/s41564-023-01483-y). However, we previously showed that TcS genes detected in a high proportion of cells are, in most cases, dispersed throughout the genome, arguing against a model in which their preferential expression results from colocalization within a small number of ubiquitously activated chromatin domains. In response to the reviewer’s suggestion, we performed a more detailed analysis of the genomic locations of these TcS genes. We found that many of them are localized within the core compartment (new Figure 5). Because the core compartment is enriched for conserved, housekeeping genes that typically display more constitutive expression (doi.org/10.1038/s41564-023-01483-y), whereas the disruptive compartment is enriched for lineage-specific multigene families associated with variable, stage-specific, and recently reported stochastic expression (doi.org/10.1038/s41467-025-64900-2), our results are consistent with a model in which compartment-specific regulatory mechanisms (in addition to post-transcriptional regulation) influence the differential cellular expression of core- versus disruptive-located TcS genes. We have incorporated these results and discussion in line 301-313 of the revised manuscript.

Reviewer #3 (Recommendations for the authors):

(1) The authors consistently refer to gene "expression" but somewhere they should acknowledge that in trypanosomes RNA abundance is less predictive of protein than in most other organisms.

We thank the reviewer for this important comment, highlighting a central challenge when studying trypanosomatid biology. We acknowledge that in most eukaryotes and particularly in T. cruzi, where there is a predominant role of post-transcriptional regulation, mRNA levels are not always directly correlated with protein abundance, as previously reported by us and others (10.1186/s12864-015-1563-8, 10.1128/msphere.00366-21, 10.1590/S0074-02762011000300002, 10.1042/bse0510031). Nevertheless, steady-state transcript levels obtained by RNA-seq remain informative for assessing differential gene expression, and this approach has been widely used as a proxy for the study of gene expression profiles in T. cruzi (10.7717/peerj.3017, 10.1371/journal.ppat.1005511, 10.1016/j.jbc.2023.104623, 10.3389/fcimb.2023.1138456, 10.1186/s13071-023-05775-4).

It's also interesting to note that recent proteomic analyses (10.1038/s41467-025-64900-2) have revealed substantial heterogeneity in the expression of surface proteins, including trans-sialidases, supporting the idea that the transcriptional heterogeneity we observe reflects a genuine biological feature that propagates to the protein level.

We have now added a sentence to the discussion acknowledging this limitation and discussed the results from Cruz-Saavedra, et al. in linea 266-271 of the revised manuscript.

(2) Line 29, in the abstract there is a strong statement that T. cruzi "does not employ antigenic variation". I don't think there is much evidence either way if we are thinking about antigenic variation in the broad sense rather than the extreme model of T. brucei VSG switching. Later in the abstract they state that "no recurrent combinations of TcS genes were observed between individual cells in the population", which sounds very much like a form of antigenic variation.

We agree with the reviewer. Indeed, we meant to state that T. cruzi does not employ an antigenic variation mechanism such as the one from T. brucei. We change this statement as suggested in lines 28 - 32.

(3) Line 29, "relies on a diverse array of cell-surface-associated proteins encoded by large multi-copy gene families (multigene families) essential for infectivity and immune evasion" and lines 55-58 "T. cruzi infection relies on a heterogeneous set of membrane proteins, encoded mainly by large multigene families ... most of which are involved in infection, tropism, and immune evasion". It would be worth adding a bit more detail on the nature and strength of the evidence that Tc "relies on" these various genes or that they are "essential" for infectivity, tropism, and immune evasion.

Because the journal’s short format imposes word limits, we strengthened the original statement by adding specific references that document genomic, transcriptomic and functional evidence linking the major multigene families to infectivity, tropism and immune evasion (doi.org/10.1371/journal.pone.0025914; doi.org/10.1038/nrmicro1351; doi.org/10.1128/iai.05329-11; doi.org/10.1093/nar/gkp172, doi.org/10.1371/journal.ppat.1006767), in line 77.

(4) Line 89, 1088 genes detected per cell - what is this as a % of genes in the genome?

We detected a mean of 1088 genes per cell. Based on the 15,319 annotated protein-coding genes in the reference genome, this represents 7.1% of the T. cruzi protein-coding gene complement detected in each cell.

Across the entire dataset, a total of 14,321 genes were detected in at least one cell, representing 93.5% of all annotated protein-coding genes. This suggests that our experiment captured a broad representation of the parasite's transcriptome.

This per-cell detection rate is characteristic of droplet-based scRNA-seq and is consistent with other trypanosomatid studies. For example, the T. brucei single-cell atlas (Hutchinson et al., 2021) reported a median detection of 1052 genes per cell. In the case of T. cruzi, the recently published pre-print of the T. cruzi single cell atlas from Laidlaw & García-Sánchez et al. reported a mean between 298 and 928 genes detected per cell (depending on the sample).

This information is now included in Methods (line 435).

(5) Line 93-94, how many cells were assigned to clusters 0 and 1?

Cluster 0 had 2201 cells and cluster 1 had 824 cells assigned.  We have now included these specific numbers in new version of the manuscript (line 114).

(6) Line 96, cluster 2 ama-trypo transitioning parasites - were these observable by microscopy?

We did not perform microscopy specifically to observe or quantify the putative ama/trypo transitioning subpopulation: microscopy was only used as a pre-experiment quality check to verify cell morphology and viability. The inference that cluster 2 reflects ama/trypo transitioning parasites is drawn from the transcriptomic profile (particularly from the pattern of stage-associated marker expression observed in that cluster) and should be considered a hypothesis generated by the data, that merits further analysis, as stated in the manuscript.

(7) Line 106-107, "As expected, single-copy gene expression is high in both amastigotes and trypomastigotes and similar on average between both cell types".

(8) Why as expected? For a broad journal it would be useful to explain this. Amastigotes are replicative and trypomastigotes are not, so would we not expect to see some differences that reflect this?

(9) What do you mean by the expression being "high"? High compared to what?

(10) "Similar on average between both cell types". This does not seem concordant with Figure 1a showing a highly significant difference between ama and trypo.

We thank the reviewer for this helpful request for clarification for broader readers and the observations regarding global expression of single copy and multigene family genes.

Figure 2a is intended as an experimental control where we show that our 10X Genomics data shows the previously reported upregulation of surface protein genes in trypomastigotes. We have now modified the text in order to highlight this (line 129). In turn, Supplementary Figure 1a is shown as a control that this upregulation is not a general feature of trypomastigote cells.

Regarding comment 9, what we meant is that single-copy genes display relatively high expression in both amastigotes and trypomastigotes compared with surface protein-coding genes (see expression values in Figures 2a and Supplementary Figure 1a).

Finally, differential expression between amastigotes and trypomastigotes at the transcriptomic level has been previously studied and has shown that most single copy genes do not show variation, explaining the overall pattern of Supplementary Figure 1a where average expression is similar between stages (mean fold change = 1.1). This is likely due to the fact that these genes are related to basic cellular functions. Genes related to stage specific functions such as replication in amastigotes or normalization effects may be causing the slight, but statistically significant increase observed in overall expression in amastigotes. This contrasts with the pattern observed for multigene families where there is a clear overexpression in trypomastigotes (mean fold change = 1.5).

As observations commented on questions 9 and 10 have been described in previous studies and are not novel nor key points in our results, we decided not to focus on them and modified the text accordingly in lines 129-135.

(11) Line 110, "with high variation". What does "high variation" mean here? Compared to what? For the two metrics (n cells +ve for each gene and total expression level) can they give an average and the SD? It would be useful to know how many parasites the "average" surface (and core) gene is expressed in, or more precisely for which the RNA is above the limit of detection.

We refer to the comparison with the expression profile observed for single-copy genes. This point has now been clarified in the text, and we have included the mean and standard deviation for both TcS multigene family genes and single-copy genes in trypomastigotes for both metrics in the Figure 2 legend. The average and distribution of the number of cells in which each gene is detected are shown in Figure 2c and Supplementary Figure 1a. We also added a reference to this panel at the point in the text where the phenomenon is first described.

(12) Line 134, Figure 2b legend needs more detail - what are num_multigene and z_multigene?

Please see our response to Reviewer 1, Question 6. We have now added a clarification to the legends of Figure 1 and Supplementary Figure 1.

(13) Figure 2c, correct the y-axis legend because it implies your values are log10 transformed. Also, it would be useful to have more markers on the y axis so the reader can better estimate the data ranges.

We thank the reviewer for this observation. We have now corrected the y-axis label and markers.

(14) If the y-axis of Figure 2D started at 0 instead of 0.8 and if Lorenz curves were provided then the reader would probably get a fuller sense of the expression heterogeneity in the dataset. The legend states the differences are statistically significant but the actual p-values are not shown.

(15) Line 142-3, more precision is needed on the p-values.

We thank the reviewer for this helpful suggestion. We agree that Lorenz curves provide a clearer representation of expression heterogeneity than the previous plot. Accordingly, we have replaced the original panel (Figure 2d) with Lorenz curves for the groups under comparison, and have made the same change in Supplementary Figure 1d. In addition, we have included gini index values and p-values for all comparisons in Supplementary Table 2.

(16) Figure 3, as in Figure 1a it would be useful to add another UMAP plot to show the two trypo subpopulations.

We thank the reviewer for this suggestion. We have now updated Figure 3 to include a UMAP plot showing the two trypomastigote subpopulations.

(17) What is the observed proportion of broad vs slender trypomastigote morphologies for Dm28c? To be consistent with the speculation at line 162 then wouldn't it need to be approximately 50-50?

The proportions of each trypomastigote subpopulation in the DM28c strain are currently unknown. The only available relevant data come from Brener, 1965 (doi.org/10.1080/00034983.1965.11686277), in which this strain was not included. In the strains analyzed in that study, the relative proportions of broad and slender trypomastigote morphologies were highly variable: across seven strains, broad forms ranged from 18.0% to 77.3%, while slender forms ranged from 2.3% to 71.6%. Given this wide variability and the lack of DM28c-specific data, we cannot assume any expected proportion for this strain.

(18) Line 170, please state how many genes are in the TcS subgroup mentioned here. This is an interesting finding - does this include mostly catalytically active trans-sialidase genes or is it a mixture from across all the subfamilies?

The TcS subgroup with a high frequency of detection comprises 31 genes, none of which belong to the catalytically active Group I trans-sialidases. Instead, this subgroup includes members of Groups II, III, IV, V, VI, and VIII. This information has been added to Supplementary Table 3 and is now stated in the revised manuscript (lines 227 - 228).

(19) Line 175-176, "Gene dropouts might favor random patterns of gene family's detection in scRNA-seq experiments, particularly affecting genes with low expression" - I'm not sure if the authors mean the detection of a gene (or not) in an individual parasite is truly random (pure luck) or whether the term stochastic would be more appropriate because they seem to be referring to randomness around a certain threshold of RNA abundance/stability? They go on to rule this out, at least for TcS genes, essentially arguing that they have something resembling an ON or OFF pattern rather than a spectrum of expression levels. This is potentially very important and could advance the field in a major way, but the fact that so many core and ribosomal genes, which 'should' be always ON, cannot be detected in most cells is a concern. A version of Figure 4B for core and ribosomal genes could be informative - do they show a different pattern to TcS?

Our results reveal a small subset of TcS genes that are frequently detected across cells, a pattern that is not compatible with random detection unless these genes were highly expressed and preferentially captured by random sampling. However, as shown in Figure 4b, many genes expressed at comparable levels are not detected at high frequencies. In line with this, Figure 4c shows that within individual cells, the detected TcS genes exhibit similar expression levels. Finally, we confirmed that this frequently detected subset shows high read counts at the bulk RNA-seq level (Supplementary Figure 2), consistent with the fact that these TcS are frequent in the population even when they are not specially highly expressed within each cell. Taken together, these findings argue against a purely random sampling of TcS genes and support the interpretation that this pattern reflects an underlying biological feature. We agree that further validation will be required. Accordingly, since the initial submission, we have been careful to frame our conclusions conservatively, explicitly noting that dropout remains a limitation of these data that could influence the observed patterns. In the revised version, we have strengthened this point by including a specific statement in the final remarks. Our interpretation is presented as a working hypothesis that is fully compatible with the observations reported here and may be informative for the field. To better reflect this reasoning, we have revised Figure 4b, expanded the discussion, and explicitly included this limitation in the final remarks of the revised manuscript.

(20) Line 238-9, Add details of removing extracellular epimastigotes after cell infections.

Only cellular trypomastigotes collected from the supernatant on day 6 were used for the secondary infection, at a 10:1 parasite-to-cell ratio. After 24 hours, the cultures were washed twice with PBS to remove any remaining extracellular parasites. Under these conditions, i.e. using exclusively trypomastigotes, at this infection ratio, and maintaining the cultures in mammalian medium, we do not expect the presence or survival of extracellular epimastigotes. We have included a sentence in the Methods section clarifying this information in the revised version of the manuscript, line 382.

(21) Line 260, was methanol used to directly resuspend the parasite pellet, or was it resuspended first e.g. in a small volume of PBS?

As described in lines 250-257 of the original manuscript, parasites were washed and resuspended in DPBS before methanol fixation. Methanol fixation was then carried out according to the 10X Genomics Methanol Fixation Protocol. We have now emphasized this more clearly in the revised text in line 400.

(22) What was the doublet rate?

We identified and removed 41 doublets, all belonging to cluster 2, and retained 3,151 singlets for downstream analysis (total cells before removal = 3,192). The resulting doublet rate was 1.28%. We have included a sentence in the Methods section clarifying this information in the revised version of the manuscript, line 439 -440.

(23) What was the frequency of rRNA and kDNA-derived reads?

Approximately 4.02% of the reads were derived from kDNA sequences, while 1.10% corresponded to rRNA-derived reads (Author response image 4).

Author response image 4.

Percentage of mitochondrial and ribosomal rRNA derived reads.

eLife Assessment

This work of fundamental significance introduces a novel statistical model of spiking activity that incorporates continuous-time gain modulation. The authors provide exceptional evidence that the model outperforms earlier approaches and alternative candidates in capturing spiking responses across multiple visual areas in the macaque. Beyond its methodological contribution, the study offers new insights into how stimulus-driven variability and internally generated gain fluctuations evolve over time and between brain areas. The framework is likely to find broad application beyond the datasets examined here.

Reviewer #1 (Public review):

Summary:

In this manuscript, Rupasinghe and co-authors introduce a new statistical model for spiking neurons. Building on earlier work, they propose to model spikes as arising from a Poisson process whereby the firing rate is the product of stimulus drive and a stimulus-independent gain signal. The critical innovation of this work is that the gain signal is modeled in continuous time. Earlier explorations of this statistical construction treated the gain-signal as constant within a trial. This innovation is elegant and important. It makes the model richer, more plausible, and more broadly applicable. The authors show that the model parameters are recoverable from realistic amounts of data and then apply the framework to previously studied datasets. They show that the new model outperforms earlier models and alternative candidates in capturing spiking data across four visual areas of the macaque monkey. Analysis of the model parameters replicates some earlier findings and uncovers several new insights. The model and fitting methods can be broadly applied to partition different types of signals and noise from spiking data and are likely to be widely adopted in the systems neuroscience community.

Strengths:

(1) Through clever use of advanced statistical techniques, the authors manage to infer critical information from single-trial single-cell data.

(2) The question of which aspect of a spike train is signal and which is noise is omnipresent in neuroscience. By improving our ability to characterize the distinct factors that shape spiking activity, this work makes a fundamental contribution to the literature.

Weaknesses:

Overall, I find the work impressive and important. I have a couple of questions and suggestions.

(1) The work is entirely focused on single-cell data. While this is a great starting point, expanding the approach to spiking activity in neural populations is an important future goal.

(2) Line 49-53: These statements seem incorrect to me. The modulated Poisson model, as introduced in Goris et al (2014), is a process model that can perfectly be used to generate spike trains (within a trial, spiking emerges from a Poisson process, which can be homogeneous or inhomogeneous). Moreover, the model contains a parameter that represents the duration of the counting window (delta t). The dependency of over-dispersion on the size of the time bins for real neurons is shown in Figure 1b (inset plot) of that paper (and shown to resemble the model prediction). This time-dependency was further explored by the same authors in Goris et al (2018 - Journal of Vision) and also in Hénaff et al (2020 - Nature Communications ). I suggest that the authors rephrase this argument (here and at some later points in the paper). They could just say that the Goris model makes the simplistic and implausible assumption that, within a given trial, gain does not fluctuate. This is clearly an important limitation and the key difference with the continuous model introduced here.

(3) Line 54-55: I think the first part of the claim is a bit misleading. There is nothing in the Goris model that would inherently limit it to homogeneous Poisson processes, as seems to be implied by this description. The model is built on the assumption that spike generation within a trial arises from a Poisson process. This may very well be an inhomogeneous Poisson process (i.e., a stimulus-dependent time-varying firing rate). Homogeneous and inhomogeneous Poisson processes both give rise to Poisson distributed spike counts (and thus a mixture of Poisson distributions across trials in the Goris model). I suggest the authors clarify this description a bit. Note that the two model variants illustrated in Figure 1b and c were also explored in Hénaff et al (2020 - Nature Communications).

(4) The extension to the continuous case is very elegant!

(5) I find the result shown in Appendix 3 critically important. The recoverability of the model for realistic amounts of data is foundational for the rest of the paper. I would consider including this analysis in the main results section. Not all readers may check Appendix 3, but they should know about this result.

(6) Figure 3: I am wondering whether the inferred gain is capturing some response fluctuations that originate from the cell's phase-selectivity. Could the authors compute the trial-averaged inferred gain (ideally, aligned to stimulus-phase at the start of the trial if this experimental parameter varied across repeats)? If they have successfully partitioned the response variance, the trial-averaged gain should have no systematic temporal structure. If it has a sinusoidal modulation, it may partially capture stimulus-drive. This could be an interesting test to run on all model fits to further validate that the partitioning into a signal and noise component succeeded as intended.

(7) One common observation that is currently not explored is the quenching of neuronal response variability following stimulus onset (Churchland et al 2010 - Nature Neuroscience), which was suggested to reflect a quenching of gain variability in Goris et al (2024 - Nature Reviews Neuroscience). Building on the previous suggestion, the authors could compute the temporal evolution of cross-trial gain variability from the inferred gain traces. Do they recognize a reduction in gain variability following stimulus onset? If so, it would be worthwhile to show this.

(8) Line 543-565: I want to make sure I understand the Baseline Poisson model and Poisson-GP correctly. For the baseline model, I had imagined that the authors would simply use the stimulus-conditioned PSTH as an estimate of the time-dependent firing rate, coupled with an inhomogeneous Poisson process assumption. But they additionally assume a Gamma prior on the firing rate to compensate for the sparseness of the data (sometimes only 5 repeats per condition). The Poisson-GP includes exactly the same model components, but now the time-dependent firing rate is modeled by a Gaussian process. Doing this massively improves the goodness-of-fit (Fig 4A). Do I understand this correctly?

Reviewer #2 (Public review):

Summary:

Neurons have varied responses to external stimuli that cannot be explained by naive Poisson models. Previous work has quantified and partitioned higher-than-Poisson variability in the brain into different components. The authors improve on these methods to infer how both the stimulus drive and internal gain dynamics impact neuronal variability continuously in time. The clean and well-reasoned model is rigorously developed and then applied to neural data across the visual hierarchy. This lends new insights into how variability is partitioned, agreeing with and extending previous work on how that variability changes from early visual areas (LGN, V1) through to higher, motion-sensitive areas (area MT). Another key contribution is that this partitioning can be fully addressed as a continuous-time process, which allows for the dissection of how the timescale of fluctuations in these two components changes across the brain's processing arc.

Strengths:

(1) The model is cleanly derived and thoroughly documented, including usable code shared in a GitHub repo. This makes the method immediately portable to other neural systems.

(2) This is a clear and well-presented piece of work. The figures and writing are clear and understandable, and all pieces of the derivations are included in the main text and supplementary information.

(3) Comparisons to other models, particularly the one from Goris et al., 2014 shows how this Continuous Modulated Poisson (CMP) model outperforms previous work.

(4) New insights about how variability partitioning changes across the visual stream from LGN to MT are revealed, including how the gain fluctuates on longer timescales in higher visual areas. Another key result about the anticorrelation between the variance in stimulus drive and gain fluctuations comports with theories about how neurons maintain efficient, reliable encoding.

(5) In addition to the results reported here, this work will serve as an excellent tutorial for students and postdocs first delving into the sources of variability in the brain.

Weaknesses:

The work is somewhat incremental, building on previous studies of the partitioning of variability in the brain, but it provides important new extensions, as noted above.

The only major gap I would suggest addressing in the Discussion is the observation of sub-Poisson variability in the brain. It seems clear that this model can extend to sub-Poisson variability and its partitioning and perhaps even show how that varies in real time, with an animal's attentional state. That is, of course, beyond the scope of the current work, but could be mentioned in the Discussion.

Author response:

Reviewer #1 (Public review):

We thank the reviewer for the thoughtful and detailed evaluation of our manuscript. We are pleased that the continuous-time formulation and its methodological contributions were viewed as elegant and broadly applicable, and that the empirical analyses provide meaningful new insights into neural variability across the visual hierarchy. We appreciate the reviewer’s constructive suggestions and clarifications, which will help us improve the precision, clarity, and scope of the manuscript. Below we respond to each point in turn and outline the revisions we will make.

(1) Extension to neural populations: We thank the reviewer for this important suggestion. We agree that extending the framework to population recordings is a natural next step. In this work, we focus on single-cell data to establish the model and validate inference. In the revised manuscript, we will expand the Discussion to outline how the framework could be generalized to population activity, for example by incorporating shared latent-variable structure.

(2) Clarification regarding the Modulated Poisson model: We thank the reviewer for pointing this out. We agree that our description was not sufficiently precise and may have been unclear. The modulated Poisson model introduced in Goris et al. (2014) is indeed a generative process model that can be used to generate spike trains, and we apologize for the inaccurate characterization of this framework. Our intended point was that the original formulation assumes gain is constant within a trial (or counting window) and does not provide a principled mechanism for modeling continuously time-varying gain fluctuations within trials. In the revised manuscript, we will clarify this distinction and revise the relevant passages accordingly. We will also cite and discuss related extensions and analyses in Goris et al. (2018) and Hénaff et al. (2020) to provide a more accurate and complete characterization of prior work.

(3) Continuous extensions of the Goris model: We thank the reviewer for this helpful clarification. We agree that the Goris model is not limited to homogeneous Poisson spiking and can incorporate a stimulus-dependent, time-varying firing rate within trials. We did not intend to imply otherwise, and we will revise the relevant text to avoid this misunderstanding. Our intended point was that, in formulating continuous-time extensions, we explicitly model the time-varying stimulus drive using a GP prior, as in the CMP framework, and then consider different assumptions about the temporal structure of the gain process, including constant and finely sampled gain. This highlights the distinction between piecewise-constant gain assumptions and the fully continuous gain process introduced in our model. We will clarify this distinction in the revised manuscript. We will also acknowledge related variants explored in Hénaff et al. (2020) and more clearly describe how our formulation differs, including the role of smoothness priors on the stimulus drive and gain processes.

(4) Continuous-time extension: We thank the reviewer for the positive comment and are pleased that the continuous-time formulation was viewed as elegant.

(5) Parameter recovery analysis: We thank the reviewer for emphasizing the importance of this result. We agree that demonstrating parameter recoverability is foundational to the paper. In the revised manuscript, we will move the Appendix 3 analysis into the main Results section and clearly illustrate how our inference procedure faithfully recovers the generative parameters in simulation studies.

(6) Validation of gain–stimulus separation: We thank the reviewer for this insightful suggestion. We agree that verifying that the inferred gain does not capture stimulus-driven structure is an important validation of the model. In the revised manuscript, we will compute the trial-averaged inferred gain, to assess whether it exhibits systematic temporal structure. This analysis will provide an additional check that the partitioning between stimulus drive and gain fluctuations operates as intended.

(7) Temporal evolution of gain variability: We thank the reviewer for this valuable suggestion. We agree that examining whether gain variability decreases following stimulus onset is an important and relevant analysis. In the revised manuscript, we will compute the temporal evolution of cross-trial gain variability from the inferred gain traces and assess whether a quenching effect is observed after stimulus onset. If present, we will report and illustrate this result.

(8) Clarification of Baseline Poisson and Poisson-GP models: We thank the reviewer for this careful reading. Yes, this understanding is correct. The Baseline Poisson model uses a stimulus-conditioned PSTH as an estimate of the time-dependent firing rate and includes a Gamma prior to regularize rate estimates in conditions with sparse repeats. The Poisson-GP model retains the same structure but models the time-dependent firing rate using a stimulus-specific Gaussian process prior, which substantially improves goodness-of-fit. In the revised manuscript, we will clarify this description. We will also highlight that Figure 4 – figure supplement 2 illustrates how introducing a GP smoothness prior on the stimulus drive markedly improves model fit, even within the Goris-style model.

Reviewer 2 (Public review):

We thank the reviewer for the thoughtful and positive assessment of our work. We are pleased that the model development, empirical analyses, and presentation were found to be clear and rigorous. We appreciate the recognition that the continuous-time formulation meaningfully extends prior variability-partitioning approaches and enables a more precise characterization of how stimulus drive and internal gain dynamics evolve across temporal scales. We are also encouraged that the cross-area analyses and model comparisons were viewed as providing new insights and clear empirical improvements. Below, we address the specific suggestions raised by the reviewer.

Positioning relative to prior work: Regarding the comment on incremental contribution, we agree that our framework builds directly on earlier variability-partitioning approaches. Our goal was to extend these models to continuous time and to develop a principled inference framework capable of characterizing how gain dynamics evolve across temporal scales. We will further clarify this positioning in the revised manuscript.

Extension to sub-Poisson variability: We thank the reviewer for this suggestion. We agree that sub-Poisson variability is an important phenomenon observed in neural data. Because the CMP model builds on a Poisson observation model with stochastic gain modulation, it naturally captures Poisson and super-Poisson variability but cannot generate sub-Poisson spike count statistics in its existing form. We will clarify this limitation in the revised manuscript and expand the Discussion to outline potential extensions that could address sub-Poisson variability, such as incorporating spike-history effects, renewal-process models, or alternative count distributions.

eLife Assessment

This valuable study demonstrates how individual taste preferences shift over time, how these changes relate to cortical activity, and how experience reshapes both. The evidence is largely solid, although additional analyses are needed to strengthen some of the conclusions. The results should be of interest to neuroscientists studying sensory physiology.

Reviewer #1 (Public review):

Summary:

Maigler et al. set out to test the hypothesis that individual differences in taste preferences are (in part) due to individual differences in central taste processing. The first tested rats' preferences for a variety of taste stimuli on multiple days. They then recorded responses of neurons in the taste cortex to the same tastes on two consecutive days.

Strengths:

The authors collected high-resolution behavioral data from the same animals across multiple days, allowing for a detailed characterization of individual variation in taste preferences. They then performed recordings from the same set of animals in response to the same stimuli, allowing them to draw parallels between behavioral and neural responses. They convincingly show that preference ranks for a variety of basic tastes change over time and that the correlation between neural responses and preferences is not stable, correlating more strongly with more recent measures of preference.

Weaknesses:

Behavioral analysis: Data presentation does not show how preferences change over the course of testing. In particular, it is unclear whether there are any systematic changes in preferences over the course of testing that could explain the observed changes in correlation with neural responses, such as changes due to learning (e.g., flavor nutrient conditioning, relief of neophobia), changes in deprivation state, or habituation to/proficiency with the BAT setup. A secondary point is whether any changes in preference are attributed to internal individual versus external contextual factors. Both types of variation (i.e., across individuals and across time within an individual) are mentioned in the introduction, but it is not clear what the authors believe about the nature or neural representation of these sources of variation.

With respect to neural data analysis, no individual animal/day data are shown, making it difficult to assess the extent to which differences in correlation match individual differences in preferences and/or changes in preference with time within individuals. The correlation analysis is also lacking control for the fact that there is a certain degree of "chance" associated with behavioral and neural measures having matching ranks.

Finally, the conclusion that correlations between final day preferences and neural responses obtained from the second recording session are the result of experience needs more justification; it is unclear to what extent changes in correlation may be attributed to overall changes in responsiveness of the neural population.

Reviewer #2 (Public review):

Summary:

The study from Maigler et al investigates how between- and within-animal differences in taste preference relate to differences in neural responsiveness. The experiments rely on an elegant combination of behavioral assays to measure preference (e.g., repeated brief access testing, BAT) and electrophysiological recordings to monitor the activity of ensembles of neurons in the gustatory cortex (GC) of rats.

BAT with distinct batteries of tastants revealed pronounced variability in preference (measured as licking bout size) across individuals. This variability across individuals persisted after repeated testing. Repeated BAT also revealed that each rat's preference for different tastants changed across time.

Electrophysiological responses of GC neurons to batteries of tastants showed that firing in the "late epoch" of taste processing (i.e., 500ms post taste delivery) correlated more strongly with the individualized rat's BAT preference rather than with a canonical preference ranking. Importantly, this correlation was stronger for the last BAT session compared to the first. Finally, the author shows that the correlation disappeared in a second, consecutive recording session, indicating that exposure to tastants reconfigures preferences.

Strengths:

(1) The experimental design allows for an unprecedented look at the relationship between individual variability in taste preferences and neural processing.

(2) The study demonstrates that taste preference variability is not mere experimental noise but reflects the dynamic nature of taste. A key strength is the clear evidence that behavioral variability is reflected in neural activity patterns, establishing a strong correlation between brain and behavior.

(3) The evidence that simple exposure to familiar tastes can reconfigure preferences and taste representations is interesting.

Weaknesses:

(1) The manuscript could use additional corollary analyses to provide a more complete picture of the phenomenon. For instance, how many neurons (per animal and in total) have significant correlations with the final BAT patterns? And with the first BAT? Can a time course of such counts be provided? Can some decoding analyses be performed at a single session level to reconstruct a rat's behavioral preference pattern from its neural activity?

(2) The manuscript could benefit from additional polishing, both in the text as well as in the figures.

Reviewer #3 (Public review):

Summary:

Maigler & Lin et al present a compelling set of behavioral and electrophysiological experiments exploring how individual differences in taste preference map onto neural responses in the gustatory cortex (GC). They go on to examine how both preferences and neural responses shift following intervening taste experience. Their experiments are strengthened by examining tastes of distinct identities and palatability (sweet, sour, salty, bitter) and corresponding each animal's individual preference to the palatability-related late phase of the neural response.

Strengths:

(1) They demonstrate a relationship between the behavioral expression of taste preference and palatability-related GC neural responses. The direct correlation of expression of taste preference with GC neural responses indicates that taste preference behavior may be less noisy than previously thought, reflecting actual neural activity.

(2) They address the stability of individual taste preference by comparing within and between session expression. This finding indicates that individual preference on any given test session can differ from canonical palatability.

(3) They provide evidence that representational drift in palatability coding may arise from sensory experience rather than from the passive passage of time. The findings are novel and potentially impactful. The results are relatively complete.

Weaknesses:

Experiments require further clarification. The interpretations would be strengthened by reorganizing the experimental design.

(1) Figures 5-6 show shifts in palatability-related GC responses from recording day 1 to recording day 2. The authors propose that this drift is due to the taste experience during recording day 1, but the study, as designed, does not directly test this idea. Without a behavioral measure collected after recording day 1 intraoral exposure, it is not possible to determine whether taste preference was altered by that experience, nor whether the neural responses collected on recording day 2 represent current or most recent palatability expression vs something else. The authors' conclusion would be strengthened by adding an intervening brief access test between recording days 1 and 2. The authors could then determine whether the behavioral preferences changed after intraoral taste exposure on recording day 1, as well as whether the new set of taste-related palatability responses correlates with the updated taste preferences.

(2) The current experimental design exposes animals to 3 distinct sets of substances. These substances differ in identity (some rats never experienced sweet, while others did not experience bitter during the recording sessions) and concentration (ranging from very aversive to slightly aversive or possibly even neutral). Because palatability is known to be comparative depending on the other substances available and concentration-dependent, this introduces challenges to interpretation.

The authors state that "no differences in effects were observed between taste batteries" (Methods), but it is not clear which analyses were performed to determine the lack of difference, especially considering that many of the analyses are within-animal. Without more clarity, it is difficult to evaluate whether the interaction of different tastes within the sets of stimuli biases the main conclusions.

(3) Responses to sweet tastes are not reported in the electrophysiology data. This is seemingly the case because rats given set 1 received no sweet stimulus while rats given set 2 received to 2 distinct sweet tastes. Finally, rats given set 3 did not receive quinine, yet quinine is reported in electrophysiology data.

(4) The choice of reporting average lick cluster size is problematic because the authors use thirsty rats with 10-second-long trials. Thirsty rats are likely to lick in relatively long clusters, especially for neutral and palatable tastes. If the rat is mid-cluster when the trial ends, the final cluster would be cut off prematurely, resulting in shorter overall average lick cluster size, disproportionately affecting neutral and palatable tastes over aversive tastes.

(5) Canonical palatability rankings may not apply to the concentrations selected in every stimulus set. This is particularly true for set 1, which included two concentrations of citric acid and quinine for the behavior. It is also not clear which concentrations are reported in Figures 3A2 and 3B2. Meanwhile, the concentrations of quinine and citric acid used for electrophysiology are quite low.

Author response:

Public Reviews:

Reviewer #1 (Public review):

…It is unclear whether there are any systematic changes in preferences over the course of testing that could explain the observed changes in correlation with neural responses, such as changes due to learning (e.g., flavor nutrient conditioning, relief of neophobia), changes in deprivation state, or habituation to/proficiency with the BAT setup.

For the revision, we will add analysis (including either additional panels for Figure 3 or as a new Figure between what are now Figures 3 & 4) testing the hypothesis that preference changes across testing days are non-random. Concretely, we will test: 1) whether the preference for palatable tastes increase with experience (a result that would make sense given research on neophobia; 2) whether the preference for aversive tastes decrease with experience; and 3) whether absolute consumption of any particular taste changes in a reliable direction from session to session.

A secondary point is whether any changes in preference are attributed to internal individual versus external contextual factors. Both types of variation (i.e., across individuals and across time within an individual) are mentioned in the introduction, but it is not clear what the authors believe about the nature or neural representation of these sources of variation.

While we assume that differences between rats are due to internal factors (given the controlled home-cage environment), we can’t be sure that some subtle, subthreshold (for us as observers) factor impacts taste preferences. Similarly, while changes across time within an individual is categorically within the individual, we cannot be sure whether some subtle facet of their experiences determines how preferences change (as opposed to it being purely internal). We will add prose to the Discussion session on this topic—including citation of Hilary Schiff’s recent work showing nurture-related preference changes as part of this new prose.

With respect to neural data analysis, no individual animal/day data are shown, making it difficult to assess the extent to which differences in correlation match individual differences in preferences and/or changes in preference with time within individuals.

The revision will include Figure panels (with analysis) showing the relationships between individual neural responses and consumption in the first and last BAT tests for 1-2 representative rats.

The correlation analysis is also lacking control for the fact that there is a certain degree of "chance" associated with behavioral and neural measures having matching ranks.

Certainly chance cannot explain our results, which consist mainly of within-rat differences in match (i.e., specific enhancement of that match for the most recent behavioral assessment)—a finding that is all the more surprising given that: 1) 2 weeks separate that behavior test and the electrophysiology session; and that 2) that 2-week gap is only 1-3 days less than the gap using the first behavioral test (that reliably correlates less well with the neural data). Nonetheless, we will add an independent, convergent analysis to the revision, testing whether the observed pattern vanishes when we shuffle the preference ranks in the behavioral data—if the result is based on chance, this shuffling should have no impact on the neural-behavioral match.

Finally, …it is unclear to what extent changes in correlation may be attributed to overall changes in responsiveness of the neural population.

We will include a new analysis in the revision testing the hypothesis that the reduction in match between the neural and behavioral rankings reflects changes in neural excitability—spontaneous and taste-driven—between the first and second electrophysiology sessions.

Reviewer #2 (Public review):

The manuscript could use additional corollary analyses to provide a more complete picture of the phenomenon. For instance, how many neurons (per animal and in total) have significant correlations with the final BAT patterns? And with the first BAT? Can a time course of such counts be provided? Can some decoding analyses be performed at a single session level to reconstruct a rat's behavioral preference pattern from its neural activity?

These are all really good ideas. We are in the process of implementing all but the last; we will attempt the last as well, but can’t promise that we have large enough ensembles to provide stable results of such a subtle decoding task (reflecting the last BAT session’s preference pattern significantly better than the first session’s pattern).

The manuscript could benefit from additional polishing, both in the text as well as in the figures.

It is being done, on the basis of suggestions made by R2 in the non-public comments.

Reviewer #3 (Public review):

Without a behavioral measure collected after recording day 1 intraoral exposure, it is not possible to determine whether taste preference was altered by that experience…The authors' conclusion would be strengthened by adding an intervening brief access test between recording days 1 and 2.

We very much appreciate Reviewer 3’s suggestion, but the primary authors involved in data collection on this project have moved on, and we won’t be able to collect the additional dataset that would be required. Instead, we will soften the conclusion that we reach in the last section, and suggest this experiment as a future direction.

The current experimental design exposes animals to 3 distinct sets of substances … [that] differ in identity … and concentration. Because palatability is known to be comparative depending on the other substances available and concentration-dependent, this introduces challenges to interpretation, [and] without more clarity, it is difficult to evaluate whether the interaction of different tastes within the sets of stimuli biases the main conclusions.

This is an interesting point. We hope that some of the work that we are undertaking in response to Reviewers 1 & 2 (see above) will shed light on whether there is any non-randomness in between-session preference changes; such non-randomness would imply that we might want to conclude that preferences change more with one battery than another. But we will perform a more direct test of this hypothesis, breaking the dataset apart and asking whether our phenomena are observed more with one battery than another. If it turns out that the magnitude of the impact of experience does depend on the nature of the taste battery (we predict not, for reasons that are in the manuscript), we shall introduce that complexity into our interpretation, and the Discussion thereof.

Responses to sweet tastes are not reported in the electrophysiology data. This is seemingly the case because rats given set 1 received no sweet stimulus while rats given set 2 received to 2 distinct sweet tastes. Finally, rats given set 3 did not receive quinine, yet quinine is reported in electrophysiology data.

We are unsure of the source of this confusion—in every case, the rat received the same tastes in the electrophysiology sessions that were delivered in the BAT preference tests—but we will modify the text to ensure: 1) that panels reflecting data from a single rat (panels that will therefore necessarily include only a subset of possible tastes) are clearly marked as such; and 2) that the nature of which taste batteries were delivered is more explicit.

The choice of reporting average lick cluster size is problematic because the authors use thirsty rats with 10-second-long trials. Thirsty rats are likely to lick in relatively long clusters, especially for neutral and palatable tastes. If the rat is mid-cluster when the trial ends, the final cluster would be cut off prematurely, resulting in shorter overall average lick cluster size, disproportionately affecting neutral and palatable tastes over aversive tastes.

We have ourselves been deeply concerned with this issue; we have recently published a paper that includes within it a direct test demonstrating that calculations of lick bout lengths from 10-sec BAT trials result in taste palatability estimates that are identical to (and less noisy than) those generated from more classically-used 15-min ad lib licking. We will cite this paper (Lin, et al., 2026) in the Methods section of the revision, along with text clarifying how we calculated lick clusters. That said, we are also planning to conduct an additional analysis that estimates taste preference after removing these “premature bouts” and will evaluate how this recalculation affects our results.

Of course, even if 10-sec BAT trial data DIDN’T provide reliable preference measures, the result of clusters being cut short by the end of a trial would be an underestimation of the preference for the palatable tastes (which drive far more licking than aversive tastes and are therefore more likely to be mid-bout at the end of a trial). Such an underestimation would in turn be expected to reduce the observed neural-behavioral correlation. This fact actually highlights the robustness of our findings.

Canonical palatability rankings may not apply to the concentrations selected in every stimulus set. This is particularly true for set 1, which included two concentrations of citric acid and quinine for the behavior. It is also not clear which concentrations are reported in Figures 3A2 and 3B2. Meanwhile, the concentrations of quinine and citric acid used for electrophysiology are quite low.

In the revision Methods section, we will explicitly motivate our reasoning behind canonical rankings for each taste battery used (the added text will include citations). We have also added to the Discussion section prose concerning the possible impact of possibly getting those rankings wrong—i.e., the impact is minimal, given that our results are largely driven by differences between rats (and day-to-day differences within rat), and the resultant fact that almost any choice of canonical rankings would poorly reflect the behavior of individual rats on individual days.

eLife Assessment

This study presents valuable insights into cellular sites of monoamine production and presence in Pristionchus pacificus, providing a comparative reference for the detailed knowledge of C. elegans, as well as using this information to compare serotonergic anatomy in 22 nematode species. Functional assays support evolved differences in monoaminergic control over certain, but not all, tested behaviors. The evidence is convincing, combining careful genetic experiments and comparative analysis that are well aligned with the conclusions. The results will serve as a basis for (comparative) structural-functional studies of nematode behavior.

Reviewer #1 (Public review):

Summary:

The authors provide extensive immunoreactivity and expression data to map monoaminergic neurotransmitter production sites in Pristionchus pacificus. This nematode is relatively distantly related to the popular model nematode Caenorhabditis elegans, for which such information is already available. They find that dopamine, tyramine, and octopamine are present in the same neurons in both species, but differences are observed for serotonin. This forms the basis for a comparison of serotonergic neurons across 22 nematode species. In addition, they evaluate monoaminergic effects on egg-laying, head movement during reversals, and nictation behavior, to find that monoaminergic control over the latter differs between C. elegans and P. pacificus. This shows that some anatomical flexibility supports similar outcomes, whereas in other cases it is the basis of evolved regulatory differences.

Strengths:

The comparative efforts are laudable and valuable, including a thorough revisiting of old data and corrections of what is judged as a historic misannotation. The expected continued value of this work is also appreciated, because nematodes have similar anatomies and behaviors, cellular-resolution data of different species permits the study of functional evolution of neurotransmitter usage in homologous neurons.

Despite the strong experimental approach, there are some points that require addressing:

(1) Not all the concepts of the introduction ('feeding behaviors', to a lesser extent also 'evolution of neurotransmitter usage in homologous neurons') are followed up upon in the results or discussion sections.

(2) The choice of nematodes ('only' 13 species) may affect what is perceived as ancestral. Also, identifying their cells based on comparisons with Ce or Ppa identifications only is understandable but mildly risky: there are many cells in the head, and mistakes would go unnoticed until detailed analysis in each species can provide conclusive evidence.

(3) It is not reported whether the nictation-defective mutants have general locomotion defects; therefore, whether the reported problem is specific to this host-finding behavior or not.

(4) The section on RIP neurons makes sense for Ppa, but not for Ce (dauers in fact have weakened IL2-to-RIP connections), and should be revised. The nictation data also do not support the breadth of the conclusions, which should either be toned down or rephrased as hypothetical.

(5) The discussion mostly reiterates the results, leaving little room for the author's interpretations and opinions. I would suggest reworking in favor of conceptual discussion.

Reviewer #2 (Public review):

Summary:

This paper makes important contributions to our understanding of how nervous systems evolve, with a particular focus on whether changes in neurotransmitter usage within homologous neurons represent a mechanism for evolutionary adaptation without large-scale changes to circuitry. Comparing the predatory nematode P. pacificus with C. elegans, this study systematically examines monoamine-producing neurons, assesses how their neurotransmitter identities differ between homologous neural types, and determines how these differences relate to behavior.

Strengths:

The major strength of this work is its breadth, rigor, and data quality. It combines multiple, independent lines of evidence to assign neurotransmitter identity for neurons with homology grounded in lineage, morphology, and connectomics, which is essential for meaningful cross-species comparisons. Additionally, by extending the analysis beyond P. pacificus and C. elegans to other nematodes, the authors convincingly argue that features observed in P. pacificus likely reflect an ancestral state. This depth greatly enhances the significance of the conclusions.

This work is likely to have a significant impact on the fields of comparative neurobiology and nervous system evolution. It demonstrates a powerful system and approach for linking molecular identity, cell-type homology, circuit context, and behavior across species. The data generated here will be a valuable resource for the community and provide a strong foundation for future mechanistic studies.

More broadly, the study reinforces the idea that evolutionary change in nervous systems can occur through modulation of chemical signaling within conserved circuits, rather than through complete rewiring. This conceptual framework is likely to influence how researchers think about neural evolution in other systems.

Weaknesses:

Given the availability of detailed connectivity information for both species, a more explicit comparison of the local circuit context of key neurons would further strengthen the link between molecular identity and circuit function.

Reviewer #3 (Public review):

Summary:

The study by Hong, Loer, Hobert, and colleagues is a comprehensive description of monoaminergic neurons in the nematode Pristionchus pacificus. The work used multiple, complementary approaches, including immunostaining and expression of genes involved in neurotransmitter synthesis or transport, to identify neurons that express a monoamine neurotransmitter. Moreover, this study characterized the phenotypes of various mutants to study their organismal function. Extensive comparisons are made to C. elegans, the nematode model that, in a way, anchors the model studied here, and new outgroup species were examined for some features so that the polarity of their evolution could be inferred. Although there is no simple or groundbreaking punchline to distill from the manuscript (i.e., other than some things are the same as in C. elegans, and some things are different), and while the study is basically descriptive in nature, the scope of the project warrants broad attention.

Strengths:

This manuscript offers a tremendous resource for those who use this species as a model, which, based on the author list alone, includes many labs. This study sets the bar for what can be done in a "satellite" model system.

Given the complementarity of approaches used, such as the position of cell bodies, the connectivity and morphology of dendrites, and a previously published atlas of the connectome for this species, the identification of specific neurons (which, as the authors point out, can be easily mistaken) is convincing throughout. Likewise, appropriate caution is observed where neuron identities are ambiguous, e.g., unlabelled cells in Figure 5, or ambiguous identities in other species, as shown in Figure 10. There was a lot of data to unpack in this manuscript, but I could not find any obvious flaws in neuron identification.

Also, the phenotypic assays were straightforward and informative.

Weaknesses:

No serious weaknesses were noted. One minor comment is that in general, I think the Methods could use some additional text to describe what the goal of any given technique was. For example, although there is a description of the HCR protocol in the methods, nowhere does it say what genes this method would be used for. In addition to what is shown in Figure 4, this information should be given in the Methods.

Author response:

Public Reviews:

Reviewer #1 (Public review):

Summary:

The authors provide extensive immunoreactivity and expression data to map monoaminergic neurotransmitter production sites in Pristionchus pacificus. This nematode is relatively distantly related to the popular model nematode Caenorhabditis elegans, for which such information is already available. They find that dopamine, tyramine, and octopamine are present in the same neurons in both species, but differences are observed for serotonin. This forms the basis for a comparison of serotonergic neurons across 22 nematode species. In addition, they evaluate monoaminergic effects on egg-laying, head movement during reversals, and nictation behavior, to find that monoaminergic control over the latter differs between C. elegans and P. pacificus. This shows that some anatomical flexibility supports similar outcomes, whereas in other cases it is the basis of evolved regulatory differences.

Strengths:

The comparative efforts are laudable and valuable, including a thorough revisiting of old data and corrections of what is judged as a historic misannotation. The expected continued value of this work is also appreciated, because nematodes have similar anatomies and behaviors, cellular-resolution data of different species permits the study of functional evolution of neurotransmitter usage in homologous neurons.

Despite the strong experimental approach, there are some points that require addressing:

(1) Not all the concepts of the introduction ('feeding behaviors', to a lesser extent also 'evolution of neurotransmitter usage in homologous neurons') are followed up upon in the results or discussion sections.

We will address the relative treatment of particular topics in the introduction and discussion in a revised version of the article.

(2) The choice of nematodes ('only' 13 species) may affect what is perceived as ancestral.

See above regarding ‘13 species’ (actually 22). Most species and genera were specifically selected previously (Loer and Rivard, 2007; Rivard et al., 2010) for broad phylogenetic coverage, representing different species and genera in 4 major clades within ‘clade V’ (Kiontke et al., 2007; Sudhaus, 2011): Anarhabditis (Caenorhabditis, including both the Elegans and Drosophilae species groups), Synrhabditis (Oscheius, Metarhabditis, Reiterina and Rhabditella), Pleiorhabditis (Teratorhabditis, Mesorhabditis, Rhomborhabditis and Pelodera), and Diplogastrids represented by P. pacificus. Among the outgroups to clade V, there are 3 distinct clades represented, each with at least two species and/or genera represented. Therefore, we believe that the determination of an ancestral condition is well-founded. We plan to add this rationale to the revised version to make this clearer.

(2, continued) Also, identifying their cells based on comparisons with Ce or Ppa identifications only is understandable but mildly risky: there are many cells in the head, and mistakes would go unnoticed until detailed analysis in each species can provide conclusive evidence.

We agree that there is a mild risk of incorrect identification but believe that appropriate caveats are noted in the text. Furthermore, the recent head EM reconstruction and complete embryonic cell lineage of the P. pacificus (Cook et al., 2025) shows a nearly 1-1 homology correspondence between head neurons (e.g., only a single head neuron is missing in the Ppa head relative to Cel due to altered apoptosis), and a quite high level of conservation of neurite morphology and soma position between Cel and Ppa suggests that identifications are likely correct when examining related nematodes. In cases for which a serotonin-immunoreactive cell is found in the predicted location (and often having apparent associated neurites), its homology to the matching Cel and Ppa cell is the most parsimonious interpretation: otherwise, one cell would have to lose expression and another nearby cell gain it.  

(3) It is not reported whether the nictation-defective mutants have general locomotion defects; therefore, whether the reported problem is specific to this host-finding behavior or not.

None of the mutants we tested for nictation behavior, including those that show severe defects in nictation (Ppa-cat-1, Ppa-tph-1, Ppa-tdc-1, Ppa-tbh-1), exhibited noticeable general locomotion defects either as dauers or non-dauers. Further clarification will be provided in a revised version of the article.

(4) The section on RIP neurons makes sense for Ppa, but not for Ce (dauers in fact have weakened IL2-to-RIP connections) and should be revised. The nictation data also do not support the breadth of the conclusions, which should either be toned down or rephrased as hypothetical.

We plan to address these concerns in a revised version of the article.

(5) The discussion mostly reiterates the results, leaving little room for the author's interpretations and opinions. I would suggest reworking in favor of conceptual discussion.

As noted above, we agree to address the relative treatment of matters in discussion in a revised version of the article.

Reviewer #2 (Public review):

Summary:

This paper makes important contributions to our understanding of how nervous systems evolve, with a particular focus on whether changes in neurotransmitter usage within homologous neurons represent a mechanism for evolutionary adaptation without large-scale changes to circuitry. Comparing the predatory nematode P. pacificus with C. elegans, this study systematically examines monoamine-producing neurons, assesses how their neurotransmitter identities differ between homologous neural types, and determines how these differences relate to behavior.

Strengths:

The major strength of this work is its breadth, rigor, and data quality. It combines multiple, independent lines of evidence to assign neurotransmitter identity for neurons with homology grounded in lineage, morphology, and connectomics, which is essential for meaningful cross-species comparisons. Additionally, by extending the analysis beyond P. pacificus and C. elegans to other nematodes, the authors convincingly argue that features observed in P. pacificus likely reflect an ancestral state. This depth greatly enhances the significance of the conclusions.

This work is likely to have a significant impact on the fields of comparative neurobiology and nervous system evolution. It demonstrates a powerful system and approach for linking molecular identity, cell-type homology, circuit context, and behavior across species. The data generated here will be a valuable resource for the community and provide a strong foundation for future mechanistic studies.

More broadly, the study reinforces the idea that evolutionary change in nervous systems can occur through modulation of chemical signaling within conserved circuits, rather than through complete rewiring. This conceptual framework is likely to influence how researchers think about neural evolution in other systems.

Weaknesses:

Given the availability of detailed connectivity information for both species, a more explicit comparison of the local circuit context of key neurons would further strengthen the link between molecular identity and circuit function.

We plan to address these concerns in a revised version of the article.

Reviewer #3 (Public review):

Summary:

The study by Hong, Loer, Hobert, and colleagues is a comprehensive description of monoaminergic neurons in the nematode Pristionchus pacificus. The work used multiple, complementary approaches, including immunostaining and expression of genes involved in neurotransmitter synthesis or transport, to identify neurons that express a monoamine neurotransmitter. Moreover, this study characterized the phenotypes of various mutants to study their organismal function. Extensive comparisons are made to C. elegans, the nematode model that, in a way, anchors the model studied here, and new outgroup species were examined for some features so that the polarity of their evolution could be inferred. Although there is no simple or groundbreaking punchline to distill from the manuscript (i.e., other than some things are the same as in C. elegans, and some things are different), and while the study is basically descriptive in nature, the scope of the project warrants broad attention.

Strengths:

This manuscript offers a tremendous resource for those who use this species as a model, which, based on the author list alone, includes many labs. This study sets the bar for what can be done in a "satellite" model system.

Given the complementarity of approaches used, such as the position of cell bodies, the connectivity and morphology of dendrites, and a previously published atlas of the connectome for this species, the identification of specific neurons (which, as the authors point out, can be easily mistaken) is convincing throughout. Likewise, appropriate caution is observed where neuron identities are ambiguous, e.g., unlabeled cells in Figure 5, or ambiguous identities in other species, as shown in Figure 10. There was a lot of data to unpack in this manuscript, but I could not find any obvious flaws in neuron identification.

Also, the phenotypic assays were straightforward and informative.

Weaknesses:

No serious weaknesses were noted. One minor comment is that in general, I think the Methods could use some additional text to describe what the goal of any given technique was. For example, although there is a description of the HCR protocol in the methods, nowhere does it say what genes this method would be used for. In addition to what is shown in Figure 4, this information should be given in the Methods.

More detailed methods will be provided in a revised version of the article.

eLife Assessment

This study presents valuable findings on how retrieval practice protects memory inferences from stress via covert memory reactivation. Via two EEG experiments manipulating stress and retrieval practice, the authors provide solid evidence supporting the conclusion. This work will be of interest to cognitive and affective neuroscientists working on the intersection between memory and stress.

Reviewer #1 (Public review):

Summary:

This manuscript examines whether retrieval practice protects memory-based inference from acute stress and proposes rapid neural reactivation of a bridging memory element as the underlying mechanism. Using a two-day associative inference paradigm combined with EEG decoding, the authors report that stress impairs inference accuracy and speed, while retrieval practice eliminates these deficits and restores neural signatures associated with bridge-element reactivation. The study addresses an important and timely question by integrating research on retrieval-based learning, stress effects on memory, and neural dynamics of inference. While the work provides promising multi-level evidence linking behavioral and neural findings, limitations in experimental design, causal interpretation, and decoding specificity weaken the strength of the mechanistic claims and suggest that further work is needed to disentangle strengthened associative memory from inference-specific protection effects

Strengths:

(1) Strong theoretical integration<br /> The study integrates three influential frameworks: memory integration through associative inference, stress-induced retrieval impairment, and the testing effect. The authors present a clear theoretical narrative linking these domains and derive testable hypotheses that retrieval practice protects inference by strengthening neural reactivation of a bridge element. The conceptual framing is well-grounded in prior literature and addresses an important gap regarding neural dynamics during inference.

(2) Multi-level evidence<br /> The study provides converging behavioral and neural evidence. The authors demonstrate that stress reduces inference accuracy and speed, while retrieval practice eliminates these deficits. EEG decoding further suggests that bridge element reactivation predicts successful inference. The combination of behavioral performance and neural decoding strengthens the overall argument.

(3) Transparent experimental implementation<br /> The procedures are described in substantial detail, including stimulus construction, stress manipulation, and decoding pipelines. Data and code availability are also strengths, facilitating reproducibility.

Weaknesses:

(1) Insufficient evidence that retrieval practice specifically protects inference rather than strengthening associative memories

A central claim of the manuscript is that retrieval practice specifically protects inference ability rather than simply strengthening underlying associative memories. However, the current data do not convincingly distinguish between these possibilities. Although the authors limited analyses to trials in which AB and BC pairs were correctly retrieved in the subsequent memory test, this procedure does not fully rule out the possibility that improved inference performance reflects stronger base associative memories rather than enhanced integrative processes.

Importantly, the direct memory retrieval test used a two-alternative forced-choice (2AFC) format, which inherently allows a substantial proportion of correct responses to arise from guessing. Consequently, trials classified as "successfully retrieved" may still include weak associative memory traces, making it difficult to conclude that failures in inference reflect deficits in integration rather than incomplete associative learning.

The authors further argue that retrieval practice does not improve inference in the absence of stress, suggesting independence between inference and associative memory strength. However, this null effect does not sufficiently rule out mediation through strengthened premise memory. A factorial design and/or mediation analysis would be necessary to determine whether inference resilience emerges independently of premise memory strength.

(2) Apparent below-chance inference performance raises interpretational concerns

One surprising aspect of the results is that inference performance across experiments and groups appears to fall below the theoretical chance level (0.33) in Figure 4A. This is particularly unexpected because analyses were restricted to trials in which participants correctly retrieved both AB and BC associations.

If performance is indeed below chance, this raises concerns regarding whether participants fully understood the task instructions or whether other methodological factors influenced performance. Additionally, below-chance performance complicates the interpretation of subsequent behavioral and neural analyses. It is possible that this reflects my misunderstanding of the figure; therefore, clarification from the authors regarding how inference accuracy is calculated and presented would be helpful.

(3) Between-experiment implementation of retrieval practice weakens causal inference

The retrieval practice manipulation was implemented as a separate experiment rather than as part of a factorial design. Experiment 2 was conducted after results from Experiment 1 were known, and the authors acknowledge this post hoc decision. This design introduces several potential confounds, including cohort differences between experiments, possible differences in participant motivation or task familiarity, and reduced ability to rigorously test interaction effects.

Although the authors combined data across experiments to test interactions between stress and retrieval practice, such post hoc aggregation cannot fully substitute for a factorial design. A within-experiment 2 × 2 design (Stress × Retrieval Practice) would provide substantially stronger causal evidence and reduce confounding influences.

(4) Lack of an appropriate comparison condition for retrieval practice limits the interpretation of the mechanism

Although acknowledged briefly in the discussion, the absence of an appropriate comparison condition for retrieval practice represents a critical limitation. Without a matched re-exposure or restudy control condition, it remains unclear whether observed benefits are attributable specifically to retrieval practice or to additional exposure to AB and BC associations.

Furthermore, it is unclear whether retrieval practice operates at the trial level or the participant level. Retrieval practice could enhance memory representations for specific practiced items, making those trials more resistant to stress, or it could induce a more global change in cognitive strategy or stress resilience across participants. One way to address this issue would be to analyze inference performance separately for trials that were successfully retrieved during the retrieval practice phase versus those that were not.

(5) Interpretation of EEG decoding as bridge-element reactivation may be overstated

The neural decoding results form the mechanistic foundation of the manuscript; however, the interpretation that decoding reflects reactivation of specific bridging memories may be overstated. The classifier distinguishes between face and building categories, and because the bridging element belongs to one of these categories, successful decoding may reflect category-level semantic activation rather than reinstatement of item-specific episodic representations.

Alternative explanations include category-level retrieval, strategic task differences, or even attentional biases. Because only two categories were used, the decoding analysis lacks the specificity necessary to distinguish between category-level and item-level reactivation. As such, conclusions regarding the reinstatement of specific bridging memories should be tempered or supported with additional analyses.

Reviewer #2 (Public review):

Summary:

Guo et al. investigate the neural and behavioral mechanisms of stress-induced impairments in memory-based inference. Across two well-powered experiments (N=136), the authors demonstrate that acute stress disrupts the rapid neural reactivation of "bridge" elements necessary for novel inferences. Crucially, they identify retrieval practice as a robust behavioral buffer that restores both inferential performance and the underlying neural signatures of memory reactivation.

Strengths:

(1) The use of two independent experiments provides high confidence in the behavioral findings.

(2) Utilizing time-resolved EEG decoding allows the authors to pinpoint the "online" moment of inferential failure, a significant advancement over the lower temporal resolution of fMRI.

Weaknesses:

(1) The authors correctly timed the inference task to begin approximately 20 minutes after the onset of the stressor. While this window aligns with the expected peak of the glucocorticoid (HPA) response, it also represents a period where the rapid adrenergic (SAM) response, confirmed by heart rate elevation, is still highly influential. As the authors acknowledge, because they did not collect saliva samples due to safety protocols, they cannot definitively separate the influence of peak cortisol from the tail-end of the adrenergic surge on the observed memory impairments.

(2) Figures 4 and 6: Without asterisks is really difficult to compare the significant group differences.

Appraisal and Impact:

This study provides high-quality evidence that acute stress impairs the rapid neural reactivation of "bridge" elements necessary for novel memory-based inferences. By leveraging the high temporal resolution of EEG decoding, the authors identify the specific neural "chokepoint" where inferential failure occurs. The research is strengthened by two independent experiments and the identification of retrieval practice as a powerful buffer that not only preserves but also enhances neural reactivation under pressure. The findings have significant implications for both cognitive neuroscience and applied learning science.

Reviewer #3 (Public review):

Summary:

In this study, Guo and colleagues investigated the effects of stress and retrieval practice on memory inference. In the first experiment, they found that memory inference was significantly worse after induced stress. Conversely, when participants received retrieval practice in the second experiment, they found no significant differences between these conditions. They monitored EEG during the inference phase and applied multivariate decoding analysis to examine evidence of neural reactivation. Complementing the behavioural findings of the first experiment, they found that they were able to decode the stimulus category of the inference item with more fidelity in the no stress condition. Surprisingly, they found the opposite direction when participants had retrieval practice, with stronger evidence of reactivation in the stress condition than in the control condition.

Strengths:

(1) The authors have carefully designed two studies investigating the effects of stress and memory retrieval on memory inference.

(2) The use of multivariate decoding on the inference phase data sheds new light on how stress and retrieval may impact the neural signatures of inference processing.

Weaknesses:

(1) There are some key gaps in the reporting of the data. In particular, data is missing on how many trials were included in the inference phase and how many were retrieved in the direct memory task. This is important to know as the main conclusions are based on inference trials proportional to the direct retrieval trials. Considering that the direct retrieval performance differs significantly between the experiments, there could be issues with floor/ceiling effects (in the behaviour) and statistical power (in the EEG results) that confound the comparisons between experiments. Without the data, it is difficult to draw conclusions.

(2) There are some relatively strong conclusions drawn without the data to support them. An important example is the title suggesting a mechanistic role of memory reactivation for these effects; however, the data instead suggest a relationship between successful inference and evidence of reactivation. Additionally, one-tailed t-tests have been used in follow-up tests, and, as I understand it, no multiple comparisons corrections have been applied to the post-hoc tests, suggesting that these findings should be interpreted with caution.

(3) In places, the structure is unclear, making the narrative difficult to follow, often making it necessary for the reader to go back and forth between the sections to understand the study and analyses. I have made some recommendations for how to improve this.

Author response:

Public reviews:

Reviewer #1 (Public review):

(1) We agree that the current design does not allow us to cleanly dissociate whether the beneficial effect of retrieval practice on AC inference under stress reflects a selective enhancement of inferential processing or, instead, stronger memory for the underlying AB and BC premise pairs that supports later inference. We plan to revise the manuscript to remove wording that could be read as claiming that retrieval practice specifically protects inference independently of associative-memory strengthening.

Our intended interpretation is more modest. As shown in Section 3.2.3, retrieval practice improved direct premise-memory performance, consistent with the well-established testing effect. In the present paradigm, successful AC inference necessarily depends on access to the AB and BC premise associations. Accordingly, strengthened premise memory is not an alternative explanation that can be excluded by our data, but rather a plausible mechanism through which retrieval practice may promote more resilient inference performance under stress.

Because AC inference in our paradigm necessarily depends on retrieving and linking the AB and BC premise pairs, strengthened premise memory is not merely a competing explanation that can be separated from inference performance in the current dataset. Rather, it is a plausible mechanism through which retrieval practice may support inference, especially under stress. We therefore will revise the manuscript to avoid implying that retrieval practice protects inferential processing independently of associative-memory strengthening, and instead interpret the effect more conservatively as reflecting enhanced premise representations and/or more effective reactivation of bridge information during inference.

We also agree that the post-inference direct memory test, which used a 2AFC format, provides only a coarse measure of premise-memory strength and allows some proportion of correct responses to arise from guessing. Therefore, restricting analyses to trials in which AB and BC were later answered correctly does not fully guarantee that those trials were supported by strong associative memories. We will acknowledge this limitation explicitly in the manuscript and have tempered our interpretation of these “successfully retrieved” premise trials accordingly. More stringent measures, such as cued recall, confidence-based memory judgments, or other continuous indices of premise-memory strength, would be better suited to this question in future work.

Finally, we agree that the absence of a retrieval-practice benefit in the non-stress condition does not by itself rule out mediation through strengthened premise memory. Because the retrieval-practice manipulation was introduced in a follow-up study after completion of Study 1, the present dataset was not designed as a single fully crossed factorial experiment. In response to the reviewer’s suggestion, we will add an exploratory mediation analysis testing whether premise-memory performance statistically accounts for the relationship between retrieval practice and inference performance. We will report this analysis cautiously, given that premise memory was assessed using a post-inference 2AFC measure, and we note in the manuscript that a future fully crossed design with more sensitive premise-memory measures will be needed for a stronger test.

(2) We apologize that the presentation of Figure 4A was not sufficiently clear and may have created the impression of below-chance inference performance. The values shown in Figure 4A do not represent raw 3-alternative forced-choice (3AFC) A-C inference accuracy, for which the theoretical chance level would be 0.33. Instead, Figure 4A plots a normalized inference index, calculated as inference performance relative to direct retrieval performance, to account for individual differences in the availability of the directly learned premise pairs. Therefore, the raw 3AFC chance level is not the appropriate reference for interpreting this measure. To avoid this confusion, we will clarify in the revised manuscript and figure legend that Figure 4A shows a normalized inference index rather than raw inference accuracy.

(3) We agree that implementing retrieval practice in a separate experiment, rather than within a single 2 × 2 factorial design, limits the strength of the causal inference regarding retrieval practice and reduces our ability to formally test the retrieval practice × stress interaction within one unified design.

In response, we will revise the manuscript to more explicitly acknowledge this limitation and to temper our interpretation throughout. Specifically, we now avoid overstating retrieval practice as definitively preventing the effects of stress, and instead describe the findings more cautiously as evidence that retrieval practice was associated with attenuation of stress-related inference impairments across experiments. We also will add a limitation statement in the Discussion noting that the current design cannot fully rule out cohort-related confounds and that a fully crossed factorial design will be necessary in future work to provide a more rigorous test of the interaction between retrieval practice and stress.

At the same time, we have clarified that the two experiments were conducted under closely matched conditions: participants were recruited using the same protocol from the same campus population, demographic characteristics were matched, and both experiments were run in the same laboratory using the same EEG system, task procedures, and experimenter team. We agree, however, that these procedural consistencies reduce but do not eliminate the concern about between-experiment confounds.

(4) We agree that the absence of a matched re-exposure/restudy control condition limits the mechanistic interpretation of the retrieval-practice effect. In the revised manuscript, we will make this limitation more explicit in the Discussion and temper our conclusions accordingly. Specifically, we clarify that the present design shows that a post-encoding retrieval-practice intervention buffered the impact of acute stress on later inference, but it does not allow us to determine whether this benefit is specific to retrieval practice per se, rather than to additional exposure to the AB and BC associations.

We also agree that it is important to distinguish whether the effect operates at the level of specific practiced items or reflects a more global participant-level effect. In the current study, however, the retrieval-practice phase in Experiment 2 was implemented as a brief timed free-recall procedure rather than a trial-by-trial cued retrieval task, and the available records do not allow us to reliably link retrieval-practice success for individual associations to specific later AC inference trials. Therefore, we cannot directly compare later inference performance for successfully versus unsuccessfully retrieved items on a trial-by-trial basis.

To address this issue as far as possible with the current dataset, we instead plan to conduct an additional item-level robustness analysis using mixed-effects models that accounted for variability across ABC associations. Specifically, we tested whether the critical stress-by-retrieval-practice effect remained after modeling triad-level variability, and whether there was evidence that this effect differed substantially across triads. This analysis does not provide a direct test of whether successfully retrieved items benefit more than unsuccessfully retrieved items, but it does help assess whether the observed effect is broadly distributed across associations or driven by only a small subset of items.

(5) We agree that our current decoding approach does not justify a strong claim of item-specific reinstatement of a unique bridge memory. The classifier was trained to discriminate stimulus categories (faces vs. buildings) in the independent localizer and then applied during the inference phase. Therefore, the present analysis is better interpreted as indexing reactivation of bridge-related category information, rather than reinstatement of an item-specific episodic representation.

Importantly, however, we believe this signal remains theoretically informative for the inferential process examined here. In our design, the bridge element B belonged to one of the trained categories, and the classifier was applied during the cue period when no face or building stimulus was physically present. Thus, successful decoding in this time window suggests that task-relevant bridge-related information was re-expressed online during inference, rather than reflecting concurrent perceptual processing. At the same time, we agree that, because only two categories were used, the decoding analysis cannot fully dissociate bridge-related category reactivation from broader category-level retrieval, strategic task differences, or attentional contributions.

To address this concern, we plan to revise the manuscript in three ways. First, we will soften the interpretation throughout the Results and Discussion to avoid claims of item-specific bridge-memory reinstatement. Second, we now refer to the decoding effect more conservatively as bridge-related or category-level mnemonic reactivation during inference. Third, we have added an explicit limitation stating that the current design does not allow us to distinguish item-specific episodic reinstatement from category-level reactivation, and that future work using more fine-grained representational analyses and/or a larger stimulus set will be needed to resolve this issue more directly.

Reviewer #2 (Public review):

(1) We agree with this important point. The inference task was scheduled to begin approximately 20 minutes after stress onset based on prior human stress literature, with the intention of probing a time window commonly associated with glucocorticoid effects. However, as the reviewer notes, this period may also still reflect residual adrenergic/SAM influences. Because salivary cortisol was not collected due to the COVID-19-related safety protocol, we cannot disentangle the relative contributions of glucocorticoid and adrenergic responses to the observed stress-related effects on inference and neural reactivation. We will revise the manuscript to make this limitation more explicit in the Discussion and to avoid attributing the effects to a specific physiological component of the stress response.

(2) In the revised manuscript, we will add asterisks (or equivalent significance annotations) to Figures 4 and 6 to improve clarity and readability.

Reviewer #3 (Public review):

(1) We thank the reviewer for highlighting this important reporting issue. We agree that the number of trials contributing to the behavioral and EEG analyses should be reported more explicitly, particularly because inference performance was analyzed in relation to direct retrieval performance and because direct retrieval differed across experiments.

In the revised manuscript, we will report, for each group and experiment, the number of trials presented in the AC inference phase, the number of trials retained for the behavioral analyses, and the number of successfully retrieved direct-memory trials in the AB and BC tasks. These values will be summarized in the revised Results section and in Supplementary Tables.

To directly address the reviewer’s concern, we will also compared trial counts across groups/experiments and evaluated whether differences in direct retrieval performance could account for the inference and EEG effects. To further address the concern about potential unequal trial numbers, we plan to repeat the analyses such as trial-count-matched subsets analyses to see whether results remained qualitatively unchanged.

(2) We thank the reviewer for this important comment. We agree that our original title and some parts of the manuscript used language that was stronger than warranted by the data. Our results show that rapid reactivation of the bridge element is associated with successful inference and is modulated by stress and retrieval practice, but they do not by themselves establish a causal mechanistic role for reactivation. We therefore plan to revise the title and softened the relevant wording throughout the manuscript to better reflect the correlational nature of this evidence.

Specifically, we plan to change the title from “Retrieval practice prevents stress-induced inference impairment by restoring rapid memory reactivation” to “for example, Retrieval practice prevents stress-induced inference impairment and preserves rapid bridge-item memory reactivation” We also revised the Abstract, Results, and Discussion to replace stronger mechanistic wording such as “prevents,” “restoring,” and “essential neural mechanism” with more cautious phrasing such as “buffers” or “attenuates,” “preserves” or “is associated with,” and “neural correlate” or “candidate process,” as appropriate. This revision will led us to temper the overall interpretation of the EEG findings: rather than claiming that reactivation is the mechanism by which retrieval practice prevents stress-related inference deficits, we now conclude that rapid bridge-item reactivation is a neural correlate of successful inference that is sensitive to stress and enhanced by retrieval practice.

We also appreciate the reviewer’s concern regarding the use of one-tailed follow-up tests and the absence of multiple-comparison correction. With respect to the one-tailed t-tests, these follow-up comparisons were conducted because the relevant hypotheses were directional a priori. Based on prior work and our theoretical framework, we specifically predicted that acute stress would impair inference-related performance and neural reactivation, and that retrieval practice would mitigate these effects. The follow-up tests were therefore not exploratory post-hoc comparisons, but planned tests used to decompose the significant omnibus effects in the predicted direction. For this reason, we considered one-tailed testing appropriate for these comparisons.

Similarly, we did not apply an additional multiple-comparison correction to these planned follow-up tests because they were limited in number, theory-driven, and conducted to evaluate specific directional predictions rather than to search broadly across many possible contrasts. Importantly, our interpretation does not depend on any isolated post-hoc comparison, but on the consistency of the results across behavioral inference measures, neural decoding of bridge-item reactivation, and theta-band analyses. We have revised the manuscript to make this rationale clearer and to ensure that the follow-up results are interpreted in the context of the full pattern of evidence.

(3) We agree that, in the previous version, parts of the manuscript were not structured clearly enough, which may have made it difficult for readers to follow the logic of the study and the sequence of analyses without moving back and forth across sections. In the revised manuscript, we will reorganize the presentation to improve the overall narrative flow and readability. Specifically, we plan to clarify the study logic and analysis sequence, strengthened transitions between sections, and revised the relevant text in line with the #reviewer3’s detailed suggestions.

eLife Assessment

In this fundamental work Horne et al present compelling evidence that YbjP is a novel binding partner of the TolC channel protein. The YbjP is characterized using cryo-EM, and its role probed using pull-down experiments, in vivo crosslinking, functional assays along with phylogenetic analysis which are all properly performed and presented and support the main conclusions. While the study does not identify a clear role for this protein, the results contribute to the understanding of this complex system and will be of interest to those working in the fields of membrane transport and antimicrobial resistance.

Reviewer #1 (Public review):

Summary:

The authors report a novel binding partner of the TolC channel protein that forms complexes with the two principal classes of transporter-based tripartite assemblies (both ABC- and RND-transporter based) and appears to modulate their function, while also anchoring TolC into the outer membrane, compensating for the lack of direct lipidation seen in other members of the OMF family.

The newly identified protein, YbjP, is comprehensively characterized from both phylogenetic and structural perspectives. Two independent cryo-EM structures (MacAB-TolC-YbjP and AcrABZ-TolC-YbjP) provide strong structural evidence for its role and are generated using peptidiscs, mimicking the membrane environment. These findings are further supported by pull-down experiments (including state-of-the-art in vivo photo crosslinking) and functional assays for a well-rounded characterisation of the protein, and a significant amount of modelling and phylogenetic analysis. This work sheds light on the function of the members of the DUF3828-containing protein family, which appear to anchor TolC to the outer membrane and influence the expression of the TnaB and YojI transporters.

Strengths:

The strengths of the manuscript are numerous, and it presents a well-rounded package of structural biology complemented by functional and computational studies.

The full assemblies of both MacAB-TolC-YbjP and AcrABZ-TolC-YbjP are reconstituted and resolved to near-atomic resolution using cryo-EM for unambiguous assignment of binding interfaces, which are then validated using a number of techniques, including ITC, in vitro and in vivo binding assays and cross-linking.

The evolutionary analysis is particularly notable, and provides genuine insight into the DUF3828-containing proteins, the function of which remains enigmatic till now. Similarly, the involvement of YbjP in trafficking of TolC and the analysis of the impact of YbjP deletion of the full E. coli proteome is commendable.

Overall, this is a very solid piece of work, competently executed and presented, which significantly advances the field.

Weaknesses:

None obvious, however the presentation and especially main-text illustrative material seems to focus disproportionately on MacAB-TolC-YbjP complex, and the AcrABZ-TolC-YbjP is relegated to supplementary data which is somewhat confusing. There is no high-resolution side view of the AcrABZ-TolC-YbjP side-by-side to MacAB-TolC-YbjP which may be helpful to spot parallels and differences in the organisation of the two systems.

Supplementary Figure 2 may also be better presented in the main text, as it shows specific displacements of residues upon binding of the YbjP relative to the apo-complexes, although this can be left at the authors' discretion.

Reviewer #2 (Public review):

This article focuses on the study of two E. coli tripartite efflux pumps both using TolC as partner in the outer membrane, namely MacAB-TolC and AcrABZ-TolC.

By preparing MacAB-TolC in Peptidiscs rather than in detergent for cryo-EM structure determination, they visualized an extra protein localized around TolC. The resolution was sufficient to build part of the structure, and using the AlphaFold2 database and DALI topology recognition program, they identified it as the lipoprotein YbjP. This protein has an anchorage in the outer membrane, and it was suggested that it could act as a support for TolC that is the only OMF that does not have an N-terminal extension anchored in the outer membrane, which is very puzzling for the community working in this field of research.

Authors used a large number of different approaches to evaluate the importance of YbjP (structure, genomic evolution, microbiology, photocrosslink in vivo, proteomic profile), but did not succeed in finding it a clear role so far, even if it could be important depending on environmental stress. Nevertheless, their results are of main interest for the comprehension of the complexity of such systems and deserve publication.

The different analyses are properly performed and presented, and support the conclusions.

My only concern is for the photocrosslink presented in Figures 3 and S3. My impression is that the bands do not migrate at the proper size after the crosslink.

A second point that could be discussed further is the comparison of the structure of the pump in the presence of the peptidoglycan with the images previously obtained by tomography. It is not totally clear to me if YbjP could have been positioned in these maps.

eLife Assessment

This useful study presents a new method to identify the activity of single motor units from intramuscular EMG recordings. Validation against state-of-the-art techniques is limited to a small sample of simulated motor units; consequently, the evidence supporting the method's accuracy remains incomplete. The manuscript would be significantly strengthened by using more unbiased simulations for validation, validating the method with experimental datasets, comparing it against more recent techniques, and investigating how muscle physiology impacts accuracy.

Reviewer #1 (Public review):

Summary:

The authors introduce EMUsort, an open-source algorithm for the automatic decomposition of high-resolution intramuscular EMG recordings. The method builds upon the Kilosort4 framework and incorporates modifications designed to better handle the spatial and temporal characteristics of intramuscular signals. The performance of EMUsort is evaluated on openly available datasets and compared against KS4 and MUEdit, demonstrating improved motor unit accuracy.

Strengths:

(1) The manuscript is clearly written, technically detailed, and well structured.

(2) The open-source software is thoroughly documented, both within the manuscript and in the accompanying repository README, facilitating adoption by the community.

(3) The availability of both code and datasets is a major strength, enabling reproducibility and independent validation.

(4) The authors provide quantitative comparisons with existing decomposition algorithms, which is essential for contextualizing the proposed method.

(5) The methodological details are sufficiently described to allow replication and further development by other researchers.

Weaknesses:

While the manuscript is strong overall, I have several suggestions that could further strengthen its impact and clarity.

(1) Benchmarking and community integration

A recent work has proposed standardized datasets and benchmarking pipelines for high-density surface EMG decomposition ("MUniverse: A Simulation and Benchmarking Suite for Motor Unit Decomposition", Mamidanna*, Klotz*, Halatsis* et al, NeurIPS 2025). A similar effort for intramuscular EMG would be highly valuable to the field. The authors may consider discussing how their dataset and algorithm could be integrated into broader benchmarking initiatives (e.g., platforms such as MUniverse), enabling systematic comparisons across multiple datasets and decomposition methods.

(2) Comparison with additional decomposition algorithms

Since the manuscript compares EMUsort with MUEdit, it would be appropriate to also include a comparison with Swarm-Contrastive Decomposition (SCD), which has been proposed for both surface and intramuscular EMG signals. Including this comparison, or explicitly discussing why it was not feasible, would strengthen the positioning of EMUsort relative to the current state of the art.

(3) Manual editing and post-processing

In practical EMG decomposition workflows, manual inspection and editing of motor units are often required after automatic decomposition. It would be useful for readers to know whether EMUsort provides (or is compatible with) a graphical interface or workflow for manual refinement, or how the authors envision this step being handled.

(4) Ablation analysis of algorithmic modifications

EMUsort is described as an extension of Kilosort4. An ablation analysis examining the impact of the main modifications introduced relative to KS4 would help clarify which changes contribute most to the observed performance improvements and under which conditions.

(5) Failure modes and limitations

A more explicit discussion of when EMUsort is likely to fail or degrade in performance would be valuable. For example, sensitivity to the number of channels, recording duration, signal quality, or motor unit density could be discussed to guide users.

(6) Generalisability to surface EMG

Given the shared methodological foundations between surface and intramuscular EMG decomposition, it would be helpful to know whether EMUsort has been tested on high-density surface EMG datasets or whether the authors expect limitations when applied outside the intramuscular domain.

(7) Applicability to human intramuscular recordings

The authors could clarify whether EMUsort has been tested on human intramuscular EMG, and discuss any expected differences in performance due to anatomical or physiological factors.

(8) Parameter sensitivity

Clustering-based methods can be sensitive to parameter choices. Reporting a parameter sensitivity analysis, or at least discussing the robustness of EMUsort to parameter variations, would increase confidence in the method's reliability and ease of use.

(9) Differences between template matching and BSS methods

Since the manuscript proposes a new template matching algorithm, but it compares its performance with a BSS one (MUedit), BSS algorithms should be described in the introduction. The differences between the methodologies should be highlighted, and the pros and cons of each described.

Conclusion:

The authors largely achieve their stated aims, and the results mostly support the main conclusions. EMUsort represents a meaningful contribution to the EMG decomposition literature, particularly for researchers working with high-resolution intramuscular recordings. With additional clarification regarding benchmarking, algorithmic ablations, and limitations, the manuscript would be further strengthened and likely to have a substantial impact on the field.

Reviewer #2 (Public review):

Summary:

This work presents a new spike sorter, EMUsort, to target the challenging task of spike sorting Motor Unit Action Potentials (MUAP). EMUsort is essentially a modified version of Kilosort, with some key extensions to target EMG data: correct for large delays due to propagation across channels, spike detection of highly overlapping and large units via multiple thresholds, an increased number of waveform templates for spike detection, and an extended representation of waveforms to grasp complex MUAP spike shapes. The results on simulated data show solid evidence that the applied modifications make a difference for EMG recordings. All in all, I believe that EMUsort will greatly improve spike sorting performance for high-density EMG data.

Strengths:

The manuscript is well written, and the methods and modifications to the Kilosort pipeline are well-motivated, well-explained, and clear. The simulation results provide strong evidence that the presented modifications make spike sorting of high-density EMG data more accurate.

Weaknesses:

The method is overall only validated on 15 simulated motor units. The monkey dataset, in particular, seems too "easy" and not challenging enough to highlight weaknesses of any of the spike sorters. A second weakness is in the distribution of the code, which is shipped with submodules for Kilosort and SpikeInterface, and makes it hard to maintain long-term, and pins to old versions of these key dependencies.

Reviewer #3 (Public review):

Summary

This paper introduces EMUsort, an extension of Kilosort4 designed to sort motor unit action potentials from high-density intramuscular EMG recordings. Using rat and monkey forelimb recordings, the authors generate realistic simulated datasets with known ground truth and demonstrate that EMUsort substantially outperforms Kilosort4 and MUedit, particularly during periods of high motor unit overlap.

Strengths

This is a timely study in light of recent advances in intramuscular muscle recording technologies and the growing interest in automated methods for decoding neural and neuromuscular signals. The work leverages state-of-the-art electrode arrays and combines them with advanced signal processing tools to address a challenging and relevant problem in motor unit analysis.

Weaknesses

There are several aspects of the study that substantially limit the interpretation of the main results and conclusions. The following major points should be carefully considered by the authors.

(1) Choice of experimental model and validation framework: The study aims to validate a new methodology for EMG decomposition, yet the rationale for the chosen experimental models is unclear. Specifically, it is not evident why the authors focused on intramuscular recordings from two animal models performing dynamic tasks. Given the extensive literature on the development and validation of EMG decomposition methods, the choice of an experimental design that substantially deviates from established approaches is insufficiently justified. In particular, it is unclear why the authors did not consider more standard validation paradigms based on (i) isometric contractions, (ii) human data, (iii) surface EMG recordings, or (iv) combinations of their recording technologies with previously validated motor unit identification methods. This methodological divergence makes it difficult to interpret the findings in the context of existing evidence.

(2) Lack of manual EMG decomposition as reference: Related to the previous point, it is unclear why standard manual EMG decomposition methods were not used to generate reference datasets for validation. Manual decomposition has been shown to be reliable under specific conditions (low contraction levels, slow dynamics, etc.) and would have substantially strengthened the validation of the proposed algorithm.

(3) Neglect of muscle deformation effects: While the manuscript discusses several factors that complicate EMG decomposition relative to brain recordings, it does not address the well-known effects of muscle deformation during contractions on motor unit action potential shapes. There is extensive literature demonstrating that dynamic muscle contractions lead to systematic changes in action potential morphology, representing a major challenge for EMG decomposition and a fundamental difference from brain recordings. Additionally, even small relative movements of intramuscular electrodes can produce waveform changes that are large relative to muscle fiber dimensions. These issues are particularly relevant given the highly dynamic tasks studied here (e.g., treadmill walking in rats), yet they are not discussed or incorporated into the analysis.

(4) Exclusive reliance on simulated data for validation: The primary validation of EMUsort is based on simulated data, which represents a major limitation of the study. This reliance should be clearly and explicitly stated in the abstract, introduction, and discussion. Moreover, the simulation approach itself raises concerns. The simulated EMG signals are generated using templates derived from the same sorting framework being validated, which introduces a potential methodological bias. The linear combination of components used to synthesize waveforms constitutes an unjustified modeling assumption that may favor template-based approaches such as EMUsort. Additionally, the spike time generation procedure appears unnecessarily complex and insufficiently justified. Previous validation studies typically modeled motor units as firing at relatively stable levels along their recruitment curves, producing long spike trains with pseudo-random relative timing and diverse overlap conditions. This framework would likely provide a more robust and interpretable validation. If the authors believe their simulation approach is superior, a stronger justification is required. Finally, the limited number of simulated motor units is difficult to reconcile with the expected level of motor unit recruitment during the studied behaviors, and this choice is not adequately justified.

(5) Incomplete reporting and visualization of experimental data: The manuscript would benefit from a clearer description of the number of rats and monkeys used, which should be reported explicitly in the abstract. In addition, visualizations of the raw multichannel EMG data across different task phases and activation levels would substantially improve transparency. Providing comprehensive visualizations of motor unit action potential shapes across all channels and identified units (for both rats and monkeys) would also help readers assess the spatiotemporal features that underpin unit identification and sorting reliability.

(6) Physiological limitations of conduction delay correction: The proposed method for correcting conduction delays across channels is physiologically suboptimal. First, motor unit conduction velocities differ substantially across units, implying that delay correction should be applied at the unit level rather than uniformly across channels. Second, conduction delays depend on fiber orientation and distance relative to electrode geometry; if fibers are oriented at different angles with respect to the array, a single delay correction becomes invalid. Additionally, the schematic in Figure 2A appears to contradict the proposed correction approach: if electrode threads are arranged perpendicular to muscle fibers, conduction delays across channels within a single thread should be minimal.

(7) Clarity issues in Figure 4: Figure 4 (panels A-D) is potentially misleading. It should be clearly stated whether the signals shown are artificial examples or derived from real recordings; ideally, real data should be used to illustrate the advantages of dynamic thresholds. In panels B-D, the depiction of overlapping action potentials is difficult to interpret due to the thickness of the traces, and it is unclear whether overlaps with neighboring action potentials are absent by design or expected to occur in real data. If overlaps are expected, one would also expect to observe contamination in the extracted waveforms, which is not evident in the figure.

(8) Concerns regarding method comparisons: The comparison with existing methods raises methodological concerns. It appears that EMUsort was carefully optimized, whereas the competing algorithms were not equivalently fine-tuned. The literature clearly shows that EMG decomposition performance depends strongly on adapting algorithms to the signal type (intramuscular vs. surface, species, electrode geometry). Furthermore, it is surprising that MUedit is reported to perform particularly poorly during periods of motor unit overlap, as blind source separation methods were explicitly developed to handle convolutive mixtures and overlapping sources, especially in surface EMG (which is an extreme case of motor unit overlapping). This discrepancy requires further explanation.

(9) Insufficient characterization of motor unit firing properties: The study does not provide sufficient information about the firing characteristics of the identified motor units in experimental data. Relevant metrics that should be reported include average, minimum, and maximum firing rates; coefficients of variation of discharge rate; signal-to-noise ratios of motor unit action potentials; potential evidence of motor unit rotation over time; and stability of firing behavior across recording intervals.

(10) Lack of theoretical framing: Given the scope and claims of the paper, it would be valuable to include a more theory-driven introduction explaining why different sorting approaches (e.g., template matching vs. blind source separation) may be more or less suitable depending on the nature of the recorded signals. A clearer conceptual rationale for why the proposed approach is expected to outperform existing methods would substantially strengthen the manuscript.

(11) Limitations of validation metrics: Some of the metrics used to evaluate performance are not ideal. For example, reporting 0% accuracy for a unit is misleading and should instead be described as a failure to identify that unit. Similarly, comparisons of total spike counts are of limited interpretive value and may be misleading, as correct spike counts do not necessarily imply correct spike identities.

(12) Clarification of computational performance claims: Finally, the discussion of computation times should clarify that some existing methods require substantial time for offline estimation of projection vectors but can operate in near real time once these vectors are learned and remain stable. This distinction is important for a fair comparison of practical usability.

eLife Assessment

This modeling study proposes important new insights into the circuit mechanisms underlying navigational control in insects. High-speed video recordings of ants are compared to detailed predictions from a new computational model, whose description is incomplete. If the model is sound, the similarities between the model and behavioral data suggest how complex behavioral motifs can emerge from a simple neural circuit. These results will be of interest to scientists studying the neural circuit basis of behavior, particularly in insects.

Reviewer #1 (Public review):

Summary:

Freas and Wystrach present a computational model of steering in insects. In this model, the central complex provides an error signal indicating the animal should turn left or right; this error signal biases the function of an oscillator composed of two mutually inhibiting self-exciting units. The output of these units generates a "steering signal" that is used both to set the direction and speed of the ant. Additionally, a separate module induces pauses, and an inverse relation between forward speed and turning speed is externally imposed. Statistics of the trajectories generated by the model are compared to the measured behaviors of ants.

Strengths:

While the model is very simple compared to state-of-the-art models, that simplicity makes it a potentially useful guide to researchers studying insect navigation. Some predictions that emerge from the model appear to be experimentally testable, although a more complete description of the model and its parameters, as well as an analysis of how this model's predictions differ from previous models' predictions, would be required to design these experiments.

Weaknesses:

I found it difficult to identify evidence in the paper supporting central elements of the abstract. Hopefully, these difficulties can be resolved with a clearer presentation and the addition of supporting detail, especially in the methods.

(1) The model is not clearly described

In the Materials and Methods, there is no description of the model, just "The computational model is presented in Figure 1." (This is probably a typo and may refer to Figure 2A-C), and a link to Matlab source code. It is inappropriate to ask readers or reviewers to examine source code in lieu of providing a method, but I attempted to do so anyway. To my eye, the source code does not match the model presented in 2A-C. For instance, in 2C, "Steering signal" inhibits "Freeze", but I couldn't find this in the source. "Freeze" is shown to inhibit "steering signal," but as "steering signal" is a signed quantity, it's not clear what this means. Literally, since "ang_speed_raw = L-R," it would seem to indicate the "freeze" would bias towards right turns. In the code, "freeze" appears to be implemented through the boolean variable "speed_inhibition_time." The logic controlled by this variable doesn't appear to inhibit the "steering signal" but instead (depending on control parameters) either reduces the movement speed and amplifies the turning rate, or it turns the angular speed output into a temporal integral of the control signal.

There are a number of parameters in the source code that aren't described at all in the paper, including the internal oscillator parameters.

Together, these limitations make it difficult to understand what is being simulated, what parts of the model are tied to biology, and where the model improves on or departs from previous work.

It is absolutely essential that authors fully describe the computational model, that they explain the meaning of all parameters of the model, and that they explain how the particular values of these parameters were chosen.

(2) The biological inspiration is unclear

A central claim of the paper is that the model is "biologically grounded." But some elements, for instance, using a signed quantity to represent left-right steering drive, are not biologically possible; at best, these are shorthand for biologically possible implementations, e.g., opposing groups of left-right driving neurons.

The mechanism that produces fixations and saccades - the "freeze" module - is not tied to any particular anatomy of the insect brain. Initiation of a freeze occurs at a specific time coded into the model by the authors; it is not generated by an internal model signal. Release of a freeze is by drawing a random variable; there is no neural mechanism proposed to generate this signal.

In some versions of the model, instead of directly controlling the signal, during fixations, the angular drive signal is integrated into a variable "cumul_drive." No neural substrate is proposed for this integrator. In the code, if cumul_drive passes a threshold, the angular heading of the ant changes (saccades), but only if this threshold is passed before the Poisson process ends the fixation. No neural substrate is proposed for any of this logic.

The model steps forward in time by a fixed increment - the actual duration (in seconds) of this time step is not specified. From Figure 4F, G, it appears a simulation time step is meant to be about 10ms. This would imply an oscillator frequency of about 2 Hz (Fig 2B), that the heading oscillates at a similar frequency (2G), and that a forward crawling ant stops moving every 500 ms (2I). Are these plausible? Can they be compared to an experiment?

Model parameters, including the ones that control the frequency of the oscillator, are non-dimensionalized. It is not possible to evaluate whether these parameters are biologically plausible or match experimental results.

(3) Claims that behaviors emerge from the model may be overstated

The abstract claims that steering correction and fixations/saccades emerge naturally from the same model. But it appears to me that fixations/saccades are externally imposed by the specification of specific times for a "freeze." Faster angular rotation during saccades than during course correction is imposed and does not emerge naturally from neural simulations.

(4) Citations to previous literature are difficult to follow, and modeling results are presented as though they are experimental data

I would ask the authors to be much clearer in their description and citation of previous work. It should be clear whether the cited work was experimental or computational. To the extent possible, the actual measurement should be described succinctly. Instead of grouping references together to support a sentence with multiple claims, references should be cited for each claim. Studies of computational models should not be presented as proving a biological result.

For example:

a) Lines 141-146:<br /> "Previous studies have established many key components of insect navigation, including .... the intrinsic oscillatory dynamics in the lateral accessory lobes (LALs) that support continuous zigzagging locomotion (Clément et al., 2023; Kanzaki, 2005; Namiki and Kanzaki, 2016; Steinbeck et al., 2020)."

The first reference is to one author's previous modeling work - it hypothesizes that oscillations in the LAL support zigzagging but includes no data that would "establish" the fact. Kanzaki et al. 2005 describes numerical modeling and simulation with a physical robot. Namiki and Kanzaki, 2016 is a review article that links the LAL to zigzagging behavior. It describes the LAL as a winner-take-all bistable network but does not describe or hypothesize that the LAL has intrinsic oscillatory dynamics. Steinbeck et al. 2020 is a more comprehensive review; it reinforces that the LAL is a winner-take-all bistable network that drives left-right steering, including during zig-zagging behavior. But in my reading, I could not find a statement that the LAL has intrinsic oscillatory dynamics (the closest is Steinbeck et al. saying the activity pattern switches regularly, as does the behavior; this doesn't imply that the LAL is intrinsically oscillatory.)

b) Lines 701-703:<br /> "In plume-tracking moths, CX output has been shown to modulate LAL flip-flop neurons driving zigzagging (Adden et al., 2022)."

This reads as though an experimental measurement was made, but in fact, this is modeling work.

c) Lines 703-706:<br /> "In ants, strong goal signals in the CX - whether elicited by the path integrator or visual familiarity (Wehner et al., 2016; Wystrach et al., 2020b, 2015) do not only sharpen directional accuracy but also increase oscillation frequency (Clément et al., 2023)."

Here again, modeling results are presented as though they were experimental data.

Reviewer #2 (Public review):

Summary:

The paper by Freas and Wystrach is an interesting computational study, exploring the detailed mechanisms of how simple neural circuits could explain complex behavioral patterns observed in navigating ants. The authors compare detailed, high-speed video recordings of Australian desert ants (Melophorus bagoti) with predictions made by their new computational model and find convincing similarities between the model and the behavioral data, at a level of detail not previously studied. Particularly interesting are emerging properties of the model, yielding behavioral motifs it was not designed to reproduce, but which occur in natural ant behavior.

Strengths:

A strength of the study is that the model is based on previous models, without making major novel explicit assumptions. It combines existing models of the insect central complex with a model of the lateral accessory lobe and adds a stochastic inhibition of forward velocity to the interaction of central complex and lateral accessory lobes. The central complex provides corrective steering signals when the goal direction and the current heading of an insect are not aligned, while the lateral accessory lobes provide an intrinsic oscillator underlying the behavioral oscillations shown by walking ants at all times. These background oscillations are modulated by the steering signals from the central complex. Depending on which phase of the intrinsic oscillations coincides with the corrective signals, and how fast the ant is moving forward during this time, a complex set of behaviors emerges. Most prominently, scanning behaviors, which are regularly carried out by the ants, are recapitulated in great detail by the model. Additionally, other behaviors, such as full loops, emerge naturally from the model. While computational models are not to be seen as definite evidence for any biological reality, they can provide strong support for particular neural implementations. The current study is an excellent example in that it provides evidence for a serial arrangement of central complex circuits upstream of the lateral accessory lobe circuits, modulated by speed-regulating input. While the latter is hypothetical, it yields a clear hypothesis that can be validated by connectomics studies and functional work in the future.

The study shows that even complex behavioral motifs do not require dedicated neural modules, but can rather emerge from the interplay of already known circuits - highlighting the efficiency of insect brains and possibly providing the path towards embodied hardware solutions of such circuits in autonomous agents.

Weaknesses:

There are several weaknesses in the paper as it is.

Firstly, the model is not described in the methods, but only found when following the link to the authors' GitHub repository. This is clearly not sufficient and prevents readers from evaluating the model's assumptions directly. Most importantly, how natural do the emerging properties indeed emerge from the model? What parameters need to be tuned to generate a match between data and model?

Second, it is often not entirely clear what is biological data and what is a computational model. This relates to figures, text, and references. As a reader, this makes it difficult to clearly judge what is new in the current paper, how it adds to previous models, and what the predictions and assumptions are for biology.

Third, while neural data from bees and flies are taken to motivate and design the computational model, the discussion and interpretation revolve almost exclusively around ants. For the most part, this is justified, as the behavioral data used to benchmark the model are taken from ants. Nevertheless, more broadly discussing the newly defined circuit in the context of flying insects would give a better idea of the broad relevance of the neural circuits predicted by the model.

eLife Assessment

This study provides important insights into the crosstalk between ATG2A with components of the early secretory pathway. While the mechanisms governing autophagic membrane expansion remain yet to be fully understood, in this study the authors employ an elegant proximity labelling approach and identify two ER-Golgi intermediate compartment (ERGIC)-localized proteins. Through a series of complementary experiments combining microscopy and biochemistry, the authors identify ARFGAP1 and Rab1A as components of early autophagic membranes, which accumulate at the periphery of pre-autophagosomal structures induced by loss of ATG2. The overall study is well executed and the evidence supporting the claims is convincing.

Reviewer #1 (Public review):

Summary:

D. Fuller et al. set out to study the molecular partners that cooperate with ATG2A, a lipid transfer protein essential for phagophore elongation, during the process of autophagy. Through a series of experiments combining microscopy and biochemistry, the authors identify ARFGAP1 and Rab1A as components of early autophagic membranes, which accumulate at the periphery of aberrant pre-autophagosomal structures induced by loss of ATG2. While ARFGAP1 has no apparent function in autophagy, the authors show that RAB1A is implicated in autophagy, although the precise mechanisms are not explored in the manuscript.

Strengths:

The work presented by Fuller et al. provides new insights into the composition of early autophagic membranes. The authors provide a series of MS experiments identifying proteins in close proximity to ATG2A, which is a valuable dataset for the field. Furthermore, they show for the first time the interaction between ATG2A and RAB1A both in fed and starved conditions, which extends the characterisation of the pre-autophagosomal structures observed in ATG2 DKO cells.

Weaknesses / Specific comments:

(1) The authors claim that Rab1A/B knockdown phenocopies the LC3-II accumulation observed in ATG2 DKO cells. While LC3-II accumulation is consistent with this interpretation, depletion of many autophagy-related proteins can give rise to a similar phenotype, even when they function at distinct stages of the autophagic cascade. Therefore, LC3-II accumulation alone is insufficient to support phenocopying in my vew. Immunofluorescence analyses demonstrating comparable cellular phenotypes-such as membrane accumulation of pre-autophagosomal structures-following Rab1 knockdown should be provided. Moreover, p62 does not accumulate upon Rab1 depletion, suggesting that loss of Rab1 does not fully phenocopy ATG2 deficiency. Consequently, it remains unclear whether Rab1A depletion truly phenocopies ATG2A depletion with respect to autophagy progression or the accumulation of pre-autophagosomal structures.

(2) Interpretation of the significance of the data

(2.1) The significance statement asserts that "this study elucidates the role of early secretory membranes in autophagosome biogenesis." While the data convincingly demonstrate an association between the RAB1A GTPase and ATG2A, the study does not provide mechanistic insight into how this interaction functionally contributes to autophagy. As presented, the findings support a correlative relationship rather than a defined role in autophagosome biogenesis.

(2.2) The title states that ATG2A "engages" Rab1A- and ARFGAP1-positive membranes during autophagosome formation. However, both Rab1A and ARFGAP1 are shown to localize to pre-autophagosomal structures independently of ATG2A. In the absence of evidence demonstrating a functional or causal dependency, the term "engages" appears overstated. A more descriptive term, such as "associates," would more accurately reflect the data.

(2.3) In the Discussion, the authors state that previous studies have reported increased LC3-II levels following knockdown of Rab1 proteins (refs. 38 and 49). However, it is unclear where this observation is documented in the cited references.

(3) Some concerns remain in specific figures, as outlined below:<br /> • Quantification is missing in Fig S2D.<br /> • The authors claim: "siRNA against ARFGAP1 had very little effect" but the quantification and blots show actually no effect.<br /> • Conclusions drawn from KD experiments in Fig. S2 should be interpreted with caution, as knockdown efficiency is very low, particularly for ARFGAP1/3 in the triple knockdown.<br /> • In New Fig. 4, the representative blot is not representative of the results showed in the quantification as previously noted.

Reviewer #2 (Public review):

The mechanisms governing autophagic membrane expansion remain incompletely understood. ATG2 is known to function as a lipid transfer protein critical for this process; however, how ATG2 is coordinated with the broader autophagic machinery and endomembrane systems has remained elusive. In this study, the authors employ an elegant proximity labeling approach and identify two ER-Golgi intermediate compartment (ERGIC)-localized proteins-Rab1 and ARFGAP1-as novel regulators of ATG2 during autophagic membrane expansion.

Their findings support a model in which autophagosome formation occurs within a specialized subdomain of the ER that is enriched in both ER exit sites (ERES) and ERGIC, providing valuable mechanistic insight. The overall study is well executed and offers an important contribution to our understanding of autophagy. I support its publication in eLife and offer the following minor comments for clarification and improvement.

Specific Comments

(1) Integration with Prior Literature<br /> The data convincingly implicate the ERES-ERGIC interface in autophagosome biogenesis. It would strengthen the manuscript to discuss previous studies reporting ERES and ERGIC remodeling and formation of ERERS-ERGIC contact sites (PMID: 34561617; PMID: 28754694) in the context of the current findings.

(2) Figure Labeling<br /> The font size in Figure 1A and Supplementary Figure S1G is too small for comfortable reading. Please consider enlarging the labels to improve clarity.

(3) Experimental Conditions<br /> In Figures 2A-C and Figure 4, it is unclear how the cells were treated. Were they starved in EBSS? Please include this information in the corresponding figure legends.

(4) LC3 Lipidation vs. Cleavage<br /> In Figure 2A, ARFGAP1 knockdown appears to reduce LC3 lipidation without affecting Halo-LC3 cleavage. Clarifying this observation would help readers better understand the functional specificity of ARFGAP1 in the pathway.

(5) Use of HT-mGFP in Figure 2C<br /> Please clarify why the assay in Figure 2C was performed in the presence of HT-mGFP. Explaining the rationale would aid interpretation of the results.

(6) FIB-SEM Imaging<br /> For the FIB-SEM images in Figures 3 and S3, directly labeling the cellular structures in the images would greatly facilitate interpretation for the reader.

(7) Supplementary Figures<br /> Many of the supplemental figures are high quality and contain key data. If space permits, I suggest moving these into the main figures. In particular, the FLASH-PAINT experiment could be presented as part of Figure 1.

(8) Text Revision for Clarity<br /> In line 242, the phrase "but protein-protein interactions appear to be limited to RAB1" would benefit from clarification. A more precise formulation could be: "but stable protein-protein interactions appear to be limited to RAB1."

(9) COPII Inhibition Strategy<br /> The authors used the dominant-active SAR1(H79G) mutant to inhibit COPII function. While this is effective in in vitro budding assays, the GDP-locked mutant SAR1(T39N) has been shown to be more effective in blocking COPII-mediated trafficking in cells. Including SAR1(T39N) in the analysis would provide stronger support for the conclusions.

Reviewer #3 (Public review):

The manuscript by Fuller et al describes a crosstalk between ARTG2A with components of the early secretory pathway, namely RAB1A and ARFGAP1. They show that ATG2A is recruited to membranes positive for RAB1A, which they also show to interact with ATG2A. In agreement with earlier findings by other groups, silencing RAB1A negatively affects autophagy. While ARFGAP1 was also found on ATG2A positive membranes, silencing ARFGAP1 had no impact autophagy. Notably, these ARFGAP1 positive membranes are not Golgi membranes.

The findings are interesting and the data are in general of good quality. I think the story is good enough to be published in eLife and I have the following questions, which the authors may attend to:

(1) Are the membranes to which ATG2A is recruited a form of ERGIC?

(2) Figure 3A/B: Is it possible to show a better example? The difference is barely detectable by eye. Since Immunoblotting is not really a quantitative method, I think that such a weak effect is prone to be wrong. Is there another tool/assay to validate this result?

(3) Is the curvature-sensitive region of ARFGAP1 required for its co-localization with ATG2A?

(4) What does Rab1A do? What is its effector? Or does the GTPase itself remodel the membrane?

(5) What about Arf1? It appears that this role of ARFGAP1 is unrelated to Arf1 and COPI? Thus, one would predict that Arf1 does not localize to these structures and does not affect ATG2A function

(6) Does ARFGAP1 promote fission of the membrane from its donor compartment?

(7) What are ARFGAP1 and Rab1A recruited to? What is the lipid composition, or protein that recruits these two players to regulate autophagy?

Comments on the latest version:

The revisions carried out by the authors are fine. The new data on ArfGAP1 and about the indirectness of the ATG2A and Rab1A interaction improve both clarity and strength of the manuscript. I have no further comments.

Author response:

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

We thank the reviewers and editors for their thoughtful comments, which substantially improved the quality and clarity of our manuscript. We have attempted to address each major concern with either new experiments or significant textual revisions.

Reviewer 1 noted that “this research is conducted exclusively in HEK293 cells… including at least one additional cell line would significantly strengthen the main findings.” To directly address this concern, we repeated our RAB1A/B double-knockdown experiments in H4 neuroglioma cells, which endogenously express a tandem fluorescent-tagged LC3B reporter. Using flow cytometry to quantify autophagic flux, we confirmed that RAB1 depletion in H4 cells recapitulates the flux defects observed in HEK293 cells, thereby validating the generality of our findings across distinct lineages.

To validate the robustness of the ATG2 DKO phenotype and the localization of ARFGAP1-positive membranes, we acquired an ATG2 double knockout HeLa cell line. We confirmed the presence of the characteristic large ATG2-deficient PAS compartment in HeLa cells, and the recruitment of ARFGAP1 membranes, but note that ARFGAP1 displays a solid distribution through the compartment in these cells, in contrast to the more peripheral enrichment observed in HEK293 cells. These data are now included and discussed in the revised manuscript.

Multiple reviewers asked for greater clarity around the interaction between ATG2A and RAB1A. Although our original data showed that these proteins co-immunoprecipitate in cells, we had not established whether their association was direct. In response, we attempted in vitro co-immunoprecipitations from purified components.  As we could not detect interactions in this simplified system, we now speculate that the ATG2A–RAB1A interaction is indirect. This clarification is now incorporated into the results section.

Multiple reviewers also raised questions regarding the nature of the membranes recruiting ARFGAP1 and the potential relationship to Arf1 and Golgi trafficking. In particular, Reviewer 3 asked: “(5) What about Arf1? … one would predict that Arf1 does not localize to these structures and does not affect ATG2A function.” To examine whether ARFGAP1 recruitment depends on Golgi integrity or Arf1-regulated trafficking, we perturbed the Golgi using three mechanistically distinct methods: Brefeldin A, mitotic entry, and SidM expression, each of which dissolves Golgi architecture. In each condition, ARFGAP1 localization to the enlarged PAS compartment in ATG2 DKO cells was unchanged. These results indicate that ARFGAP1 recruitment is independent of Golgi structure and provide indirect support for the notion that Arf1 does not participate in this process. Reviewer 3 also asked: “Is the curvature-sensitive region of ARFGAP1 required for its co-localization with ATG2A?” To address this, we generated ARFGAP1 mutants lacking either GAP catalytic activity or the ALPS curvature-sensing domain. When expressed in ATG2 DKO cells, all mutants retained full recruitment to the PAS compartment. Thus, neither GAP activity nor ALPS-mediated curvature sensing is required for ARFGAP1 localization in this context.

Response to Reviewer 3 -“(2) Figure 3A/B: … is there another tool/assay to validate this result?”—we quantified autophagic flux following SAR1B(H79G) overexpression using the flow-cytometry tandem-fluorescent LC3 assay. These experiments confirmed that SAR1B(H79G) causes only a modest reduction in autophagic flux, consistent with partial inhibition of COPII, thereby supporting our original interpretation.

We also took steps to improve the integration of our findings with prior literature. Reviewer 2 requested that we strengthen the manuscript by incorporating studies on ERES–ERGIC remodeling (“It would strengthen the manuscript to discuss previous studies…”). We now cite and discuss the studies corresponding to PMIDs 34561617 and 28754694, aligning our observations with mechanistic models of early secretory pathway remodeling. More broadly, Reviewer 1 commented that our discussion “overlooks some important aspects,” and Reviewer 3 asked, “Are the membranes to which ATG2A is recruited a form of ERGIC?” In response, we substantially rewrote the discussion, expanding our integration of existing literature and explicitly addressing models in which ATG2A acts at an ERGIC-derived membrane.

eLife Assessment

This study presents valuable findings on the ability of a state-of-the-art method, Temporally Delayed Linear Modelling (TDLM), to detect the replay of sequences in human memory. The investigation provides compelling evidence that TDLM has significant limitations in its sensitivity to detect replay in extended (minutes-long) rest periods. The work will be of strong interest to researchers investigating memory reactivation in humans, especially using iEEG, MEG, and EEG.

Reviewer #1 (Public review):

Summary:

Participants learned a graph-based representation, but, contrary to the hypotheses, failed to show neural replay shortly after. This prompted a critical inquiry into temporally delayed linear modeling (TDLM)--the algorithm used to find replay. First, it was found that TDLM detects replay only at implausible numbers of replay events per second. Second, it detects replay-to-cognition correlations only at implausible densities. Third, there are concerning baseline shifts in sequenceness across participants. Fourth, spurious sequences arise in control conditions without a ground truth signal. Fifth, the revised manuscript adapts a previously published synthetic simulation to show that previous validations/support of TDLM may have overestimated TDLM sensitivity because synthetic assumptions can produce unrealistically high pattern separability and reduced baseline confounds.

Strengths:

- This work is meticulous and meets a high standard of transparency and open science, with preregistration, code and data sharing, external resources such as a GUI with the task and material for the public.

- The writing is clear, balanced, and matter-of-fact.

- By injecting visually evoked empirical data into the simulation, many surface-level problems are avoided, such as biological plausibility and questions of signal-to-noise ratio.

- The investigation of sequenceness-to-cognition correlations is an especially useful add-on because much of the previous work uses this to make key claims about replay as a mechanism.

- In the revised version, the authors foreshadow ways to improve sequenceness detection by introducing a sign-flipping analysis.

Weaknesses:

Many of the weaknesses are not so much flaws in the analyses, but shortcomings when it comes to interpretation and a lack of making these findings as useful as they could be. Furthermore, as I will explain below, some weaknesses have been partially improved in the last round of revisions.

- I found the bigger picture analysis to be lacking, though improved in the latest version. Let us take stock: in other work during active cognition, including at least one study from the Authors, TDLM shows significant sequenceness. But the evidence provided here suggests that even very strong localizer patterns injected into the data cannot be detected as replay except at implausible speeds. How can both of these things be true? Assuming these analyses are cogent, do these findings not imply something more destructive about all studies that found positive results with TDLM? In the revisions, the manuscript concentrates a bit more on criteria that influence detection of sequences, though it is still not entirely clear what consequences there are for previous work.

- All things considered, TDLM seems like a fairly vanilla and low assumption algorithm for finding event sequences. Although the authors have improved their discussion of "boundary conditions" or factors for why TDLM might fail, it remains not fully clear to what extent the core problem is TDLM on an algorithmic/mathematical level (intrinsic factor), vs data quality, power, window size (extrinsic factors).

- The new sign-flip analysis underscores the authors' goal of being solution-oriented, though it is important to emphasize that a comprehensive way forward is not yet provided. This is fine, but the manuscript could be improved further through a concrete alternative or a revised version of the original approach.

Reviewer #2 (Public review):

Summary:

Kern et al. investigated whether temporally delayed linear modeling (TDLM) can uncover sequential memory replay from a graph-learning task in human MEG during an 8 minute post-learning rest period. After failing to detect replay events, they conduct a simulation study in which they insert synthetic replay events, derived from each participants' localizer data, into a control rest period prior to learning. The simulations suggest that TDLM only reveals sequences when replay occurs at very high densities (> 80 per minute) and that individual differences in baseline sequenceness may lead to spurious and/or lacklustre correlations between replay strength and behavior.

Strengths:

The approach is extremely well documented and rigorous. The authors have done an excellent job re-creating the TDLM methodology that is most commonly used, reporting the different approaches and parameters that they used, and reporting their preregistrations. The hybrid simulation study is creative and provides a new way to assess the efficacy of replay decoding methods, and its comparison to earlier published TDLM simulations is particularly useful. The authors remain measured in the scope/applicability of their conclusions, constructive in their discussion, and end with a useful set of recommendations for how to best apply TDLM in future studies. I also want to commend this work for not only presenting a null result, but thoroughly exploring the conditions under which such a null result is expected. I think this paper is interesting and will be generally quite useful for the field.

In the revised version, the authors have adequately addressed each of the weaknesses I raised previously. In brief, they:

(i) Added new power analyses of sequenceness for bootstrapped sample sizes, along with a new permutation test (Supplemental Fig 11),

(ii) Qualified their conclusions with added limitations and clarified several points that I found previously unclear,

(iii) Added several new analyses to the Appendices

(iv) Demonstrated that previous simulations validating TDLM overestimated TDLM sensitivity relative to the hybrid simulation.

(v) Added a new and extensive appendix on the relationship between TDLM and replay characteristics.

Weaknesses:

The remaining weaknesses of the work relate primarily to explaining the cause of measured non-random fluctuations in the simulated correlations between replay detection and performance at different time lags, as well as a lack of general recommendations of parameter choices for applying TDLM in future work. But these are minor weaknesses that can be left to future work.

Reviewer #3 (Public review):

Summary:

Kern et al. critically assess the sensitivity of temporally delayed linear modelling (TDLM), a relatively new method used to detect memory replay in humans via MEG. While TDLM has recently gained traction and been used to report many exciting links between replay and behavior in humans, Kern et al. were unable to detect replay during a post-learning rest period. To determine whether this null result reflected an actual absence of replay or sensitivity of the method, the authors ran a simulation: synthetic replay events were inserted into a control dataset, and TDLM was used to decode them, varying both replay density and its correlation with behavior. The results revealed that TDLM could only reliably detect replay at unrealistically (not-physiological) high replay densities, and the authors were unable to induce strong behavior correlations. These findings highlight important limitations of TDLM, particularly for detecting replay over extended, minutes long time periods.

Strengths:

Overall, I think this is an extremely important paper, given the growing use of TDLM to report exciting relationships between replay and behavior in humans. I found the text clear, the results compelling, and the critique of TDLM quite fair: it is not that this method can never be applied, but just that it has limits in its sensitivity to detect replay during minutes long periods. Further, I greatly appreciated the authors efforts to describe ways to improve TDLM: developing better decoders and applying them to smaller time windows.

The power of this paper comes from the simulation whereby the authors inserted replay events and attempted to detect them using TDLM. Regarding their first study, there are many alternative explanations or possible analysis strategies that the authors do not discuss; however, none of these are relevant if replayed, under conditions where it is synthetically inserted, cannot be detected.

Further, the authors provide a simulation and series of analyses aimed at replicating previous TDLM-based replay studies. They demonstrate methodological flaws, and show that previous simulations greatly overestimated the sensitivity of TDLM. This work emphasizes the need to cast a critical eye over both past and future studies applying TDLM to detect replay.

Finally, the authors are relatively clear about which parameters they chose, why they chose them, and how well they match previous literature (they seem well matched); and provide suggestions for how others can determine the best parameters for TDLM within their own experimental contexts.

Comments on revisions:

The authors thoroughly addressed my previous comments; the added analyses and discussion significantly strengthen the paper's clarity, utility, and impact.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

(1) I found the bigger picture analysis to be lacking. Let us take stock: in other work, during active cognition, including at least one study from the Authors, TDLM shows significance sequenceness. But the evidence provided here suggests that even very strong localizer patterns injected into the data cannot be detected as replay except at implausible speeds. How can both of these things be true? Assuming these analyses are cogent, do these findings not imply something more destructive about all studies that found positive results with TDLM?

Our focus here is on advancing methodology. Given the diversity of tasks and cognitive states in the TDLM literature, replay could exceed detection thresholds under specific conditions—especially when true event durations align with short analysis windows. While a comprehensive re-analysis of prior datasets is beyond our scope, we agree a concise synthesis can strengthen the paper.

The previous TDLM literature uses a diverse set of tasks and addresses a broad spectrum of cognitive constructs/processes. As we acknowledge, it is perfectly possible that replay bursts in short time windows are well detectable by TDLM. However, we acknowledge that some commentary on this is warranted and have added the following paragraph to the discussion that addresses “improving TDLMs sensitivity”:

“Finally, what do our simulations imply for the broader MEG replay literature? Our implementation successfully detects replay when boundary conditions are met, as shown in the simulation. But sensitivity depends critically on high fidelity between the analysis window and the density of replay events. A systematic evaluation of these conditions as they apply to prior studies remains beyond the scope of the current paper. Instead, our focus is on delineating boundary conditions that we hope will motivate conduct of power analyses in future work as well as inclusion of simulations that approximate realistic experimental conditions.”

(2) All things considered, TDLM seems like a fairly 'vanilla' and low-assumption algorithm for finding event sequences. It is hard to see intuitively what the breaking factor might be; why do the authors think ground truth patterns cannot be detected by this GLM-based framework at reasonable densities?

We agree with the overall sentiment of the referee. Our intuition is that one of the principal shortcomings of the method relates to spurious sequenceness induced by unknown factors at baseline, and poor transfer of the decoder to other modalities. and have a rough understanding of how they occur, we are currently not in a position to identify their nature. Note that we believe that these confounders are not exclusive to TDLM but are potentially threatening to all kinds of sequenceness analysis of longer time series that rely on decoders. Indeed, we suspect that classifier training is another bottleneck, as we don’t know the exact nature of the representations that are replayed, including the degree of overlap there is with a commonly used visual localizer. That said, this is not of relevance for the simulation in so far as we insert patterns that exceed the pattern strength in the localizer.

Finally, a potential major drawback is the permutation test for significance testing. As the original authors of TDLM have noted, the current test which permutes states is overly conservative. It measures fixed effects and as it only considers the group level mean it is accordingly easily biased by individual outliers. This we have tried to account for by z-scoring sequenceness scores. We have also conferred on this with some of the authors of TDLM and discussed a yet unpublished method that aims to address this exact issue. The proposed new method uses a sign-flip permutation test at a group level and therefore implements a random-effects model of the data. This significance test has markedly increased power while still controlling for FWER. However, while we show in our power analysis that the new method is indeed more sensitive, it does not materially change the interpretation of the data. We have included this novel method in the paper and added it into the main analysis and most of the simulations.

(3) Can the authors sketch any directions for alternative methods? It seems we need an algorithm that outperforms TDLM, but not many clues or speculations are given as to what that might look like. Relatedly, no technical or "internal" critique is provided. What is it about TDLM that causes it to be so weak?

We believe there are several shortcomings and bottlenecks within TDLM that need to be evaluated and improved. While we highlight these issues in the discussion section titled “Improving TDLMs sensitivity,” we agree that we should provide a clearer outline of its current shortcomings. We have now added to the discussion to expand on that we think needs improvement (‘fixed time lag’) and also add a summary statement at the end of the relevant paragraph to recap the main issues needed for an improved successor method. The new paragraphs read:

“Lastly, there are certain assumptions that TDLM makes that might not hold (see Methods Study II): Current implementations look for a fixed time lag that is the same across all participants and between all reactivation events. If time lags differ across participants, TDLM will fail to find them. Similarly, TDLM assumes a fixed sequence order and is not robust against slight within-sequence permutations or in-sequencemissing reactivation events. However, from other data sources., such as hippocampal place cell recordings, it is known that such permutations can occur where some states are skipped or fail to decode during replay. Similarly, it is assumed that each reactivation event lasts between 10-30 milliseconds, but the true temporal evolution of reactivation measured by TDLM is currently unknown. Future method development might focus on improving invariance to these assumptions.

[…]

In summary, there are several areas where TDLM might be improved, including a restriction in its search space, improvement in classifiers, a validation of localizer representation transfer to other domains (e.g. memory representations), and the extension of TDLM to render it more robust against violations of its core assumptions.”

Reviewer #2 (Public review):

Weaknesses:

The sample size is small (n=21, after exclusions), even for TDLM studies (which typically have somewhere between 25-40 participants). The authors address this somewhat through a power analysis of the relationship between replay and behavioural performance in their simulations, but this is very dependent on the assumptions of the simulation. Further, according to their own power analysis, the replay-behaviour correlations are seriously underpowered (~10% power according to Figure 7C), and so if this is to be taken at face value, their own null findings on this point (Figure 3C) could therefore just reflect under sampling as opposed to methodological failure. I think this point needs to be made more clearly earlier in the manuscript.

We agree with the referee that our sample is smaller than previous studies due to participant exclusion criteria. However, the take-away message from our behavioural simulation and bootstrapping is that even with larger sample sizes, it is difficult to overcome baseline fluctuations of sequenceness, even if very strong replay patterns were detectable and sample sizes were of similar size to that of previous studies. Therefore, we are not convinced that that our null findings are fully explained by the smaller sample size compared to that of previous studies, Additionally, we show that even within the range of other studies, similar power would have been expected (Supplement Figure 11). However, it is true that in general null findings can be explained by under-sampling, under the assumption that an effect is present. To amplify this point, we have added the following to the Figure 3C:

“[…]. NB, however, as our simulation shows, correlations of sequenceness with behavioural markers are likely to be underpowered and occur only with very high replay rates or much higher sample size. See our simulation discussion for a more detailed explanation on how correlations may be inherently biased, where fluctuations in baseline sequenceness overshadow individual scaling with behavioural markers.”

Furthermore, we have added the following paragraph to the discussion to highlight this point and refer to a power analysis we have now added to the supplement (see next answer):

“Sample sizes in previous TDLM literature usually range between 20 to 40 participants. A bootstrap power analysis shows that even at those sample sizes, power would remain low unless unrealistically high replay rates are assumed (Supplement Figure 11). Our bootstrap simulation shows that a correlation analysis between sequenceness and behaviour would in these cases be drastically underpowered, even under an assumption of high replay densities.”

Finally, we have added a remark about the sample size to the limitations section, as naturally, an increase in sample size would yield higher power:

“Finally, while initially planning for thirty participants, due to exclusion criteria, our study featured fewer participants than most previous studies using TDLM (i.e. usually 25-40, but 21 in our study). While we are confident that our simulation results hold under these sample sizes, as sample sizes of other studies show comparable power to ours (Fehler! Verweisquelle konnte nicht gefunden werden.), we cannot fully rule out a possibility that our null-findings are explained by a lack in power alone.”

Relatedly, it would be very useful if one of the recommendations that come out of the simulations in this paper was a power analysis for detecting sequenceness in general, as I suspect that the small sample size impacts this as well, given that sequenceness effects reported in other work are often small with larger sample sizes. Further, I believe that the authors' simulations of basic sequenceness effects would themselves still suffer from having a small number of subjects, thereby impacting statistical power. Perhaps the authors can perform a similar sort of bootstrapping analysis as they perform for the correlation between replay and performance, but over sequenceness itself?

We agree with the referee that this, in principle, is a great idea. However, the way that significance thresholds are calculated poses a conceptual problem for such an analysis: as for significance threshold we are defining the maximum sequenceness value across all participants, all time lags and all permutations. This sequenceness value is compared against the mean of all participants, disregarding the standard deviation. This maximum threshold would not change if we bootstrapped some of our samples. Additionally, the 95% would also not change significantly. To illustrate this point, we have added this analysis to the supplement, as Supplement Figure 10. However, the new sign-flip permutation test we now include allows for such a comparison, as it takes variance between participants into account as well! We have included all three variants of the power analysis and the figure description now reads:

“Supplement Figure 11 Power analysis of sequenceness significance for bootstrapped samples sizes. A) Powermap for state-permutation thresholds. However, here the bootstrap approach suffers from a conceptual problem: significance thresholds are defined by the permutation maximum and/or 95-percentile of the maximums across all sequence-permutations across participants. If we resample bootstrap-participants from our existing pool, the maximum thresholds computed will remain relatively stable across resampled participants, as it only compares against the mean and disregards the standard deviation. B) The newly presented statistical approach is significantly more sensitive at higher sample sizes. Note that even then, 80% power is only reached with replay density of higher than 50 min-1 at a sample size of 60 participants. Additionally, the sign-flip permutation test assumes that the mean is at zero. As we observed a non-zero mean due to spurious oscillations, we subtracted the mean sequenceness of the baseline condition from each participant before permuting to achieve a null distribution with mean zero, as otherwise, we would have found significant replay effects in the baseline condition at increasing sample size. Nevertheless, due to the higher sensitivity, the new sign-flip test is recommended over the previous sequence-permutation-based test. Colours indicate the power from 0 to 1 for different bootstrapped sample sizes and densities. 80% power thresholds are outlined in black.”

The task paradigm may introduce issues in detecting replay that are separate from TDLM. First, the localizer task involves a match/mismatch judgment and a button press during the stimulus presentation, which could add noise to classifier training separate from the semantic/visual processing of the stimulus. This localizer is similar to others that have been used in TDLM studies, but notably in other studies (e.g., Liu, Mattar et al., 2021), the stimulus is presented prior to the match/mismatch judgment. A discussion of variations in different localizers and what seems to work best for decoding would be useful to include in the recommendations section of the discussion.

We agree and thank the referee for raising this issue. Note, we acknowledge we forgot to mention that these trials were excluded from classifier training. Our rationale of presenting the oddball during stimulus presentation, and not thereafter, was an assumption that by first presenting the audio and then the visual cue we would create more generalized representations that would be less modalitydependent. However, importantly, we excluded all trials that were oddballs from localizer training. Therefore we assume that this particular design choice will not greatly affect the decoder training. If some motor-preparation activity is present during the stimulus presentation, then it should be present equally across all trials and hence be ignored by the classifier as we balanced the transitions between images. We now added this information to the main text:

“In each trial, a word describing the stimulus was played auditorily, after which the corresponding stimulus was shown. In ~11% of cases, there was a mismatch between word and image (oddball trials), and these trials were excluded from the localizer training.” Additionally in the methods section: “These oddball-trials were excluded from all further analysis and decoder training.”

Nevertheless, we agree that the extant variety in localizer designs is underdiscussed where many assumptions of classifier training are not, as yet, fully validated. We have added a sentence highlighting different oddball paradigms to the section on the discussion of localizers and also add a summary statement with recommendations. The passage now reads:

“Additionally, a wide variety of oddballs has been used (e.g. upside-down, scrambled, or mismatched images, cues presented visually, as words, auditorily, etc), and at this time it is unclear if these affect the representations that the classifier learns [...] In summary, we would expect a multimodal categorical localizer, and a classifier that isn’t trained on a specific timepoint, to generalize best.”

Second, and more seriously, I believe that the task design for training participants about the expected sequences may complicate sequence decoding. Specifically, this is because two images (a "tuple") are shown together and used for prediction, which may encourage participants to develop a single bound representation of the tuple that then predicts a third image (AB -> C rather than A -> B, B -> C). This would obviously make it difficult to i) use a classifier trained on individual images to detect sequences and ii) find evidence for the intended transition matrix using TDLM. Can the authors rule out this possibility?

We thank the reviewer for raising a possibility we have not considered! While there is some evidence that a single bound representation would have overlap with its constituents (especially before long term-consolidation) and therefore be detectable by the classifiers, we acknowledge the possibility that individual classifiers would fail to be sensitive to such a compound representation. In fact we find in the retrieval data some evidence for a combined replay of representations (where representations are replayed seemingly at the same time, see Kern 2024). We have added such a possibility to the interims-discussion of Study 1 as a qualification . However, this does not change the results or interpretation of our simulation which we consider is a key message of the paper.

The relevant segment in the discussion section now reads:

“Additionally, given that the stimuli were presented in combined triplets, participants may have formed a singular representation of associated items and subsequently replayed these (e.g., AB→C), instead of replaying item-by-item transitions (A→B→C). Under such a scenario, a classifier trained on individual items may fail to detect these newly formed bound representations, particularly if they diverge strongly from the single-item patterns. In our previous study where we address retrieval (Kern et al., 2024) we found that states were to varying extent co-reactivated, yet classifiers trained on single items retained sensitivity to detect these combined reactivation events. Consistent with this, prior work suggests that unified representations retain overlap with their constituent item representations (Dennis et al., 2024; Liang et al., 2020), however, there’s also evidence that different brain regions are involved if representational unitization occurs (Staresina & Davachi, 2010), potentially confusing classifiers. Therefore, we cannot exclude that rest-related consolidation replays engendered unitized representations that were insufficiently captured by our singleitem classifiers.“

Participants only modestly improved (from 76-82% accuracy) following the rest period (which the authors refer to as a consolidation period). If the authors assume that replay leads to improved performance, then this suggests there is little reason to see much taskrelated replay during rest in the first place. This limitation is touched on (lines 228-229), but I think it makes the lack of replay finding here less surprising. However, note that in the supplement, it is shown that the amount of forward sequenceness is marginally related to the performance difference between the last block of training and retrieval, and this is the effect I would probably predict would be most likely to appear. Obviously, my sample size concerns still hold, and this is not a significant effect based on the null hypothesis testing framework the authors employ, but I think this set of results should at least be reported in the main text.

We disagree that an absence or presence of replay might be inferred from an absolute memory enhancement. While consolidation can lead to absolute improvement of performance in, for example, motor memory domains one formulation is that in declarative learning tasks replay stabilizes latent memory traces, and in such a scenario would not necessarily lead to a boosted performance. While many declarative consolidation studies report an increase of performance compared to a control condition (i.e. without a consolidation window), this does not necessarily entail an absolute performance increase, as replay might just act to protect against loss of memory traces. Therefore, the modest increase we observe does not inference as to the presence of absence of replay absent a proper control condition.

We did expect to find a correlation between replay and individual behavioural. Indeed, a weak correlation with performance and sequenceness can be detected. However, as we also show any such correlation is overshadowed by baseline fluctuations in sequenceness such that its overall validity is questionable, even under very high replay rates. We are therefore circumspect about this correlation, even if it was significant. Therefore, in the discussion, we chose to refrain from putting much focus on this correlation. Nevertheless, we do add a short statement to the corresponding figure label, discussing this precise issue. The segment now reads:

“While we found a non-significant relation between a memory performance enhancement and post-learning forward sequenceness we are cautious not to overinterpret these results. As in the section “Correlation with behaviour only present at high replay speeds” the noted correlational measure oscillates heavily with baseline sequenceness fluctuations, and any true replay effect is likely to be overshadowed by such fluctuations.”

I was also wondering whether the authors could clarify how the criterion over six blocks was 80% but then the performance baseline they use from the last block is 76%? Is it just that participants must reach 80% within the six blocks *at some point* during training, but that they could dip below that again later?

We thank the reviewer for highlighting this point: The first block wherein participants reached >80% ended the learning blocks. After a maximum of six blocks the learning session was ended regardless of performance. Therefore, some participants’ learning blocks were ended after six blocks and without them reaching a performance of 80%.. While we described this in the Methods section, it was missing from the Results Study I section, which now contains:

“[...] Participants then learned triplets of associated items according to a graph structure. Within the learning session, participants performed a maximum of six learning blocks, but the session was stopped if participants reached 80% memory performance (criterion learning,, up to a memory performance criterion of 80% (see Methods for details)”

The Figure 2 description now contains

“[...] Participants’ completed up to six blocks of learning trials. After reaching 80% in any block, no more learning blocks were performed (criterion learning) [...]”

Lastly, there was a mistake in the Behavioural results section, which stated “All thirty participants, except one, [..] to criterion of 80%.” This is an error. In our preregistration, we defined to only include participants that successfully learned anything at all above chance. Here,we meant that only one participant failed to reach a criterion that we defined as “successful learning”. We fixed it and it now reads

“with an accuracy above 50% (which we preregistered beforehand as an exclusion criterion for “successful learning above chance”).”

Additionally, we have noted this for clarity in the methods section and excuse this mistake:

“Additionally, as successful above-chance learning was necessary for the paradigm, we ensured all remaining participants had a retrieval performance of at least 50% (one participant had to be excluded, but was already excluded due to low decoding performance).”

Because most of the conclusions come from the simulation study, there are a few decisions about the simulations that I would like the authors to expand upon before I can fully support their interpretations. First, the authors use a state-to-state lag of 80ms and do not appear to vary this throughout the simulations - can the authors provide context for this choice? Does varying this lag matter at all for the results (i.e., does the noise structure of the data interact with this lag in any way?)

This was a deliberate choice but we acknowledge the reasoning behind this was not detailed in our initial submission. We chose a lag of 80 millisecond for three reasons: first, it is distant from the 9-11 Hz alpha oscillations we observed in our participants and does not share a harmonic with the alpha rhythm; second, we wanted to get a clear picture of the effect of simulated replay that is as isolated as possible from spurious sequenceness confounders present in the baseline condition. Thus, we chose a lag in which the sequenceness score was close to zero in the baseline condition; thirdly , in this revision, we subtracted the mean sequenceness value of the baseline such that any simulation effects would start, on average, at zero sequenceness. In this way, we could attribute any increase in sequenceness to the experimentally inserted replay, that was independent of spurious oscillations. Finally (but less importantly), as we observed that a correlation of sequenceness with behaviour was fluctuated strongly, for the reason detailed above, we chose a lag in which a correlation was as close as possible to zero. If we had not chosen a lag that adhered to these conditions, we were at risk of measuring simulated replay plus spurious sequenceness confounders.

We have added a sentence to the main text detailing this justification:

“We chose this timepoint (80 msec state to state lag) as its sequenceness value was close to zero in the baseline condition as well as being distant to the observed alpha rhythms of the participants (which varied between ~9-11 Hz). Additionally, we subtracted the mean sequenceness value of the baseline at 80 milliseconds lag such that any simulation effects would, on average, start at zero sequenceness “

Additionally, we now add a more detailed explanation to the methods section.

“This time lag (80 msec) was chosen in order to isolate precisely an effect of the experimentally inserted sequenceness. Thus, we chose a lag at which the mean baseline sequenceness was close to zero and where the correlation with behaviour was low. Additionally, we subtracted the mean sequenceness value (at 80 milliseconds) at baseline from the specific lag recorded for each participant, such that simulation effects would be initialized at zero sequenceness on average enabling any effects to be attributed purely to inserted replay. Additionally, we excluded time lags too close to the alpha rhythms of participants (which varied between ~9-11 Hz) or lags which would have a harmonic with the rhythm.”

Second, it seems that the approach to scaling simulated replays with performance is rather coarse. I think a more sensitive measure would be to scale sequence replays based on the participants' responses to *that* specific sequence rather than altering the frequency of all replays by overall memory performance. I think this would help to deliver on the authors' goal of simulating an "increase of replay for less stable memories" (line 246).

The referee makes an excellent point and our simulations could be rendered more realistic by inserting the actual tuples that participants answered correctly. If we understand the point correctly, there are two different ways replay might be impacted by performance: First, we can conjecture that there is greater replay if memory performance is not saturated. Second, replay only occurs for content that has actually been encoded!

The main reasons why we chose to simulate the entire sequence being replayed for each participant is based on the following. TDLM is implemented such that the amount of replay alone is relevant, and actual transitions are not affecting the results beyond noise. Under the assumption that class-specific classifiers perform equally well, simulating A->B, B->C or simulating A->B, A->B yields equivalent results. However, results can differ if this assumption is violated. By drawing from the entire space of classes we insert, we minimize the risk of some classifiers being worse than others for some participants. For example, if we simulated only A->B for some participant instead of the whole sequence, and by chance classifier A performs suboptimally, we would then introduce additional unwanted variance into our results.

Secondly, from our reading of the literature we infer that replay is increased generally (i.e. density of learning-specific replay is increased) for less stable memories. However, we do not have indicators of memory strength, but only a binary “remembered or not”. As TDLM is invariant to the actual transitions being replayed and only indexes the number of transitions, we chose to ignore which transitions we insert and only scaled the amount of replay.

We have added an analysis to the Appendix that discusses this specific aspect of our study where we show that results are equivalent if we simulate replay of “A->B B->C C->D” or only “A->B A->B A->B A->B”. As we do not know how replay density interacts with memory trace stability, we opted to leave the current simulation as is. The corresponding paragraph and figure description now read:

“From literature we know that replay is increased after learning and that less stable memories are replayed more often. We simulated this effect by scaling our replay density inversely with performance. However, for simplicity, in our simulation, we inserted sampled transitions from all valid transitions given by the graph structure, i.e., the following transitions were valid: However, this meant that some participants would have transitions inserted that they didn’t actually remember. To show that this would not change results, we simulated two scenarios: In the full sequence scenario, all valid graph transitions are inserted (i.e. all participant’s replay is sampled from 'A->B, B->C, C->D, D->E, E->F, F->G, G->E, E->H, H->I, I->B, B->J, J->A'). In the second scenario (memorized transitions) we only replayed transitions that the participant actually retrieved correctly during the post-resting state testing sessions (i.e. a participant’s replay would have been sampled from ‘A->B, B->C, G->E, E->H, H>I’, if those were the ones he remembered). In both scenarios, the number of events is kept constant. The results are equivalent as can be seen in Appendix A Figure 3. NB this only holds under the assumptions that classifiers are equally good at decoding each class.”

[…]

“TDLM is insensitive towards which transitions are replayed and only sensitive to how many transitions are detected in total. Here we simulate transitions either sampled from the full graph (light orange/green) or participant-specific transitions of trials that participants correctly remembered (dark orange/green). Shaded areas denote the standard error across participants.”

On the other hand, I was also wondering whether it is actually necessary to use the real memory performance for each participant in these simulations - couldn't similar goals (with a better/more full sampling of the space of performance) be achieved with simulated memory performance as well, taking only the MEG data from the participant?

The decision to use real memory performance is indeed arbitrary. We could have also used randomly sampled values. However, as we wanted to understand our nullresults better we opted to use real performance to adhere as close as possible to the findings we previously reported. Using uniformly sampled memory performance would be less explanatory w.r.t to our actual results of the resting state data that are reported in the first study we report in the manuscript (Study I).

Nevertheless, our current implementation already presents an approach that samples the entire performance range for the sub-analysis focusing on the correlation with behaviour. Here, in the section on “best-case”-scenario, we implement this such that it spans factors from 1 to 0 (i.e., a participant with 100% performance gets a replay scale factor of 0 and hence no replay simulated, and the worst performing participant with 50% performance has a replay rate multiplied by 1). We scale the amount of replay with this factor. As a correlation is invariant to linear scaling, statistically this is equivalent to stretching the performance distribution from 0 to 100%. We have added a sentence to the methods to provide further focus on this point:

“To assess how performance might affect replay in our specific dataset, we chose to use the original participants’ performance values instead of uniformly sampling the performance space (which ranged from 50 to 100%). However, for the correlation analysis, we additionally added a “best-case” scenario, in which we scale replay from 0 to 1, an approach that is statistically equivalent to scaling values to the full space of possible performance (0 to 100%) (see Results Study II: Simulation).”

Finally, Figure 7D shows that 70ms was used on the y-axis. Why was this the case, or is this a typo?

Thanks, this is indeed a typo, we fixed it.

Because this is a re-analysis of a previous dataset combined with a new simulation study on that data aimed at making recommendations about how to best employ TDLM, I think the usefulness of the paper to the field could be improved in a few places. Specifically, in the discussion/recommendation section, the authors state that "yet unknown confounders" (line 295) lead to non-random fluctuations in the simulated correlations between replay detection and performance at different time lags. Because it is a particularly strong claim that there is the potential to detect sequenceness in the baseline condition where there are no ground-truth sequences, the manuscript could benefit from a more thorough exploration of the cause(s) of this bias in addition to the speculation provided in the current version.

We are currently working on a theoretical basis to explain these spurious sequenceness confounders in the baseline condition. Indeed, in our preliminary work, in certain contexts we can induce significant sequenceness in the absence of any replay signal during baseline. However, this work is at an early stage and we still have some conceptional problems to solve before we are confident enough with these data. We believe at present it would be premature to add these data to the current manuscript. Nevertheless, we now mention these spurious sequenceness confounders to raise awareness for the field and also add greater context to the discussion, highlighting one of the issues that we think is of importance:

“[…] For example, if two classifiers’ probabilities oscillate at 10 Hz but at a different phase, a spurious time lag can be found reflecting this phase shift. We speculate that more complex interactions between classifiers oscillating at different phases are also conceivable.”

In addition, to really provide that a realistic simulation is necessary (one of the primary conclusions of the paper), it would be useful to provide a comparison to a fully synthetic simulation performed on this exact task and transition structure (in addition to the recreation of the original simulation code from the TDLM methods paper).

Thank you for this suggestion! We have now added a synthetic simulation, trying to keep as close as possible to the original simulation code in Liu et al. (2021), while also incorporating our current means of simulating the data (i.e. scaling by performance). We think this synthetic simulation greatly improves the paper and gives weight to our suggestion about the superiority of a hybrid approach. Additionally, it prompted us to look closer at patterns that are inserted in the synthetic simulation and perform a comparative analysis. We have now added the simulation to the main text, together with a methodological explanation of how we simulated the data in the methods section. We also added a discussion on the results and why we think a hybrid approach is currently superior to synthetic approach. The whole new section is too long to paste here – it is found after the main simulation section in the manuscript. We have also added another sentence to the abstract referring to this new inclusion.

Finally, I think the authors could do further work to determine whether some of their recommendations for improving the sensitivity of TDLM pan out in the current data - for example, they could report focusing not just on the peak decoding timepoint but incorporating other moments into classifier training.

While we do understand the desire to test further refinement to TDLM on the data directly, we intentionally do not include such analyses in the current paper. Our experience also informs us that there is an enormous branching factor of parameters when applying TDLM, with implications for significance of results in one or other direction. However, as there are currently only limited ways to know how well parameter changes actually improve the sensitivity to replay versus exacerbate potential underlying confounders that induce spurious sequenceness (e.g., we can get significant replay in the control condition with some parameter changes). To exclude such false positive findings, we opt for a relatively strict adherence to previously published approaches. Thus, in the current paper, we limit ourselves to assessing the reliability and robustness of previous approaches.

Furthermore, while training on a later timepoint might increase sensitivity for a classifier when transferring between different modalities (e.g. visual to memory representation), this approach does not transfer well in our simulations, as the inserted patterns are from the same modality. We consider other, more bespoke studies, are better suited to improve classifier training. NB also see our recently started Kaggle challenge to tackle this problem: https://www.kaggle.com/competitions/the-imagine-decoding-challenge

However, we have added a note about this dilemma to the improvement section. The section now includes:

“Nevertheless, as the considerable branching factor poses a threat of increased falsepositive findings we opt to focus the current simulations on previously published pipelines and parameters. Future studies should systematically evaluate parameter choices on TDLM under different conditions, something that is beyond the remit of the current study.”

Lastly, I would like the authors to address a point that was raised in a separate public forum by an author of the TDLM method, which is that when replays "happen during rest, they are not uniform or close." Because the simulations in this work assume regularly occurring replay events, I agree that this is an important limitation that should be incorporated into alternative simulations to ensure the lack of findings is not because of this assumption.

The temporal distribution of replay throughout the resting state should not matter, as TDLM is invariant w.r.t to how replay events are distributed within the analysis window. Specifically, it does not matter if replay events occur in bursts or are uniformly distributed. Only the number of transitions is relevant, where they occur or if they are close to each other is not relevant to the numerical results (as long as the refractory window is kept, too short distances will lead to interactions between events and reduce sensitivity).). To emphasize this point, we have added another simulation which is shown in Appendix A.1 and Appendix A Figure 1. We have referenced it in the text and added the following paragraph in the Methods section

Additionally, the timepoints of inserting replay within the resting state are sampled from a uniform distribution. Even though TDLM tracks reactivation events over time, at a macro-scale the algorithm is invariant to the temporal distribution. At each time step, the GLM regresses onto a future time step up to the maximum time lag of interest, yielding a predictor per lag. However, these predictors within the GLM are independently assessed, and hence, TDLM is, outside of the time lag window, relatively invariant to the temporal distribution of replay. To demonstrate our claim, we simulated uniform replay vs “bursty” replay that only occurs in some parts of the resting state, both yield equivalent sequenceness results (see Appendix A.1).

Reviewer #3 (Public review):

(1) I am still left wondering why other studies were able to detect replay using this method. My takeaway from this paper is that large time windows lead to high significance thresholds/required replay density, making it extremely challenging to detect replay at physiological levels during resting periods. While it is true that some previous studies applying TDLM used smaller time windows (e.g., Kern's previous paper detected replay in 1500ms windows), others, including Liu et al. (2019), successfully detected replay during a 5-minute resting period. Why do the authors believe others have nevertheless been able to detect replay during multi-minute time windows?

(Due to similarity, we combined our responses with the first question of Reviewer 1)

We are reluctant to make sweeping judgments in relation to previous literature as we wanted to prioritize on advancing methodology instead. The previous TDLM literature uses a diverse set of tasks and cognitive processes. As we state ourselves, it is possible that replay bursts in short time windows are well detectable by TDLM. We were intentionally cautious to directly critique previous studies without detailed re-analysis of their work and wanted to leave such a conclusion up to the reader. However, we realize that such a “thought-starter” might be warranted and improve the paper. Therefore, we have added the following paragraph to the discussion about “improving TDLMs sensitivity”:

“Finally, what do our simulations imply for the broader MEG replay literature? Our implementation successfully detects replay when boundary conditions are met, as shown in the simulation. But sensitivity depends critically on high fidelity between the analysis window and the amount of replay events. A systematic evaluation of these conditions across prior studies is beyond the scope of this paper, so we do not want to adjudicate earlier findings and leave this assessment up to the reader. Instead, we delineate the boundary conditions and urge future work to conduct power analyses where possible and include simulations that approximate realistic experimental conditions.”

For example, some studies using TDLM report evidence of sequenceness as a contrast between evidence of forwards (f) versus backwards (b) sequenceness; sequenceness was defined as ZfΔt - ZbΔt (where Z refers to the sequence alignment coefficient for a transition matrix at a specific time lag). This use case is not discussed in the present paper, despite its prevalence in the literature. If the same logic were applied to the data in this study, would significant sequenceness have been uncovered? Whether it would or not, I believe this point is important for understanding methodological differences between this paper and others.

This approach was first introduced as part of a TDLM-predecessor that utilized crosscorrelations (Kurth-Nelson 2016), where this step is a necessity to extract any sequenceness signal at all by subtracting signals that are present in both (akin to an EEG reference). However, its validity is less clear when fwd and bkw are estimated separately, as is in the GLM case. The rationale behind subtracting here is the same as for autocorrelations: there are oscillatory confounds present in the data that introduce spurious sequenceness in both directions alike, i.e. at the same time lag, that can simply be removed by subtracting. However, this assumption only holds if the sole confounder is auto-correlations caused by a global signal that oscillates at all sensors at the same phase. In our own experience, and mentioned in the discussion, we do not think this assumption holds. Arguably, there are more complex interactions at play that cannot be removed by such a subtraction such as an increase in false positives if confounders are in an opposite direction at a specific time lag. This assumption-violation can be seen in our baseline condition, where other spurious sequenceness diverges in opposite directions for some time lags (e.g. at ~90 ms where forward sequenceness is negative and backward sequenceness is positive). We reasoned that oscillatory confounds are more stable when comparing pre vs post for the same direction than comparing within session between forward minus backward.

Finally, we note issues introduced by the various ways that sequenceness has been analysed in previous papers: normalization of sequenceness (z-scoring across time lags or across participants or not at all), normalization of probabilities (taking raw decision scores, z-scoring, soft-max, dividing by mean, subtracting mean), taking a windowed approach and summing sequenceness scores, not to mention the various classifier choices that can be made, and all of this can be applied before subtracting conditions from each other or before subtraction. In our experience there is insufficient regard to control for multiple comparison when running all these analyses risking selectivity in reporting.

Nevertheless, subtracting forward from backward replay is probably as valid as post minus pre. Therefore, we have added fwd-bkw plots to the supplement and explained some of the reasoning for not reporting them in the main text in the figure label. The figure label and reference now read:

“Finally, we report forward minus backward sequenceness and our motivation for using an across-session post-pre comparison instead of within-session forwardbackward in Supplement Figure 10.”

[…]

“Forward minus backward sequenceness within each resting state session. Previous papers often report subtraction of backward from forward sequenceness (fwd-bkw) as a means to remove oscillatory confounds that impact both sequenceness directions in synchrony. While required in early cross-correlation approaches (KurthNelson et al., 2016), its validity in GLM-based frameworks depends on an assumption that confounds are global and in-phase across sensors. We observed this assumption is violated in our baseline data, where spurious sequenceness occasionally diverges in opposite directions at specific time lags (e.g., ~90 ms). In such instances, subtraction would increase the false-positive rate rather than suppress noise. In Figure 3B, we prioritized the comparison of pre-task versus post-task sequenceness within the same direction, as oscillatory confounds appeared more stable across time within a single direction, as opposed to across directions within a single session. However, we consider both approaches are valid. We now provide the fwd-bkw plots for completeness and comparison with previous literature. A) forward minus backwards sequenceness for Control (left) and Post-Learning resting-state (right). B) T-value distribution of the sign-flip permutation test for Control (left) and Post-Learning resting-state (right)”

(2) Relatedly, while the authors note that smaller time windows are necessary for TDLM to succeed, a more precise description of the appropriate window size would greatly improve the utility of this paper. As it stands, the discussion feels incomplete without this information, as providing explicit guidance on optimal window sizes would help future researchers apply TDLM effectively. Under what window size range can physiological levels of replay actually be detected using TDLM? Or, is there some scaling factor that should be considered, in terms of window size and significance threshold/replay density? If the authors are unable to provide a concrete recommendation, they could add information about time windows used in previous studies (perhaps, is 1500ms as used in their previous paper a good recommendation?).

We currently do not have an empirical estimate of which window sizes are appropriate. While we used 1500ms in our previous paper, this was solely given by the experiment design which had a 1.5s wait period before the next stimulus. Our recommendation for best guidance on this matter would be to investigate related intracranial literature for SWR rate increases under similar experimental conditions. We have added the following paragraph to the discussion:

“At this stage we cannot offer a general recommendation for window sizes as they are likely to depend on details of the research paradigm. However, intracranial recordings can be used as proxy to estimate the duration of replay bursts, for example as reported in (Norman et al., 2019) where increased SWRs were seen up to 1500 ms after retrieval cue onset”

(3) In their simulation, the authors define a replay event as a single transition from one item to another (example: A to B). However, in rodents, replay often traverses more than a single transition (example: A to B to C, even to D and E). Observing multistep sequences increases confidence that true replay is present. How does sequence length impact the authors' conclusions? Similarly, can the authors comment on how the length of the inserted events impacts TDLM sensitivity, if at all?

Good point! So far, most papers do not seem to include multi-step TDLM and in our experience rightfully, as it is conceptionally difficult to define clear significance thresholds while keeping in mind that shorter sub-sequences are contained within a longer sequence (e.g. ABC contains both AB and BC and a longer dependency of AC) that renders it difficult to define the correct way to create a null distribution for the permutation test. Therefore, we tried to stay as close as possible to previous approaches and only looked for single-step transitions. Nevertheless, we have added an analysis to the supplement comparing how TDLM behaves if we simulate A->B->C or A->B and separate B->C. It shows that TDLM is only sensitive to the number of transitions present in the data, and it does not matter if they are chained or chunked. The segment reads:

“We intentionally designed our study to encourage replay of triplets. However, this begs the question as to whether it matters if triplets or individual chunks of a sequence are replayed at different time points? Here, we simulated two scenarios. In one, we inserted replay of single transitions alone with a refractory period, e.g. A->B and separate B->C transitions. In a second scenario, we simulate replay of chained triplets, e.g. A->B->C, with a distance of 80 milliseconds each. Importantly, we kept the number of transitions constant (i.e., A->B, … B->C and where A->B->C would both have 2 transitions. This creates a context wherein a four-minute resting state would have ~100 events of A->B->C inserted and ~200 events of A->B or B->C, such that in both cases this results in the same number of single step transitions. We found both are equivalent, with TDLM agnostic to the length of sequence trains, i.e., it does not matter if replay is chunked or chained under the assumption that the number of transitions remains fixed, as can be seen in Appendix A Figure 2”

And the reference Figure description reads:

“TDLM is invariant to the length of sequence replay trains under an assumption that the number of target transitions (e.g. single steps) is fixed. We simulated replay either as two temporally separate A->B, B->C events (light orange/green) or as a single A>B->C event (dark orange/green), both yielding equivalent sequenceness. Shaded areas denote the standard error across participants”

For example, regarding sequence length, is it possible that TDLM would detect multiple parts of a longer sequence independently, meaning that the high density needed to detect replay is actually not quite so dense? (example: if 20 four-step sequences (A to B to C to D to E) were sampled by TDLM such that it recorded each transition separately, that would lead to a density of 80 events/min).

Indeed, this is an interesting proposal. We intentionally kept our simulation close to the way previous simulations were set-up (i.e. Liu & Dolan et al 2021, Liu & Mattar 2021) by simulating one-step transitions and simulated them such that there is no overlap between separate events (e.g. by defining a refractory period). If the duration of replay is increased then we would also need to increase the length of the refractory period, resulting in a reduced upper limit of how much replay can occur in a 1-minute time window. This in turn would approximate roughly the same number of transitions that can be inserted into the resting state and, as detailed above, would yield the same results. Nevertheless, as we chose to use replay density and not transition density as a marker, the density would be reduced, even if the number of transitions stay the same. We have added an analysis using multi-step replay to the supplement and discuss its implications and caveats. In the main discussion we have added the following segment:

“Similarly, in our simulation, for simplicity and to keep consistency with previousstimulations, we restricted replay events to span two reactivation events. While the characteristics of replay as measured by TDLM are unknown, it is conceivable that several steps can be replayed within one replay event. We show that the vanilla version of TDLM is fundamentally sensitive to the number of single-step transitions alone, and disregards if these are replayed chained or chunked (Appendix A.2 and Appendix A Figure 2). Nevertheless, if the number of reactivation events chained within a replay event increases, TDLMs sensitivity is increased relative to the replay density and thresholds are reached earlier (see Appendix A Figure 4). See Appendix A.4 for a simulation of multi-step replay events and our discussion of the caveats.”

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Please label the various significance thresholds in the legend of Figure 3.

We have labelled all the thresholds in the figure legends.

Reviewer #2 (Recommendations for the authors):

I think that some of the clarity is hampered because there is a bit too much reliance on explanations from the previous paper using this task, which hampers clarity in the paper. For example, Figure 1 is not particularly useful for understanding the study in its current form; I found myself relying almost exclusively on Supplementary Figure 1 (which is from the previous paper). I'd recommend presenting some version of SF1 in the main text instead. Another example of this overreliance on the previous paper is that, as far as I can tell, the present paper never explicitly states which transitions are being tested in TDLM. In the prior work, it states "all allowable graph transitions", and so I assumed this was the same here, but the paper should standalone without having to go back to the other study. I'd recommend that the authors revise the paper in these and other places where the previous paper is mentioned.

Thanks for raising this point! We were uncertain ourselves how to deal with the overlap in content and did not want to bloat the paper or plagiarize ourselves too much. On the advice of the referee have implemented the following to improve the manuscript and reduce a reliance on the previous paper:

Supplement Figure 1 is indeed crucial to understanding the experiment. We have moved it to the methods section under Methods: Procedure

Added more stimulus description to the Methods: Localizer section

Included more details about the localizer and graph learning that were missing before

We have added the note about which transitions we were looking for in the Methods section. Additionally, we have added this information to the Results section of Study 1.

There are also a few typos I noticed:

(1) Line 73: "during in the context of."

(2) Line 287: " to exploring the."

We fixed the typos.

Reviewer #3 (Recommendations for the authors):

(1) Why did the authors choose an 80ms state-to-state time lag for their simulation? I believe they should make the reason for this decision clear in the main text.

Indeed, this point was also raised by the other reviewer. We have added a sentence to the main text about the rationale behind this decision:

“We chose this timepoint (80 millisecond state-to-state lag) as its sequenceness value was close to zero in the baseline condition as well as being distant to the observed alpha rhythms of the participants (which varied between ~9-11 Hz). Additionally, we subtracted the mean sequenceness value of the baseline at 80 millisecond lag such that any simulation effects would, on average, start at zero sequenceness.“

Additionally, we have added some further explanation to the Methods section.

“This time lag (80 msec) was chosen in order to isolate precisely an effect of the experimentally inserted sequenceness. Thus, we chose a lag at which the mean baseline sequenceness was close to zero and where the correlation with behaviour was low. Additionally, we subtracted the mean sequenceness value (at 80 milliseconds) at baseline from the specific lag recorded for each participant, such that simulation effects would be initialized at zero sequenceness on average enabling any effects to be attributed purely to inserted replay. Additionally, we excluded time lags too close to the alpha rhythms of participants (which varied between ~9-11 Hz) or lags which would have a harmonic with the rhythm.“

(2) Line 168: Can the authors define what these conservative and liberal criteria are in the text?

We have added definitions of the criteria in the text. The text now reads:

“[..] significance thresholds (conservative, i.e. the maximum sequenceness across all permutations and timepoints or liberal criteria, i.e. the 95% percentile of aforementioned sequenceness).”

(3) Line 478: "calculate" instead of "calculated".

(4) Figure 7 D: y-axis is labeled "70 ms" I believe it should be labeled 80 ms.

Thanks, we fixed the two typos.

(5) With replay defined as sequential reactivation at a compressed temporal timescale, many of the iEEG citations (lines 54-55) do not demonstrate replay (they show stimulus reinstatement or ripple activity, but not sequential replay). Replay studies in humans using intracranial methods have been mostly limited to those measuring single-unit activity, a good example being Vaz et al., 2020 (https://www.science.org/doi/10.1126/science.aba0672).

We agree that, under a strict definition articulated by Genzel et al. that defines replay as sequential reactivation, many prior human iEEG studies are better described as stimulus reinstatement or ripple-related activity rather than true sequence replay. We have revised the text accordingly and now highlight the few intracranial microelectrode studies that demonstrate replay of firing sequences at the cellular/ensemble level in humans (Eichenlaub et al., 2020; Vaz et al., 2020), distinguishing these from macro-scale iEEG work providing indirect evidence alone.

The revised paragraph now reads:

“Replay has been shown using cellular recordings across a variety of mammalian model organisms (Hoffman & McNaughton, 2002; Lee & Wilson, 2002; Pavlides & Winson, 1989). Replay studies in humans using intracranial recordings are few, but include work demonstrating compressed replay of firing-pattern sequences in motor cortex during rest (Eichenlaub et al., 2020) as well as single-unit replay of trialspecific cortical spiking sequences during episodic retrieval (Vaz et al., 2020). By contrast, most iEEG studies report stimulus-specific reinstatement or ripple-locked activity changes without explicit demonstration of temporally compressed sequential replay (Axmacher et al., 2008; Staresina et al., 2015). As these methods are only applied under restricted clinical circumstances, such as during pre-operative neurosurgical assessments, this limits opportunities to investigate human replay. Therefore, this gives urgency to efforts aimed at developing novel methods to investigate human replay non-invasively.”

(6) The expectations about replay frequency are grounded in literature on hippocampal replay sequences. However, MEG captures signals from across the entire brain, and the hippocampal contribution is likely relatively weak compared to all other signals. This raises an important question: is TDLM genuinely unable to detect replay at physiological (i.e., hippocampal) levels, or is it instead detecting a different form of sequential reactivation - possibly involving cortex or other regions - that may occur more frequently? More broadly, when we have evidence of replay from TDLM, do we believe it is the same thing as replay of CA1 place cell spiking sequences, as detected in rodents? Commenting on this distinction would help further develop theories of replay and what TDLM is measuring.

This is indeed an important point that has garnered relatively little attention. While there is some evidence of a relation to hippocampal replay in form of high-frequency power increase in the hippocampus, ultimately it is not possible to know without intracranial recordings, as signal strength from those regions is rather poor in MEG.

We have added the following segment to the manuscript that discusses these issues:

“However, while we are using indices of SWRs as a proxy for replay density estimation, the relationship between hippocampal replay and replay detected by TDLM remains uncertain. While current decoding approaches measure replay-like phenomena on cortical sites, previous papers have reported a power increase in hippocampal areas coinciding with replay episodes as detected by TDLM. Nevertheless, it is conceivable that cortical replay found by TDLM could occur independently of hippocampal replay and SWRs and be generated by different mechanisms. Some TDLM-studies find a replay state-to-state time lag of above 100 ms, much slower than e.g. previously reported place cell replay. Future studies should employ simultaneous intracranial and cortical surface recordings to establish the relationship between hippocampal replay and replay found by TDLM.”

eLife Assessment

This study presents an assessment of the effect of lactate dehydrogenase (LDH) inhibition on the activity of glycolysis and tricarboxylic acid cycle. The data were collected and analyzed using solid and validated methodology. This paper makes a useful contribution to the field as it considers a control analysis of LDH flux. The findings differ from other published findings likely due to the time course of the incubations used to assess metabolism. While such comparative studies were not presented in the manuscript, the manuscript should be interpreted in light of this critical distinction.

Reviewer #1 (Public Review):

Summary:

Zeng et al. have investigated the impact of inhibiting lactate dehydrogenase (LDH) on glycolysis and the tricarboxylic acid cycle. LDH is the terminal enzyme of aerobic glycolysis or fermentation that converts pyruvate and NADH to lactate and NAD+ and is essential for the fermentation pathway as it recycles NAD+ needed by upstream glyceraldehyde-3-phosphate dehydrogenase. As the authors point out in the introduction, multiple published reports have shown that inhibition of LDH in cancer cells typically leads to a switch from fermentative ATP production to respiratory ATP production (i.e., glucose uptake and lactate secretion are decreased, and oxygen consumption is increased). The presumed logic of this metabolic rearrangement is that when glycolytic ATP production is inhibited due to LDH inhibition, the cell switches to producing more ATP using respiration. This observation is similar to the well-established Crabtree and Pasteur effects, where cells switch between fermentation and respiration due to the availability of glucose and oxygen. Unexpectedly, the authors observed that inhibition of LDH led to inhibition of respiration and not activation as previously observed. The authors perform rigorous measurements of glycolysis and TCA cycle activity, demonstrating that under their experimental conditions, respiration is indeed inhibited. Given the large body of work reporting the opposite result, it is difficult to reconcile the reasons for the discrepancy. In this reviewer's opinion, a reason for the discrepancy may be that the authors performed their measurements 6 hours after inhibiting LDH. Six hours is a very long time for assessing the direct impact of a perturbation on metabolic pathway activity, which is regulated on a timescale of seconds to minutes. The observed effects are likely the result of a combination of many downstream responses that happen within 6 hours of inhibiting LDH that causes a large decrease in ATP production, inhibition of cell proliferation, and likely a range of stress responses, including gene expression changes.

Strengths:

The regulation of metabolic pathways is incompletely understood, and more research is needed, such as the one conducted here. The authors performed an impressive set of measurements of metabolite levels in response to inhibition of LDH using a combination of rigorous approaches.

Weaknesses:

Glycolysis, TCA cycle, and respiration are regulated on a timescale of seconds to minutes. The main weakness of this study is the long drug treatment time of 6 hours, which was chosen for all the experiments. In this reviewer's opinion, if the goal was to investigate the direct impact of LDH inhibition on glycolysis and the TCA cycle, most of the experiments should have been performed immediately after or within minutes of LDH inhibition. After 6 hours of inhibiting LDH and ATP production, cells undergo a whole range of responses, and most of the observed effects are likely indirect due to the many downstream effects of LDH and ATP production inhibition, such as decreased cell proliferation, decreased energy demand, activation of stress response pathways, etc.

Comments on revisions:

Based on the response to comments that the authors have submitted, I do not think I need to make any changes to my review, as the time course experiment that could have explained the difference between reported results and extensive prior literature has not been performed.

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

Summary:

Zeng et al. have investigated the impact of inhibiting lactate dehydrogenase (LDH) on glycolysis and the tricarboxylic acid cycle. LDH is the terminal enzyme of aerobic glycolysis or fermentation that converts pyruvate and NADH to lactate and NAD+ and is essential for the fermentation pathway as it recycles NAD+ needed by upstream glyceraldehyde-3-phosphate dehydrogenase. As the authors point out in the introduction, multiple published reports have shown that inhibition of LDH in cancer cells typically leads to a switch from fermentative ATP production to respiratory ATP production (i.e., glucose uptake and lactate secretion are decreased, and oxygen consumption is increased). The presumed logic of this metabolic rearrangement is that when glycolytic ATP production is inhibited due to LDH inhibition, the cell switches to producing more ATP using respiration. This observation is similar to the well-established Crabtree and Pasteur effects, where cells switch between fermentation and respiration due to the availability of glucose and oxygen. Unexpectedly, the authors observed that inhibition of LDH led to inhibition of respiration and not activation as previously observed. The authors perform rigorous measurements of glycolysis and TCA cycle activity, demonstrating that under their experimental conditions, respiration is indeed inhibited. Given the large body of work reporting the opposite result, it is difficult to reconcile the reasons for the discrepancy. In this reviewer's opinion, a reason for the discrepancy may be that the authors performed their measurements 6 hours after inhibiting LDH. Six hours is a very long time for assessing the direct impact of a perturbation on metabolic pathway activity, which is regulated on a timescale of seconds to minutes. The observed effects are likely the result of a combination of many downstream responses that happen within 6 hours of inhibiting LDH that causes a large decrease in ATP production, inhibition of cell proliferation, and likely a range of stress responses, including gene expression changes.

Strengths:

The regulation of metabolic pathways is incompletely understood, and more research is needed, such as the one conducted here. The authors performed an impressive set of measurements of metabolite levels in response to inhibition of LDH using a combination of rigorous approaches.

Weaknesses:

Glycolysis, TCA cycle, and respiration are regulated on a timescale of seconds to minutes. The main weakness of this study is the long drug treatment time of 6 hours, which was chosen for all the experiments. In this reviewer's opinion, if the goal was to investigate the direct impact of LDH inhibition on glycolysis and the TCA cycle, most of the experiments should have been performed immediately after or within minutes of LDH inhibition. After 6 hours of inhibiting LDH and ATP production, cells undergo a whole range of responses, and most of the observed effects are likely indirect due to the many downstream effects of LDH and ATP production inhibition, such as decreased cell proliferation, decreased energy demand, activation of stress response pathways, etc.

We thank reviewer for the careful reading of our manuscript, the accurate summary of the prevailing model, and the positive assessment of the rigor of our measurements. We agree that much prior literature reports increased oxygen consumption following LDH inhibition, and we recognize that our finding—coordinated suppression of glycolysis, the TCA cycle, and OXPHOS—differs from this prevailing interpretation. We address below the reviewer’s main concern regarding the 6-hour time point and clarify the conceptual scope of our study.

(1) Scope: steady-state metabolic regulation versus immediate transient effects

The reviewer raises an important point that many metabolic perturbations can trigger rapid, transient responses within seconds to minutes, whereas our measurements were performed after sustained LDH inhibition. We agree that very early time points would be required if the primary goal were to isolate the most immediate, proximal consequence of LDH inhibition before downstream propagation. However, the objective of our study is different: we aim to characterize the metabolic steady state re-established after sustained inhibition of LDH activity, because this adapted steady state is more relevant for understanding long-term metabolic consequences and therapeutic outcomes of LDH inhibition in cancer cells.

(2) Genetic LDHA/LDHB knockout: comparison of two steady states

A related point applies to the LDHA/LDHB knockout models. We fully agree that the knockout process necessarily involves a temporal perturbation during cell line generation and adaptation. Nevertheless, the experimental comparison in our study is explicitly between two steady states: the baseline steady state of control cells and the steady state achieved after stable genetic disruption of LDHA or LDHB. The observation that LDHA or LDHB knockout alone had minimal effects on glycolysis and respiration indicates that partial reduction of LDH activity can be compensated in a steady-state manner, consistent with the exceptionally high catalytic capacity of LDH in cancer cells relative to upstream rate-limiting enzymes.

(3) LDH-activity-dependent quantitative relationships support stable metabolic states

Importantly, our conclusions do not rely on a single inhibitor condition at a single time point. Rather, we established quantitative steady-state relationships between residual LDH activity and pathway behavior across a wide range of LDH inhibition. These LDH-activity-dependent data strongly support that the system resides in stable metabolic states at different degrees of LDH activity, rather than reflecting non-specific collapse due to prolonged stress.

Specifically, we observed that when LDH activity was reduced from 100% to approximately ~9% (e.g., by genetic perturbation and partial pharmacologic inhibition), glucose consumption and lactate production remained essentially unchanged, indicating maintenance of a steady-state glycolytic flux despite substantial LDH inhibition. Only when LDH activity was further reduced below this threshold did glycolytic flux decrease in a graded manner, consistent with a nonlinear control structure (Figure 8 A & B)).

Likewise, the isotope tracing results showed distinct LDH-activity-dependent transitions in TCA cycle labeling patterns. Over the range in which LDH activity decreased from 100% to ~9%, the [¹³C₆]glucose-derived labeling pattern of citrate remained largely unchanged, whereas deeper inhibition led to a decrease in m2 citrate with a compensatory rise in higher-order citrate isotopologues, consistent with altered flux entry versus cycling/retention in the TCA cycle (Figure 8C). Similarly, [¹³C₅]glutamine tracing revealed that deeper LDH inhibition reduced the direct m5 contribution, accompanied by corresponding shifts in other isotopologues (Figure 8D). These graded, quantitative transitions—rather than an abrupt global failure—support the interpretation of distinct metabolic steady states across LDH activity levels, linking LDH inhibition to changes in both glycolysis and mitochondrial metabolism.

(4) Reconciling discrepancies with prior studies

We agree that multiple prior studies have reported increased oxygen consumption or enhanced oxidative metabolism following LDH inhibition in cancer cells. However, we note that this prevailing notion often persists because LDH inhibition is frequently discussed by analogy to the classical Pasteur and Crabtree effects, in which cells toggle between fermentation and respiration depending on oxygen and glucose availability. We believe this analogy can be misleading.

In the Pasteur effect, the metabolic shift is primarily driven by oxygen limitation, i.e., restriction of the terminal electron acceptor for the mitochondrial electron transport chain, which enforces reliance on fermentation. In the Crabtree effect, high glucose availability suppresses respiration through regulatory mechanisms while glycolysis is strongly activated. Both phenomena are fundamentally controlled by oxygen availability and respiratory capacity, rather than by inhibition of a specific cytosolic enzyme.

By contrast, LDH inhibition is mechanistically distinct: it directly perturbs cytosolic redox recycling by limiting NADH-to-NAD⁺ regeneration and can therefore constrain upstream glycolytic flux (particularly at GAPDH) and reshape pathway thermodynamics. Under conditions where LDH inhibition sufficiently limits effective NAD⁺ availability and reduces glycolytic flux into pyruvate, the downstream consequence is reduced carbon input into the TCA cycle and suppressed OXPHOS—consistent with our experimental measurements. We therefore suggest that divergent outcomes reported across studies likely reflect differences in residual LDH activity, cell-type–specific metabolic wiring, and the extent to which glycolytic flux remains sustained versus becoming redox-limited upstream, rather than a universal Pasteur/Crabtree-like “switch” from fermentation to respiration. Accordingly, interpreting LDH inhibition as a Pasteur/Crabtree-like toggle may oversimplify the biochemical consequences of disrupting cytosolic NAD⁺ regeneration.

We have revised the Discussion to clarify this conceptual distinction and to avoid relying on comparisons that are not mechanistically equivalent to LDH inhibition.

Reviewer #2 (Public Review):

Summary:

Zeng et al. investigated the role of LDH in determining the metabolic fate of pyruvate in HeLa and 4T1 cells. To do this, three broad perturbations were applied: knockout of two LDH isoforms (LDH-A and LDH-B), titration with a non-competitive LDH inhibitor (GNE-140), and exposure to either normoxic (21% O2) or hypoxic (1% O2) conditions. They show that knockout of either LDH isoform alone, though reducing both protein level and enzyme activity, has virtually no effect on either the incorporation of a stable 13C-label from a 13C6-glucose into any glycolytic or TCA cycle intermediate, nor on the measured intracellular concentrations of any glycolytic intermediate (Figure 2). The only apparent exception to this was the NADH/NAD+ ratio, measured as the ratio of F420/F480 emitted from a fluorescent tag (SoNar).

The addition of a chemical inhibitor, on the other hand, did lead to changes in glycolytic flux, the concentrations of glycolytic intermediates, and in the NADH/NAD+ ratio (Figure 3). Notably, this was most evident in the LDH-B-knockout, in agreement with the increased sensitivity of LDH-A to GNE-140 (Figure 2). In the LDH-B-knockout, increasing concentrations of GNE-140 increased the NADH/NAD+ ratio, reduced glucose uptake, and lactate production, and led to an accumulation of glycolytic intermediates immediately upstream of GAPDH (GA3P, DHAP, and FBP) and a decrease in the product of GAPDH (3PG). They continue to show that this effect is even stronger in cells exposed to hypoxic conditions (Figure 4). They propose that a shift to thermodynamic unfavourability, initiated by an increased NADH/NAD+ ratio inhibiting GAPDH explains the cascade, calculating ΔG values that become progressively more endergonic at increasing inhibitor concentrations.

Then - in two separate experiments - the authors track the incorporation of 13C into the intermediates of the TCA cycle from a 13C6-glucose and a 13C5-glutamine. They use the proportion of labelled intermediates as a proxy for how much pyruvate enters the TCA cycle (Figure 5). They conclude that the inhibition of LDH decreases fermentation, but also the TCA cycle and OXPHOS flux - and hence the flux of pyruvate to all of those pathways. Finally, they characterise the production of ATP from respiratory or fermentative routes, the concentration of a number of cofactors (ATP, ADP, AMP, NAD(P)H, NAD(P)+, and GSH/GSSG), the cell count, and cell viability under four conditions: with and without the highest inhibitor concentration, and at norm- and hypoxia. From this, they conclude that the inhibition of LDH inhibits the glycolysis, the TCA cycle, and OXPHOS simultaneously (Figure 7).

Strengths:

The authors present an impressively detailed set of measurements under a variety of conditions. It is clear that a huge effort was made to characterise the steady-state properties (metabolite concentrations, fluxes) as well as the partitioning of pyruvate between fermentation as opposed to the TCA cycle and OXPHOS.

A couple of intermediary conclusions are well supported, with the hypothesis underlying the next measurement clearly following. For instance, the authors refer to literature reports that LDH activity is highly redundant in cancer cells (lines 108 - 144). They prove this point convincingly in Figure 1, showing that both the A- and B-isoforms of LDH can be knocked out without any noticeable changes in specific glucose consumption or lactate production flux, or, for that matter, in the rate at which any of the pathway intermediates are produced. Pyruvate incorporation into the TCA cycle and the oxygen consumption rate are also shown to be unaffected.

They checked the specificity of the inhibitor and found good agreement between the inhibitory capacity of GNE-140 on the two isoforms of LDH and the glycolytic flux (lines 229 - 243). The authors also provide a logical interpretation of the first couple of consequences following LDH inhibition: an increased NADH/NAD+ ratio leading to the inhibition of GAPDH, causing upstream accumulations and downstream metabolite decreases (lines 348 - 355).

Weaknesses:

Despite the inarguable comprehensiveness of the data set, a number of conceptual shortcomings afflict the manuscript. First and foremost, reasoning is often not pursued to a logical conclusion. For instance, the accumulation of intermediates upstream of GAPDH is proffered as an explanation for the decreased flux through glycolysis. However, in Figure 3C it is clear that there is no accumulation of the intermediates upstream of PFK. It is unclear, therefore, how this traffic jam is propagated back to a decrease in glucose uptake. A possible explanation might lie with hexokinase and the decrease in ATP (and constant ADP) demonstrated in Figure 6B, but this link is not made.

We appreciate the reviewer's critical comment. In Figure 3C, there is no accumulation of F6P or G6P, which are upstream of PFK1. This is because the PFK1-catalyzed reaction sets a significant thermodynamic barrier. Even with treatment using 30 μM GNE-140, the ∆G_PFK1 (Gibbs free energy of the PFK1-catalyzed reaction) remains -9.455 kJ/mol (Figure 3D), indicating that the reaction is still far from thermodynamic equilibrium, thereby preventing the accumulation of F6P and G6P.

We agree with the reviewer that hexokinase inhibition may play a role, this requires further investigation.

The obvious link between the NADH/NAD+ ratio and pyruvate dehydrogenase (PDH) is also never addressed, a mechanism that might explain how the pyruvate incorporation into the TCA cycle is impaired by the inhibition of LDH (the observation with which they start their discussion, lines 511 - 514).

We agree with the reviewer’s comment. In this study, we did not explore how the inhibition of LDH affects pyruvate incorporation into the TCA cycle. As this mechanism was not investigated, we have titled the study:

"Elucidating the Kinetic and Thermodynamic Insights into the Regulation of Glycolysis by Lactate Dehydrogenase and Its Impact on the Tricarboxylic Acid Cycle and Oxidative Phosphorylation in Cancer Cells."

It was furthermore puzzling how the ΔG, calculated with intracellular metabolite concentrations (Figures 3 and 4) could be endergonic (positive) for PGAM at all conditions (also normoxic and without inhibitor). This would mean that under the conditions assayed, glycolysis would never flow completely forward. How any lactate or pyruvate is produced from glucose, is then unexplained.

This issue also concerned me during the study. However, given the high reproducibility of the data, we consider it is true, but requires explanation. The PGAM-catalyzed reaction is tightly linked to both upstream and downstream reactions in the glycolytic pathway. In glycolysis, three key reactions catalyzed by HK2, PFK1, and PK are highly exergonic, providing the driving force for the conversion of glucose to pyruvate. The other reactions, including the one catalyzed by PGAM, operate near thermodynamic equilibrium and primarily serve to equilibrate glycolytic intermediates rather than control the overall direction of glycolysis, as previously described by us (J Biol Chem. 2024 Aug8;300(9):107648).

The endergonic nature of the PGAM-catalyzed reaction does not prevent it from proceeding in the forward direction. Instead, the directionality of the pathway is dictated by the exergonic reaction of PFK1 upstream, which pushes the flux forward, and by PK downstream, which pulls the flux through the pathway. The combined effects of PFK1 and PK may account for the observed endergonic state of the PGAM reaction.

However, if the PGAM-catalyzed reaction were isolated from the glycolytic pathway, it would tend toward equilibrium and never surpass it, as there would be no driving force to move the reaction forward.

Finally, the interpretation of the label incorporation data is rather unconvincing. The authors observe an increasing labelled fraction of TCA cycle intermediates as a function of increasing inhibitor concentration. Strangely, they conclude that less labelled pyruvate enters the TCA cycle while simultaneously less labelled intermediates exit the TCA cycle pool, leading to increased labelling of this pool. The reasoning that they present for this (decreased m2 fraction as a function of DHE-140 concentration) is by no means a consistent or striking feature of their titration data and comes across as rather unconvincing. Yet they treat this anomaly as resolved in the discussion that follows.

GNE-140 treatment increased the labeling of TCA cycle intermediates by [¹³C₆]glucose but decreased the OXPHOS rate, we consider the conflicting results as an 'anomaly' that warrants further explanation. To address this, we analyzed the labeling pattern of TCA cycle intermediates using both [¹³C₆]glucose and [¹³C₅]glutamine. Tracing the incorporation of glucose- and glutamine-derived carbons into the TCA cycle suggests that LDH inhibition leads to a reduced flux of glucose-derived acetyl-CoA into the TCA cycle, coupled with a decreased flux of glutamine-derived α-KG, and a reduction in the efflux of intermediates from the cycle. These results align with theoretical predictions. Under any condition, the reactions that distribute TCA cycle intermediates to other pathways must be balanced by those that replenish them. In the GNE-140 treatment group, the entry of glutamine-derived carbon into the TCA cycle was reduced, implying that glucose-derived carbon (as acetyl-CoA) entering the TCA cycle must also be reduced, or vice versa.

This step-by-step investigation is detailed under the subheading "The Effect of LDHB KO and GNE-140 on the Contribution of Glucose Carbon to the TCA Cycle and OXPHOS" in the Results section in the manuscript.

In the Discussion, we emphasize that caution should be exercised when interpreting isotope tracing data. In this study, treatment of cells with GNE-140 led to an increase labeling percentage of TCA cycle intermediates by [¹³C₆]glucose (Figure 5A-E). However, this does not necessarily imply an increase in glucose carbon flux into TCA cycle; rather, it indicates a reduction in both the flux of glucose carbon into TCA cycle and the flux of intermediates leaving TCA cycle. When interpreting the data, multiple factors must be considered, including the carbon-13 labeling pattern of the intermediates (m1, m2, m3, ---) (Figure 5G-K), replenishment of intermediates by glutamine (Figure 5M-V), and mitochondrial oxygen consumption rate (Figure 5W). All these factors should be taken into account to derive a proper interpretation of the data.

Reviewer #3 (Public Review):

Hu et al in their manuscript attempt to interrogate the interplay between glycolysis, TCA activity, and OXPHOS using LDHA/B knockouts as well as LDH-specific inhibitors. Before I discuss the specifics, I have a few issues with the overall manuscript. First of all, based on numerous previous studies it is well established that glycolysis inhibition or forcing pyruvate into the TCA cycle (studies with PDKs inhibitors) leads to upregulation of TCA cycle activity, and OXPHOS, activation of glutaminolysis, etc (in this work authors claim that lowered glycolysis leads to lower levels of TCA activity/OXPHOS). The authors in the current work completely ignore recent studies that suggest that lactate itself is an important signaling metabolite that can modulate metabolism (actual mechanistic insights were recently presented by at least two groups (Thompson, Chouchani labs). In addition, extensive effort was dedicated to understanding the crosstalk between glycolysis/TCA cycle/OXPHOS using metabolic models (Titov, Rabinowitz labs). I have several comments on how experiments were performed. In the Methods section, it is stated that both HeLa and 4T1 cells were grown in RPMI-1640 medium with regular serum - but under these conditions, pyruvate is certainly present in the medium - this can easily complicate/invalidate some findings presented in this manuscript. In LDH enzymatic assays as described with cell homogenates controls were not explained or presented (a lot of enzymes in the homogenate can react with NADH!). One of the major issues I have is that glycolytic intermediates were measured in multiple enzyme-coupled assays. Although one might think it is a good approach to have quantitative numbers for each metabolite, the way it was done is that cell homogenates (potentially with still traces of activity of multiple glycolytic enzymes) were incubated with various combinations of the SAME enzymes and substrates they were supposed to measure as a part of the enzyme-based cycling reaction. I would prefer to see a comparison between numbers obtained in enzyme-based assays with GC-MS/LC-MS experiments (using calibration curves for respective metabolites, of course). Correct measurements of these metabolites are crucial especially when thermodynamic parameters for respective reactions are calculated. Concentrations of multiple graphs (Figure 1g etc.) are in "mM", I do not think that this is correct.

We thank the reviewer’s comment and the following are clarification of the conceptual framework, the quantitative methodology, and the experimental basis supporting our conclusions.

(1) “It is well established that glycolysis inhibition or forcing pyruvate into the TCA cycle… leads to upregulation of TCA/OXPHOS… (authors claim lowered glycolysis leads to lower TCA/OXPHOS)”

This framing is not accurate in the context of our study. PDK inhibition and LDH inhibition are fundamentally different perturbations. PDK inhibition directly promotes mitochondrial pyruvate oxidation by enabling PDH flux, whereas LDH inhibition primarily perturbs cytosolic redox balance (free NADH/NAD⁺) and thereby constrains upstream glycolytic reactions, particularly the GAPDH step. Therefore, the metabolic outcomes of these interventions are not expected to be identical and should not be treated as interchangeable.

Importantly, we do not “ignore” prior studies proposing increased OXPHOS after LDH inhibition; we explicitly cite and summarize this prevailing interpretation in the Introduction. Our study was motivated precisely because this interpretation does not resolve key quantitative inconsistencies, including (i) the large mismatch between glycolytic flux and mitochondrial oxidative capacity, and (ii) the exceptionally high catalytic capacity of LDH relative to upstream rate-limiting glycolytic enzymes. These constraints raise a mechanistic question: how does LDH inhibition actually suppress glycolytic flux in intact cancer cells, and what are the consequences for TCA cycle and OXPHOS?

Our central contribution is the identification of a biochemical mechanism supported by integrated measurements of fluxes, metabolite concentrations, redox state, and reaction thermodynamics: LDH inhibition increases free NADH/NAD⁺, decreases free NAD⁺ availability, inhibits GAPDH, drives accumulation/depletion patterns in glycolytic intermediates, shifts Gibbs free energies of near-equilibrium reactions (PFK1–PGAM segment), suppresses pyruvate production, and consequently reduces carbon input into TCA cycle and OXPHOS. These analyses are not provided by most prior work and directly address the mechanistic gap.

(2) Lactate signaling (Thompson/Chouchani) and metabolic modeling (Titov/Rabinowitz)

These research directions are valuable, but they address questions that are different from the one investigated here. Our manuscript focuses on steady-state biochemical control of metabolic flux by LDH inhibition through redox-linked kinetics and pathway thermodynamics.

(3) Pyruvate in RPMI

Pyruvate in standard medium does not invalidate our conclusions. All experimental comparisons were performed under identical conditions across groups, and the major conclusions rely on orthogonal measurements including glycolytic flux (glucose consumption/lactate production), OCR profiling, and isotope tracing with [¹³C₆]glucose and [¹³C₅] glutamine, which directly quantify carbon entry into lactate and TCA cycle intermediates. These tracer-based results are not confounded by unlabeled extracellular pyruvate in a way that would reverse the mechanistic conclusions.

(4) LDH activity assay in homogenates and “many enzymes can react with NADH”

This concern is overstated. In the LDH assay, substrates are pyruvate + NADH, and the measured signal reflects NADH oxidation coupled to pyruvate reduction. In cell lysates, LDH is uniquely abundant and catalytically efficient for this reaction pair, and the inhibitor-response behavior matches the known LDHA/LDHB selectivity of GNE-140 and the cellular phenotypes. Thus, the assay is mechanistically specific in this context.

(5) Enzyme-coupled metabolite assays and request for LC–MS validation

The reviewer’s implication that enzyme-coupled assays are intrinsically unreliable is incorrect. Enzymatic cycling assays are a widely used quantitative approach when performed with proper specificity and calibration, and they are particularly useful for labile glycolytic intermediates that are challenging to quantify reproducibly by MS without specialized quenching, derivatization, and isotope dilution standards.

We agree that MS-based quantification is valuable, and we have developed LC–MS methods for selected metabolites. However, absolute quantification of these intermediates remains technically difficult due to the inherent limitation of this method and, in our hands, did not provide uniformly robust performance for all intermediates required for thermodynamic analysis.

(6) Units (“mM”)

The metabolite concentration units are correct.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

If the goal is to investigate the direct impact of LDH inhibition, then in my opinion, most of these experiments need to be repeated at a very early time point immediately after or a few minutes after LDH inhibition. I understand that this is a tremendous amount of work that the authors might not want to pursue. I do want to highlight that the quality of the experiments performed in this work is impressive. I hope the authors continue investigating this subject and look forward to reading their future manuscripts on this topic.

We thank the reviewer for this thoughtful and constructive comment and for the positive assessment of the experimental quality of our work.

We fully agree that measurements at very early time points after LDH inhibition would be required if the goal were to isolate an immediate, proximal molecular event occurring before downstream propagation. However, the primary objective of our study is not to dissect a single instantaneous biochemical consequence of LDH inhibition, but rather to characterize the metabolic steady state that is re-established after sustained suppression of LDH activity, which we believe is more relevant for understanding the long-term metabolic and therapeutic consequences of LDH inhibition in cancer cells.

(1) Scope: steady-state metabolic regulation versus immediate transient effects

The reviewer raises an important point that many metabolic perturbations can trigger rapid, transient responses within seconds to minutes, whereas our measurements were performed after sustained LDH inhibition. We agree that very early time points would be required if the primary goal were to isolate the most immediate, proximal consequence of LDH inhibition before downstream propagation. However, the objective of our study is different: we aim to characterize the metabolic steady state re-established after sustained inhibition of LDH activity, because this adapted steady state is more relevant for understanding long-term metabolic consequences and therapeutic outcomes of LDH inhibition in cancer cells.

(2) Genetic LDHA/LDHB knockout: comparison of two steady states

A related point applies to the LDHA/LDHB knockout models. We fully agree that the knockout process necessarily involves a temporal perturbation during cell line generation and adaptation. Nevertheless, the experimental comparison in our study is explicitly between two steady states: the baseline steady state of control cells and the steady state achieved after stable genetic disruption of LDHA or LDHB. The observation that LDHA or LDHB knockout alone had minimal effects on glycolysis and respiration indicates that partial reduction of LDH activity can be compensated in a steady-state manner, consistent with the exceptionally high catalytic capacity of LDH in cancer cells relative to upstream rate-limiting enzymes.

(3) LDH-activity-dependent quantitative relationships support stable metabolic states

Importantly, our conclusions do not rely on a single inhibitor condition at a single time point. Rather, we established quantitative steady-state relationships between residual LDH activity and pathway behavior across a wide range of LDH inhibition. These LDH-activity-dependent data strongly support that the system resides in stable metabolic states at different degrees of LDH activity, rather than reflecting non-specific collapse due to prolonged stress.

Specifically, we observed that when LDH activity was reduced from 100% to approximately ~9% (e.g., by genetic perturbation and partial pharmacologic inhibition), glucose consumption and lactate production remained essentially unchanged, indicating maintenance of a steady-state glycolytic flux despite substantial LDH inhibition. Only when LDH activity was further reduced below this threshold did glycolytic flux decrease in a graded manner, consistent with a nonlinear control structure.

Likewise, the isotope tracing results showed distinct LDH-activity-dependent transitions in TCA cycle labeling patterns. Over the range in which LDH activity decreased from 100% to ~9%, the [¹³C₆]glucose-derived labeling pattern of citrate remained largely unchanged, whereas deeper inhibition led to a decrease in m2 citrate with a compensatory rise in higher-order citrate isotopologues, consistent with altered flux entry versus cycling/retention in the TCA cycle. Similarly, [¹³C₅]glutamine tracing revealed that deeper LDH inhibition reduced the direct m5 contribution, accompanied by corresponding shifts in other isotopologues. These graded, quantitative transitions—rather than an abrupt global failure—support the interpretation of distinct metabolic steady states across LDH activity levels, linking LDH inhibition to changes in both glycolysis and mitochondrial metabolism.

Reviewer #2 (Recommendations For The Authors):

All in all, the authors would benefit from collaboration with a group more well-versed in quantitative aspects of metabolism (such as Metabolic Control Analysis) and modelling methods (such as flux analysis) to boost the interpretation and impact of their really nice data set.

We sincerely thank the reviewer for this insightful and constructive suggestion. We fully agree that collaboration with groups specializing in quantitative metabolic analysis, such as Metabolic Control Analysis and flux modeling, would further expand the interpretative depth and broader impact of this work.

The primary objective of the present work, however, was not to construct a global mathematical model, but to experimentally dissect the biochemical mechanism by which LDH inhibition coordinately suppresses glycolysis, the TCA cycle, and OXPHOS, integrating enzyme kinetics with thermodynamic constraints at steady state. Within this scope, we focused on experimentally demonstrable relationships between LDH activity, redox balance, GAPDH perturbation, thermodynamic shifts in near-equilibrium reactions, and emergent flux suppression.

We fully recognize the power of MCA and related modeling approaches in formalizing control coefficients and system-level sensitivities, and we view our dataset as particularly well suited to support such future analyses. We therefore see this work as providing a robust experimental platform upon which more comprehensive quantitative modeling can be built, either in future studies or through collaboration with specialists in metabolic modeling.

Reviewer #3 (Recommendations For The Authors):

We sincerely thank the reviewer for the important suggestions.

(1) I strongly disagree that "regulation of glycolytic flux".. "remained largely unexplored.”

Our original wording was meant to emphasize not the absence of prior work on glycolytic flux regulation, but rather that the specific biochemical mechanism by which LDH regulates glycolytic flux—particularly through the integrated effects of enzyme kinetics, redox balance, and thermodynamic constraints within the pathway—has not been fully elucidated.

To avoid any ambiguity or overstatement, we have revised the relevant text to more precisely reflect this intent. The revised wording now reads:

“This study elucidates a biochemical mechanism by which lactate dehydrogenase influences glycolytic flux in cancer cells, revealing a kinetic–thermodynamic interplay that contributes to metabolic regulation.”

We believe this revised phrasing more accurately acknowledges prior work while clearly defining the specific mechanistic contribution of the present study.

(2) Very confusing in the Introduction section: "If LDH is inhibited at the LDH step..”

We sincerely thank the reviewer for pointing out the potential confusion caused by the phrase “If LDH is inhibited at the LDH step” in the Introduction.

Our intention was to contrast two conceptual models of LDH inhibition. The first is the conventional view, in which the effect of LDH inhibition is assumed to be confined to the LDH-catalyzed reaction itself, leading primarily to local accumulation of pyruvate and its redirection toward mitochondrial metabolism. The second, which is supported by our data, is that LDH inhibition initiates a system-wide biochemical response, perturbing redox balance, upstream enzyme kinetics, and the thermodynamic state of the glycolytic pathway, ultimately resulting in coordinated suppression of glycolysis, the TCA cycle, and OXPHOS.

We agree that the original phrasing was ambiguous and potentially misleading. To improve clarity, we have revised the text as follows:

“If the effect of LDH inhibition were confined solely to its catalytic step…”

(3) The entire introduction part when the authors attempt to explain how decreased glycolysis will lead to decreased mitochondrial respiration is confusing.

We would like to clarify that the Introduction does not attempt to explain how decreased glycolysis leads to decreased mitochondrial respiration. Rather, the final paragraph of the Introduction is intended to highlight an unresolved conceptual inconsistency in the existing literature and to motivate the central question addressed in this study.

Specifically, we summarize the prevailing view that LDH inhibition redirects pyruvate toward mitochondrial metabolism and enhances oxidative phosphorylation, and then point out that this interpretation is difficult to reconcile with quantitative considerations, such as the large disparity between glycolytic and mitochondrial flux capacities and the excess catalytic activity of LDH relative to upstream glycolytic enzymes. These observations are presented to emphasize that the biochemical mechanism linking LDH inhibition to changes in glycolysis and mitochondrial respiration has not been fully resolved.

Importantly, the Introduction does not propose a mechanistic explanation for the observed suppression of mitochondrial respiration; rather, it poses this as an open question, which is then systematically addressed through experimental analysis in the Results section.

(4) Line 144: "which is 81(HeLa-LDHAKO) -297(HeLa-Ctrl) times"- here and in many other places wording is confusing to the reader.

Our intention was to emphasize the significant redundancy of LDH activity relative to hexokinase (HK), the first rate-limiting enzyme in the glycolysis pathway, in cancer cells.

Specifically, we wanted to express that in HeLa-Ctrl cells, the total LDH activity is 297 times that of HK activity; while in HeLa-LDHAKO cells, although the total LDH activity decreased, it was still 81 times that of HK activity. This data comes from supplement Table 1 in the paper and aims to provide quantitative evidence for "why knocking out LDHA or LDHB alone is insufficient to significantly affect glycolysis flux," because the remaining LDH activity is still far higher than the HK activity at the pathway entrance, sufficient to maintain flux.

Based on your suggestion, we rewrite it in the revised draft with a more specific statement: "...the total activity of LDH in HeLa cells is very high, which is 297-fold higher than the first rate-limiting enzyme HK activity in HeLa-Ctrl cells and 81-fold higher in HeLa-LDHAKO cells.”

(5) Line 153: "in the following four aspects:"- but what are these aspects, the text below has no corresponding subtitles, etc.

Our intention was to indicate that after LDHA or LDHB knockout alone failed to affect the glycolysis rate, we further explored its potential impact on the glycolytic pathway from four deeper perspectives: the glucose carbon to pyruvate and lactate, the glucose carbon to subsidiary branches of glycolysis, the concentration of glycolytic intermediates and the thermodynamic state of the pathway, and the redox state of cytosolic free NADH/NAD⁺.

Following your valuable suggestion, we have now added the aforementioned clear subtitles to these four aspects in the revised manuscript.

(6) Lines 193, another example of the very confusing statement: "The results suggested that the loss of total LDH concentration was compensated.."

The actual catalytic activity (reaction rate) of LDH is determined by both its enzyme concentration and substrate concentration (pyruvate and NADH). When the total LDH protein concentration (enzyme amount) in the cell is reduced through gene knockout, the reaction equilibrium is disrupted. To maintain sufficient lactate production flux to support a high glycolysis rate, the cell compensates by increasing the concentration of one of the substrates—free NADH (as shown in Figure 1I). This results in an increased substrate concentration, despite a reduction in the amount of enzyme, thus partially maintaining the overall reaction rate.

We have revised the original statement to more accurately describe this kinetic equilibrium process: "The decrease in total LDH concentration was counterbalanced by a concomitant increase in the concentration of its substrate, free NADH, thereby maintaining the reaction velocity.”

(7) Line 222-223: "did not or marginally significantly affect....”

Our intention is to reflect the complexity of the data in Figure 1. Specifically: Regarding "did not affect": This means that there were no statistically significant differences in most key parameters, such as glycolytic flux (glucose consumption rate, lactate production rate). Regarding "or marginally significantly affected": This means that in a few indicators, although statistical calculations showed p-values less than 0.05, the absolute value of the difference was very small, with limited biological significance.

To clarify this, we rewrite it as: "...did not significantly affect glucose-derived pyruvate entering into TCA cycle, neither significantly affect mitochondrial respiration, although statistically significant but minimal changes were observed in a few specific parameters (e.g., m3-pyruvate% in medium).”

(8) It is very confusing to use the same colors for three GNE-140 drug concentrations (Figure 2a-b) and for 3 different cell lines right next to each other (Figure 2c-d).

The figures have been revised accordingly.

(9) Lines 263-273: nothing is new here as oxidized NAD+ is required for run glycolysis and LDH inhibition/KO leads to a high NADH/NAD+ ratio; Also below it is well known that reductive stress blocks serine biosynthesis;

It is well established that oxidized NAD⁺ is required for glycolysis, that LDH inhibition or knockout increases the NADH/NAD⁺ ratio, and that reductive stress can suppress serine biosynthesis. We did not intend to present these observations as novel.

The key point of this section is not the qualitative requirement of NAD⁺ for GAPDH, but rather the mechanistic alignment between LDH inhibition, changes in free NAD⁺ availability, and the emergence of GAPDH as a flux-controlling step within the glycolytic pathway under steady-state conditions. Previous studies have largely treated the increase in NADH/NAD⁺ following LDH inhibition as a correlative or downstream effect, without directly demonstrating how this redox shift quantitatively propagates upstream to reorganize glycolytic flux distribution and thermodynamic driving forces.

In our study, we explicitly link LDH inhibition to (i) an increase in free NADH/NAD⁺ ratio, (ii) inhibition of GAPDH activity in intact cells, (iii) accumulation of upstream glycolytic intermediates, (iv) suppression of serine biosynthesis from 3-phosphoglycerate, and critically, (v) coordinated shifts in the Gibbs free energies of reactions between PFK1 and PGAM. This integrated kinetic–thermodynamic framework goes beyond the established qualitative understanding of NAD⁺ dependence and provides a pathway-level mechanism by which LDH activity controls glycolytic flux.

(10) Lines 368-370: "... we reached an alternative interpretation of the data.."- does not provide much confidence.

Our intention was to prudently emphasize that we proposed a new interpretation based on detailed data, differing from conventional views. Our interpretation is grounded in key and consistent evidence from dual isotope tracing experiments using [¹³C₆]glucose and [¹³C₅]glutamine: The [¹³C₆]glucose tracing data: the labeling pattern of citrate, the starting product of TCA cycle, showed a significant decrease in m+2 %. This directly reflects a reduction in the flux of newly generated acetyl-CoA from glucose entering the TCA cycle. Simultaneously, the sum of other isotopologues % (m+1/ m+3/ m+4/m+5/m+6) increased, indicating a longer retention time of the labeled carbon in the cycle, implying a simultaneous decrease in the flux of cycle intermediates effluxed for biosynthesis. [¹³C₅]Glutamine tracing data: the labeling pattern of α-ketoglutarate showed a decrease in m+5 %, indicating a reduction in glutamine replenishment flux. The pattern of change in the total percentage of other isotopologues % (m+1/ m+2/ m+3/m+4) also supports the conclusion of reduced intermediate product efflux.

These two sets of data corroborate each other, pointing to a unified conclusion: LDH inhibition not only reduces carbon source inflow into the TCA cycle but also decreases intermediate product efflux, leading to a decrease in overall cycle activity. Therefore, our "alternative interpretation" is a well-supported and more consistent explanation of our overall experimental results. We revise the original wording to: "Integrated analysis of dual isotope tracing data demonstrates that LDH inhibition reduces both influx and efflux of the TCA cycle..."

(11) Lines 418-421: This entire discussion on how TCA cycle activity is decreased upon LDH inhibition is very confusing. I also would like to see these tracer studies when ETC is inhibited with different inhibitors.

We would like to clarify that the mitochondrial respiration rate data presented in Figure 5W are based on studies using different ETC inhibitors, and the cell treatment conditions (including culture time, etc.) for these oxygen consumption measurements are consistent with the conditions for the [¹³C₆]glucose and [¹³C₅]glutamine isotope tracing experiments (Figure 5A-V). Therefore, the changes in TCA cycle flux revealed by the tracing data and the inhibition of OXPHOS rate shown by the respiration measurements are mutually corroborating evidence from the same experimental conditions.

(12) Figure 6F, G - very limited representation of growth curves, why not perform these experiments with all corresponding cell lines and over multiple days. Especially since proliferation arrest vs cell death was implicated.

We have provided the growth curves of the HeLa-Ctrl and HeLa-LDHAKO cell lines under the corresponding treatments in Figure 6—figure supplement 1, as a supplement to Figure 6F, G (HeLa-LDHBKO cells). The choice of 48 hours as the cutoff observation point is based on clear biological evidence: under the stress of hypoxia (1% O₂) combined with GNE-140 treatment, HeLa-LDHBKO cells experienced substantial death within 24 to 48 hours, at which point the differences in the growth curves were already very significant.

(13) Move most of the Supplementary tables into an Excel file - so values can be easily accessed.

We have compiled the tables into an Excel file and submitted it along with the revised manuscript as supplementary material.

(14) Consider changing colors to more appealing- especially jarring is a bright blue, red, black combination on many bar graphs.

We have adjusted the color scheme of the figures (especially the bar graphs) in the paper, and have submitted them with the revised manuscript.

(15) Double check y-axis on multiple graphs it says "mM".

We have checked y-axis, the unit (mM) is correct.

(16) Instead TCA cycle use the TCA cycle.

In the revised manuscript, TCA cycle is used.

eLife Assessment

This valuable study aims to determine mechanisms underlying breast cancer initiation and tumour progression. The manuscript includes a solid set of transcriptomic and proteomic datasets from tumour samples and examines mitochondrial function within the tumours. While the underlying mechanisms linking expression changes to functional effects remain speculative. This paper provides a resource for researchers working on breast cancer and/or HER2-driven bioenergetics changes.

Reviewer #1 (Public review):

Summary:

In this manuscript, Frangos at al. used a transcriptomic and proteomic approach to characterise changes in HER2-driven mammary tumours compared to healthy mammary tissue in mice. They observed that mitochondrial genes, including OXPHOS regulators, were among the most down-regulated genes and proteins in their datasets. Surprisingly, these were associated with higher mitochondrial respiration, in response to a variety of carbon sources. In addition, there seems to be a reduction in mitochondrial fusion and an increase in fission in tumour tissues compared to healthy tissues.

Strengths:

The data are clearly presented and described.

The author reported very similar trends in proteomic and transcriptomic data. Such approaches are essential to have a better understanding of the changes in cancer cell metabolism associated with tumorigenesis.

The authors provided a direct link between HER2 inhibition and OXPHOS, strengthening the mechanistic aspect of the work.

Weaknesses:

The manuscript would have benefited from more ex-vivo approaches to further dissect mechanistic links and resolve the contradiction of elevated respiration with reduced expression of most associated proteins (but these points are clearly articulated in the discussion).

The results presented support the authors' conclusions, and limitations are addressed in the discussion. This work will likely impact the progression of the field, and the provided data will benefit the scientific community.

Comments on revisions:

The authors addressed all my concerns.

Reviewer #2 (Public review):

Frangos et al present a set of studies aiming to determine mechanisms underlying initiation and tumour progression. Overall, this work provides some useful datasets, further establishing mitochondrial dysfunction during the cellular transformation process.

A key strength is the coordinated analysis of transcriptomics and proteomics from tumour samples derived from a Neu-dependent mouse model for breast cancer. This analysis provides rigorous datasets that show robust patterns, including down-regulation across many components of mitochondrial OXPHOS that were generally consistent at both the mRNA and protein level. Parallel analysis of corresponding tumour samples thereby clearly shows the opposite trend of increased mitochondrial function, which is unexpected. As such, this work further establishes altered mitochondrial phenotypes in tumour contexts and further illustrates that mitochondrial function is not necessarily always tightly correlated with mitochondrial gene expression patterns.

Several key weaknesses remain. It remains unclear how increased mitochondrial function is being sustained despite wide decreases in mRNA and protein levels of OXPHOS components. In terms of mechanism, the study confirmed that pharmacologic EGFR inhibition decreases OXPHOS in a EGFR-dependent breast cancer line. However, it remains unclear if the cell culture system recapitulates other key observations of the tumour model (namely decreased expression with increased function).

Therefore, the mechanistic basis of increased mitochondrial function in light of decreased mitochondrial content remains speculative, as does the role of these changes for tumour initiation or progression.

Comments on revisions:

We agree with the overall findings of the study and appreciate that the claims in text and title have been appropriately toned down.

As additional suggestions eg for presentation, many of the graphics/labels are still too small to be useful. It would be interesting to see if this cell line is similar to the tumours in terms of all the phenotypes. The lapatinib experiment was good. I wonder how quick this drug affects the mitochondria. Also it would be interesting to see if these cells have higher OXPHOS than other non-transformed breast epithelial cells.

The WB on oxphos components is good with ab110413 but this looks like many subunits are detected so this should be made clear.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

In this manuscript, Frangos et al. used a transcriptomic and proteomic approach to characterise changes in HER2-driven mammary tumours compared to healthy mammary tissue in mice. They observed that mitochondrial genes, including OXPHOS regulators, were among the most down-regulated genes and proteins in their datasets. Surprisingly, these were associated with higher mitochondrial respiration, in response to a variety of carbon sources. In addition, there seems to be a reduction in mitochondrial fusion and an increase in fission in tumours compared to healthy tissues.

Strengths:

The data are clearly presented and described.

The author reported very similar trends in proteomic and transcriptomic data. Such approaches are essential to have a better understanding of the changes in cancer cell metabolism associated with tumourigenesis.

Weaknesses:

(1) This study, despite being a useful resource (assuming all the data will be publicly available and not only upon request) is mainly descriptive and correlative and lacks mechanistic links.

We appreciate this point. While the primary goal of our study was to assess mitochondrial adaptations with HER2-driven tumorigenesis, we agree strengthening the mechanistic interpretation would improve the impact of the data. To address this, we have provided experiments demonstrating HER2 inhibition in NF639 cells with lapatinib supresses respiratory capacity, directly supporting the interpretation that HER2 activity regulates respiratory function (Figure 10). We have expanded the discussion appropriately (lines 378-394). Both raw RNA-seq and proteomic data were deposited through GEO and the PRIDE repositories (accession numbers included in Data Availability Statement).

(2) It would be important to determine the cellular composition of the tumour and healthy tissue used. Do the changes described here apply to cancer cells only or do other cell types contribute to this?

We thank the reviewer for this suggestion; we have added experiments that have directly addressed this concern.

Cell type composition analysis by immunofluorescence was added (Figure 6) where we quantified epithelial, mesenchymal, endothelial, immune and stromal populations in our benign mammary tissue and tumor samples. We found no major shift in the dominant cell types that would confound transcriptomic data in whole tissues.

We integrated immunofluorescence data with a publicly available scRNA-seq dataset from human breast tumors which allowed us to estimate cell-type-specific expression of OXPHOS genes in our own samples. Despite the possibility of species differences, this is the only dataset of its kind, and we used this to generate an estimate of cell type weighted OXPHOS mRNA expression (Figure 6). This revealed that epithelial cells are likely the dominant contributors to OXPHOS gene expression for CIIV. All calculations are delineated in the Methods section.

(3) Are the changes in metabolic gene expression a consequence of HER2 signalling activation? Ex-vivo experiments could be performed to perturb this pathway and determine cause-effects.

Thank you for this suggestion – we have included an experiment directly testing this concept. We assessed mitochondrial respiration in NF639 HER2-driven mammary tumor epithelial cells in the presence or absence of the well-described dual tyrosine kinase inhibitor lapatinib. Lapatinib reduced basal, CI-linked and CI+II linked respiration without compromising mitochondrial integrity or coupling, demonstrating that HER2 activation regulates respiration in our model. This data is presented in Figure 10, and a new section has been added to the discussion describing the implications of this finding in the context of the current literature (lines 378-394).

(4) The data of fission/fusion seem quite preliminary and the gene/protein expression changes are not so clear cut to be a convincing explanation that this is the main reason for the increased mitochondria respiration in tumours.

We agree mitochondrial morphology and dynamics alone cannot fully account for the observed respiratory phenotype – this was emphasized in the discussion but has since been further clarified (lines 365-377). We retained the TEM and dynamics gene/protein data because they do support morphological differences consistent with enhanced fission. However, we have revised the tone of our interpretation to more explicitly acknowledge that these findings are correlative, and the updated discussion now emphasizes that the increased respiratory capacity in tumors is likely driven by multiple converging mechanisms.

Reviewer #2 (Public review):

Frangos et al present a set of studies aiming to determine mechanisms underlying initiation and tumour progression. Overall, this work provides some useful insights into the involvement of mitochondrial dysfunction during the cellular transformation process. This body of work could be improved in several possible directions to establish more mechanistic connections.

(5) The interesting point of the paper: the contrast between suppressed ETC components and activated OXPHOS function is perplexing and should be resolved. It is still unclear if activated mitochondrial function triggers gene down-regulation vs compensatory functional changes (as the title suggests). Have the authors considered reversing the HER2-derived signals e.g. with PI3K-AKT-MTOR or ERK inhibitors to potentially separate the expression vs. functional phenotypes? The root of the OXPHOS component down-regulation should also be traced further, e.g. by probing into levels of core mitochondrial biogenesis factors. Are transcript levels of factors encoded by mtDNA also decreased?

We appreciate this insight and agree that the discordance between mitochondrial content and function is fascinating and have addressed the concerns above in the following manner:

- We have altered the title – we agree we cannot definitively say that the enhanced respiratory capacity observed is compensatory.

- We have added experiments in NF639 cells in the presence of lapatinib, a tyrosine kinase inhibitor to interrogate whether HER2 is necessary for our functional outcome of interest – the enhanced respiratory capacity in the tumors. Lapatinib significantly suppressed respiration (Figure 10) demonstrating HER2 signaling directly regulates mitochondrial respiration.

- We have expanded the discussion to provide further comment on potential explanations for increased respiratory function and low mitochondrial content.

(6) The second interesting aspect of this study is the implication of mitochondrial activation in tumours, despite the downregulation of expression signatures, suggestive of a positive role for mitochondria in this tumour model. To address if this is correlative or causal, have the authors considered testing an OXPHOS inhibitor for suppression of tumorigenesis?

Previous studies have eloquently highlighted that directly or indirectly inhibiting mitochondria can supress growth in HER2-driven breast cancer (PMID:31690671) or alternatively, amplification of mt-HER2 enhances tumorigenesis (PMID: 38291340). In many solid tumors, this is the concept of preclinical and clinical studies using IACS-010759 or similar inhibitors of OXPHOS which do suppress growth but have significant off target effects in healthy tissues (PMID: 36658425, 3580228We have expanded the discussion to ensure the reader is aware of these previous contributions and highlighted the importance of future work delineating the role of enhanced respiratory function in HER2-driven mammary cancer (lines 378-394).

(7) A number of issues concerning animal/ tumour variability and further pathway dissection could be explored with in vitro approaches. Have the authors considered deriving tumourderived cell cultures, which could enable further confirmations, mechanistic drug studies and additional imaging approaches? Culture systems would allow alternative assessment of mitochondrial function such as Seahorse or flow cytometry (mitochondrial potential and ROS levels).

We thank the reviewer for this suggestion – we have addressed this in part by using the NF639 HER2driven tumor epithelial line which demonstrated that HER2 regulates our observed respiratory response. Unfortunately, the addition of tumor derived cell cultures was not feasible or within the scope of our study. Animal and tumor variability has been clarified in the Methods section (lines 424-429). Mitochondrial respiration experiments were performed in paired tissue (benign and tumor from same mouse). Transcriptomic, proteomic and histological analyses were performed on tumors and benign samples from different mice due to tissue limitations.

(8) The study could be greatly improved with further confirmatory studies, eg immunoblotting for mitochondrial components with parallel blots for phospho-signalling in the same samples. It would be interesting if trends could be maintained in tumour-derived cell cultures. It is notable that OXPHOS protein/transcript changes are more consistent (Figure 5, Supplementary Figure 4) than mitochondrial dynamics /mitophagy factors (Figure 8). Core regulatory factors in these pathways should be confirmed by conventional immunoblotting.

We thank the reviewer for this thoughtful comment. While we agree that additional confirmatory studies can be valuable, due to tissue quantity constraints and the number of assays required for our multi-omics analysis, extensive additional blots were not feasible. However, we had sufficient protein to provide select OXPHOS proteins to verify the proteomic data (now provided in S-Fig.4H). Furthermore, we have plotted the fold change of genes and proteins detected in both datasets and added this to Figure 4 (4A, B), further highlighting the consistency between our transcriptomic and proteomic findings. We believe that the highly consistent and concordant nature of our datasets collectively provides strong support for our central objective - determining whether mitochondrial content and respiratory function correlate in HER2-driven mammary tumors. The reproducibility of OXPHOS-related changes reinforces the robustness of our observations. We also appreciate the reviewer’s insight that OXPHOS alterations appear particularly consistent. In response, we have edited the discussion to further emphasize this point, especially in relation to the distinctive pattern observed for Complex V, which showed greater preservation relative to Complexes I–IV across several methods (lines 348-364). We comment on how this stoichiometric shift may contribute to intrinsic respiratory activation despite reduced mitochondrial content.

Recommendations for the authors:

Reviewer #2 (Recommendations for the authors):

Further Minor points.

(9) It would be helpful to know further details regarding the source of the tumour samples, particularly for the proteomics (N=5) and transcriptomics (N=6) datasets, since the exact timepoint of tissue harvest and number of tumours/mouse varied, according to the methods section. Were all samples from the omics studies from different mice (ie 11 mice)? B4 and B6 seem like outliers in mitochondrial transcriptomes. Are these directly paired eg with T4 and T6? Are the side-by-side pairs of Ben and Tum samples for blots in Figure 1 and Supplementary Figure 1 from the same mouse.

This has been clarified in the Methods section (lines 424-429). Mitochondrial respiration experiments were performed in paired tissue (benign and tumor from same mouse). Transcriptomic, proteomic and histological analyses were performed on tumors and benign samples from different mice due to tissue limitations.

(10) Further references and details are needed to support the methodology of the mitochondrial function tests (eg. nutrients vs pairing with complexes). What was the time point of nutrient supplementation? It would seem that the lipid substrates should take longer to activate OXPHOS than pyruvate/malate or succinate. Is this the case? Is there speculation as to why succinate supplementation is much more active than pyruvate+malate? What is +MD in Figure 6? The rationale for pooling data for Figure 7A is unclear since the categories appear to overlap: (pyruvate, malate, ADP) vs. (palmitoyl-carnitine, malate, ADP).

Thank you for this comment. We have expanded the methods (lines 515-531) to provide additional detail on the mitochondrial respiration protocol. Briefly, permeabilized tissues were exposed to substrates delivered at supraphysiological concentrations in a sequential protocol lasting ~30–60 minutes. Under these conditions, mitochondrial respiration reflects the maximal capacity to utilize each substrate rather than the physiological time course of substrate mobilization or uptake that would occur in vivo with the influence of blood flow and transport/substrate availability limitations.

(11) Many of the figures were blurry (Figure 1F, 2B) or had labels that were too small to be effective (Figures 1G, H, 2D-G, 3E-G, 5E-I, 7C, 8B).

The font size of figure labels has been increased where possible and all figures have been exported to maximize resolution.

eLife Assessment

This study presents an important methodological advance-Liver-CUBIC combined with multicolor metallic nanoparticle perfusion-that enables high-resolution 3D visualization of the liver's complex multi-ductal architecture. The identification of the Periportal Lamellar Complex (PLC) as a novel perivascular structure with distinct cellular composition and low-permeability characteristics is convincing, supported by rigorous imaging data. The observed scaffolding role during fibrosis offers intriguing biological insights, though the functional claims would benefit from direct experimental validation.

Reviewer #1 (Public review):

Summary:

In this manuscript, Chengjian Zhao et al. focused on the interactions between vascular, biliary, and neural networks in the liver microenvironment, addressing the critical bottleneck that the lack of high-resolution 3D visualization has hindered understanding of these interactions in liver disease.

Strengths:

This study developed a high-resolution multiplex 3D imaging method that integrates multicolor metallic compound nanoparticle (MCNP) perfusion with optimized CUBIC tissue clearing. This method enables the simultaneous 3D visualization of spatial networks of the portal vein, hepatic artery, bile ducts, and central vein in the mouse liver. The authors reported a perivascular structure termed the Periportal Lamellar Complex (PLC), which is identified along the portal vein axis. This study clarifies that the PLC comprises CD34⁺Sca-1⁺ dual-positive endothelial cells with a distinct gene expression profile, and reveals its colocalization with terminal bile duct branches and sympathetic nerve fibers under physiological conditions.

Comments on revisions:

The authors very nicely addressed all concerns from this reviewer. There are no further concerns and comments.

Reviewer #3 (Public review):

Xu, Cao and colleagues aimed to overcome the obstacles of high-resolution imaging of intact liver tissue. They report successful modification of the existing CUBIC protocol into Liver-CUBIC, a high-resolution multiplex 3D imaging method that integrates multicolor metallic compound nanoparticle (MCNP) perfusion with optimized liver tissue clearing, significantly reducing clearing time and enabling simultaneous 3D visualization of the portal vein, hepatic artery, bile ducts, and central vein spatial networks in the mouse liver. Using this novel platform, the researchers describe a previously unrecognized perivascular structure they termed Periportal Lamellar Complex (PLC), regularly distributed along the adult liver portal veins.<br /> Using available scRNAseq data, the authors assessed the CD34⁺Sca-1⁺ cells' expression profile, highlighting mRNA presence of genes linked to neurodevelopment, bile acid transport, and hematopoietic niche potential. Different aspects of this analysis were then addressed by protein staining of selected marker proteins in the mouse liver tissue. Next, the authors addressed how the PLC and biliary system react to CCL4-induced liver fibrosis, implying PLC dynamically extends, acting as a scaffold that guides the migration and expansion of terminal bile ducts and sympathetic nerve fibers into the hepatic parenchyma upon injury.

The work clearly demonstrates the usefulness of the Liver-CUBIC technique and the improvement of both resolution and complexity of the information, gained by simultaneous visualization of multiple vascular and biliary systems of the liver. The identification of PLC and the interpretation of its function represent an intriguing set of observations that will surely attract the attention of liver biologists as well as hepatologists. The importance of the CD34+/Sca1+ endothelial cell population and claims based on transcriptomic re-analysis require future assessment by functional experimental approaches to decipher the functional molecules involved in PLC formation, maintenance, and the involvement in injury response before establishing their role in biliary, arterial, and neural liver systems.

Strengths:

The authors clearly demonstrate an improved technique tailored to the visualization of the liver vasulo-biliary architecture in unprecedented resolution.<br /> This work proposes a new morphological feature of adult liver facilitating interaction between the portal vein, hepatic arteries, biliary tree, and intrahepatic innervation, centered at previously underappreciated protrusions of the portal veins - PLCs.

Weaknesses:

The importance of CD34+Sca1+ endothelial cell sub-population for PLC formation and function was not tested and warrants further validation.

Comments on revisions:

I appreciate the author's effort to revise the text so it more rigorously adheres to the presented evidence. Following a thorough read of the revised text, a few remaining minor issues were identified in the Discussion.

(1) From where comes the hard evidence for PLC being the stem cell niche in the following sentence?<br /> for the two following statements:

This suggests that the PLC may not only provide structural support but also serve as a perivascular stem cell niche specific to the portal region, potentially involved in hematopoiesis and tissue regeneration.

The PLC serves as a directional scaffold for ductal growth, a specialized stem cell niche, and a potential site of neurovascular coupling.

(2) In the following paragraph, I lack references to the previously published evidence of liver innervation guidance mechanisms, such as the mesenchyme-mediated guidance (CD31- population) Gannoun et al., 2023 https://doi.org/10.1242/dev.201642, an important context for your finding.

Further analysis showed significant upregulation of genes involved in neurodevelopment and axonal guidance in the CD34⁺Sca-1⁺ cluster, along with activation of neuronal signaling pathways. Immunostaining confirmed the presence of TH⁺ sympathetic nerve fibers wrapping around the PLC in a "beads-on-a-string" pattern (Fig. 6), consistent with a classic neurovascular unit(Adori et al., 2021). Previous studies have shown that sympathetic nerves enter the liver along collagen fibers of Glisson's capsule and interact with hepatic arteries, portal veins, and bile duct epithelium, supporting the PLC as a scaffold for intrahepatic neurovascular integration.

(3) Several sentences have issues with a lack of space between words.

Author response:

The following is the authors’ response to the previous reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

In this manuscript, Chengjian Zhao et al. focused on the interactions between vascular, biliary, and neural networks in the liver microenvironment, addressing the critical bottleneck that the lack of high-resolution 3D visualization has hindered understanding of these interactions in liver disease.

Strengths:

This study developed a high-resolution multiplex 3D imaging method that integrates multicolor metallic compound nanoparticle (MCNP) perfusion with optimized CUBIC tissue clearing. This method enables the simultaneous 3D visualization of spatial networks of the portal vein, hepatic artery, bile ducts, and central vein in the mouse liver. The authors reported a perivascular structure termed the Periportal Lamellar Complex (PLC), which is identified along the portal vein axis. This study clarifies that the PLC comprises CD34⁺Sca-1⁺ dual-positive endothelial cells with a distinct gene expression profile, and reveals its colocalization with terminal bile duct branches and sympathetic nerve fibers under physiological conditions.

Comments on revisions:

The authors very nicely addressed all concerns from this reviewer. There are no further concerns or comments.

We sincerely thank the reviewer for the positive evaluation of the revised manuscript.

Reviewer #2 (Public review):

Summary:

The present manuscript of Xu et al. reports a novel clearing and imaging method focusing on the liver. The Authors simultaneously visualized the portal vein, hepatic artery, central vein, and bile duct systems by injected metal compound nanoparticles (MCNPs) with different colors into the portal vein, heart left ventricle, vena cava inferior and the extrahepatic bile duct, respectively. The method involves: trans-cardiac perfusion with 4% PFA, the injection of MCNPs with different colors, clearing with the modified CUBIC method, cutting 200 micrometer thick slices by vibratome, and then microscopic imaging. The Authors also perform various immunostaining (DAB or TSA signal amplification methods) on the tissue slices from MCNP-perfused tissue blocks. With the application of this methodical approach, the Authors report dense and very fine vascular branches along the portal vein. The authors name them as 'periportal lamellar complex (PLC)' and report that PLC fine branches are directly connected to the sinusoids. The authors also claim that these structures co-localize with terminal bile duct branches and sympathetic nerve fibers and contain endothelial cells with a distinct gene expression profile. Finally, the authors claim that PLC-s proliferate in liver fibrosis (CCl4 model) and act as scaffold for proliferating bile ducts in ductular reaction and for ectopic parenchymal sympathetic nerve sprouting.

Strengths:

The simultaneous visualization of different hepatic vascular compartments and their combination with immunostaining is a potentially interesting novel methodological approach.

Weaknesses:

This reviewer has some concerns about the validity of the microscopic/morphological findings as well as the transcriptomics results, and suggests that the conclusions of the paper may be critically viewed. Namely, at this point, it is still not fully clear that the 'periportal lamellar complex (PLC)' that the Authors describe really exists as a distinct anatomical or functional unit or these are fine portal branches that connect the larger portal veins into the adjacent sinusoid. Also, in my opinion, to identify the molecular characteristics of such small and spatially highly organized structures like those fine radial portal branches, the only way is to perform high-resolution spatial transcriptomics (instead of data mining in existing liver single cell database and performing Venn diagram intersection analysis in hepatic endothelial subpopulations). Yet, the existence of such structures with a distinct molecular profile cannot be excluded. Further research with advanced imaging and omics techniques (such as high resolution volume imaging, and spatial transcriptomics/proteomics) are needed to reproduce these initial findings.

We thank the reviewer for the thoughtful and constructive comments. In response to the reviewer’s concerns regarding the anatomical and molecular definition of the periportal lamellar complex (PLC), we have further clarified the scope and methodological boundaries of the present study in the revised manuscript.

Regarding the key question raised by the reviewer—namely, whether the PLC represents an independent anatomical or functional unit, or merely small portal venous branches connecting larger portal veins to adjacent sinusoids—we provide below a more detailed explanation of the criteria used to define the PLC in this study. The identification of the PLC is primarily based on periportal structures that can be reproducibly recognized by three-dimensional imaging across multiple mice, exhibiting a relatively consistent spatial distribution within the periportal region. The PLC could be stably observed across different MCNP dye color assignments and independent experimental batches. In addition, three-dimensional CD31 immunofluorescence consistently revealed vascular-associated signal distributions in the same periportal region, indirectly supporting its spatial association with the periportal vascular system.

At the morphological level, the PLC appears as a periportal vasculature-associated structure distributed around the main portal vein trunk and maintains a relatively consistent spatial proximity to portal veins, bile ducts, and neural components in three-dimensional space. This highly conserved spatial organization across multiple tissue systems supports the anatomical positioning of the PLC as a relatively distinct structural tissue unit within the periportal region.

The present study primarily focuses on a descriptive characterization of the three-dimensional anatomical organization and spatial relationships of the PLC based on volumetric imaging and vascular labeling strategies. As a complementary exploratory analysis, we reanalyzed endothelial cell populations potentially associated with the PLC using existing liver single-cell transcriptomic datasets. This analysis was intended to provide molecular-level information consistent with the structural observations and to offer preliminary clues to its potential biological functions, rather than to independently define the PLC at the spatial level or to functionally validate it.

We fully acknowledge the value of spatial transcriptomic and spatial proteomic technologies in revealing molecular heterogeneity within tissue architecture. However, under current technical conditions, these approaches are largely dependent on thin tissue sections and are limited by spatial resolution and signal mixing effects, which still pose challenges for resolving periportal structures with pronounced three-dimensional continuity, such as the PLC. In the future, further integration of high-resolution volumetric imaging with spatial omics technologies may enable a more refined understanding of the molecular features and potential functions of the PLC at higher spatial resolution.

Reviewer #3 (Public review):

Summary:

In the revised version of the manuscript authors addressed multiple comments, clarifying especially the methodological part of their work and PLC identification as a novel morphological feature of the adult liver portal veins. Tet is now also much clearer and has better flow.

The additional assessment of the smartSeq2 data from Pietilä et al., 2025 strengthens the transcriptomic profiling of the CD34+Sca1+ cells and the discussion of the possible implications for the liver homeostasis and injury response. Why it may suffer from similar bias as other scRNA seq datasets - multiple cell fate signatures arising from mRNA contamination from proximal cells during dissociation, it is less likely that this would happen to yield so similar results.

Nevertheless, a more thorough assessment by functional experimental approaches is needed to decipher the functional molecules and definite protein markers before establishing the PLC as the key hub governing the activity of biliary, arterial, and neuronal liver systems.

The work does bring a clear new insight into the liver structure and functional units and greatly improves the methodological toolbox to study it even further, and thus fully deserves the attention of the Elife readers.

Strengths:

The authors clearly demonstrate an improved technique tailored to the visualization of the liver vasulo-biliary architecture in unprecedented resolution.

This work proposes a new morphological feature of adult liver facilitating interaction between the portal vein, hepatic arteries, biliary tree, and intrahepatic innervation, centered at previously underappreciated protrusions of the portal veins - the Periportal Lamellar Complexes (PLCs).

Weaknesses:

The importance of CD34+Sca1+ endothelial cell subpopulation for PLC formation and function was not tested and warrants further validation.

We thank the reviewer for the careful and constructive comments regarding the functional validation of cell populations associated with the PLC. The central aim of this study is to establish and validate a novel volumetric imaging and vascular labeling strategy and to apply it to the periportal region of the liver, thereby revealing previously underappreciated structural organizational patterns at the three-dimensional level, rather than to perform a systematic functional validation of specific cellular subpopulations.

We agree that the precise roles of the CD34⁺Sca-1⁺ endothelial cell subpopulation in the formation and function of the periportal lamellar complex (PLC) have not been directly addressed through functional intervention experiments in the present study. Our conclusions are primarily based on three-dimensional imaging and spatial distribution analyses, which reveal a stable and consistent spatial association between this cell population and the PLC structure, but are not intended to independently support causal or functional inferences. The underlying functional mechanisms remain to be elucidated in future studies using genetic or functional perturbation approaches.

In light of these considerations, we have further refined the relevant statements in the revised manuscript to more clearly define the functional scope and limitations of the current study in the Discussion section, and to avoid functional interpretations that extend beyond the direct support of the data. At the same time, we consider functional validation of the PLC to be an important and promising direction for future investigation.

It should be emphasized that the present study is not primarily designed to provide direct functional validation, but rather to systematically characterize the three-dimensional structural features of the periportal lamellar complex (PLC) and its cellular associations using volumetric imaging and vascular labeling approaches. At this stage, we mainly provide spatial and histological evidence for the organizational relationship between the PLC structure and the CD34⁺Sca-1⁺ endothelial cell population, while their specific roles in PLC formation and functional regulation await further investigation.

Recommendations for the authors:

Reviewer #2 (Recommendations for the authors):

I highly appreciate the Authors' endeavors to improve the manuscript. I am enlisting those points (from my original review) where I still have further comments.

(2) I would suggest this sentence:

"...the liver has evolved a highly complex and densely organized ductal vascular-neuronal network in the body, consisting primarily of the portal vein system, central vein system, hepatic artery system, biliary system, and intrahepatic autonomic nerve network [6, 7]."

We thank the reviewer for the valuable suggestion. We have revised the relevant sentence accordingly, and the revised wording is as follows:

“The liver has evolved a highly complex and densely organized vascular–biliary–neural network, primarily composed of the portal venous system, central venous system, hepatic arterial system, biliary system, and the intrahepatic autonomic neural network.”

(3) I suggest renaming 'clearing efficiency' to 'clearing time', and revise the last sentence like:

'...The results showed that the average transmittance increased by 20.12% in 1mm-thick cleared tissue slices.'

We thank the reviewer for this helpful suggestion. Accordingly, we have replaced the term “clearing efficiency” with “clearing time” and revised the final sentence to reflect this change. The revised wording is as follows:

“The results showed that the average transmittance increased by 20.12% in cleared tissue slices with a thickness of 1 mm.”

(4) While the dye perfusion was indeed on full lobe, FigS1F also seems to be rather a thick section instead of a full 3d reconstruction. This is OK, but please, be clear and specific about this in the respective part of the ms.

We thank the reviewer for the careful review and detailed comments. We would like to clarify that Fig. S1F shows whole-lobe imaging of the mouse left liver lobe obtained after dye perfusion at the whole-liver scale, rather than an image derived from a thick tissue section. Although this image does not represent a three-dimensional reconstruction, it does reflect imaging of the entire left liver lobe at the macroscopic level.

In addition, for the reviewer’s reference, we have provided in this response a representative image of a 200 μm-thick liver tissue section to directly illustrate the morphological differences between thick-section imaging and whole-lobe imaging. We note that the third and fourth panels in Fig. 1G of the main text already show local imaging results from 200 μm-thick sections; in contrast, the comparative image provided here presents a larger field of view and overall morphology. To avoid redundancy, this additional image is included solely for clarification in the present response and has not been incorporated into the revised manuscript or the supplementary materials.

(11) Regarding the 'transmission quantification':

'Regarding the comparative quantification of different clearing methods, as the reviewer noted, nearly all aqueous or organic solvent based clearing techniques can achieve relatively uniform transparency in 1 mm thick tissue sections, so differences at this thickness are limited.'

So, based on all these, I think, measuring/comparisons of clearing efficacy in the present form are kind of pointless --- one may consider omitting this part.

We thank the reviewer for the valuable comments. The purpose of the transmittance quantification in this study was not to provide a comprehensive comparison among different tissue-clearing methods, but rather to serve as a quantitative reference supporting the optimization of the Liver-CUBIC protocol. Accordingly, we have narrowed and clarified the relevant statements in the revised manuscript to define their scope and avoid overinterpretation.

The revised text now reads as follows:

“Importantly, Liver-CUBIC treatment did not induce significant tissue expansion (Figure 1B–D). In addition, quantitative transmittance measurements in 1-mm-thick cleared tissue slices showed an average increase of 20.12% (P < 0.0001; 95% CI: 19.14–21.09; Figure 1E).”

Author response image 1.

(16) It is OK, but please, indicate this clearly in the Methods/Results because in its present form it may be confusing for the reader: which color means what.

We thank the reviewer for this helpful request for clarification. We agree that the previous wording may have caused confusion regarding the meaning of different MCNP colors. Accordingly, we have revised the Methods section and the relevant figure legends to clearly state that the color assignment of MCNP dyes is not fixed across different experiments or figures. The use of different colors serves solely for visualization and presentation purposes, facilitating the distinction of anatomical structures in multichannel and three-dimensional imaging, and does not indicate any fixed or intrinsic correspondence between a specific color and a particular vascular or ductal system. We believe that this clarification will help prevent misinterpretation and improve the overall clarity of the manuscript.

(17) Still I think the hepatic artery is extremely shrunk, while the portal vein is extremely dilated. Please, note that in the referring figure (from Adori et al), hepatic artery and portal vein are ca 50 micrometers and 250 micrometers in diameter, respectively. In your figure, as I see, ca. 9-10 micrometers and 125 micrometers, respectively. This means 5x (Adori) vs. 13-14x differences (you). I would not say that this is necessarily problematic --- but may reflect some perfusion issues that may be good to consider.

We thank the reviewer for the careful comparison and acknowledge the quantitative differences pointed out. Compared with the study by Adori et al., the diameter ratio between the hepatic artery and the portal vein in our images does indeed differ to some extent. We believe that this discrepancy primarily arises from methodological differences in imaging and analysis strategies between the two studies.

In the work by Adori et al., periportal vasculature identification and three-dimensional segmentation were mainly based on 488 nm autofluorescence signals acquired from inverted tissues. This signal predominantly reflects the overall outline of periportal tissue regions rather than direct imaging of the vascular lumen itself. Consequently, the measured “vessel diameter” largely represents a spatial domain delineated by surrounding periportal structures, and does not necessarily correspond to the actual or functional luminal diameter of the vessel.

In contrast, the present study employed fluorescent MCNP dye perfusion under low perfusion pressure, combined with tissue clearing and three-dimensional optical imaging. Under these experimental conditions, the measured vessel diameters more closely reflect the perfusable luminal space of vessels in a fixed state, rather than their maximally dilated diameter, and are not defined by the morphology of surrounding tissues. This distinction is particularly relevant for the hepatic artery: as a high-resistance, smooth muscle–rich vessel, its diameter is highly sensitive to perfusion pressure and post-excision changes in vascular tone. In comparison, the portal vein exhibits greater compliance and is relatively less affected by these factors.

Based on these methodological differences, the observation of relatively smaller apparent hepatic arterial diameters—and consequently a higher arterial-to-portal vein diameter ratio—under dye perfusion–based optical imaging conditions is an expected outcome. Importantly, the primary focus of the present study is the identification and characterization of the periportal lamellar complex (PLC) as a three-dimensional lamellar tissue structure that can be stably and reproducibly recognized across different samples and imaging conditions, rather than absolute comparisons of vascular diameters.

(21) After the presented documentation, I still have some concerns that the 'periportal lamellar complex (PLC)' that the Authors describe is really a distinct anatomical or functional unit. The confocal panel in Fig. 4F is nice and high quality. However, as far as I see, it shows that CD34+/Sca-1+ immunostaining is not specific for the presumptive PLCs in the peri-portal region. Instead, Sca-1 immunoreactivity is highly abundant also in the midzone --- to which the supposed PLCs do not extend, according to the cartoon shown in panel D, same figure. Notably, this questions also the specificity of the single cell analysis.

We thank the reviewer for this detailed and important comment regarding the specificity of CD34⁺/Sca-1⁺ markers and the definition of the periportal lamellar complex (PLC).

It should be emphasized that the PLC is not defined on the basis of any single molecular marker, but rather by a reproducible periportal lamellar anatomical structure consistently revealed by three-dimensional imaging across multiple samples. The co-expression of CD34 and Sca-1 is interpreted within this clearly defined anatomical context and is used to characterize the molecular features of endothelial cells associated with the PLC structure.

As shown in Fig. 4F, the co-expression of CD34 and Sca-1 delineates a continuous, lamellar endothelial structure surrounding the portal vein. In contrast, outside the periportal region—including the midlobular areas—Sca-1 or CD34 expression can also be detected, but these signals appear scattered and discontinuous, lacking an organized lamellar topology.

In the single-cell transcriptomic analysis, we treated CD34⁺/Sca-1⁺ endothelial cells as an operational population to explore molecular features that may be enriched in the microenvironment of the periportal lamellar complex (PLC). Importantly, this analysis was intended to provide molecular clues associated with the PLC, rather than to precisely assign spatial locations or identities to individual cells.

Occasional isolated Sca-1⁺ signals detected outside the periportal region do not affect the anatomical definition of the PLC, nor do they alter the interpretation of the single-cell analysis. These analyses serve to provide supportive and exploratory molecular information for the structural identification of the PLC, rather than constituting decisive spatial evidence.

(23) '....In the manuscript, we have carefully stated that this analysis is exploratory in nature and have avoided overinterpretation. In future studies, high-resolution spatial omics approaches will be invaluable for more precisely delineating the molecular characteristics of these fine structures.'

I do not find these statements either in the Discussion or in the Results. I must reiterate my opinion that the applied methodical approach in the single cell transcriptomics part has severe limitations, and the readers must be aware of this.

We thank the reviewer for this further comment. We understand and acknowledge the reviewer’s concerns regarding the methodological limitations of single-cell transcriptomic analyses, and we agree that these limitations should be clearly communicated to readers in the main text.

We acknowledge that in the previous version of the manuscript, the exploratory nature of the single-cell transcriptomic analysis and its methodological boundaries were discussed only in the response to reviewers and were not explicitly stated in the manuscript itself. We thank the reviewer for pointing out this omission. In the revised manuscript, we have now added explicit clarifications in the main text to prevent potential overinterpretation of these results.

In the present study, our primary effort is focused on the descriptive characterization of the three-dimensional anatomical organization and spatial relationships of the PLC using volumetric imaging and vascular labeling strategies. As a complementary exploratory analysis, we reanalyzed existing liver single-cell transcriptomic datasets to examine endothelial cell populations exhibiting PLC-associated features, and performed differential gene expression and Gene Ontology enrichment analyses. Importantly, these results are intended to provide molecular-level support for the structural identification of the PLC and to offer preliminary insights into its potential biological functions. Accordingly, we have narrowed the presentation and interpretation of the single-cell analysis in both the Results and Discussion sections of the revised manuscript.

In addition, we have expanded the Discussion to address the limitations of current spatial transcriptomic approaches in validating a continuous three-dimensional structure such as the PLC. Most existing spatial transcriptomic methods rely on two-dimensional tissue sections of 8–10 μm thickness, whereas identification of the PLC depends on three-dimensional imaging of tissue volumes with thicknesses of ≥200 μm, making reliable reconstruction of its spatial continuity from single sections challenging. Furthermore, because each spatial transcriptomic capture spot often encompasses multiple adjacent cells, signal mixing effects further limit precise resolution of specific periportal microstructures.

Overall, we agree with the reviewer’s central point that the limitations of single-cell transcriptomic analyses should be clearly understood by readers. By explicitly clarifying the methodological boundaries and refining the related statements in the main text, we believe this concern has now been adequately addressed in the revised manuscript. We thank the reviewer for identifying this omission, which has helped to improve the rigor and clarity of the study.

Reviewer #3 (Recommendations for the authors):

(1) While interesting observations, suitable for discussion, the following sections are speculations, given that no functional characterization of PLC importance has been performed yet. This is the most felt when commenting on the role in hematopoiesis, which transiently takes place in the liver during embryogenesis (Khan et al 2016) but ceases to exist after ligation of the umbilical inlet. Adult Liver hematopoiesis remains controversial, and more solid evidence would need to be presented to support its existence in PLC regions.

265 - These findings suggest that the Periportal Lamellar Complex (PLC) is not only a morphologically and spatially distinct, low-permeability vascular unit surrounding the portal vein, but also likely serves as a critical nexus connecting the portal vein, hepatic artery, and liver sinusoids. Thus, the PLC constitutes a key node within the interactive vascular network of the mouse liver.

We thank the reviewer for the comments and suggestions regarding the potential functional interpretation of the periportal lamellar complex (PLC), particularly its possible association with hematopoietic function. We would like to clarify that the statement on page 265 was intended solely to describe the structural characteristics and spatial organization of the PLC within the periportal vascular network. Specifically, the original wording aimed to summarize the morphological features of the PLC and its spatial relationships among the portal vein, hepatic artery, and hepatic sinusoids.

Nevertheless, to minimize potential misunderstanding, we have revised this section to avoid unnecessary functional implications. The revised text now reads:

“These results suggest that the periportal lamellar complex (PLC) is a morphologically and spatially distinct vascular structure that surrounds the portal vein and may serve as a key organizational node coordinating the spatial relationships among the portal vein, hepatic artery, and hepatic sinusoids. Accordingly, the PLC represents an important structural element within the interactive vascular network of the mouse liver.”

This revision preserves the structural significance of the PLC while avoiding overinterpretation of its functional roles.

(2) The same is true also for this section, following Figure 3 - no functional experiment tested this. For example, diphtheria toxin is expressed in the CD34+Sca1+ population. Or at least a careful mapping of the developing liver, which would indicate if the PLC precedes or follows the BD development.

356 as a spatial positional cue guiding bile duct growth and branching but also as a regulatory node involved in coordinating bile drainage from the hepatic lobule into the biliary network.

To avoid potential misunderstanding, we have further refined and revised the statements in the manuscript regarding the functional interpretation of the periportal lamellar complex (PLC) and its relationship to bile duct development. We agree that cell ablation strategies are of great importance for functional validation studies. However, it should be noted that CD34 and Sca-1 are relatively broadly expressed markers during liver development, labeling multiple endothelial, mesenchymal, and progenitor cell populations, and their expression is not restricted to the PLC. Owing to this broad expression pattern, ablation of CD34⁺Sca-1⁺ cell populations would likely exert widespread effects on vascular and stromal structures, thereby complicating the distinction between direct PLC-specific effects and secondary developmental alterations. As such, this strategy may present technical limitations for specifically dissecting the role of the PLC in bile duct development. At the same time, given that the primary objective of this study is the systematic characterization of the three-dimensional anatomical features and spatial organization of the PLC, we have correspondingly revised the manuscript to restrict statements regarding the relationship between the PLC and bile ducts to spatial associations supported by the current data. Specifically, our results show that primary bile ducts run along the main portal vein trunk, secondary bile ducts exhibit directed branching toward the PLC region, and terminal bile duct branches tend to spatially cluster in the vicinity of the PLC, thereby forming a reproducible periportal spatial arrangement. Based on these observations, the PLC delineates a relatively conserved anatomical microenvironment within the portal region, whose spatial position is closely associated with the organization and terminal distribution of the intrahepatic bile duct network.

We believe that these revisions more accurately reflect the experimental evidence and the defined scope of the present study.

(3) The following statement ought to be rephrased or skipped, considering that CD34 and Sca1 (Ly6a) are markers of periportal endothelial cells (Pietilä et al., 2025, Gómez-Salinero et al., 2022) and as shown by the authors in their own Fig. 6D. In this context and the context of the CCL4 experiments, a "simple" proliferative progenitor portal vein endothelial cell phenotype, suggested also by the presence of DLL4 (Fig5A) and JAG1 (Pietilä et al., 2025) (Benedito et al., 2009) ought to be considered.

409 Notably, CD34 and Sca-1 (Ly6a) were co-expressed exclusively within PLC structures surrounding the portal vein, but absent from central vein ECs and midzonal LSECs (Figure 4F).

We thank the reviewer for pointing out the potential imprecision in this wording. We agree that both CD34 and Sca-1 (Ly6a) are well-established markers of periportal endothelial cells, as previously reported (Pietilä et al., 2025; Gómez-Salinero et al., 2022), and as also illustrated in Fig. 4F of our study.

Accordingly, the original statement suggesting that CD34 and Sca-1 are co-expressed exclusively within the PLC structure may indeed represent an overinterpretation. Following the reviewer’s suggestion, we have revised the relevant text on page 409 by removing the exclusive phrasing (“only in”) and by emphasizing instead that CD34⁺Sca-1⁺ endothelial cells are enriched in periportal regions associated with the PLC, rather than being specific to or confined within the PLC.

In addition, in the context of the CCl₄-induced liver fibrosis model, we agree with the reviewer that the observed expression of DLL4 and JAG1 under fibrotic conditions is more appropriately interpreted as reflecting an activated or proliferative periportal endothelial progenitor–like phenotype, rather than defining a novel endothelial lineage. The corresponding statements in the revised manuscript have been adjusted accordingly.

(4) Again, these concluding sentences are based on correlative evidence of mRNA expression and literature but not experimental evidence.

436 These findings suggest that this unique endothelial cell subset in the periportal region may possess dual regulatory functions in both metabolic and hematopoietic modulation

441 results suggest that PLC endothelial cells may not only regulate periportal microcirculatory blood flow but also help establish a specialized microenvironment that potentially supports periportal hematopoietic regulation, contributing to stem cell recruitment, vascular homeostasis, and tissue repair.

We thank the reviewer for this thoughtful comment. We agree that these statements are primarily based on transcriptomic correlation analyses and support from previous literature, rather than direct functional experimental evidence.

Accordingly, in the revised manuscript, we have appropriately toned down and adjusted the relevant concluding statements to more accurately reflect their inferential nature. The revised wording emphasizes associations and potential involvement, rather than definitive functional roles. These changes preserve the overall scientific interpretation while aligning the level of inference more closely with the available evidence.

The revised text now reads:

“Finally, we found that the main trunk of the PLC is primarily composed of CD34⁺Sca-1⁺CD31⁺ endothelial cells (Fig. 4J). These CD34⁺Sca-1⁺ double-positive cells are mainly distributed in the basal region of the PLC structure and exhibit molecular features associated with hematopoiesis. Taken together, these results suggest that PLC endothelial cells may contribute to the establishment of a local microenvironment related to periportal hematopoietic regulation and may play potential roles in stem cell recruitment and maintenance of vascular homeostasis.”

(5) The following part is speculative and based on re-analysis from the dataset that was gathered after 6 more weeks of CCL4 treatment (12weeks Su et al., 2021), then in the linked experiments from the manuscript. And should be moved to discussion or removed.

504 Moreover, single-cell transcriptomic re-analysis revealed significant upregulation of bile duct-related genes in the CD34⁺Sca-1⁺ endothelium of PLC in fibrotic liver, with notably high expression of Lgals1 (Galectin-1) and Hgf (Figure 5G). Previous studies have shown that Galectin-1 is absent in normal liver parenchyma but highly expressed in intrahepatic cholangiocarcinoma (ICC), correlating with tumor dedifferentiation and invasion (Bacigalupo, Manzi, Rabinovich, & Troncoso, 2013; Shimonishi et al., 2001). Additionally, hepatocyte growth factor (HGF), particularly in combination with epidermal growth factor (EGF) in 3D cultures, promotes hepatic progenitor cells to form bile duct-polarized cystic structures (N. Tanimizu, Miyajima, & Mostov, 2007). Together, these findings suggest the PLC endothelium may act as a key regulator of bile duct branching and fibrotic microenvironment remodeling in liver fibrosis.

Collectively, our results demonstrate that the PLC, situated between the portal vein and periportal sinusoidal endothelium, constitutes a critical vascular microenvironmental unit. It may not only colocalize with bile duct branches under normal physiological conditions, but also through its basal CD34⁺Sca-1⁺ double-positive endothelial cells, potentially orchestrate bile duct epithelial proliferation, branching morphogenesis, and bile acid transport homeostasis via multiple signaling pathways. Particularly during liver fibrosis progression, the PLC exhibits dynamic structural extension, serving as a spatial scaffold facilitating terminal bile duct migration and expansion into the hepatic parenchyma (Figure 5H). These findings highlight the PLC endothelial cell population and the vascular-bile duct interface as key regulatory hubs in bile duct regeneration, tissue repair, and pathological remodeling, providing novel cellular and molecular insights for understanding bile duct-related diseases such as ductular reaction, cholangiocarcinoma, and cholestatic disorders, and offering potential targets for therapeutic intervention.

We thank the reviewer for this careful and thought-provoking comment. We understand and agree with the reviewer’s assessment that this section involves a degree of inference, as the analysis is based on a re-analysis of a previously published single-cell transcriptomic dataset from a CCl₄-induced liver fibrosis model (Su et al., 2021), rather than on experimental data directly generated in the present study.

In response to the reviewer’s suggestion, we have carefully re-examined and revised the relevant paragraphs. Without altering the overall structure of the manuscript, we have appropriately moderated the wording to clarify that these results primarily describe the transcriptional features of PLC-associated CD34⁺Sca-1⁺ endothelial cells under fibrotic conditions, and their associations with bile duct–related gene expression, rather than providing direct functional evidence for their roles in bile duct branching or microenvironmental remodeling.

In addition, we have explicitly clarified in the main text the data source and methodological limitations of the single-cell transcriptomic analysis, and emphasized that these findings should be interpreted in conjunction with the spatial information revealed by three-dimensional imaging. Through these revisions, we aim to retain the value of this analysis in providing complementary molecular insight into PLC characteristics, while avoiding potential over-interpretation of its functional implications.

Formal suggestions:

(6) The following sentence would benefit from being more clearly written.

263 - The formation of PLC structures in the adventitial layer may participate in local blood flow regulation, maintenance of microenvironmental homeostasis.

We thank the reviewer for this helpful suggestion. The sentence has been revised to improve clarity by correcting the parallel structure and refining the wording.

The formation of PLC structures in the adventitial layer may participate in local blood flow regulation and the maintenance of microenvironmental homeostasis.

(7) The following sentence is misleading as it implies cell sorting, and "subsetted" rather than "sorted" should be used.

414 Based on this, we sorted CD34⁺Sca-1⁺ endothelial populations from the total liver EC pool (Figure 4G).

Thank you for your comment.

We have revised the term as suggested. This avoids the misleading implication of physical sorting, as our operation was analytical subsetting of the target subpopulation.

We appreciate your careful review.

(8) Correct typos, especially in the results section related to Fig. 6. and formatting issues in the discussion.

730 Morphologically, the PLC shares features with previously described telocytes (TCs)- 731 a recently identified class of interstitial cells in the liver observed via transmission electron

We thank the reviewer for pointing out this textual error. In the submitted version, the sentence describing the morphological similarity between the PLC and previously reported telocytes was inadvertently interrupted due to a punctuation issue. This has now been corrected to ensure sentence integrity and consistent formatting.

eLife Assessment

This study now provides solid evidence for a role of EndoA3-mediated trafficking of ICAM-1 to the immune synapse with T cells. The study will be valuable to those studying cell-cell communication in the immune system, and opens additional questions regarding the mechanisms involved and how other adhesion ligands are regulated.

Reviewer #1 (Public review):

Summary:

This study by Xu et al. investigates how clathrin-independent endocytosis in cancer cells influences T cell activation. Using a combination of biochemical approaches and imaging, the authors identify ICAM1, the ligand for the T cell integrin LFA-1, as a novel cargo of EndoA3-mediated endocytosis.

The authors then explore the functional consequences of EndoA3 depletion in cancer cells on T cell function using cytokine measurements, surface marker analyses, cytotoxicity assays and imaging. Loss of EndoA3 results in reduced T cell cytokine production, while expression of activation and exhaustion markers such as TIM-3, PD-1, and CD137 remains largely unchanged. EndoA3 knockout is associated with reduced ICAM1 surface levels and increased ALCAM levels in cancer cells. Imaging experiments further reveal directional transport of ICAM1 toward the immunological synapse, seemingly slightly reduced ICAM1 levels at the synapse upon EndoA3 depletion and an enlarged contact area between T cells and cancer cells.

Based on these observations, the authors propose a model in which EndoA3-mediated endocytosis and retrograde trafficking of ICAM1 (and ALCAM) supplies the immunological synapse with ligands for adhesion molecules. In the absence of EndoA3, T cells are suggested to compensate for suboptimal ICAM1 availability by enlarging the synaptic contact area, altering synapse architecture, leading to reduced cytokine secretion but modestly enhanced cytotoxicity.

Overall, the study provides convincing evidence for a modulatory role of EndoA3-mediated endocytosis in regulating T cell-cancer cell interactions. However, the choice of cellular model systems, the limited number of biological replicates and insufficiently supported mechanistic interpretations weaken the manuscript and weaken the strength of its conclusions.

Strengths:

The authors employ a rigorous and innovative experimental strategy that convincingly identifies ICAM1 as a novel cargo of EndoA3-mediated endocytosis with convincing visualization of directional ICAM1 transport toward the immunological synapse. In addition, the study provides a comprehensive characterization of how EndoA3 depletion in cancer cells affects T cell cytokine production, activation, proliferation and cytotoxic function, representing a valuable contribution to our understanding of how membrane trafficking pathways in target cells can modulate immune responses.

Comments on revised version:

Thank you very much for submitting your revised manuscript. I appreciated your efforts to answer all of the reviewers questions. While in my opinion the manuscript truly improved I think there are still lingering questions, in particular regarding the following points:

(1) Limited biological replication:

The LB33-MEL system remains problematic, as also noted by other reviewers. While it clearly represents an improvement over highly derived model systems such as Jurkat or Raji cells, it nevertheless effectively restricts the study to a single biological replicate. In this context, it may be more appropriate to compare the chosen approach to more state-of-the-art systems, such as expression of HLA-A*02:01, peptide loading (e.g. NY-ESO), and introduction of the matching TCR into donor-derived primary T cells. Such an approach would allow the use of multiple T cell donors and would substantially strengthen the generalizability of the conclusions.

(2) Expression levels of ICAM1:

Based on available database information (e.g. UniProt) and published literature (PMID: 9371813), ICAM1 appears to be expressed at relatively low levels in both HeLa and LB33-MEL cells. While the effects on T cells are initially discussed in terms of broader changes in EndoA3-mediated recycling of multiple surface proteins, including ICAM1 and ALCAM (and potentially others), the focus of the manuscript increasingly shifts toward ICAM1 as the primary driver of the observed phenotypes. Given the comparatively low endogenous expression of ICAM1 in the chosen model systems, it is unclear whether this emphasis is fully justified. In addition, if ICAM1 polarization toward the immunological synapse was assessed using ICAM1 overexpression, whereas other phenotypes (such as enlarged contact area) were analyzed under endogenous expression conditions, this further complicates the interpretation. As a first step toward clarifying these issues, it would be helpful to include representative flow cytometry histograms showing surface expression levels of ICAM1 and ALCAM, rather than only normalized quantifications.

(3) Cell-cell contact dynamics:

The manuscript suggests that altered contact dynamics may underlie the observed increase in cytotoxicity upon EndoA3 depletion. However, these claims are not directly tested. Such effects could be addressed with relatively straightforward experiments, for example by directly measuring T cell-cancer contact duration in co-culture assays.

Reviewer #2 (Public review):

The manuscript by Xu et al. studies the relevance of endophilin A3-dependent endocytosis and retrograde transport of immune synapse components and in the activation of cytotoxic CD8 T cells. First, the authors show that ICAM1 and ALCAM, known component of immune synapses, are endocytosed via endoA3-dependent endocytosis and retrogradely transported to the Golgi. The authors then show that blocking internalization or retrograde trafficking reduces the activation of CD8 T cells. Moreover, this diminished CD8 T cells activation resulted the formation of an enlarged immune synapse with reduced ICAM1 recruitment.

Comments on revisions:

The authors have addressed all my comments adequately.

Reviewer #3 (Public review):

Shiqiang Xu and colleagues have examined the importance of ICAM-1 and ALCAM internalization and retrograde transport in cancer cells on formation of a polarized immunological synapse with cytotoxic CD8+ T cells. They find that internalization is mediated by Endophilin A3 (EndoA3) while retrograde transport to the Golgi apparatus is mediated by the retromer complex. Perturbing these trafficking pathways reduces cytokine release, but increases cytolytic killing. The paper is building on previous findings from corresponding author Henri-François Renard showing that ALCAM is an EndoA3 dependent cargo in clathrin-independent endocytosis.

The work is interesting as it describes a novel mechanism by which cancer cells might influence CD8+ T cell activation and immunological synapse formation, and the authors have used a variety of cell biology and immunology methods to study this. The authors have also made substantial efforts to address the reviewers comments to the first version of the paper. However, there are still some points which could be further improved to underpin their conclusions:

The movies and the related micrographs of EndoA3-mediated ICAM-1 endocytosis could be more convincing. Is the invagination of large membrane patches visible by volumetric imaging (e.g. confocal z-stacks) or brightfield microscopy?

There is still a lack of quantitative evidence for polarized transport of ICAM-1 positive vesicles towards the immunological synapse. Only one example is shown and the authors state that the data is from a single movie representative of two independent experiments. If there are multiple cells per experiment, the number of cells should be stated and more examples should be included.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

This study by Xu et al. focuses on the impact of clathrin-independent endocytosis in cancer cells on T cell activation. In particular, by using a combination of biochemical approaches and imaging, the authors identify ICAM1, the ligand for T cell-expressed integrin LFA-1, as a novel cargo for EndoA3-mediated endocytosis. Subsequently, the authors aim to identify functional implications for T cell activation, using a combination of cytokine assays and imaging experiments.

They find that the absence of EndoA3 leads to a reduction in T cell-produced cytokine levels. Additionally, they observe slightly reduced levels of ICAM1 at the immunological synapse and an enlarged contact area between T cells and cancer cells. Taken together, the authors propose a mechanism where EndoA3-mediated endocytosis of ICAM1, followed by retrograde transport, supplies the immunological synapse with ICAM1. In the absence of EndoA3, T cells attempt to compensate for suboptimal ICAM1 levels at the synapse by enlarging their contact area, which proves insufficient and leads to lower levels of T cell activation.

Strengths:

The authors utilize a rigorous and innovative experimental approach that convincingly identifies ICAM1 as a novel cargo for Endo3A-mediated endocytosis.

Weaknesses:

The characterization of the effects of Endo3A absence on T cell activation appears incomplete. Key aspects, such as surface marker upregulation, T cell proliferation, integrin signalling and most importantly, the killing of cancer cells, are not comprehensively investigated.

We agree with the reviewer that the effects of EndoA3 depletion on T cell activation were not characterized enough. In new data presented in Fig.S4G-J, we explored additional activation markers and proliferation parameters. We didn’t observe any difference for the surface markers PD-1, CD137 and Tim-3 between LB33-MEL EndoA3+ cells treated with control and EndoA3 siRNAs. Regarding proliferation (Fig. S4J), although the proliferation index seems slightly lower upon EndoA3 depletion, we didn’t observe any significant difference either. Degranulation has also been monitored (Fig. S4K), but we didn’t observe any significant differences. In the new Fig. 3F however, we performed chromium release assays to assess the killing of cancer cells. Very interestingly, we observed an ~15% higher lysis of LB33-MEL EndoA3+ cells after EndoA3 depletion, when compared to the control condition at a ratio of 3:1 T cells:target cells (where the maximal effect is observed). These data are further discussed in the discussion section (new §6-9).

As Endo- and exocytosis are intricately linked with the biophysical properties of the cellular membrane (e.g. membrane tension), which can significantly impact T-cell activation and cytotoxicity, the authors should address this possibility and ideally address it experimentally to some degree.

Evaluating changes in the biophysical properties of cancer cell plasma membrane upon EndoA3 depletion is not trivial. An indirect way to address this question is by observing the area and shape of cells after siRNA treatment. In the new data added in the new Fig. S4B-D, we compared the area, aspect ratio and roundness of LB33-MEL EndoA3+ cells treated with negative control or EndoA3 siRNAs. While we observed a slight cell area reduction upon EndoA3 depletion, no significant changes were observed regarding the aspect ratio and the roundness. Hence, we think that the biophysical properties of cancer cells are not drastically modified by EndoA3 depletion.

Crucially, key literature relevant to this research, addressing the role of ICAM1 endocytosis in antigen-presenting cells, has not been taken into consideration.

We thank the reviewer for this important point. We have now considered and cited the relevant literature (Discussion, Page no.9).

Reviewer #2 (Public review):

Summary:

The manuscript by Xu et al. studies the relevance of endophilin A3-dependent endocytosis and retrograde transport of immune synapse components and in the activation of cytotoxic CD8 T cells. First, the authors show that ICAM1 and ALCAM, known components of immune synapses, are endocytosed via endoA3-dependent endocytosis and retrogradely transported to the Golgi. The authors then show that blocking internalization or retrograde trafficking reduces the activation of CD8 T cells. Moreover, this diminished CD8 T cell activation resulted in the formation of an enlarged immune synapse with reduced ICAM1 recruitment.

Strengths:

The authors show a novel EndoA3-dependent endocytic cargo and provide strong evidence linking EndoA3 endocytosis to the retrograde transport of ALCAM and ICAM1.

Weaknesses:

The role of EndoA3 in the process of T cell activation is shown in a cell that requires exogenous expression of this gene. Moreover, the authors claim that their findings are important for polarized redistribution of cargoes, but failed to show convincingly that the cargoes they are studying are polarized in their experimental system. The statistics of the manuscript also require some refinement.

We fully acknowledge that the requirement for exogenous expression of EndoA3 in our immunological model represents a limitation of our study. Unfortunately, it remains challenging to identify cancer cell lines for which autologous CD8 T cells are available and that endogenously express all molecular players investigated (in particular EndoA3). At this stage, we do not have access to any other cancer cell line/autologous CD8⁺ T cell pairs that are sufficiently well characterized. In future studies, it would be valuable to investigate tumor types with high endogenous EndoA3 expression (such as glioblastomas, gliomas, and head and neck cancers) for which autologous CD8 T cells could be obtained, but this remains technically challenging.

To address the reviewer’s second point regarding polarized redistribution of cargoes, we have added new data in the new Figure 4 and Movies S8-9. Using high-speed spinningdisk live-cell confocal microscopy, we captured the movement of ICAM1-positive tubulovesicular carriers in cancer cells at the moment of contact with CD8 T cells. Capturing such events is technically challenging, as T cell–cancer cell contacts form randomly and transiently. Successful imaging requires that the cancer cell be well spread and express ICAM1–GFP at an optimal level (as it is transiently expressed as a GFP-tagged construct), while acquisition must occur precisely at the moment when the T cell initiates contact. Despite these technical constraints, we successfully imaged early stages of immune synapse formation, enabling visualization of ICAM1 vesicular transport.

The data reveal a flux of ICAM1-positive carriers emerging from the perinuclear region (corresponding to the Golgi area) and moving toward the contact site with the CD8 T cell, with fusion events of vesicles occurring at the developing immune synapse. AI-based segmentation and tracking analyses showed that ICAM1-positive carrier trajectories were predominantly oriented toward the forming immune synapse, whereas carriers moving toward other cellular regions were markedly less frequent. These results provide direct evidence for polarized ICAM1 transport via vesicular trafficking toward the immune synapse.

Reviewer #3 (Public review):

Summary:

Shiqiang Xu and colleagues have examined the importance of ICAM-1 and ALCAM internalization and retrograde transport in cancer cells on the formation of a polarized immunological synapse with cytotoxic CD8+ T cells. They find that internalization is mediated by Endophilin A3 (EndoA3) while retrograde transport to the Golgi apparatus is mediated by the retromer complex. The paper is building on previous findings from corresponding author Henri-François Renard showing that ALCAM is an EndoA3dependent cargo in clathrin-independent endocytosis.

Strengths:

The work is interesting as it describes a novel mechanism by which cancer cells might influence CD8+ T cell activation and immunological synapse formation, and the authors have used a variety of cell biology and immunology methods to study this. However, there are some aspects of the paper that should be addressed more thoroughly to substantiate the conclusions made by the authors.

Weaknesses:

In Figure 2A-B, the authors show micrographs from live TIRF movies of HeLa and LB33MEL cells stably expressing EndoA3-GFP and transiently expressing ICAM-1-mScarlet. The ICAM-1 signal appears diffuse across the plasma membrane while the EndoA3 signal is partially punctate and partially lining the edge of membrane patches. Previous studies of EndoA3-mediated endocytosis have indicated that this can be observed as transient cargo-enriched puncta on the cell surface. In the present study, there is only one example of such an ICAM-1 and EndoA3 positive punctate event. Other examples of overlapping signals between ICAM-1 and EndoA3 are shown, but these either show retracting ICAM1 positive membrane protrusions or large membrane patches encircled by EndoA3. While these might represent different modes of EndoA3-mediated ICAM-1 internalization, any conclusion on this would require further investigation.

We agree with the reviewer that the pattern of cargoes during endocytosis (puncta vs large patches) as observed by live-cell TIRF microscopy may be confusing. Actually, a punctate pattern has been observed quasi systematically when we monitored the uptake of endogenous cargoes via antibody uptake assays (whatever the imaging approach: TIRF, spinning-disk, classical confocal or lattice light-sheet microscopy). For example:

- ALCAM: Fig.1e-h, Supplementary Figure 5 and Supplementary Movies 1-3 and 6 in Renard et al. 2020, https://doi.org/10.1038/s41467-020-15303-y; Fig.1D and Movie 2 in Tyckaert et al. 2022, https://doi.org/10.1242/jcs.259623.

- L1CAM: Fig.2 and 3D, Movies S1-4 in Lemaigre et al. 2023, https://doi.org/10.1111/tra.12883.

In rare examples, bigger clusters of antibodies were observed, where EndoA3 was observed to surround them, delineate them in a “lasso-like” pattern, and the clusters were progressively taken up:

- ALCAM: Supplementary Movie 4 in Renard et al. 2020, https://doi.org/10.1038/s41467-020-15303-y.

However, bigger patches of cargoes were more often observed when uptake was observed using transient expression of GFP-/mCherry-tagged versions of cargoes. In these cases, EndoA3 was predominantly observed to delineate cargo patches as a “lasso-like” pattern, progressively triming those patches leading to endocytosis. For example:

- L1CAM: Fig.3E, Movie S5-7 in Lemaigre et al. 2023, https://doi.org/10.1111/tra.12883.

- We also observed this pattern with CD166-GFP (unpublished).

The fact that we observed rather patches than punctate patterns upon transient expression of fluorescently-tagged constructs of cargoes is likely due to the elevated expression level of the cargoes.

Therefore, the patchy pattern observed for ICAM1 and ALCAM, transiently expressed in fusion with fluorescent proteins, and surrounded by EndoA3 in Fig.2A-B and old Movies S1-3, is not surprising. Of note, upon anti-ALCAM antibody uptake, we observed a more punctate pattern (Fig.2C), as previously described. Unfortunately, the lower quality of commercial anti-ICAM1 antibody did not allow us to proceed to uptake assays as for ALCAM.

Regarding Fig.S2 and old Movies S4-5, we agree with the reviewer that these data may be misleading, as they represent phenomena happening at protrusions and contact zones between two adjacent cells. We have now replaced these images with other examples where we avoid contact zones (Fig.S2 and new Movies S5-7).

These different patterns (patches vs dots) are still unexplained at the current stage, and may indeed represent different modes of endocytosis. We think these various patterns may depend on the abundance/expression level of cargoes and their degree of clustering. This will be investigated in future studies. Still, whatever the pattern, these data demonstrate and confirm the association between EndoA3 and cargoes (such as ICAM1 or ALCAM), even in the absence of antibodies.

Moreover, in Figure 2C-E, uptake of the previously established EndoA3 endocytic cargo ALCAM is analyzed by quantifying total internal fluorescence in LB33-MEL cells of antibody labelled ALCAM following both overexpression and siRNA-mediated knockdown of EndoA3, showing increased and decreased uptake respectively. Why has not the same quantification been done for the proposed novel EndoA3 endocytic cargo ICAM-1? Furthermore, if endocytosis of ICAM-1 and ALCAM is diminished following EndoA3 knockdown, the expression level on the cell surface would presumably increase accordingly. This has been shown for ALCAM previously and should also be quantified for ICAM-1.

As correctly pointed by the reviewer, anti-ICAM1 antibody uptake assays would have been great. We have tried to do them many times. Unfortunately, all commercial antibodies we tested did not yield satisfying results in uptake experiments. Either the labeling was too week/non-specific, or the antibody was not effectively stripped from the cell surface by acid washes, i.e. the acid-wash conditions required for efficient stripping were too harsh for the cells to tolerate. We have tried other approaches using the same commercial antibody which do not require acid washes (loss of surface assays by FACS, or uptake assays using surface protein biotinylation) or based on insertion of an Alfa-tag in the extracellular part of ICAM1 by CRISPR-Cas9 and detection of ICAM1 with an antiAlfa-tag nanobody (unpublished approach; collaboration with the lab of Prof. Leonardo Almeida-Souza, University of Helsinki, who developed the approach), but without success. However, we were more successful with the SNAP-tag-based approach to follow retrograde transport, for which the commercial anti-ICAM1 antibody worked properly. In Fig. 1F, we could show that retrograde transport of ICAM1 (and thus most likely its endocytosis step) was significantly decreased upon EndoA3 depletion in HeLa cells, indirectly demonstrating that ICAM1 is effectively an EndoA3-dependent cargo.

Regarding the fact that surface level of ICAM1 should increase upon perturbation of EndoA3-mediated endocytosis, we agree with the reviewer that this could be an expected result. However, this is not necessarily systematic, as the surface level of a protein cargo is always the result of a balance between its endocytosis, recycling to plasma membrane, and lysosomal degradation. We also have to take into account the neosynthesized protein flux. One must also consider that multiple endocytic mechanisms exist in parallel, and that the perturbation of one mechanism (EndoA3-mediated CIE, here) may be partially compensated by others, as cargoes can often be taken up via multiple endocytic doors. Hence, an increased abundance at the cell surface is not always guaranteed upon endocytosis perturbation. Anyway, we measured the cell surface level of both ICAM1 and ALCAM in LB33-MEL EndoA3+ cells treated with negative control or EndoA3 siRNAs (Fig. S4E-F). Only minor differences were observed.

In Figure 4A the authors show micrographs from a live-cell Airyscan movie (Movie S6) of a CD8+ T cell incubated with HeLa cells stably expressing HLA-A*68012 and transiently expressing ICAM1-EGFP. From the movie, it seems that some ICAM-1 positive vesicles in one of the HeLa cells are moving towards the T cell. However, it does not appear like the T cell has formed a stable immunological synapse but rather perhaps a motile kinapse. Furthermore, to conclude that the ICAM-1 positive vesicles are transported toward the T cell in a polarized manner, vesicles from multiple cells should be tracked and their overall directionality should be analyzed. It would also strengthen the paper if the authors could show additional evidence for polarization of the cancer cells in response to T-cell interaction.

A similar point was raised by reviewer #2. We have revised this section accordingly. In the new Fig. 4 and Movies S8-9, we replaced the live-cell Airyscan confocal data with highspeed spinning-disk confocal imaging data, enabling a more accurate analysis of cargo polarized redistribution and at a higher time resolution.

Using this approach, we captured the movement of ICAM1-positive tubulo-vesicular carriers in cancer cells at the moment of contact with CD8 T cells. Capturing such events is technically challenging, as T cell–cancer cell contacts form randomly and transiently. Successful imaging requires that the cancer cell be well spread and express ICAM1–GFP at an optimal level (as it is transiently expressed as a GFP-tagged construct), while acquisition must occur precisely at the moment when the T cell initiates contact. Despite these technical constraints, we successfully imaged early stages of immune synapse formation, enabling visualization of ICAM1 vesicular transport.

The data reveal a flux of ICAM1-positive carriers emerging from the perinuclear region (corresponding to the Golgi area) and moving toward the contact site with the CD8 T cell, with fusion events of carriers occurring at the developing immune synapse.

AI-based segmentation and tracking analyses showed that ICAM1-positive carrier trajectories were predominantly oriented toward the forming immune synapse, whereas carriers moving toward other cellular regions were markedly less frequent. These results provide direct evidence for polarized ICAM1 transport via vesicular trafficking toward the immune synapse.

Finally, in Figures 4D-G, the authors show that the contact area between CD8+ T cells and LB33-MEL cells is increased in response to siRNA-mediated knockdown of EndoA3 and VPS26A. While this could be caused by reduced polarized delivery of ICAM-1 and ALCAM to the interface between the cells, it could also be caused by other factors such as increased cell surface expression of these proteins due to diminished endocytosis, and/or morphological changes in the cancer cells resulting from disrupted membrane traffic. More experimental evidence is needed to support the working model in Figure 4H.

Regarding the cell surface expression of both ICAM1 and ALCAM, as already explained above, only minor differences were observed (Fig. S4E-F). Regarding morphological changes of cancer cells upon EndoA3 depletion (Fig. S4B-D), we compared the area, aspect ratio and roundness of LB33-MEL EndoA3+ cells treated with negative control or EndoA3 siRNAs. While we observed a slight cell area reduction upon EndoA3 depletion, no significant changes were observed regarding the aspect ratio and the roundness. Cancer cell morphology is thus not drastically modified by EndoA3 depletion. All these new data are now discussed in the manuscript.

Recommendations for the authors:

Reviewing Editor Comments:

The reviewers discussed the paper and all agreed it was incomplete in supporting the conclusions. Additional data needed to support the conclusions were:

(1) Better characterisation of Endo3A-expressing and knock-down cells such as morphology, ICAM-1, and ALCAM surface levels to name two parameters.

As discussed above, we have now added new data addressing these points:

- Morphology: Fig. S4B-D

- ICAM1 and ALCAM surface levels: Fig. S4E-F These new data are discussed in the main text.

(2) Better characterisation of the ICAM-1 polarisation process. Does this require interaction with LFA-1 can ICAM-1 be delivered to the synapse without this?

As discussed above, we have now added new data better addressing the characterization of ICAM1 polarized trafficking to the immune synapse, that can be found in the new Fig. 4 (high-speed spinning-disk confocal imaging of ICAM1 trafficking upon conjugate formation between CD8 T cell and cancer cell). The text has been modified accordingly. The dependency on LFA-1 has not been addressed directly, but we may suppose it is indeed important as (i) it has already been addressed in other cellular systems by previous studies (Jo et al. 2010), and (ii) we observed a denser flux of ICAM1-positive carriers in the cancer cell toward regions involved in immune synapses with CD8 T cells, than other regions. As we didn’t address this question more directly in our study, we briefly mentioned this point in the Discussion section.

(3) Better characterisation of T cell response- activation markers, cytotoxicity assays.

As discussed above, we have now added new data addressing these points:

- Cell surface activation markers: Fig. S4G-I

- Proliferation: Fig. S4J

- Degranulation: Fig. S4K

- Cytotoxic activity: Fig. 3F

These new data are discussed in the main text.

(4) Citing relevant literature.

The relevant literature (in particular the paper by Jo et al. 2010) is now cited and discussed.

(5) Number of donors evaluated - is it true there was only one blood donor? For human studies better to have key results on >4 donors.

Our immunological working model indeed originates from a single patient (Baurain et al., 2000), from whom both a cancer cell line (LB33-MEL) and autologous CD8 T cells were derived. These CD8 T cells specifically recognize an HLA molecule presenting a defined antigenic peptide (MUM-3) on the surface of the cancer cells. This provides us with a unique and fully natural experimental system that allows us to faithfully reconstitute cytotoxic T lymphocyte (CTL)-mediated killing of cancer cells in vitro.

Using CD8 T cells from other donors would not be meaningful in this context, as they would not recognize the LB33-MEL cells. Conversely, testing the same CD8 T cells on other cancer cell lines requires engineering these lines to express the appropriate HLA molecule and to be exogenously pulsed with the correct antigenic peptide – which is precisely what we did with the HeLa cell line.

Therefore, increasing the number of donors would require obtaining both cancer cell lines and CD8 T cells from each donor, ideally with evidence that the donor’s T cells recognize their own tumor cells. This is technically challenging and not trivial, although it would indeed be highly valuable to diversify immunological models in future studies.

Importantly, the high specificity of our autologous co-culture system, where cancer cells interact with their naturally matched CD8 T cells, offers clear advantages over commonly used in vitro models such as Jurkat (T) and Raji (B) cell lines, which rely on artificial stimulation with a superantigen to enforce immunological synapse formation and T cell activation.

(6) How does the binding of antibodies to ICAM-1 and ALCAM impact their trafficking?

As IgG antibodies are bivalent and can bind two target antigens, they may induce clustering, which could in turn affect endocytosis. To address this concern, we performed an uptake assay based on surface protein biotinylation using a cleavable biotin reagent (with a reducible linker). Briefly, after allowing endocytosis for different time intervals, cell surface–exposed biotins were removed by treatment with the cellimpermeable reducing agent MESNA, while internalized (endocytosed) biotinylated proteins remained protected. These internalized proteins were then recovered by affinity purification on streptavidin resin and analyzed by Western blot to detect the protein of interest.

Importantly, this uptake assay can be performed in the absence or presence of an anticargo antibody, allowing assessment of its potential influence on endocytosis. Author response image 1 shows the results for ALCAM uptake in HeLa cells, with and without anti-ALCAM antibody:

Author response image 1.

Antibody binding to an extracellular epitope of ALCAM increases its endocytosis. HeLa cellsurface proteins were biotinylated on ice using EZ-Link Sulfo-NHS-SS-Biotin (Pierce) and then incubated at 37 °C for the indicated times to allow endocytosis. Internalization was assessed in the absence or presence of an anti-ALCAM antibody (Ab) added to the extracellular medium. Endocytosis was stopped by returning the cells to ice, and surface-exposed biotin was removed by treatment with the cell-impermeable reducing agent MESNA. Internalized, MESNA-resistant biotinylated proteins were affinity-purified on streptavidin resin and analyzed by Western blot to detect ALCAM. The “unstripped” condition shows the total amount of ALCAM at the cell surface at the beginning of the experiment (signal at ~95 kDa). Quantification of the time course (normalized to the no-antibody condition) shows increased ALCAM endocytosis in the presence of antibody at 15 and 30 min. Blot is representative of two independent experiments; quantifications include data from both experiments.

We observed that the anti-ALCAM antibody slightly enhanced ALCAM uptake. A similar experiment was attempted for ICAM1, but we were unable to detect the protein by Western blot using the available commercial antibody.

Although this outcome was expected, it highlights a potential caveat in using antibodies to monitor endocytosis. Alternative tools such as nanobodies, while monovalent and theoretically less perturbing, are not yet available for many cargo proteins and may still influence cargo conformation or dynamics. Therefore, antibodies remain the current gold standard in endocytosis studies. Nevertheless, data obtained with antibodies should always be validated by complementary approaches that do not rely on antibody binding, as we have done in this study (e.g. live-cell imaging of fluorescently tagged proteins).

The work is of interest and we look forward to your response/revision.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Thank you for submitting your manuscript which I had the pleasure to review. While I enjoyed your work, I feel that it would strongly benefit by addressing the following points:

(1) In-depth characterization of T cell responses upon Endo3A depletion: The characterization should be expanded to include surface marker upregulation, T cell proliferation, and, most importantly, tumor cell cytotoxicity. I was wondering if the incomplete characterization of T-cell responses is due to limited supplies of antigenspecific T-cells? My understanding is that these cells have been derived from a single patient. This also raises concerns in terms of reproducibility as all data are practically from a single biological replicate. My suggestion would be to use an additional system of specific cell-cell contacts to complement the current findings. For instance, HeLa cells could be transfected to express CD19 or EpCAM, for both of which bispecific T cell engagers (Invivogen) exist that would allow specific contact formation, thereby allowing the study of the effect of Endo3A depletion across T cells from different donors and through a more complete set of assays.

We refer the reviewer to our responses above, where these points have been addressed in detail. We sincerely thank the reviewer for the excellent suggestion of transfecting HeLa cells with CD19 or EpCAM and using bispecific T-cell engagers. However, after careful consideration, we concluded that this approach falls outside the scope of the present study, which was specifically designed to investigate the most natural system, cancer cells and their autologous CD8 T cells. We nevertheless appreciate this insightful suggestion and will certainly consider it for future studies.

(2) Alterations in membrane tension as an alternative explanation: Endo- and exocytosis have been found to influence the biophysical properties of cells, such as membrane tension (e.g., Djakbaravo et al., 2021, PMID: 33788963), which in turn influences their susceptibility to cytotoxic T cells with lower tension corresponding to reduced cytotoxicity (e.g., Basu & Whitlock, 2016, PMID: 26924577). Thus, interference with endocytic pathways could arguably lead to changes in membrane tension that could contribute to the observed effects. These possible effects should be discussed and addressed experimentally to a degree. While measuring membrane tension directly requires specialized expertise (e.g., tether pulling experiments) and is not within the scope of this study, membrane tension affects cell spreading and actin organization. Thus, I would suggest conducting a thorough comparative phenotypical and morphological characterization of the Endo3A+ and Endo3A- cancer cells to estimate the possible effect of changes in membrane tension (if any) on the results.

We refer the reviewer to our responses above, where these points have been addressed in detail. New data have been added and the text of our manuscript has been modified accordingly.

(3) Citation and consideration of earlier work: Jo & Kwon et al., 2010 (PMID: 20681010) have previously shown that ICAM1 undergoes clathrin-independent recycling and repolarization to the immunological synapse in APCs. Furthermore, they provided evidence that actin-based transport, but not lateral diffusion, together with recycling is crucial for the repolarization of ICAM1 to the immunological synapse. This important earlier work has to be cited. Actin-based transport on the cell surface has not been considered in the current manuscript. In light of these earlier findings, it is unclear in Figure 4A if ICAM1 is delivered to the T cell from within- or from the surface of the cancer cell. I would suggest changing the imaging modalities in this experiment to be able to differentiate cell surface from internal ICAM1, e.g., by detaching the cancer cells from the surface as has been done in Fig. 4B, E, and F.

We refer the reviewer to our responses above, where these points have been addressed in detail. New data have been added and the text of our manuscript has been modified accordingly.

Reviewer #2 (Recommendations for the authors):

Major comments:

(1) The authors should be more careful with their claims about the importance of their results for cell polarity as their evidence for this is scarce (i.e. The live-cell imaging in Figure 4A is not quantified and the ICAM1 polarization effect shown in figure 4B-C is, albeit significant, small and not very convincing).

We refer the reviewer to our responses above, where these points have been addressed in detail. New data have been added and the text of our manuscript has been modified accordingly.

(2) The absence (or very low expression) of EndoA3 on the LB33-MEL cell suggests that EndoA3-mediated recycling of immune synaptic components is not required for T-cell activation. The fact that EndoA3 exogenous expression in LB33-MEL cells leads to increased cytokine production in T cells is, however, interesting.

We fully agree with the reviewer’s observation. Although EndoA3 is not expressed in some cellular contexts, its cargoes may still be present. It is therefore reasonable to assume that alternative endocytic mechanisms can compensate for its absence. It is now widely accepted that many cargoes can be internalized through multiple endocytic routes, and that the relative contribution of each pathway depends strongly on the cellular and physiological context.

For example, we have shown that ALCAM and L1CAM, although primarily internalized via clathrin-independent pathways, present a minor fraction (< 25%) undergoing clathrinmediated endocytosis (Renard et al., 2020; Lemaigre et al., 2023). Moreover, we observed that inhibition of macropinocytosis enhances EndoA3-mediated endocytosis of ALCAM, indicating a crosstalk between specific EndoA3-mediated clathrin-independent endocytosis (CIE) and non-specific macropinocytosis (Tyckaert et al., 2022).

Thus, even in the absence of EndoA3, its cargoes are likely internalized through alternative endocytic routes. Nonetheless, our data clearly demonstrate that EndoA3 expression markedly enhances the endocytosis and intracellular trafficking of its cargoes, ultimately leading to modified CD8 T cell responses.

(3) For the statistics in bar graphs (graphs 1C, D, E &F; 3E, 3F, S1C-I, and S3C), one cannot have all values for controls simply normalized to 1. This procedure hides the variance for the controls between each replicate and makes any statistics meaningless.

We thank the reviewer for this important remark. Regarding Figures 1C–F, S1C–I, and S3C, which correspond to quantifications from Western blots, it is standard practice to normalize the quantification to a control condition set to 1 (or 100%). Absolute signal intensities cannot be directly compared across different blots due to the variability inherent to this semi-quantitative technique. For this reason, we chose to keep the data presented in normalized form. However, we agree that this type of data require the careful choice of a convenient statistical analysis approach. Here, we choose one-sample T tests, allowing to test the hypothesis that the various siRNA conditions are different from 100% (the normalized value of the siCtrl condition). We adapted the statistical analysis accordingly in the different figures mentioned.

Regarding old Figures 3E–F (now Fig. 3E and 3G), which correspond to IFNγ secretion assays, we agree that representing IFNγ secretion as a fold change relative to a control condition may obscure inter-experimental variability. However, this format was intentionally chosen to facilitate data interpretation, as IFNγ secretion was quantified by ELISA and also displayed inter-experimental variability. For completeness, we now provide below the corresponding graphs showing absolute IFNγ concentrations, which retain the information on inter-experimental variability (Author response image 2). As you can see, the overall conclusions remain unchanged.

Author response image 2.

IFNg secretion data corresponding to Fig. 3E and 3G, expressed in absolute values (pg/mL)

Minor comments:

(1) What happens to surface and total levels of ICAM1 and ALCAM in the retromer or EndoA3 knockdown/overexpression conditions? This information would put the effects described into context.

We refer the reviewer to our responses above, where these points have been addressed in detail. New data have been added and the text of our manuscript has been modified accordingly.

(2) The authors should clearly indicate that BFA means bafilomycin A in the figure legend or methods.

BFA corresponds to Brefeldin A. We have now clarified this information in legends and methods.

(3) In the sentence: "These data demonstrate that retromer-mediated retrograde transport is critical for trafficking ALCAM and ICAM1 to the Golgi and that this process requires the full secretory capacity of the TGN." What do the authors mean by full secretory capacity?

We have modified the sentence: “Together, these data demonstrate that retromermediated retrograde transport is critical for trafficking ALCAM and ICAM1 to the Golgi and that this process requires efficient secretion from the TGN (as evidenced by the involvement of Rab6).”

(4) The method used for retrograde transport seems to be a variation of the original protocol (reference 43). The manuscript would benefit from a thorough explanation of this assay, rather than citing the original protocol.

We did not modify the original SNAP-tag–based protocol used to monitor retrograde transport. A comprehensive methodological paper has been published (ref. 44), and we have followed it strictly. Additionally, we briefly summarized the rationale of the approach in Figure 1A and in the first paragraph of the Results section.

eLife Assessment

This important study investigates how infestation by the small brown planthopper (Laodelphax striatellus) reshapes rice carbohydrate allocation and demonstrates that host-derived glucose enhances insect fecundity and imidacloprid tolerance, through the activation of conserved nutrient-sensing and endocrine pathways. Across extensive and complementary approaches, including plant manipulations, glucose supplementation, RNAi, pharmacological inhibition, rescue experiments, and biochemical assays, the authors provide convincing evidence that glucose activates the TOR-juvenile hormone-vitellogenin axis to promote reproduction and co-regulates GST-mediated detoxification via both TOR-JH signaling and GCL-GSH metabolism. The mechanistic framework is coherent and well supported by hierarchical validation and functional assays. Some limitations remain regarding the generality of the findings across other pest species and insecticides, and aspects of the evolutionary framing would benefit from more cautious interpretation; nonetheless, the work substantially advances our understanding of how plant-derived nutrients interface with conserved insect signaling pathways to shape fitness-related traits, and will be of broad interest to researchers studying plant-insect interactions, insect physiology, and pest management.

Reviewer #1 (Public review):

Summary:

The authors investigate how infestation of rice plants by the small brown planthopper (Laodelphax striatellus), an important pest in rice cultivation, alters host plant carbohydrate metabolism and how these changes affect insect physiology and fitness. They show that planthopper infestation leads to a density-dependent increase in glucose levels in rice plants, which the authors suggest results from a redistribution of carbohydrates from roots to shoots. Elevated glucose levels in plants are reflected by increased glucose contents in the insects themselves, an effect that is particularly pronounced in gravid females and associated with enhanced fecundity.

In addition, the authors demonstrate that increased glucose availability enhances tolerance of the small brown planthopper to the neonicotinoid insecticide imidacloprid. These findings suggest that insect-mediated changes in plant carbohydrate allocation may benefit insect fitness in multiple ways, including increased reproductive output and enhanced tolerance to insecticides, both of which are relevant for understanding insect population dynamics in agroecosystems.

Beyond these physiological observations, the authors aim to elucidate the underlying molecular mechanisms. They propose that glucose functions not only as a nutritional resource but also as a signaling molecule. Specifically, they show that increased glucose availability is associated with activation of the Target Of Rapamycin (TOR) pathway, a conserved nutrient-sensing signaling pathway regulating growth and metabolism across eukaryotes. Activation of TOR signaling is linked to increased juvenile hormone levels, which in turn stimulate vitellogenesis and likely contribute to increased fecundity. Furthermore, elevated juvenile hormone levels are associated with increased expression of glutathione S-transferases, suggesting a mechanism contributing to enhanced detoxification capacity. Independent of this pathway, increased glucose availability also leads to higher expression of glutamate-cysteine ligase, the rate-limiting enzyme in glutathione synthesis. Together, these mechanisms provide a non-exclusive explanation for the observed increase in imidacloprid tolerance and form the basis of the authors' proposed mechanistic framework linking glucose availability to reproduction and detoxification.

Strengths:

A major strength of the manuscript is its substantial mechanistic depth and the extensive use of complementary experimental approaches that converge on a coherent mechanistic interpretation. The authors combine plant manipulations, dietary supplementation, injection assays, RNAi-mediated gene silencing, pharmacological inhibition, and rescue experiments to systematically test the role of glucose as a signaling molecule linking plant-derived nutrition to insect reproduction and insecticide tolerance. Results obtained from independent experimental strategies are highly consistent, and the different datasets collectively support the central conclusions of the study.

The role of glucose is supported by multiple lines of evidence demonstrating that increased glucose availability, whether induced by prior planthopper feeding, dietary supplementation, or direct injection, consistently results in elevated glucose levels in insects, increased oviposition, and enhanced expression of vitellogenesis-related genes (LsVg and LsVgR). The specificity of this effect is further strengthened by experiments using alternative carbohydrates that release glucose upon enzymatic cleavage, as well as inhibitor and rescue experiments, supporting the interpretation that glucose acts beyond a purely nutritional role.

The authors further establish a mechanistic link between glucose availability, TOR signaling, juvenile hormone regulation, and vitellogenesis. Activation of TOR signaling by glucose, demonstrated at the level of protein phosphorylation, together with RNAi knockdown and pharmacological inhibition, allows causal placement of TOR upstream of juvenile hormone signaling. Consistent reductions in juvenile hormone titers, vitellogenesis-related gene expression, and oviposition following TOR inhibition, as well as rescue of reproductive output by juvenile hormone analog treatment, provide strong functional support for a glucose-TOR-juvenile hormone axis regulating fecundity. The absence of additive effects following combined knockdown of TOR and juvenile hormone synthesis components further supports the interpretation that these factors act within the same signaling cascade.

Similarly, the authors provide a detailed mechanistic analysis of glucose-mediated effects on imidacloprid tolerance. Functional assays demonstrate that glutathione S-transferases contribute to detoxification in this species and that increased glucose availability enhances GST activity, glutathione synthesis, and overall glutathione levels. Transcriptomic analyses and targeted RNAi experiments further identify specific GSTs contributing to insecticide tolerance and indicate that glucose enhances detoxification through both TOR-dependent and TOR-independent mechanisms. The combined knockdown experiments, which produce additive effects on mortality, provide particularly strong support for the involvement of multiple interacting glucose-dependent pathways.

Weaknesses:

While I am impressed by the mechanistic depth of the study and the clarity with which the authors dissect the underlying physiological pathways, I am less convinced by the current conceptual framing of the phenomenon as a sophisticated adaptive strategy "co-opted" by the small brown planthopper. The data convincingly demonstrate that glucose availability activates conserved nutrient-sensing and endocrine pathways, including TOR signaling and juvenile hormone regulation, which in turn affect reproduction and detoxification capacity. However, these pathways are deeply conserved and likely operate in many insects in response to nutritional status. As such, the results may reflect a general physiological response to elevated carbohydrate availability rather than a species-specific, evolved strategy. Relatedly, herbivory-induced changes in plant carbohydrate allocation appear to be relatively common across plant-insect systems, and it would be helpful to discuss how specific (or general) the observed phenomenon is likely to be.

In particular, I encourage the authors to more clearly distinguish between (i) a conserved nutrient-responsive signaling cascade and (ii) an adaptive mechanism that evolved specifically under selection imposed by insecticide exposure. The presented data strongly support the former interpretation, whereas evidence for the latter is less clear. The increased tolerance to imidacloprid appears to arise as a consequence of enhanced metabolic and detoxification capacity under elevated glucose conditions, rather than as a trait shaped directly by insecticide-driven selection. Framing this phenomenon as an adaptation to insecticide stress may therefore overextend the conclusions that can be drawn from the data. A more cautious discussion acknowledging that glucose-mediated activation of conserved metabolic and endocrine pathways may incidentally enhance insecticide tolerance, without necessarily having evolved under insecticide selection, would strengthen the conceptual clarity of the manuscript.

Reviewer #2 (Public review):

Summary:

Zhang and colleagues investigate the molecular mechanisms by which the small brown planthopper (SBPH, Laodelphax striatellus) manipulates host rice carbohydrate metabolism to enhance its own fitness. Using a combination of molecular, pharmacological, and biochemical approaches, they demonstrate that SBPH infestation induces systemic glucose reallocation in rice, as evidenced by the upregulation of glucose levels in aerial tissues and a simultaneous reduction in root glucose levels. Notably, host-derived glucose acts as a central signaling molecule, driving two key adaptive traits: enhanced fecundity via the glucose-TOR-JH-Vg signaling cascade, and increased imidacloprid tolerance through synergistic metabolic (GCL-GSH) and regulatory (TOR-JH-GST) pathways targeting GST activity. These findings uncover a sophisticated resource-manipulation strategy in SBPH and identify nutrient-sensing and detoxification pathways as potential targets for pest control.

Strengths:

(1) The study addresses a gap in plant-insect coevolution research by identifying glucose as a dual-function signaling molecule that coordinates SBPH reproduction and insecticide tolerance, providing valuable insights into how herbivores exploit host nutritional signals.

(2) The experimental design is well structured and multifaceted, integrating RNAi, RT-qPCR, Western blotting, pharmacological inhibition, and biochemical assays. The use of appropriate controls (e.g., osmotic controls with mannitol and hydrolase-inhibitor rescue experiments) strengthens the causal interpretation of the results.

(3) The mechanistic framework is clear and well-supported. The authors delineate two interconnected molecular cascades (glucose-TOR-JH-Vg for fecundity and GCL-GSH/TOR-JH-GST for tolerance) with hierarchical validation (e.g., rescue experiments with JHA), ensuring the reliability of conclusions.

Weaknesses:

(1) The study focuses exclusively on SBPH without validating whether the observed phenomena and mechanisms are conserved in closely related planthopper species (e.g., brown planthopper Nilaparvata lugens). This limitation restricts the generalizability of the findings to other economically important rice pests.

(2) The specific upstream signals that trigger glucose reallocation in rice (e.g., SBPH salivary effectors or oviposition-associated factors) are not identified. Although this represents a complex and independent research direction, the absence of such information limits the depth and completeness of the mechanistic framework and leaves open questions regarding the initiation of host metabolic manipulation.

(3) Insecticide tolerance assays are limited to imidacloprid. Extending these analyses to one or two additional commonly used insecticides (e.g., thiamethoxam) would help determine whether the glucose-mediated detoxification pathway is specific to imidacloprid or reflects a broader resistance mechanism, thereby strengthening conclusions regarding the generality of the GST activation cascade.

(4) Given the study's potential implications for pest management, the manuscript would benefit from a brief discussion of possible practical applications, such as manipulating rice glucose metabolism through breeding strategies or developing small-molecule inhibitors targeting the TOR-JH axis. Including such perspectives would enhance the translational relevance of the work by linking mechanistic insights to real-world pest control strategies.

eLife Assessment

This manuscript presents a valuable investigation of the peptidoglycan (PG) recycling pathway in Caulobacter crescentus. The authors showed that PG recycling in C. crescentus is essential not only for β-lactam (ampicillin) resistance but also for cell morphology, efficient division, and overall fitness. The study is comprehensive and compelling.

Reviewer #1 (Public review):

Summary:

In their manuscript, Richter and colleagues comprehensively investigate the cell wall recycling pathway in the model alphaproteobacterium Caulobacter crescentus using biochemical, imaging, and genetic approaches. They clearly demonstrate that this organism encodes a functional peptidoglycan recycling pathway and demonstrate the activities of many enzymes and transporters within this pathway. They leverage imaging and growth assays to demonstrate that mutants in peptidoglycan recycling have varying degrees of beta-lactam sensitivity as well as morphological and cell division defects. They propose that, rather than impacting the levels or activity of the major beta-lactamase, BlaA, defects in PG recycling lead to beta-lactam sensitivity by limiting the availability of new cell wall precursors. The findings will be of interest to those in the field of bacterial cell wall biochemistry, antibiotics and antibiotic resistance, and bacterial morphogenesis.

Strengths:

Overall the manuscript is laid out logically, and the data are comprehensive, quantitative, and rigorous. The mutants and their phenotypes will be a valuable resource for Caulobacter researchers, and the findings may be relevant to cell wall recycling in other organisms.

Weaknesses:

No major weaknesses are noted.

Comments on revisions:

The authors addressed all of our concerns with the initial submission.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

In their manuscript, Richter and colleagues comprehensively investigate the cell wall recycling pathway in the model alphaproteobacterium Caulobacter crescentus using biochemical, imaging, and genetic approaches. They clearly demonstrate that this organism encodes a functional peptidoglycan recycling pathway and demonstrate the activities of many enzymes and transporters within this pathway. They leverage imaging and growth assays to demonstrate that mutants in peptidoglycan recycling have varying degrees of beta-lactam sensitivity as well as morphological and cell division defects. They propose that, rather than impacting the levels or activity of the major beta-lactamase, BlaA, defects in PG recycling lead to beta-lactam sensitivity by limiting the availability of new cell wall precursors. The findings will be of interest to those in the field of bacterial cell wall biochemistry, antibiotics and antibiotic resistance, and bacterial morphogenesis.

Strengths:

Overall, the manuscript is laid out logically, and the data are comprehensive, quantitative, and rigorous. The mutants and their phenotypes will be a valuable resource for Caulobacter researchers.

Thank you for this positive evaluation. Previous work has mostly focused on the role of PG recycling in the regulation of ampC expression. However, our study and recent work in A. tumefaciens (Gilmore & Cava, 2022) and C. crescentus (Modi et al, 2025) demonstrates that β-lactam resistance is heavily influenced by PG recycling and the metabolic state of the cell, even in the presence of high levels of β-lactamase activity. It is likely that these effects are not limited to the two alpha­proteo­bacterial species investigated to date but may be more widely applicable. Therefore, we believe that our results are relevant beyond the Caulobacter field and may help to stimulate similar analyses in other, medi­cally more relevant species.

Weaknesses:

The only major missing piece is the complementation of mutants to demonstrate that loss of the targeted gene is responsible for the observed phenotypes.

In our initial manuscript, we showed that the replacement of the native AmiR and NagZ genes with mutant alleles encoding catalytically inactive variants of the two proteins gave rise to the same pheno­types as gene deletions. This finding indicates that the defects observed were due to the loss of AmiR or NagZ activity, respectively. To rule out artifacts from polar effects, we have now also conducted the requested complementation analysis for the ΔampG, ΔamiR and ΔnagZ mutants. The results obtained show that deletion mutants carrying an ectopically expressed wild-type gene copy behave essentially like the wild-type strain, thereby verify­ing the validity of our conclusions (new Figure 4-figure supple­ment 1).

Reviewer #2 (Public review):

Summary:

Pia Richter et al. investigated the peptidoglycan (PG) recycling metabolism in the alpha-proteobacterium Caulobacter crescentus. The authors first identified a functional recycling pathway in this organism, which is similar to the Pseudomonas route, and they characterized two key enzymes (NagZ, AmiR) of this pathway, showing that AmiR differs in specificity from the AmpD counterpart of E. coli. Further, they studied the effects of deletions within the PG recycling pathway (ampG, amiR, nagZ, sdpA, blaA, nagA1, nagA2, amgK, nagK mutants), showing filamentation and cell widening, thereby revealing a link between PG recycling and cell division. Finally, they provide a link between PG recycling and beta-lactam sensitivity in C. crescents that is not caused by activation of a beta-lactamase, but rather is a result of reduced supply of PG building blocks increasing the sensitivity of penicillin-binding proteins.

Strengths:

This work adds to the understanding of the role of PG recycling in alpha-proteobacteria, which significantly differ in their mode of cell wall growth from the better studied gamma-proteobacteria.

Thank you for pointing out the relevance of our work. As mentioned above, we believe that our work goes beyond understanding the PG recycling pathway in alphaproteobacteria. Importantly, together with previous work, our results demonstrate a so-far largely neglected critical role of PG recycling in β-lactam resistance that goes beyond the mere regula­tion of β-lactamase gene expression. It will be interesting to determine the conservation of this phenomenon among other bacteria and to see whether blocking PG recycling could represent a potential strategy to combat β-lactam resistant pathogens.

Weaknesses:

The findings are not entirely novel as recent studies by Modi et al. 2025 mBio (studying C. crescentus) and Gilmore & Cava 2022 Nat. Commun. (studying Agrobacterium tumefaciens) came to similar conclusions.

Gilmore & Cava have made the seminal finding that blocking anhydro-muropeptide import affects cell wall integrity in a manner that is partly independent of its effect on ampC expression. We now extend this finding by investigating various critical steps in the PG recycling pathway of C. cres­centus, a species lacking an AmpC homolog. Interestingly, by characterizing a variety of different mutants, we show that the morphol­ogical and ampicillin resistance defects they exhibit are not strictly con­nected and vary substantially between strains, suggesting that different steps in PG recycling differ in their importance for cellular fitness and cell wall integrity. This finding suggests that the phenotypes observed are not simply determined by the efficiency of PG recycling but likely result from a combination of factors. Based on the results obtained, we propose a model that highlights the different factors that may be at play and suggests a mechanism explaining their effects on β-lactam resistance and cell division. Our findings partly overlap with the recent study by Modi et al., but there are various points in which we disagree with their findings and conclusions. The need to rigorously validate our differing results led to a signi­ficant delay in the submission of our manuscript.

Reviewer #1 (Recommendations for the authors):

Major Comment

Genetic complementation is lacking for deletion mutants throughout. Could you please provide complemented strains for mutants in key figures where deletion phenotypes are central to the conclusions (e.g., Figure 4 and related supplements).

As explained above, we have not performed the requested comple­mentation experiments and included the data as Figure 4-figure supplement 1.

Other minor comments:

(1) Figure 1

(a) This is a busy schematic; please consider visually separating PG biosynthesis vs. recycling (e.g., a faint divider line or shaded boxes).

We have now simplified the schematic and visually separated the PG recycling and de novo biosyn­thesis pathways.

(b) Please label "Fructose-6-phosphate" and "Glucosamine-6-phosphate (GlcN-6-P)" on the figure, since they are referenced in the caption (line 1410).

The symbols for fructose, glucosamine and phosphate are given in the legend on the right. For consistency, we would therefore prefer not to additionally label these compounds in the figure.

(c) Define all abbreviations in the caption: CM, GTase, TPase; and clarify the legend conventions (e.g., bold vs. regular font; red vs. black text).

The structure of PG and the different lytic enzymes have now been removed from Figure 1. All remaining abbreviations have now been defined in the legend.

(2) Figure 2 - Figure Supplement 2

(a) Panel B: Please include the full chromatogram (it seems to be cropped at 10 min?). For AmiR in particular, it is important to show there are no nearby peaks at earlier retention times (eg GlcNAc).

The region before 10 min is cropped in many published muropeptide profiles because the peaks contained in it are known to correspond to salts, i.e., borate from the reduction step and phos­phate, which are poorly retained on the C18 column (Figure 2–figure supplement 2). As the reviewer stated, free GlcNAc would elute in this region and would not be recognized if it were produced by AmiR. However, AmiR cleaves free anhydro-muropeptides between anhMurNAc and the peptide, and the experiment in Figure 2–figure supplement 2 shows that it does not cleave the bond between MurNAc and peptides in intact peptidoglycan.

(b) Caption line 1439: with AmiR OR the catalytically...

Done.

(3) Figure 3

Panel A: Label the products as NagZ-treated.

In this analysis, we quantify specific intermediates from the total cellular pool of PG recycling inter­mediates. Since the products were not specifically treated with NagZ, we would prefer to keep the figures as it is.

(4) Figure 4 (and Fig. 4-Figure Supplement 1, 2)

(a) Please add complemented strains for ΔampG, ΔamiR, and ΔnagZ under the same conditions.

As described in more detail above, we have now performed the requested complementation analysis.

(b) Figure 4 - Figure S1 - Please include images of all strains quantified in B (e.g. control WT).

Done.

(c) Figure 4 - Figure S2: A. Please include images of all strains quantified in B. Please include spotting dilutions on minimal medium to assess the importance of PG recycling under nutrient limitation, especially given apparent lysis in ΔamiR and ΔampG.

The length distributions of cells grown in PYE medium are taken from Figure 3 and only shown for comparison (as mentioned in the figure legend). To avoid the duplication of images, we would prefer to keep panel A as it is.

We have now performed the requested serial-dilution spot assay on minimal (M2G) medium. The results show that ampicillin resistance de­creases even more dramatically for all strains in this condi­tion. The new data are presented in Figure 4-figure supplement 3C.

(d) Figure 4 - Figures S3: A and B. Please include WT control.

We have now added images of the wild-type strain to panel B of this figure. The serial dilution spot assays shown in panel A were performed on the same plates as those depicted in Figure 4 (as men­tioned in the figure legend). To avoid the duplication of images, we would prefer to keep this panel as it is.

(5) Figure 5

A, C - please include images of WT control.

We have now added images of the wild-type strain to panel A of this figure. The serial dilution spot assays shown in panel C were performed on the same plates as those depicted in Figure 4 (as men­tioned in the figure legend). To avoid the duplication of images, we would prefer to keep this panel as it is.

(6) Figure 6:

(a) A, C - please include images of WT control.

We have now added images of the wild-type strain to panel A of this figure. The serial dilution spot assays shown in panel C were performed on the same plates as those depicted in Figure 4 (as men­tioned in the figure legend). To avoid the duplication of images, we would prefer to keep this panel as it is.

(b) It would be informative to test ΔamgK and ΔanmK on minimal medium (spotting and/or growth curves) to position these steps within the nutrient-dependent fitness landscape.

We have now analyzed the ampicillin sensitivity of the ΔamgK, ΔnagK and ΔamgK ΔnagK strains on minimal medium (see Author response image 1). Consistent with the results obtained for other mutants in the PG recycling pathway, growth on minimal (M2G) medium plates leads to increased ampicillin sensi­tivity of the ΔamgK mutant. By contrast, ΔnagK and, to a lesser extent, ΔamgK ΔnagK cells show an in­creased tolerance to ampicillin under these conditions compared to growth on PYE plates.

This phenomenon may be explained by the strong stimulatory effect of GlcNAc-6-P on NagB acti­vity. In the absence of NagK, GlcNAc-6-P levels drop, leading to reduced activation of NagB1/2. This effect, combined with abundant glucose to support central carbon metabolism may promote the GlcN-6-P biosynthesis through GlmS, thereby increasing the flux of meta­bol­ites into the de novo PG biosynthesis pathway and thus boosting ampicillin tolerance. However, more re­search is required to fully under­stand the molecular basis of this effect. Given that the results are likely to reflect complex interactions bet­ween dysregulated enzyme activity and altered metabolite pools caused by increased glucose avail­ability, they provide only limited insight into the role of PG recycling in ampicillin resistance. We therefore propose excluding this experiment from the present manuscript to avoid confusion.

Author response image 1.

Serial-dilution spot assay investigating the ampicillin resistance of the indicated mutant strains on minimal (M2G) medium plates.

(c) Could Figures 6 and 7 be combined for better comparison and since there is no WT control? If so, could you also include the MurNAc cytoplasmic level quantification for the double mutant (Figure 7)?

We would prefer to keep the two figures separated to avoid creating an overly large figure that contains a total of nine panels. However, we have now included an additional panel in Figure 7 show­ing the levels of MurNAc in the double mutant.

(7) Figure 7. A, C

Please include images of WT control.

We have now added images of the wild-type strain (now panel B). The serial dilution spot assays (now panel D) were performed on the same plates as those depicted in Figure 4 (as men­tioned in the figure legend). To avoid the duplication of images, we would prefer to keep this panel as it is.

(8) Figure 8-S1D, F

Please include images of WT control.

Panel F of this figure already contains a wild-type control.

(9) Figure 10 A, C

Please include images of WT control and ∆amiR (A).

Done.

(10) Figure 11

Consider adding or highlighting in this figure (in a simplified manner) the major PG recycling differences in Caulobacter? The current model doesn't really show any difference that is unknown.

This figure presents a model of the mechanism underlying the increased β-lactam sensitivity of PG recycling-deficient cells. Since the PG recycling pathway of C. crescentus is already presented in detail in Figure 1, we would like to keep this figure simple and thus leave it as it is.

(11) Comments by lines:

(a) Line 192: Clarify that NagZ is also part of the rate-limiting step since there is no difference between AmiR or NagZ order of hydrolysis?

We have now omitted the statement that AmiR catalyzes the rate-limiting step in the PG recycling process, because our data do not allow definitive conclusions on this point.

(b) Line 201: Define "considerable fraction" since this is known, please and cite original reference(s).

Done.

(c) Line 203: Please also cite the primary papers where they have found that disruption of the PG recycling pathway in E. coli and P. aeruginosa doesn't result in morphological defects.

Since there are a number of papers that report PG recycling-deficient mutants of E. coli and P. aeru­ginosa, we would like to keep citing reviews to support this statement. However, we have now addi­tionally included a review by Park & Uehara (2008), which provides a detailed overview of PG recycling in bacteria.

(d) Line 220-223: Though there are no obvious morphological defects, several mutants (e.g., ΔamiR, ΔampG) appear to be lysing or stressed under minimal conditions. Could you include spotting assays and/or growth curves on minimal medium (Figure 4, Figure S2) to quantify fitness under nutrient limitation?

Have performed the requested serial dilution spot assays on minimal (M2G) medium plates and now present the data obtained in Figure 4-figure supplement 3C.

(e) Line 224: PG recycling has been found to contribute to the regulation of B-lactam resistance in several organisms, not just those two. Perhaps add "including C. freundii and P. aeruginosa"

Done.

(12) Typographical errors:

(a) Line 284: "caron" should be carbon.

Done.

(b) Line 323: "Figure C" needs a figure number.

Done.

(c) Line 33: "regulaton" should be regulation.

Done.

Reviewer #2 (Recommendations for the authors):

(1) The study is well conducted and describes a number of experiments that significantly deepen previous findings. The conclusions of this paper are mostly well supported by data, but some experiments and data analysis may need to be clarified and extended.

Thank you for this positive evaluation.

(2) The data presented in Figures 2B and 2C show activities of AmiR and NagZ using LTase-cleaved cell wall preparations. Unfortunately, the preparations tested with the two enzymes should be identical, but apparently are not. Why aren't identical preparations used?

We are sorry for the confusion. As stated in the Methods section (page 28, lines 757 and 773), the AmiR activity assays used LT products from PG sacculi isolated from E. coli D456, whereas the NagZ activity assays used LT-products from PG sacculi isolated from E. coli CS703-1. Both strains have a higher penta­peptide content than wild-type E. coli D456 lacks PBPs 4, 5 and 6 and has a moderate level of pentapeptides. CS703-1 lacks PBPs 1a, 4, 5, 6, 7 as well as AmpC and AmpH, and is known to have a higher pentapeptide content than D456. These differences are the reason for the distinct muro­peptide profiles in panel B and C of Figure 2.

(3) I am missing a control experiment where muropeptides treated with NagZ were further digested with AmiR? This would show whether AmiR is able or not to cleave MurNAc-peptides. This is not evident from the provided experiments.

We have now tested the activity of AmiR towards anhMurNAc-tetrapeptide in vitro. The results show that AmiR efficiently cleaves this GlcNAc-free anhydro-muropeptide species, verifying that it can also act on turnover products that have been previously processed by NagZ. The new data are shown in Figure 2–figure supplement 5.

(4) The claim that PG recycling is critical, particularly upon transition to the stationary phase and under nutrient limitation, is not justified. It conflicts with the obvious morphological effects also in the exponential phase and with the absence of morphological defects in minimal medium: pronounced defects in rich PYE medium (Figure 4A/B) disappear in minimal M2G medium (Figure 4_figure supplement 2). It seems that catabolite repression effects apply here. Is the morphological effect in rich PYE medium reversed by adding glucose?

We agree that PG recycling is not considerably more important in stationary phase and have removed this statement. Interestingly, while PG recycling-deficient mutants show no obvious mor­phol­ogical defects in minimal (M2G) medium, their ampicillin sensitivity even increases under this condi­tion (new Figure 4-figure supplement 3C), confirming that morphological and resistance defects are not strictly coupled. Preliminary data indicate that the morphological defects of the mutant cells are also abolished upon growth in PYE+glucose medium. High glucose availability may promote increased de novo synthesis of PG precursors, thereby partially restoring the PG precursor pool. We propose that the morphological and resistance phenotypes develop at different degrees of PG precursor depletion. However, future research is required to clarify the precise molecular basis of this phenomenon.

(5) Figure 4: Why is the contribution of AmpG to ampicillin resistance much lower than for amiR or nagZ, despite ampG mutants showing the largest morphological defects? Does the accumulation of UDP-MurNAc or UDP-MurNAc-peptide correlate with ampicillin resistance, whereas the morphological effects correlate with the lack of precursors?

The exact reason why the ΔampG mutant shows such a strong discrepancy in the severity of its morphol­ogical and resistance defects compared to the ΔamiR and ΔnagZ mutants remains unclear, because all of these deletions completely block the recycling of anhydro-muropeptides. The major difference in the ΔampG mutant is its inability to import anhydro-muropeptides, causing their accu­mu­lation in the periplasm. We propose that periplasmic anhydro-muropeptides, in particular the penta­peptide-containing species, can interact with the substrate-binding sites of PG metabolic enzymes, thereby interfering with proper PG biosyn­thesis. Conversely, by interacting with transpep­tidases, they may reduce their accessibility to ampicillin and thus preserve their acti­vity under β-lactam stress, particularly under conditions in which low PG precursor availability reduces binding site occupancy and thus facilitates antibiotic association.

eLife Assessment

This important study introduces an advance in multi-animal tracking by reframing identity assignment as a self-supervised contrastive representation learning problem. It eliminates the need for segments of video where all animals are simultaneously visible and individually identifiable, and significantly improves tracking speed, accuracy, and robustness with respect to occlusion. This innovation, which is supported through compelling evidence, has implications beyond animal tracking, potentially connecting with advances in behavioral analysis and computer vision.

Reviewer #3 (Public review):

Summary:

The authors propose a new version of idTracker.ai for animal tracking. Specifically, they apply contrastive learning to embed cropped images of animals into a feature space where clusters correspond to individual animal identities. By doing this, they address the requirement for so-called global fragments - segments of the video, in which all entities are visible/detected at the same time. In general, the new method reduces the long tracking times from the previous versions, while also increasing the average accuracy of assigning the identity labels.

Comments on revisions:

I have no additional comments, the authors have responded to all the points I raised previously.

Author response:

The following is the authors’ response to the previous reviews

eLife Assessment

This important study introduces an advance in multi-animal tracking by reframing identity assignment as a self-supervised contrastive representation learning problem. It eliminates the need for segments of video where all animals are simultaneously visible and individually identifiable, and significantly improves tracking speed, accuracy, and robustness with respect to occlusion. This innovation has implications beyond animal tracking, potentially connecting with advances in behavioral analysis and computer vision. The strength of support for these advances is compelling overall, although there were some remaining minor methodological concerns.

To tackle “minor methodological concerns” mentioned in the Editorial assessment and Reviewer 3, the new version of the manuscript includes the following changes:

a) The new ms does not anymore use the word “accuracy” but “IDF1 scores”. See, for example, Lines 46, 161, 176, and 522 for our new wording as “IDF1 scores”.

b) Instead of comparing softwares using mean accuracy over the benchmark, Reviewer 3 proposes to use medians or even boxplots. We now provide boxplot results with mean, median, percentiles and outliers (Figure 1- figure Supplement 2).

Additionally, we also include in the text the other recommendations from Reviewer 3:

a) We now more explicitly describe the problems of the original idtracker.ai v4 in the benchmark (lines 66-68). Around half of the videos had a high accuracy in the original dtracker.ai (v4) but the other half of the videos had lower accuracies (Figure 1a, blue). The new idtracker.ai has high accuracy values for all the videos (Figure 1a, magenta).

Also, the videos with high accuracy in the old idtracker.ai had very long tracking times (Figure 1b, blue) and the new version does not (Figure 1b, magenta). So the benchmark allows us to distinguish the new idtracker.ai as having a better accuracy for all videos and lower tracking times, making it a much more practical system than previous ones. 

b) We further clarified the occlusion experiment (lines 188-190 and 277-290).

c) We explain why we measure accuracies with and without animal crossings (lines 49-62).

d) We added a Discussion section (lines 223-244).

We believe the new version has clarified the minor methodological concerns.

Reviewer #3 (Public review):

The authors have reorganized and rewritten a substantial portion of their manuscript, which has improved the overall clarity and structure to some extent. In particular, omitting the different protocols enhanced readability. However, all technical details are now in appendix which is now referred to more frequently in the manuscript, which was already the case in the initial submission. These frequent references to the appendix - and even to appendices from previous versions - make it difficult to read and fully understand the method and the evaluations in detail. A more self-contained description of the method within the main text would be highly appreciated.

In the new ms, we have reduced the references to the appendix by having a more detailed explanation in one place, lines 49-62.

Furthermore, the authors state that they changed their evaluation metric from accuracy to IDF1. However, throughout the manuscript they continue to refer to "accuracy" when evaluating and comparing results. It is unclear which accuracy metric was used or whether the authors are confusing the two metrics. This point needs clarification, as IDF1 is not an "accuracy" measure but rather an F1-score over identity assignments.

We thank the reviewer for noticing this. Following this recommendation, we changed how we refer to the accuracy measure with “IDF1 score” in the entire ms. See, for example, lines 46, 161, 176, and 522.

The authors compare the speedups of the new version with those of the previous ones by taking the average. However, it appears that there are striking outliers in the tracking performance data (see Supplementary Table 1-4). Therefore, using the average may not be the most appropriate way to compare. The authors should consider using the median or providing more detailed statistics (e.g., boxplots) to better illustrate the distributions.

We thank the reviewer for asking for more detailed statistics. We added the requested box plot in Figure 1- figure Supplement 2 to provide more complete statistics in the comparison.

The authors did not provide any conclusion or discussion section. Including a concise conclusion that summarizes the main findings and their implications would help to convey the message of the manuscript.

We added a Discussion section in lines 223-244.

The authors report an improvement in the mean accuracy across all benchmarks from 99.49% to 99.82% (with crossings). While this represents a slight improvement, the datasets used for benchmarking seem relatively simple and already largely "solved". Therefore, the impact of this work on the field may be limited. It would be more informative to evaluate the method on more challenging datasets that include frequent occlusions, crossings, or animals with similar appearances.

Around half of the videos also had a very high accuracy in the original dtracker.ai (v4) but the other half of the videos had lower accuracies (Figure 1a, blue). For example, we found IDF1 scores of 94.47% for a video of 100 zebrafish with thousands of crossings (z_100_1), 93.77% for a video of 4 mice (m_4_2) and 69.66% for a video of 100 flies (d_100_3). The new idtracker.ai has high accuracy values for all the videos (Figure 1a, magenta).

Importantly, the tracking times for the majority of videos was very high in the original idtracker.ai (Figure 1b, blue), making the use of the tracking system limited in practice. The new system manages both a high accuracy in all videos (Figure 1a, magenta) and much lower tracking times (Figure 1b, magenta), making it a much more practical system..

We have added a sentence of the limitations of the original idtracker.ai as obtained from the benchmark, lines 66-68.

The accuracy reported in the main text is "without crossings" - this seems like incomplete evaluation, especially that tracking objects that do not cross seems a straightforward task. Information is missing why crossings are a problem and are dealt with separately.

We have now added an explanation on why we measure accuracy without crossings and why we separated it from the accuracy for all the trajectory in lines 49-62. The reason is that the identification algorithm being presented in this ms only identifies animal images outside the crossings. This algorithm makes robust animal identifications through the video despite the thousands of animal crossings typically existing in each of our videos used in the benchmark. It is a second algorithm (that hasn’t changed since the first idTracker in 2014) the one that assigns animal positions during crossings once the first algorithm has made animal identifications before and after the crossings.

There are several videos with a much lower tracking accuracy, explaining what the challenges of these videos are and why the method fails in such cases would help to understand the method's usability and weak points.

Some videos had low accuracy on previous versions (Figure 1a, blue), but the new idtracker.ai has high accuracy in all of them (Figure 1a, magenta).

Reviewer #3 (Recommendations for the authors):

(1) As described before, the authors claim to use IDF1 as their metric in the whole manuscript (lines 414-436) but only refer to accuracy when presenting the results. It is not clear, whether accuracy was used as a metric instead of IDF1 or the authors are confusing these metrics.

Following this recommendation, we replaced “accuracy” with “IDF1 score” , see lines 46, 161, 176, and 522.

(2) In the introduction, a brief explanation why crossings need to be dealt with separately would help to understand the logic of the method design.

We added such an explanation in lines 49-62.

(3) Figure 3: We asked about how the tracking accuracy is being assessed with occlusions. The authors responded with that only the GT points inside the ROI are taken into account when computing the accuracy. Does this mean, that the occluded blobs are still part of the CNN training and the clustering? This questions the purpose of this experiment, since the accuracy performance would therefore only change, if the errors, that their approach is doing either way, are outside the ROI and, therefore, not part of the metric evaluation.

The occluded blobs are not part of any training because they are erased from the video, they do not exist. We made this more clear in lines 188-190 and 277-290.

(4) Figure 1: The fact that datasets are connected with a line is misleading - there is no connection between the data along the x-axis. A line plot is not an appropriate way to present these results.

The new ms clarifies that the lines are for ease of visualization, see last line in the caption of Figure 1.

(5) Lines 38-39: It is not clear how the CNN can be pretrained for the entire video if there are no global segments or only short ones. Here, the distinction between "no segments", "only short segments" and "pretraining on the entire video" is not explained.

This pretraining protocol is not used in the version of the software we present, so details of this are not as relevant.

(6) Figure 2a: The authors are showing "individual fragments" and individual fragments in a global fragment." However, it seems there are a few blue borders missing. In the text (l. 73-79), they note, that they are displaying them as "examples" but the absence of correct blue borders is confusing.

In the new ms, we have replaced the label “Individual fragments in a global fragment” with “Individual fragments in an example global fragment” in the legend of Figure 2.

(7) Lines 61-63, 148-151, and 162-164: Could the authors clarify why they used the average instead of median when comparing the speedups of the new version and the old ones?

We thank the reviewer for asking for more detailed statistics. We added the requested box plot in Figure 1- figure Supplement 2 to provide more complete statistics in the comparison of accuracies and tracking times for old and new systems.

(8) Lines 140-144: The post-processing steps are not clear. The authors should rather state clearly which processes of the old versions they are using. Then the authors could shortly explain them.

We removed this paragraph and explained in more detail in lines 49-62 which parts of the software are new and which ones are not.

(9) Lines 239-251: Here, the authors are clarifying on a section 1-2 pages before. This information should be directly in that section instead.

Following this recommendation, we clarified the occlusion experiment in the main text (lines 188-191) to make it more self-contained. Still, the flow of the main text is better with some details in Methods.

(10) Line 38: It is not clear how the CNN can be pretrained for the entire video if there are no global segments or only short ones. Here, the distinction between "no segments"

"only short segments" and "pretraining on the entire video" is a bit misleading/underexplained.

See number 5.

(11) Figure 2a: The authors are showing "individual fragments" and individual fragments in a global fragment." However, it seems there are a few blue borders missing. In the text (l. 73-79), they note, that they are displaying them as "examples" but the absence of correct blue borders is confusing.

See number 6.

(12) Figure 2c and line 115-118: "Batches" itself is not meaningful without any information of the batch size. The authors should rather depict the batch size and then the number of epochs. The Figure 2 contains the info 400 positive and 400 negative pairs of images per batch. However, there is no information about the total number of images.

Furthermore, these metrics are inappropriate here, since training is carried out from scratch (or already pre-trained) for every new video, each video has different number of animals, different number of images.

Following this recommendation, we clarified the number of images in each batch (Figure 1c caption and lines 134-138), why we do not work with epochs (lines 700-702), and the idea that the clusters in Figure 2 represent an example and the number of batches needed for the clusters to form depends on the video details.

Appendix 1-figure 1: why do the methods fail? It looks that for certain videos the method is fairly unreliable. What is the reason for the methods to crash and how to avoid this?

Those failures are only for the old idtracker.ai and Trex, not for the method presented here. Our new contrastive algorithm does not fail in any of the videos in the benchmark.

We thank the reviewer for the detailed suggestions. We believe we have incorporated all of them in the new version of the ms.

eLife Assessment

Karimian et al. present a valuable new model to explain how gamma-band synchrony (30-80 Hz) can support human visual feature binding by selectively grouping image elements, countering recent criticisms that the stimulus dependence of gamma oscillations limits their functional role. Grounded in the theory of weakly coupled oscillators the model captures behavioural patterns observed in human psychophysics, offering support for the potential role of synchrony-based mechanisms in feature-binding. The development of the model in alignment with primate electrophysiology convincingly supports the paper's claims that gamma synchrony may be the underlying mechanism. While the paper does not present electrophysiological results that directly link gamma oscillations to figure-ground segregation in the presented task, the model makes several predictions that can be tested experimentally.

Reviewer #1 (Public review):

Summary:

This paper by Karimian et al proposes an oscillator model tuned implementing binding by (gamma) synchrony principles in a visual task. The authors set out to show how well these principles explain human behavior in a figure-ground segregation tasks. The model is inspired by electrophysiological findings in non-human primates suggesting that gamma oscillations in early visual cortex implement feature-binding through a synchronization of feature-selective neurons. The psychophysics experiment involves the identification of a figure consisting of gabor annuli, presented on a background of gabor annuli. The participants' task is to identify the orientation of the figure. The task difficulty is varied based on the contrast and density of the gabor annuli that make up the figure. The same figures are used as inputs to the oscillator model. The authors report that both the discrimination accuracy in the psychophysics experiment and the synchrony of the oscillators in the proposed model follow a similar "Arnold Tongue" relationship when depicted as a function of the texture-defining features of the figure. This finding is interpreted as evidence for gamma synchrony being the underlying mechanism of the figure-ground segregation.

Strengths:

The design of the proposed model is well-informed by electrophysiological findings, and the idea of using computational modeling to bridge between intracranial recordings in non-human primates and behavioral results in human participants is interesting. Previous work has criticized the gamma synchrony theories based on the observation that synchronization in the gamma-band is highly localized and the frequency of the oscillation depends on the visual features of the stimulus. I appreciate how the authors demonstrate that frequency-dependence and local synchronization can be features of gamma synchrony, and not contradictory to the theory. As such, I feel that this work has the potential to contribute meaningfully to the debate on whether binding by gamma synchrony is a biophysically realistic model of feature-binding in visual cortex.

I also acknowledge the additional simulations the authors present in this version of the manuscript, showing that the model is able to segregate figure from ground.

Weaknesses:

The authors have addressed my previous concerns regarding the quantification of effect sizes. I also appreciate the authors argument that the results support the idea of feature-binding through synchronization in the gamma-band, as the model's parameters were informed by electrophysiological recordings from non-human primates. Personally, I would have been curious to see if the intrinsic frequencies of the model are indeed in the gamma-band, I don't believe the authors include a figure on that. Weaknesses are still the absence of electrophysiological recordings to support the frequency-specificity of the claims, e.g. in the form of EEG/MEG recordings, but I understand that these may be difficult to obtain, as gamma oscillations are relatively weak in response to static gratings. As the authors emphasize in this updated version, they present one possible mechanism of feature binding that is not contrasted to alternative mechanisms such as binding by increased firing rates. Understandably, implementing a second model would be out of scope.

The presented simulations and behavioural results support the authors aim of presenting an oscillator model informed by gamma synchronization in V1 that supports figure-ground segregation.

Likely impact:

This work makes several predictions about the degree of synchronization for different visual properties of the figure, that could be tested with electrophysiological methods. I therefore believe that the paper has the potential to motivate interesting follow-up studies to understand how visual cortex solves the binding problem.

Comment on revised version:

In this reviewed version of the manuscript, the authors present several follow-up simulations and clarifications that address previously outlined weaknesses.

Reviewer #2 (Public review):

The authors aimed to investigate whether gamma synchrony serves a functional role in figure-ground perception. They specifically sought to test whether the stimulus-dependence of gamma synchrony, often considered a limitation, actually facilitates perceptual grouping. Using the theory of weakly coupled oscillators (TWCO), they developed a framework wherein synchronization depends on both frequency detuning (related to contrast heterogeneity) and coupling strength (related to proximity between visual elements). Through psychophysical experiments with texture discrimination tasks and computational modeling, they tested whether human performance follows patterns predicted by TWCO and whether perceptual learning enhances synchrony-based grouping.

Strengths:

(1) The theoretical framework connecting TWCO to visual perception is innovative and well-articulated, providing a potential mechanistic explanation for how gamma synchrony might contribute to both feature binding and separation.

(2) The methodology combines psychophysical measurements with computational modeling, with a solid quantitative agreement between model predictions and human performance.

(3) In particular, the demonstration that coupling strengths can be modified through experience is remarkable and suggests gamma synchrony could be an adaptable mechanism that improves with visual learning.

(4) The cross-validation approach, wherein model parameters derived from macaque neurophysiology successfully predict human performance, strengthens the biological plausibility of the framework.

Likely Impact and Utility:

This work offers a fresh perspective on the functional role of gamma oscillations in visual perception. The integration of TWCO with perceptual learning provides a novel theoretical framework that could influence future research on neural synchrony.

The computational model, with parameters derived from neurophysiological data, offers a useful tool for predicting perceptual performance based on synchronization principles. This approach might be extended to study other perceptual phenomena and could inspire designs for artificial vision systems.

The learning component of the study may have a particular impact, as it suggests a mechanism by which perceptual expertise develops through modified coupling between neural assemblies. This could influence thinking about perceptual learning more broadly, but also raises questions about the underlying mechanism.

Additional Context:

Historically, the functional significance of gamma oscillations has been debated, with early theories of temporal binding giving way to skepticism based on gamma's stimulus-dependence. This study reframes this debate by suggesting that stimulus-dependence is exactly what makes gamma useful for perceptual grouping.

The successful combination of computational neuroscience and psychophysics is a significant strength of this study.

The field would benefit from future work extending (if possible) these findings to more naturalistic stimuli and directly measuring neural activity during perceptual tasks. Additionally, studies comparing predictions from synchrony-based models against alternative mechanisms would help establish the specificity of the proposed framework.

Comments on revised version:

The authors now soften their claim. However, the paper demonstrates that TWCO-derived predictions quantitatively match human figure-ground perception in texture stimuli, and that a synchrony-based readout provides a viable mapping from stimulus to behavior. Given that they cite (and do not show in this paper) the link to synchrony, what they actually establish is that this particular transformation of stimulus features maps better onto behavior. That's meaningful, but it is not a demonstration of mechanism.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

This paper by Karimian et al proposes an oscillator model tuned to implement binding by synchrony (BBS*) principles in a visual task. The authors set out to show how well these BBS principles explain human behavior in figure-ground segregation tasks. The model is inspired by electrophysiological findings in non-human primates, suggesting that gamma oscillations in early visual cortex implement feature-binding through a synchronization of feature-selective neurons. The psychophysics experiment involves the identification of a figure consisting of gabor annuli, presented on a background of gabor annuli. The participants' task is to identify the orientation of the figure. The task difficulty is varied based on the contrast and density of the gabor annuli that make up the figure. The same figures (without the background) are used as inputs to the oscillator model. The authors report that both the discrimination accuracy in the psychophysics experiment and the synchrony of the oscillators in the proposed model follow a similar "Arnold Tongue" relationship when depicted as a function of the texture-defining features of the figure. This finding is interpreted as evidence for BBS/gamma synchrony being the underlying mechanism of the figure-ground segregation.

Note that I chose to use "BBS" over gamma synchrony (used by the authors) in this review, as I am not convinced that the authors show evidence for synchronization in the gamma-band.

We thank the reviewer for their careful assessment of our manuscript and useful comments that we believe have served to strengthen our work.

Strengths:

The design of the proposed model is well-informed by electrophysiological findings, and the idea of using computational modeling to bridge between intracranial recordings in non-human primates and behavioral results in human participants is interesting. Previous work has criticized the BBS synchrony theory based on the observation that synchronization in the gamma-band is highly localized and the frequency of the oscillation depends on the visual features of the stimulus. I appreciate how the authors demonstrate that frequency-dependence and local synchronization can be features of BBS, and not contradictory to the theory. As such, I feel that this work has the potential to contribute meaningfully to the debate on whether BBS is a biophysically realistic model of feature-binding in visual cortex.

Weaknesses:

I have several concerns regarding the presented claims, assessment of meaning and size of the presented effects, particularly with regard to the absence of a priori defined effect sizes.

Firstly, the paper makes strong claims about the frequency-specificity (i.e., gamma synchrony) and anatomical correlates (early visual cortex) of the observed effects. These claims are informed by previous electrophysiological work in non-human primates but are not directly supported by the paper itself. For instance, the title contains the word "gamma synchrony", but the authors do not demonstrate any EEG/MEG or intracranial data in from their human subjects supporting such claims, nor do they demonstrate that the frequencies in the oscillator model are within the gamma band. I think that the paper should more clearly distinguish between statements that are directly supported by the paper (such as: "an oscillator model based on BBS principles accounts for variance in human behavior") and abstract inferences based on the literature (such as "these effects could be attributed to gamma oscillations in early visual cortex, as the model was designed based on those principles").

We thank the reviewer for this helpful comment and agree that the scope of our claims should be clearly delineated between what is directly supported by our data and what is theoretically inferred from prior literature.

We revised the Abstract, Introduction, and early Discussion to moderate the strength of our statements and make the distinction explicit. The revised title now emphasizes that our study tests principles derived from prior work on gamma synchrony rather than directly demonstrating gamma activity in humans. Throughout the text, we use more cautious phrasing that highlights potential mechanisms and theoretical predictions. The intention of our study was not to position synchrony as the only viable mechanism of figure–ground perception. Rather, our goal was to reinvigorate it as a potential contender by showing that features often cited as limitations of synchrony-based binding may in fact be essential properties of the mechanism. We updated phrasing throughout the manuscript to make this clearer and avoid overstating the study’s contribution.

Importantly, our model is not agnostic with respect to frequency band. Oscillator frequencies exhibited by model units are within the gamma range by design. Frequency emerges directly from the contrast within each oscillator’s receptive field, following an empirically established relationship between stimulus contrast and gamma frequency. To our knowledge, such a robust, quantitative relationship between stimulus features to exact oscillation frequency has not been consistently demonstrated for other frequency bands. This relationship yields gamma-band frequencies for all contrasts used in our simulations. The model is thus indeed a gamma oscillator model of V1, not a generic instantiation of Binding by Synchrony (BBS) principles.

That said, we fully agree with the reviewer that our study cannot demonstrate a direct link between gamma synchrony in visual cortex and human behavior. Our behavioral and modeling results instead show that synchronization principles derived from gamma-band physiology in V1 can predict perceptual performance patterns. We now make this distinction explicit throughout the revised manuscript.

Secondly, unlike the human participants, the model strictly does not perform figure-ground segregation, as it only receives the figure as an input.

We thank the reviewer for the opportunity to clarify our modeling approach. We chose not to model the background to reduce computational cost, since including it requires a substantially larger number of oscillators without changing the model’s predictions. The model thus indeed only receives the figure region as input. We aimed to test the local grouping mechanism predicted by TWCO, rather than to simulate a full figure–ground segregation process including a read-out stage. Our model therefore isolates the conditions under which local synchrony emerges within the figure region, assuming that a downstream read-out mechanism (not explicitly modeled here) would detect regions of coherent activity. The exact nature of such a read-out mechanism was beyond the scope of our work.

To confirm that our simplified model is a valid proxy, we ran additional simulations including the background and found that a coherent figure assembly reliably emerges, as can be seen in the phase-locking patterns relative to a reference oscillator at the center of the figure. This validates that the principles of local grouping we studied in isolation hold even when the figure is embedded in a noisy surround. We have added an explicit note in the Results (paragraph 2) that we only simulate the figure and added Supplementary Figure S1 showing the additional simulations.

Finally, it is unclear what effect sizes the authors would have expected a priori, making it difficult to assess whether their oscillator model represents the data well or poorly. I consider this a major concern, as the relationship between the synchrony of the oscillatory model and the performance of the human participants is confounded by the visual features of the figure. Specifically, the authors use the BBS literature to motivate the hypothesis that perception of the texture-defined figure is related to the density and contrast heterogeneity of the texture elements (gabor annuli) of the figure. This hypothesis has to be true regardless of synchrony, as the figure will be easier to spot if it consists of a higher number of high-contrast gabors than the background. As the frequency and phase of the oscillators and coupling strength between oscillators in the grid change as a function of these visual features, I wonder how much of the correlation between model synchrony and human performance is mediated by the features of the figure. To interpret to what extent the similarity between model and human behavior relies on the oscillatory nature of the model, the authors should find a way to estimate an empirical threshold that accounts for these confounding effects. Alternatively, it would be interesting to understand whether a model based on competing theories (e.g., Binding by Enhanced Firing, Roelfsema, 2023) would perform better or worse at explaining the data.

We thank the reviewer for these insightful and constructive comments, which have prompted additional analyses that we believe substantially strengthen our work. The reviewer raises two main points: (1) the need for a benchmark to assess our model’s performance, and (2) the concern that the relationship between model synchrony and behavior might be a non-causal “confound” of the visual features. We address each point below.

(1) Benchmarking model performance

We agree that it is important to assess how well our model performs relative to the data and included this in the original manuscript. We did not predefine an absolute good fit threshold because absolute agreement depends on irreducible noise and inter-subject variability, making a universal cutoff arbitrary. Instead, we had benchmarked model performance in two complementary ways. First, the noise ceiling shown in Figure 5 provides an empirical benchmark for the maximum fit any model could achieve on our data. Simulated Arnold tongues (based on synchrony) approach this ceiling achieving 89% of possible similarity for correlation and 79% of possible similarity for weighted Jaccard similarity, respectively. Second, the parameter sweep (Figure 3) situates our model’s performance within the broader parameter space. It shows that the model, whose key parameters were fixed a priori from independent macaque neurophysiological data, lies close to the optimal regime for explaining the human data. It also provides an estimate of the lower bound (worst-performing point) on the fit that a misspecified model implementing the identical mechanism would achieve. Our model with fixed a priori parameters does 1.41 times better than a misspecified model for the correlation fit metric and 3 times better for weighted Jaccard similarity.

(2) Synchrony as mechanism vs. potential confound

We appreciate the reviewer’s suggestion to test whether synchrony explains behavior beyond stimulus features. In our framework, synchrony is a near-deterministic function of the manipulated stimulus features given fixed model parameters. As a result, synchrony and the stimulus features are collinear (R²≈0.8) leaving no independent variance for synchrony to explain once stimulus features are included. Adding both into one statistical model yields unstable coefficients and no out-of-sample improvement.

Mechanistically, we believe the relevant question is not whether synchrony explains behavior beyond stimulus features but whether synchrony is the correct transformation of the stimulus features to reproduce the behavioral pattern. Please note that in our design we ensured that mean contrast and luminance are identical in the figure and the background such that there are not more high-contrast Gabors in the figure than in the background. We did this with the aim to render mean contrast not a relevant feature. However, there are more high-contrast Gabors in the background, and it is conceivable that the absence of such high contrasts in the figure drives the detection/discrimination of the figure. We therefore agree that testing alternative models would further clarify the unique explanatory value of the synchrony mechanism. To that end, we derived two alternative rate-based readouts from the same V1 simulations of our model from which we derived synchrony. First, average firing rates inside the figure and second, the difference between average firing rates inside the figure and average firing rates in the background (rate difference). We analyzed each individually as predictors of behavior and performed a model comparison based on out-of-sample predictions. While rate difference (but not average firing) showed meaningful associations with performance when considered alone, the synchrony readout had a larger effect size and was favored by the model comparison. We added a new subsection comparing synchrony to rate-based alternatives in the Results (paragraphs 7-9), including additional Bayesian analyses and LOO-CV model comparison. Please note that the model comparison we added to the manuscript provides an additional benchmark beyond the map-level ceiling analysis. It indicates that the mapping from stimulus features to behavior via synchrony generalizes best without requiring an a priori good-fit threshold.

We agree that formally comparing our model to a sophisticated rate-based alternative, such as an instantiation of the Binding by Enhanced Firing model, is an important direction for future work. However, it remains an open and non-trivial question whether such a model could quantitatively reproduce the precise shape of the behavioral Arnold tongue that emerges from the systematic manipulation of our stimulus parameters. Implementing and parameterizing such a model in a comparable, biologically grounded framework is a substantial undertaking that lies beyond the scope of the current study. Therefore, our goal here was not to claim exclusivity for synchrony-based mechanisms, but rather to re-evaluate their plausibility by showing that features often seen as limitations (stimulus dependence and frequency heterogeneity) are, in fact, essential characteristics of the TWCO framework that can predict complex behavioral outcomes.

We would also like to clarify that our stimulus features were derived from theory rather than psychophysical literature. Starting from the principles of TWCO, we mapped frequency detuning and coupling strength onto known anatomical and physiological properties of early visual cortex, and only then derived the corresponding stimulus manipulations (contrast heterogeneity and grid coarseness). Demonstrating that these features predict behavior is therefore not trivial but constitutes a first empirical confirmation that the core TWCO variables match perception.

Apart from adding analyses of additional rate-based readouts of our model, we also refined our discussion of the relationship between these and a synchrony-based mechanism.

Reviewer #2 (Public review):

The authors aimed to investigate whether gamma synchrony serves a functional role in figure-ground perception. They specifically sought to test whether the stimulus-dependence of gamma synchrony, often considered a limitation, actually facilitates perceptual grouping. Using the theory of weakly coupled oscillators (TWCO), they developed a framework wherein synchronization depends on both frequency detuning (related to contrast heterogeneity) and coupling strength (related to proximity between visual elements). Through psychophysical experiments with texture discrimination tasks and computational modeling, they tested whether human performance follows patterns predicted by TWCO and whether perceptual learning enhances synchrony-based grouping.

We thank the reviewer for their thoughtful and constructive review. We believe the comments have served to improve our work.

Strengths:

(1) The theoretical framework connecting TWCO to visual perception is innovative and well-articulated, providing a potential mechanistic explanation for how gamma synchrony might contribute to both feature binding and separation.

(2) The methodology combines psychophysical measurements with computational modeling, with a solid quantitative agreement between model predictions and human performance.

(3) In particular, the demonstration that coupling strengths can be modified through experience is remarkable and suggests gamma synchrony could be an adaptable mechanism that improves with visual learning.

(4) The cross-validation approach, wherein model parameters derived from macaque neurophysiology successfully predict human performance, strengthens the biological plausibility of the framework.

Weaknesses:

(1) The highly controlled stimuli are far removed from natural scenes, raising questions about generalisability. But, of course, control (almost) excludes ecological validity. The study does not address the challenges of natural vision or leverage the rich statistical structure afforded by natural scenes.

We agree with the reviewer that the insights of the present study are limited to texture stimuli and have made adjustments in the Discussion (final two paragraphs) to avoid claiming generalizability to natural stimuli. We have also adjusted the title to specifically limit our results to texture stimuli. To establish the principles of TWCO, we needed tight control over the stimulus, but are intrigued by the idea to investigate natural scenes. We have added to our Discussion (paragraph 9) that future should evaluate to what extent the principles we investigate here apply to natural scenes. Synchrony-based mechanisms have been successfully used for image segmentation tasks in machine vision, showing that the proposed mechanism can in principle work for natural scenes.

(2) The experimental design appears primarily confirmatory rather than attempting to challenge the TWCO framework or test boundary conditions where it might fail.

We thank the reviewer for this important point. Our primary motivation was to address the neurophysiological properties of gamma synchrony that have been suggested to severely challenge the binding by synchrony mechanism. Particularly the strong dependence of gamma oscillations and synchrony on stimulus features. Our goal was to show that from the perspective of TWCO, these challenges become expected components of the mechanism. In essence, we wanted to promote a conceptual shift that converts what pushes a theory to its limit into something that is actually its central tenet. To facilitate this shift, we designed the experiment to directly test this core tenet.

While our approach was designed to test a central prediction of TWCO rather than explicitly challenge its boundaries, we respectfully argue that it was far from a simple confirmatory experiment. The design incorporated high-risk elements that provided considerable room for both the theory and our model to fail. First, the core prediction itself was non-obvious and highly specific. We did not simply test whether contrast heterogeneity and grid coarseness affect perception. We tested the stronger hypothesis that they would reflect a specific, interactive trade-off (the behavioral Arnold tongue) as specified by TWCO. Second, our modeling approach was deliberately constrained to provide a further stringent test. We did not post-hoc optimize the model's key parameters to fit our behavioral data. Instead, we fixed them a priori based on independent neurophysiological data from macaques. This was a high-risk choice, as a mismatch between a priori model predictions and the human data would have seriously challenged the framework's generalizability.

We agree that future research should further challenge TWCO. For instance, by using stimuli that require segregating several objects simultaneously or objects that cover more extensive regions of the visual field.

(3) Alternative explanations for the observed behavioral effects are not thoroughly explored. While the model provides a good fit to the data, this does not conclusively prove that gamma synchrony is the actual mechanism underlying the observed effects.

We agree that our results do not conclusively show that gamma synchrony is the actual mechanism underlying figure-ground segregation. We admit that the original phrasing used throughout the manuscript was too strong and gave the impression that we wanted to establish exactly that. However, the goal of our work was only to reinvigorate gamma synchrony as a potential contender by showing that features often cited as limitations of synchrony-based binding may in fact be essential properties of the mechanism. We have revised the title and made adjustments throughout the manuscript to better reflect this more moderate goal.

Additionally, we added tests of alternatives (Results, paragraphs 7–9) to clarify the unique explanatory value of the synchrony mechanism. To that end, we derived two alternative rate-based readouts from the same V1 simulations of our model. First, we extracted average firing rates inside the figure. Second, we computed the difference between average firing rates inside the figure and average firing rates in the background (rate difference). We analyzed each individually as predictors of behavior and performed a model comparison between these two and synchrony based on out-of-sample predictions. While the rate difference (but not average firing) showed meaningful associations with performance when considered alone, the synchrony readout had a larger effect size and was favored by the model comparison.

(4) Direct neurophysiological evidence linking the observed behavioral effects to gamma synchrony in humans is absent, creating a gap between the model and the neural mechanism.

We agree that the model only provides a how-possibly account linking stimulus features to performance. Showing that the brain actually relies on this mechanism would require showing that cortical synchrony mediates the effect of stimulus features on behavior beyond firing rates. Collecting such data would constitute a major effort that would go beyond the scope of this study. We acknowledge the need for electrophysiological data and the mediation analysis in the updated Discussion.

Achievement of Aims and Support for Conclusions:

The authors largely achieved their primary aim of demonstrating that human figure-ground perception follows patterns predicted by TWCO principles. Their psychophysical results reveal a behavioral "Arnold tongue" that matches the synchronization patterns predicted by their model, and their learning experiment shows that perceptual improvements correlate with predicted increases in synchrony.

The evidence supports their conclusion that gamma synchrony could serve as a viable neural grouping mechanism for figure-ground segregation. However, the conclusion that "stimulus-dependence of gamma synchrony is adaptable to the statistics of visual experiences" is only partially supported, as the study uses highly controlled artificial stimuli rather than naturalistic visual statistics, or shows a sensitivity to the structure of experience.

Likely Impact and Utility:

This work offers a fresh perspective on the functional role of gamma oscillations in visual perception. The integration of TWCO with perceptual learning provides a novel theoretical framework that could influence future research on neural synchrony.

The computational model, with parameters derived from neurophysiological data, offers a useful tool for predicting perceptual performance based on synchronization principles. This approach might be extended to study other perceptual phenomena and could inspire designs for artificial vision systems.

The learning component of the study may have a particular impact, as it suggests a mechanism by which perceptual expertise develops through modified coupling between neural assemblies. This could influence thinking about perceptual learning more broadly, but also raises questions about the underlying mechanism that the paper does not address.

Additional Context:

Historically, the functional significance of gamma oscillations has been debated, with early theories of temporal binding giving way to skepticism based on gamma's stimulus-dependence. This study reframes this debate by suggesting that stimulus-dependence is exactly what makes gamma useful for perceptual grouping.

The successful combination of computational neuroscience and psychophysics is a significant strength of this study.

The field would benefit from future work extending (if possible) these findings to more naturalistic stimuli and directly measuring neural activity during perceptual tasks. Additionally, studies comparing predictions from synchrony-based models against alternative mechanisms would help establish the specificity of the proposed framework.

Recommendations for the authors:

Reviewing Editor Comments:

In a joint discussion to integrate the peer reviews and agree on the eLife recommendations, both reviewers agreed that the work is valuable, but they were on the fence about whether the strength of evidence was incomplete or solid, eventually settling on incomplete. The reviewers make several recommendations for improving these ratings, which I (Reviewing Editor) have organised into 3 points below, with point 1 of particular importance. Underneath the summary, please see the individual recommendations of the reviewers.

(1) Strengthen evidence for the unique role of gamma synchrony in explaining the data, and ensuring claims are directly supported by relevant data:

Reviewers 2 and 3 both note the lack of direct evidence for gamma involvement, and reviewer 2 observes that the fit with behaviour may trivially be explained by a relationship between contrast heterogeneity and grid coarseness without need for oscillation. The reviewers felt that the approach of fitting the model to human data could be strengthened to help address this issue - and they offer various solutions, e.g., more principled a-priori criteria around good vs bad fit of the model to both main task and training data, and comparison to alternative binding models (Reviewer 2), identifying and testing boundary conditions of the model (Reviewer 3). There is also the possibility of collecting direct human neurophysiological evidence linking the behavioural data to neural mechanisms. Our discussion also highlighted the need to weaken claims (including in the title) where links are not directly demonstrated by methods from the present study, e.g., resting on indirect comparisons to primate literature.

We agree with the editor and reviewers that this was a critical point. To address it, we have made several major revisions.

As suggested, we have weakened claims where the links are not directly demonstrated by our data. The title has been revised to be more specific, and we have carefully edited the abstract, introduction, and discussion to distinguish between our model's predictions and direct neurophysiological evidence.

To address the concern that our model's fit might be trivially explained by visual features, we have performed a new analysis comparing the synchrony-based readout to two alternative rate-based readouts from the same V1 simulations. This new comparison shows that the synchrony readout provides a superior out-of-sample prediction of human behavior.

While a full implementation of a competing theory like "Binding by Enhanced Firing" would be a valuable next step, we note that parameterizing such a model in a comparably grounded framework is a substantial undertaking beyond the scope of the present study. Our new analysis provides an important first step in this direction.

(2) Make explicit and address the limitations of the stimuli:

Include that the model is not extracting the figure from the background, and the controlled stimuli may limit generalizability.

To address the concern that our model was not performing true figure-ground extraction, we performed a new set of simulations that included both the figure and the immediate background. The results confirm that synchrony dynamics within the figure region are not affected by the presence of the background. We added these validation results as supplementary materials. We have additionally made the modeling choice and its justification more explicit in the Results and Methods sections.

We have revised the Discussion to be more explicit about the limitations of using highly controlled texture stimuli. We now clearly state that our findings are specific to this context and that further research is required to determine if these principles generalize to the segregation of objects in natural scenes.

(3) Some clarifications to make more accessible:

Include the figure explaining the framework (Reviewers 1&2), and also the model details (Reviewer 2).

We have revised Figure 1 and its caption to more clearly illustrate the links from TWCO principles to their neural implementation in V1 and the resulting behavioral predictions.

We have expanded the Methods section to provide a more detailed and accessible description of the model's construction. We now clarify precisely how the oscillator grid was defined in visual space, how eccentricity-dependent receptive field sizes were implemented, and how these were mapped onto a retinotopic cortical surface to determine coupling strengths.

Reviewer #1 (Recommendations for the authors):

(A) Major concerns:

(1) My main concern:

My main concern is the repeated claims that the observed findings can be attributed to gamma synchrony in the early visual cortex. I find this claim misleading as the authors do not report any electrophysiological data that directly supports such claims. As stated in my public review, I feel that the authors should be clear about direct evidence versus more abstract inferences based on the literature.

In particular, I recommend changing claims about "gamma synchrony" to "Binding by Synchrony" That being said, the authors can outline that the model was built under the assumption that this synchrony is mediated by gamma in early visual cortex, but I don't think it should be part of their main conclusions.

We appreciate that TWCO’s general principles are frequency-agnostic and can be viewed as binding by synchrony in a broad sense. Our work, however, specifically instantiates these principles in V1 gamma: the model reflects TWCO dynamics together with V1 anatomy/physiology and the well-established contrast–frequency relationship in the gamma range (which, to our knowledge, has not been demonstrated with comparable specificity for other bands). In that sense, it is a gamma oscillator model of V1, rather than a generic BBS instantiation. Moreover, stimulus dependencies often cited as challenges to BBS have been used in particular to argue against gamma; showing that these very dependencies are integral to the TWCO mechanism is central to our contribution, and we therefore keep our conclusions focused on the gamma-specific instantiation tested here.

(2) Mediation of the observed effects by the visual features of the figure:

The authors motivate the hypothesis that BBS predicts that the perception of texture-defined objects depends on the density of texture elements and their contrast heterogeneity. This hypothesis seems trivial as those are the features that distinguish figure from ground. I think it would be important to clarify how this hypothesis is unique to BBS and not explained by competing theories, such as Binding by Enhanced Firing (Roelfsema, 2023). The authors should be clear about what part of the hypothesis is not trivial based on the task and clearly attributable to oscillators and synchrony.

Our stimulus features were derived from theory rather than psychophysical literature. Starting from the principles of TWCO, we mapped frequency detuning and coupling strength onto known anatomical and physiological properties of early visual cortex, and only then derived the corresponding stimulus manipulations (contrast heterogeneity and grid coarseness). We agree that grid coarseness (element distance) is an established facilitator of figure–ground perception. By contrast, contrast heterogeneity (feature variance) is less commonly emphasized as a figure–ground cue, compared to mean-based cues, but follows directly from TWCO’s frequency detuning. Importantly, mean contrast and luminance were matched exactly between figure and background in our stimuli. Demonstrating that contrast heterogeneity and grid coarseness not only independently affect figure-ground perception, but reflect a trade-off where higher heterogeneity needs to counteracted by reduced grid coarseness in the way TWCO specifies is therefore non-obvious and provides an initial empirical indication that the core TWCO variables might shape perception. We also agree that alternative models would further clarify the unique explanatory value of synchrony. In the revised manuscript, we compare rate-based readouts (mean figure rate; figure–background rate difference) with the synchrony readout from the same simulations. Rate difference indeed constitutes a predictor of performance, but the synchrony readout showed a larger effect and was preferred by out-of-sample model comparison.

Using a linear model, the authors assess the relationship between discrimination accuracy and synchrony. Did the authors also include the factors grid coarseness and contrast heterogeneity in this model? Again, as both the task performance (as shown by the GEE analysis) and oscillatory synchrony depend on these features, the relationship between model and behavioral performance will be mediated by the visual features.

Thank you for raising this. In our framework, detuning (via contrast heterogeneity) and coupling (via grid coarseness) are the inputs, synchrony is the proposed mechanistic mediator, and behavior is the output. Because synchrony in our model is a (near-)deterministic function of the manipulated features under fixed parameters, a joint features+synchrony regression is statistically ill-posed (perfect multicollinearity up to numerical error) and cannot add information. A proper mediation test would require trial-wise neural measurements of synchrony in the same task, which we do not have and acknowledge as a limitation in the Discussion. Accordingly, we show that both the features themselves (reflecting TWCO principles) and model-derived synchrony (realizing the proposed pathway) account for behavior.

We agree this does not establish a unique contribution of synchrony. To probe alternatives, we added rate-based readouts and a model comparison to the revised manuscript. These additional analyses indicate that synchrony outperforms simple rate-based mappings. We do not claim this rules out more sophisticated rate-based mechanisms. Our aim is to demonstrate that synchrony is a viable, behaviorally informative readout for downstream processing. We do not assert it is the only mechanism the brain uses. Synchrony had been discounted due to its stimulus dependence; our results are intended to rule it back in. We have made changes throughout the manuscript to better reflect this more modest aim.

(3) Goodness of fit measures are not established a prior:

I have described this concern in my public review. It is hard to assess what the authors would have interpreted as a good or a bad fit, especially without accounting for the confound in the relationship between oscillator synchrony and behavior. Similarly, when assessing the similarity between the behavioral and dynamic Arnold Tongues across different coupling parameters, the authors found that the chosen parameters (based on macaque data) were not optimal. They offer the explanation that the human cortex has a lower coupling decay than the macaque cortex, and the similarity is higher for lower values of coupling decay. While this explanation is not entirely implausible, it is unclear where an oscillator model with human values would be in the presented plot, as the authors didn't estimate those values from the human studies. Moreover, the task used in the Lowet et al., 2017 paper is very different from the task presented here, which could also account for differences. Overall, the explanation appears hand-wavy considering the lack of empirically defined goodness of fit measures.

Thank you for these concerns.

We did indeed not provide a priori thresholds for what would be considered good fit. Instead, we used two complementary benchmarks; namely noise ceilings and parameter exploration. The former provides an upper bound on what any model (not just ours but based on completely different mechanisms) could achieve given our data. The parameter sweep provides an indication how well our concrete model can maximally fit the data and how bad it can be based on possible parameters. These benchmarks are more informative than a fixed a-priori cutoff, which would depend on unknown noise and inter-subject variability. Both the noise ceiling and the parameter exploration indicate that our model, using a priori fixed parameters, performs well. Additionally, we redid all our statistical analyses after z-normalizing every predictor to provide easier interpretation of effect sizes.

Regarding the reason that key model parameters were not optimal, we believe our interpretation to be plausible. We agree that we currently do not have data to estimate the exact human decay factor and hence cannot establish how much model fit would be affected. However, the parameter exploration in Figure 3 shows that small to modest reductions in decay would improve model fit. We discuss this now in the revised manuscript.

The reviewer’s suggestion is intriguing. While Lowet et al. (2017) used a different task, the parameters we took from their work (decay rate and maximum coupling) are intended to reflect anatomical properties and thus should not be task-dependent. That said, Lowet et al. ‘s data carry uncertainty, so our estimates may not be exact; we note this explicitly in the revised Discussion. Whether a different task would have yielded better parameter estimates is difficult to determine, but we considered Lowet’s paradigm appropriate because it was designed to target the same V1 anatomical and physiological properties that map onto TWCO.

I have concerns about a similar confound in the training effects. If I'm not mistaken, the Hebbian Learning rule encourages synchronization between the oscillators in the grid. As such, it causes synchronization to increase over several simulations. Clearly, the task performance of the participants also improves over the sessions. Again, an empirical threshold would be required to assess whether the similarity in learning between model and performance goes beyond what is expected based on learning alone. How much of these effects can be attributed to the model being oscillatory?

The reviewer is correct that, in our framework, learning operates via changes in coupling that increase synchrony. Enhanced synchrony is the proposed (and in our model also the actual) pathway by which learning impacts behavior. We agree that learning could, in principle, act through pathways other than synchrony. Demonstrating this would not be achieved by a mediation analysis here, because that requires independent, trial-level neural measurements of the candidate pathways (synchrony and alternatives). In the absence of such data, the appropriate approach would be model comparison between competing mechanistic readouts. We have added such a model comparison for a synchrony readout versus two rate-based readouts derived from the same simulations for the first session; i.e., focusing on the pathway from stimulus features to behavior. However, a similar model comparison is not possible for learning. As we show in the supplementary materials, rate-based readouts of our V1 model are not at all affected by coupling strength. As such, they are insensitive to changes in coupling and are thus not viable as alternative mechanisms to explain performance changes due to learning. A fair test of rate-based alternatives would require building a detailed rate-based figure–ground segregation model that predicts session-wise changes. We agree that this is an important next step but it is also substantial undertaking beyond the scope of the present study.

(4) Similarly, for the comparison of the Arnold Tongue in the transfer session and the early session:

In the first part of the Results section, it says: "Our model rests on the assumption that learning-induced structural changes in early visual cortex are specific to the retinotopic locations of the trained stimuli. We evaluated whether this assumption holds for our human participants using the transfer session following the main training period. [...] If learning is indeed local, participants' performance in the transfer session should resemble that of early training sessions, indicating a reset in performance for the new retinal location."

The authors find that a model fit to session 3 explains the data in the transfer session best and consider this as evidence for the above-stated expectation. Again, it is unclear where the cutoff would have been for a session to be declared as early or late. For instance, had the participants only performed 4 sessions, would the performance be best explained by session 3 or session 1?

A high number of statistical tests are used, which, firstly, need to be corrected for multiple comparisons (did the authors do this?). Secondly, I feel that the regression models could be improved. For instance, the authors fit one model per session and then assess how well each model explains the variance in the transfer session. I think the authors might want to opt for one model with the regressors contrast heterogeneity, grid coarseness, and session (and their interaction). Using this approach, the authors would still be able to assess which session predicts the data best. Similarly, interindividual variability could be accounted for by adding participant-specific random effects to the model (and using a mixed model), instead of fitting individual models per participant.

We agree the “early vs late” cutoff was underspecified. In the revision, we predefine Session 2 as the early-learning reference, excluding Session 1 to avoid familiarization/response–mapping effects. We then fit a single Bayesian hierarchical model with contrast heterogeneity, grid coarseness, and session, plus a transfer indicator, and participant-level random effects. This allows us to place the transfer session on the same scale as training and to test a) whether the transfer session precedes the state in session 2 via the posterior contrast P(βtransfer<βSess2) and b) whether it is indistinguishable from the state in session two using an equivalence test derived from the fitted model. We find that the transfer session is equivalent to session 2. We added this updated analysis of the transfer session in the Results (paragraph 15).

In response to the suggestion to use a hierarchical regression model for analyzing the transfer session, we have decided to use such a model for all our analyses in a Bayesian framework. In this Bayesian framework, inference is based on the joint posterior (credible intervals/equivalence) of all predictors in a model and additional post-hoc multiplicity corrections are not required.

(5) Questions regarding the model:

What does it mean that the grid was "defined in visual space"? How biologically plausible with regard to the retinotopy and organization of the oscillators do the authors claim the model to be?

We are happy to clarify this point. We have a total of 400 oscillators reflecting neural assemblies in V1. We start by defining a regular, 20x20, grid of the receptive field (RF) centers of these oscillators inside the figure region. Each oscillator is then also assigned a RF size based on the eccentricity of its RF center. We use the threshold-linear relationship between RF eccentricity and RF size reported in [1] to assign RF sizes. Each oscillator thus has an individual, eccentricity-dependent, RF size.

For the coupling between oscillators, we need to know their cortical distances. We obtain these by first determining the cortical location of each oscillator through a complex-logarithmic topographic mapping of neuronal receptive field coordinates onto the cortical surface [2,3]. For this mapping, we use human parameter values estimated by [4]. From these cortical locations, we then compute pairwise Euclidean distances.

The model thus captures realistic retinotopy, eccentricity-dependent RF sizes, and distance-dependent coupling on the cortical surface. We have adjusted our Methods to make these steps clearer.

(1) Freeman, J., & Simoncelli, E. P. (2011). Metamers of the ventral stream. Nature neuroscience, 14(9), 1195-1201.

(2) Balasubramanian, M., & Schwartz, E. L. (2002). The isomap algorithm and topological stability. Science, 295(5552), 7. https://doi.org/10.1126/science.1066234

(3) Schwartz, E. L. (1980). Computational anatomy and functional architecture of striate cortex: a spatial mapping approach to perceptual coding. Vision Research, 20(8), 645–669. http://www.sciencedirect.com/science/article/pii/0042698980900905

(4) Polimeni, J. R., Hinds, O. P., Balasubramanian, M., van der Kouwe, A. J. W., Wald, L. L., Dale, A. M., & Schwartz, E. L. (2005). Two-dimensional mathematical structure of the human visuotopic map complex in V1, V2, and V3 measured via fMRI at 3 and 7 Tesla. Journal of Vision, 5(8), 898. https://doi.org/10.1167/5.8.898

Similarly, do the authors claim that each gabor annuli stimulates a single receptive field in V1?

We hope that with the additional explanation above, it is clearer that there is not a one-to-one mapping. Each oscillator samples the local image by pooling over all Gabor annuli that overlap its receptive field (partially or fully) and computes the average contrast within its RF. Conversely, a single annulus typically overlaps multiple RFs and contributes to each in proportion to the overlap.

I am unsure how the oscillators were organized, if not retinotopically. How is the retinotopic input fed into the non-retinotopically arranged oscillators?

We hope that with the additional explanation above, it is clearer that the network is strictly retinotopic.

The frequency of each oscillator changes according to ω=2πv with ν=25+0.25C. How were the values for the linear regression in v chosen? Reference?

The slope and intercept parameters for this equation were first reported in [5]. We added the reference to the Methods.

(5) Lowet, E., Roberts, M., Hadjipapas, A., Peter, A., van der Eerden, J., & De Weerd, P. (2015). Input-dependent frequency modulation of cortical gamma oscillations shapes spatial synchronization and enables phase coding. PLoS computational biology, 11(2), e1004072.

(6) Hebbian Learning Rule:

I am confused about how the effective learning rate E= ∈t is calculated. It is said that it is estimated based on the similarity between the second experimental session and the distribution of synchrony after letting the model learn. How can the model learn without knowing epsilon and t?

We agree with the reviewer that our procedure to estimate the effective learning rate requires further clarification. We performed a nested grid search. Essentially, we let the model learn between session 1 and 2 with each of 25 candidate effective learning rates and evaluate how well each of them allow the model to fit performance in session 2. We then select the best effective learning rate and create a new, smaller, grid around this value and repeat that procedure. In total we perform 5 nested grids to arrive at the final effective learning rate. We expanded the explanation in the Methods.

(B) Minor concerns:

(1) Small N: 2/3 of the studies that were cited to justify the small sample were notably different from the current experiment, i.e., Intoy 2020 is an eye movement task, Lange 2020 is a memory task (Tesileanu 2020 is more similar). I think a power analysis would be great to support, as the sample size seems quite low

Our study uses a within-subject design with ~750 trials per session (≈6,000 total) per participant, analyzed with a hierarchical model that pools information across trials and participants. To assess adequacy, we ran a simulation-based design analysis using the fitted hierarchical model (i.e., post hoc, based on the observed variance components). This analysis indicated a detection probability >90% for all key effects. We now report the results of this design analysis in the (Supplementary Table 1) and note this in the Results (paragraph 1).

Regarding the literature context, we agree the cited studies are not identical to ours; we referenced them to illustrate a common practice (small N with many trials) when targeting low-level, early-visual mechanisms. Intoy (pattern/contrast sensitivity) and Lange (perceptual learning in early vision) share that focus, while Tesileanu is methodologically closest.

(2) Figure 1 could be more informative and better described in the text. The authors often don't refer to the panels in Figure 1. Maybe it would help to swap a and b to describe the Arnold tongue first? It might also be a good idea to add the coupling strength and frequency detuning axes

We have swapped panels a and b and now refer to each panel in the main text to enhance clarity.

(3) Values of rho (distance - is this degrees visual angle)? Do the authors assume that the size of the stimuli corresponds to receptive fields in V1? If so, how is this justified?

The center-to-center distance between any pair of neighboring annuli is indeed expressed in degrees of visual angle. Rho is a scaling factor for this distance. With rho=1, the center-to-center distance corresponds to the diameter of the annuli; i.e., they touch but do not overlap each other. We do not assume any relation between the size of receptive fields and the size of the annuli. Receptive field sizes in our model are purely determined by their eccentricity and each oscillator can have several annuli within its receptive field while each annulus can fall within several overlapping receptive fields of different oscillators. We believe that the schematic illustration in Figure 1 might have given the impression that each oscillator sees exactly one annulus and added a note that this is not the case and merely an oversimplification to illustrate the relationship between contrast and intrinsic frequency.

(4) Some equations are embedded in the text, and some are not. It might be easier to find the respective equation if they all have an index. For instance, the authors mention the psychometric function that relates model synchrony and performance in the results section. It would be easier to find if it had an index that the authors could refer to.

We moved this equation as well as the contrast intrinsic frequency mapping from inline to displayed and numbered them.

(5) Is there a reference for "Our model rests on the assumption that learning-induced structural changes in early visual cortex are specific to the retinotopic locations of the trained stimuli"? (If so, it should be cited.)

We added references supporting this assumption.

(6) Figure 2b: colorbar missing label.

We added the label.

Reviewer #2 (Recommendations for the authors):

Cool work!

(1) The reader would benefit from (a single) comprehensive figure that visually explains the entire conceptual framework-from TWCO principles to neural implementation to behavioural predictions-accessible to readers without specialised knowledge of oscillatory dynamics. This will give the paper a greater impact.

We have adjusted Figure 1 in accordance with suggestions made by reviewer 1 and added further explanations to the caption and the Introduction to enhance clarity on how the principles of TWCO relate to neural implementation.

(2) I think this paper would benefit from the audience eLife provides, but the paper could move closer to the audience.

(3) Pride comes before the fall, but I am not the most uninformed reader, and it took me some effort to process everything.

Thank you, we took this to heart. In the Introduction, we now state more explicitly how each variable is operationalized and how these map onto TWCO with improved reference to relevant panels in the schematic figure. We agree the framework is conceptually dense. TWCO principles reach the stimuli through specific V1 anatomy and physiology, so there are several links to keep in mind. Our goal with the revised introduction and figure is to make those links better visible.

(4) You could consider discussing potential implications for understanding perceptual disorders characterized by altered neural synchrony (e.g., schizophrenia, autism) and how your learning paradigm might inform perceptual training interventions.

Thank you for this suggestion. We have added that TWCO might provide a new lens to study perceptual disorders to the Discussion. We provide a concrete example of the relation between grouping, gamma synchrony (in light of TWCO) and lateral connectivity in schizophrenia

(5) I think this paper has real strength, but rather than dispersing limitations throughout the discussion, create a dedicated section that systematically addresses ecological validity, alternative explanations, and generalisability concerns. This will also preempt criticism.

We appreciate the suggestion. Our preference is to discuss limitations in context, next to the specific results they qualify, so readers see why each limitation matters and how it affects interpretation. Nevertheless, paragraph 7 on page 20 summarizes most limitations in a single paragraph.

eLife Assessment

This study establishes a methodology (machine vision and gaze pose estimation) and behavioral apparatus for examining social interactions between pairs of marmoset monkeys. It has been difficult to study social interactions using artificial stimuli rather than genuine interactions between unrestrained animals. This study makes a fundamental contribution to social neuroscience research in a laboratory setting. Their results are convincing showing that the study of unrestrained social interactions is possible with detailed quantification of position and gaze. The methodology presented here will be broadly useful for research in social neuroscience, neuroethology, and primatology.

Reviewer #1 (Public review):

Summary:

The current study by Xing et al. establishes the methodology (machine vision and gaze pose estimation) and behavioral apparatus for examining social interactions between pairs of marmoset monkeys. Their results enable unrestrained social interactions under more rigorous conditions with detailed quantification of position and gaze. It has been difficult to study social interactions using artificial stimuli, as opposed to genuine interactions between unrestrained animals. This study makes an important contribution for studying social neuroscience within a laboratory setting that will be valuable to the field.

Strengths:

Marmosets are an ideal species for studying primate social interactions due to their prosocial behavior and the ease of group housing within laboratory environments. They also predominantly orient their gaze through head-movements during social monitoring. Recent advances in machine vision pose estimation set the stage for estimating 3D gaze position in marmosets but requires additional innovation beyond DeepLabCut or equivalent methods. A six point facial frame is designed to accurately fit marmoset head gaze. A key assumption in the study is that head-gaze is a reliable indicator of the marmoset's gaze direction, which will also depend on the eye position. Overall, this assumption has been well supported by recent studies in head-free marmosets. Thus the current work introduces an important methodology for leveraging machine vision to track head-gaze and demonstrates its utility for use with interacting marmoset dyads as a first step in that study.

Comments on revisions:

I thank the authors for their careful revisions of the manuscript. It has addressed all of my comments.

One final suggestion would be to add a scale bar in Supplemental Figure 2A so the size of the video/image stimuli is clear (in cm of monitor size) and also to report a range for how far away was the marmoset in viewing these stimuli (in cm). This will enable calculation of the rough accuracy in visual degrees.

Reviewer #2 (Public review):

Summary:

The manuscript describes novel technique development and experiments to track the social gaze of marmosets. The authors used video tracking of multiple cameras in pairs of marmoset to infer head orientation and gaze, and then studied gaze direction as a function of distance between animals, relationships, and social conditions/stimuli.

Strengths:

Overall the work is interesting and well done. It addresses an area of growing interest in animal social behavior, an area that has largely been dominated by research in rodents and other non-primate species. In particular, this work addresses something that is uniquely primate (perhaps not unique, but not studied much in other laboratory model organisms), which is that primates, like humans, look at each other, and this gaze is an important social cue of their interactions. As such, the presented work is an important advance and addition to the literature that will allow more sophisticated quantification of animal behaviors. I am particularly enthusiastic about how the authors approach the cone of uncertainty in gaze, which can be both due to some error in head orientation measurements as well as variable eye position

Weaknesses:

While there remains some degree of uncertainty in the precise accuracy of the gaze measure, the authors have done an excellent job accounting for these as well as they can, and appropriately acknowledge the limitations of their approach.

Comments on revisions:

I have no further recommendations. The authors addressed my previous suggestions or acknowledged them as topics for future investigation. This is excellent work.

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

The current study by Xing et al. establishes the methodology (machine vision and gaze pose estimation) and behavioral apparatus for examining social interactions between pairs of marmoset monkeys. Their results enable unrestrained social interactions under more rigorous conditions with detailed quantification of position and gaze. It has been difficult to study social interactions using artificial stimuli, as opposed to genuine interactions between unrestrained animals. This study makes an important contribution for studying social neuroscience within a laboratory setting that will be valuable to the field.

Strengths:

Marmosets are an ideal species for studying primate social interactions due to their prosocial behavior and the ease of group housing within laboratory environments. They also predominantly orient their gaze through head movements during social monitoring. Recent advances in machine vision pose estimation set the stage for estimating 3D gaze position in marmosets but require additional innovation beyond DeepLabCut or equivalent methods. A six-point facial frame is designed to accurately fit marmoset head gaze. A key assumption in the study is that head gaze is a reliable indicator of the marmoset's gaze direction, which will also depend on the eye position. Overall, this assumption has been well supported by recent studies in head-free marmosets. Thus the current work introduces an important methodology for leveraging machine vision to track head gaze and demonstrates its utility for use with interacting marmoset dyads as a first step in that study.

Weaknesses:

One weakness that should be easily addressed is that no data is provided to directly assess how accurate the estimated head gaze is based on calibrations of the animals, for example, when they are looking at discrete locations like faces or video on a monitor. This would be useful to get an upper bound on how accurate the 3D gaze vector is estimated to be, for planned use in other studies. Although the accuracy appears sufficient for the current results, it would be difficult to know if it could be applied in other contexts where more precision might be necessary.

Please see our detailed responses to the reviewer comments below.

Reviewer #2 (Public review):

Summary:

The manuscript describes novel technique development and experiments to track the social gaze of marmosets. The authors used video tracking of multiple cameras in pairs of marmosets to infer head orientation and gaze and then studied gaze direction as a function of distance between animals, relationships, and social conditions/stimuli.

Strengths:

Overall the work is interesting and well done. It addresses an area of growing interest in animal social behavior, an area that has largely been dominated by research in rodents and other non-primate species. In particular, this work addresses something that is uniquely primate (perhaps not unique, but not studied much in other laboratory model organisms), which is that primates, like humans, look at each other, and this gaze is an important social cue of their interactions. As such, the presented work is an important advance and addition to the literature that will allow more sophisticated quantification of animal behaviors. I am particularly enthusiastic with how the authors approach the cone of uncertainty in gaze, which can be both due to some error in head orientation measurements as well as variable eye position.

Weaknesses:

There are a few technical points in need of clarification, both in terms of the robustness of the gaze estimate, and possible confounds by gaze to non-face targets which may have relevance but are not discussed. These are relatively minor, and more suggestions than anything else.

Please see our detailed responses to the reviewer comments below.

Reviewer #1 (Recommendations for the authors):

Major comments:

(1) It appears that the accuracy of the estimated gaze angle must be well under the size of the gaze cone (+/- 10 degrees), but I can't find any direct estimate of the accuracy even if it is just a ballpark figure. On Lines 219-233 is where performance is described for viewing images and video on a monitor, where it should be possible to reconstruct the point of gaze on the monitor while images and video are shown, in order to evaluate the accuracy of the system for where the marmoset is looking? Would you see eye position traces that would show fixation clusters around those images or videos with stationary points on the monitor much like that seen for head-fixed animals looking at faces on a screen (Mitchell et al, 2014)? If so, what is the typical spread of those clusters during fixations on an image, both in terms of the precision by RMS error during a fixation epoch and the spread around the images at different locations (accuracy of projection)? For example, if gaze clusters were always above the displayed images one would have an idea that the face plane is slightly offset above the true gaze direction. It is not completely clear how well the face plane and corresponding gaze cone do in describing gaze direction in space, but the monitor stimuli could be used as an initial validation of it.

We thank the reviewer for this important suggestion regarding the quantitative validation of gaze accuracy. We agree that, when animals view stimuli presented on a monitor, the estimated gaze direction can be evaluated by examining the spatial distribution of gaze–monitor intersection points relative to stimulus locations.

To address this, we generated a new figure (Fig. S2A) analyzing gaze behavior following the onset of video stimuli presented at different locations on the monitor. Specifically, we selected video clips in which human annotators verified that the marmosets were looking at the monitor. Consistent with prior work in head-fixed marmosets (Mitchell et al., 2014), we observe clustering of gaze–monitor intersection centers within and around the corresponding stimulus locations after stimulus onset. These clusters provide an empirical validation that the estimated gaze direction aligns with stimulus position in space.

Importantly, unlike the head-fixed preparation used in Mitchell et al. (2014), marmosets in our study were freely moving. As a result, they do not exhibit prolonged, stationary fixations on the monitor, and fixation clusters are therefore more diffuse. This increased spread reflects natural head and body motion rather than limitations of the gaze estimation method itself. Despite this, gaze intersection points remain spatially localized to the vicinity of the presented stimuli across different monitor locations.

We did observe small offsets in some gaze clusters relative to stimulus centers; however, these offsets were not systematic across stimulus locations or animals. Crucially, there was no consistent bias (e.g., clusters appearing uniformly above or below stimuli) that would indicate a systematic misalignment of the face plane or gaze cone relative to true gaze direction. Together, these observations support the conclusion that the face-plane-based gaze cone provides an accurate estimate of gaze direction in space, with precision well within the ±10° aperture of the gaze cone.

While the freely moving component of the behavior precludes direct estimation of fixation RMS error comparable to head-fixed paradigms, the observed stimulus-locked clustering serves as an initial validation of both the accuracy and practical utility of our approach under naturalistic conditions.

(2) A second major comment is about clarity in the writing of the results and discussion. At the end of the manuscript, a major takeaway is the difference between familiar and unfamiliar dyads, that males show more interest in viewing females including unfamiliar females, but for familiar females, this distinction is also associated with being likely to look at them if they look at the male, and then to engage in joint gaze with them after looking at them, which indicates more of a social interaction than simply monitoring them when they are unfamiliar. Those aspects of the results could be emphasized more in the topic sentences of paragraphs presenting data to support those features of the gaze data (at present is buried at the ends of results paragraphs and back in the discussion).

We thank the reviewer for this insightful suggestion. We have restructured the Results and Discussion sections to lead with the primary social takeaways rather than technical descriptions (Tracked changes in Word). Specifically, we now emphasize the distinction between "social monitoring" (characteristic of unfamiliar dyads) and "active social coordination" (characteristic of familiar dyads).

(1) Topic Sentences: We revised the topic sentences of all Results paragraphs to immediately highlight the findings regarding male interest and the influence of familiarity on reciprocation.

(2) Conceptual Framework: We added a conceptual distinction in the Discussion, explaining that while unfamiliar marmosets maintain high social attention through "peripheral monitoring" and proximity-dependent joint gaze, familiar pairs exhibit sophisticated, distance-independent coordination and gaze reciprocation.

(3) Clarification of Male Interest: We explicitly stated that while male interest in females is high regardless of familiarity, it manifests as persistent monitoring in unfamiliar pairs versus a more aware, reciprocal state in familiar pairs.

Minor comments:

(1) Methods:

a) Lines 522-539: The 200 continuous frames used for validation of the model containing two marmosets are sufficient to test how well it generalizes to other animals outside the training set? The RMSE reported, does it vary for animals inside vs outside the training set? To what extent does the RMSE, in image pixels, translate into accuracy in estimating the gaze direction, for example, as assessed by estimating error when marmosets look at images or video on the monitor?

To address the reviewer’s concern regarding generalization and the translation of pixel RMSE to angular accuracy, we emphasize that the six facial features selected are prominent, high-contrast features across the species. Consequently, we observed that the RMSE remained consistent for marmosets both inside and outside the training set. To quantify how pixel-level tracking error translates into gaze estimation accuracy, we performed a sensitivity analysis. We simulated landmark (i.e., feature) jitter by sampling perturbations from circular distributions based on our empirical data (2.4 pixels for eyes; 2.1 pixels for the central blaze). Our results, illustrated in uthpr response image 1, show that 90% of the resulting head gaze deviations fall within 10°, which is consistent with the angular threshold used for our gaze cone model. This confirms that the reported RMSE provides sufficient precision for reliable gaze estimation.

Author response image 1.

Probability distribution of gaze angular deviation under circular perturbation. The histogram (blue) represents the change in reconstructed gaze angle (degrees) following stochastic perturbation of facial features. To simulate real-world variance, noise was sampled from circular distributions with radii of 2.4 pixels (eyes) and 2.1 pixels (central blaze). The red curve represents an exponential fit to the empirical data (y=ae^bx, a=0.9591, b=0.1813. Approximately 90% of the reconstructed gaze deviations remain below 10°, indicating the model’s localised stability under pixel level coordinate jitter.

b) Line 542-43: Is there any difference between a rigid model fit to the six facial points, versus using the plane defined by the two eyes and central blaze in terms of direction accuracy (in the ground truth validation)? How does the "semi-rigid" set of six points (mentioned also in lines 201-203) constrain the fit of the three points (two eyes and central blaze) that define the normal plan for the gaze cone?

We thank the reviewer for the opportunity to clarify our geometric model. The plane used to define the gaze cone's origin was indeed determined by the two eyes and the central blaze. However, a plane defined by only three points was insufficient to determine a unique gaze direction, as the normal vector was ambiguous (it could point forward through the face or backward through the head).

To resolve this, we utilized the relative positions of the two ear tufts. Because the tufts are anatomically situated behind the eyes and blaze, these additional points provide the necessary spatial context to orient the gaze vector correctly. In our validation, we found that the mouth does not alter the angular accuracy compared to a 3-point fit, supporting that the facial features are correctly identified.

We use the term 'semi-rigid' to describe the six-point constellation because their relative spatial configurations remain stable across individuals and expressions, imposing a biological constraint on the model. This prevents unphysical warping of the face frame during 3D reconstruction and ensures the gaze cone remains anchored to the animal's true midline.

(2) Results:

a) Lines 203-205: What is the distinction between gaze orientation (defined by facial plane, 3D vector) and gaze direction (defined by ear tufts) ... is gaze direction in the 2D x-y plane? Why are two measures needed or different? It does not appear gaze orientation is used further in the manuscript and perhaps could be omitted.

We appreciate the reviewer’s comment regarding the terminology. We have replaced all instances of ‘gaze orientation’ with ‘gaze direction’ to ensure consistency throughout the manuscript.

To clarify, both terms referred to the same 3D unit vector. The ear tufts were not used to define a separate 2D measure; rather, they served as posterior anatomical anchors to resolve the 3D polarity of the normal vector (ensuring the vector points 'forward' from the face rather than 'backward'). Gaze direction was calculated in 3D space and was not restricted to a 2D x-y plane. We have clarified this in the revised Methods section (Lines 203–205) to avoid further ambiguity.

b) Line 215-216: why is head-gaze velocity put in normalized units instead of degrees visual angle per second? How was the normalization performed (lines 549-557)? It would be simpler to see velocity as an angular speed (degrees angle per second) rather than a change in norms.

We thank the reviewer for this suggestion. We agree that the expression is misleading.

(1) We have replaced "face norm" with "face normal vector" (N) throughout the manuscript to clarify that we are referring to the 3D unit vector perpendicular to the facial plane.

(2) Lines 224-225 and the corresponding Methods section (Lines 599-609) have been updated to reflect this change in units and terminology.

We chose to use the change in the face normal vector in normalized units for our primary calculations because it allows for efficient spatiotemporal smoothing and is computationally robust at the very low thresholds required for our stability analysis. However, to address the reviewer's concern regarding interpretability, we have verified that our threshold of 0.05 normalized units corresponds to an angular velocity of 2.87 degrees/frame duration [33ms]. Since we are operating at very small angular changes, the Euclidean distance between unit vectors is a near-linear proxy for the angular displacement in radians.

c) Lines 215-216: How do raw gaze traces appear over time ... are there gaze saccades and then stable fixations, or does it vary continuously? A plot of the gaze trace might be useful besides just showing velocity with a threshold, to evaluate to what extent stable fixation vs shifts are distinct.

Author response image 2.

Time course of gaze, angular velocity and stability, thresholding. The plot illustrates the temporal dynamics of the face normal vector velocity used to define stable gaze states. The blue trace represents the raw gaze velocity calculated in normalised units. The red dashed line demotes the empirical cut off threshold of 0.05 units per frame.

To clarify the temporal dynamics of marmoset head movements, we have provided a representative time course of head gaze velocity as shown in Author response image 2. The data clearly show a "saccade-and-fixate" pattern: large, distinct spikes in velocity (representing rapid head redirections) are separated by periods of relative stability.

While minor high-frequency fluctuations in the raw trace (blue) may be attributed to facial feature detection noise, they remain significantly below our stability threshold (red dashed line). By applying this threshold, we successfully isolated biologically relevant "stable fixations" from "head saccades," ensuring that our subsequent social gaze analysis is based on periods of intentional head gaze direction.

d) Lines 237-286: The writing in this section does not emphasize the main results. There seem to be three takeaway points that could be emphasized better in the topic sentences of each of the paragraphs: i) Marmosets tended to spend most of their time on either end of the elongated box, not in the middle, ii) Males spent more time near the front of the box near the other animal than females, iii) Familiar pairs spent more time closer to each other.

To address this comment, we have reorganized this section to lead with the three key behavioral findings:

(1) We now state clearly in the topic sentence that marmosets preferred the ends of the arena over the middle.

(2) We have highlighted the finding that males spend significantly more time near the inner edge (closer to the partner) than females, irrespective of familiarity.

(3) We emphasized that familiar pairs maintain closer and more dynamic social distances over time, whereas unfamiliar pairs tend to move further apart as a session progresses.

e) Line 303: It would be useful to see time traces of head velocity of each member of the pair and categorization over time of the gaze event types. A stable epoch must be brief on the order of 100-200ms. It is unclear how distinct the stable fixation epochs are from the moments when the gaze is shifting. Also, the state transition analysis treats each stable epoch like one event, and then following a gaze movement by either of the pair, the state is defined again, is that correct?

We defined stable epochs as continuous periods where the face normal vector velocity remained below 0.05 normalized units for both animals. This ensures that a "gaze state" is only categorized when both marmosets have relatively fixed head orientations. As shown in the provided time traces in Author response image 2), the velocity profile is characterized by sharp peaks (head saccades) and clearly defined troughs (fixations). Further, we generated a probability histogram of stable head-gaze epoch durations (Author response image 3). The median duration of these stable epochs is 200ms, which aligns with biological expectations for fixation durations in primates and confirms that these states are distinct from the high-velocity shifts.

The reviewer’s interpretation is correct. Our Markov chain model treats each stable epoch as a single event. A transition occurs when at least one animal moves (exceeding the velocity threshold), resulting in a new stable epoch where the relative gaze state is re-evaluated. This approach allows us to model the sequence of social interactions as a series of discrete behavioral decisions.

Author response image 3.

Temporal characteristics of stable gaze, head gaze, epochs. The histogram illustrates the probability distribution of the duration (ms) of stablegaze behaviour epochs. A minimum duration threshold of 100 ms was applied to exclude transient, non-purposeful head gazes.

f) Lines 316-326: Some general summarizing statements to lead this paragraph would be useful. It seems that familiar pairs are more likely to participate in joint gaze, especially when close to each other, and perhaps, that males tended to gaze at females more than the reverse. Is there any notion that males were following the gaze of females?

We thank the reviewer for these suggestions. We have revised the topic sentences of this section to lead with a summary of the social takeaways, specifically highlighting the higher level of male interest and the shift toward reciprocal coordination in familiar pairs.

The reviewer correctly identified an important dynamic. Our transition analysis (Fig. 4D) confirms that males in both familiar and unfamiliar dyads frequently follow the female's gaze. This is evidenced by a robust transition probability (~17%) from "Male-to-Female Partner Gaze" (blue node) to "Joint Gaze" (green node). We found that this gaze-following behavior was a general feature of the dyads and did not differ significantly by familiarity, which is why it was not previously emphasized. However, we have now added a statement to the Results (Lines 358-365) to explicitly describe this male-led gaze-following behavior.

g) Lines 328-337: Can these findings in this paragraph be summarized more generally? It seems males view unfamiliar females longer, whereas for familiar females they are more likely to reciprocate viewing if being viewed by them and then to join in joint gaze with them. Would that event, viewing a female and then a transition to joint gaze, not be categorized as a gaze-following event?

We have now summarized the paragraph to emphasize the transition from vigilant monitoring in unfamiliar pairs to reciprocal awareness in familiar pairs.

Regarding "longer" viewing: We have clarified the text to specify that males' interest in unfamiliar females is persistent and robust rather than simply "longer" in a single duration. The high recurrence probability signifies that males consistently re-orient their gaze back to the unfamiliar female even if the interaction is briefly interrupted by movement.

Regarding gaze following and joint gaze: The reviewer asks if the transition from viewing a female to joint gaze constitutes gaze following. We agree that a transition from "male-to-female gaze" to "joint gaze" is indeed a gaze-following event (as noted in our previous response regarding Fig. 4D). However, the specific transition discussed in this paragraph (female-to-male gaze to male-to-female gaze) is different: it describes a "reciprocal" event where the male responded to being looked at by looking back at the female, while the female simultaneously shifted her gaze away. Since the two gaze cones did not intersect on an external object or on each other's faces simultaneously at the end of this transition, it was not categorized as joint gaze or gaze following.

h) Lines 339-351: It is not clear why gazing at the region surrounding a female's face (as opposed to the face itself) reflects "gaze monitoring tied to increased social attention (Dal Monte et l, 2022). This hypothesis could be expanded to make the prediction clear in this paragraph.

We thank the reviewer for identifying the need to clarify the hypothesis regarding the region surrounding the face. We have expanded this paragraph to explain why gazing at the peripheral facial region reflects social monitoring.

In many primate species, direct and sustained eye contact can be often interpreted as a threat or a challenge, particularly between unfamiliar individuals. Peripheral monitoring (looking at the area immediately surrounding the face) can strategically allow an animal to stay highly attentive to the partner's head orientation, gaze direction, and facial expressions—all critical for anticipating future actions—while minimizing the risk of social conflict. By demonstrating that unfamiliar marmosets utilize this peripheral strategy significantly more than familiar ones, we provide evidence that social attention in novel dyads is characterized by a social monitoring strategy that balances the need for information with social caution.

i) Lines 354-373: This section seems to suggest again that in a familiar male/female pair, the male is more likely to follow the female gaze and establish a joint gaze, and this occurs less with the unfamiliar pair only when closer in distance. Some summary sentences to begin the paragraph could help frame what to expect from the results.

We have added summarizing topic sentences to this section to clarify the relationship between familiarity and the spatial distribution of joint gaze.

(3) Discussion:

Lines 380-463: This section reads more clearly than most of the results, where it is often hard to connect the data plots to their significance for behavior. Overall, I believe the manuscript could be improved by setting up a hypothesis before presenting results in the paragraphs demonstrating the data. Some of the main findings appear in text from lines 413-419 (somewhat hidden even in discussion).

We sincerely appreciate the reviewer’s positive feedback on the clarity of the latter sections of our Discussion. We have taken the suggestion to heart and have performed a comprehensive restructuring of the Results and Discussion sections.

(1) We have moved the key takeaways, specifically the distinction between vigilant monitoring in unfamiliar pairs and reciprocal coordination in familiar pairs, from the end of the Discussion to the topic sentences of the relevant Results paragraphs.

(2) We established a unified framework throughout the manuscript that connects pixel-level tracking stability to the biological "saccade-and-fixate" movement pattern, and ultimately to the social dimensions of sex and familiarity.

(4) A couple of additional questions to address in the discussion:

a) Can you speculate why in this behavioral context the marmosets do not engage in reciprocal gaze where both are simultaneously looking at each other (lines 297-301)? How low is the incidence of this event, numerically, in comparison to the other events (1 in 1000 events, etc)?

We appreciate the reviewer’s interest in the lack of reciprocal gaze (mutual eye contact).

Numerically, reciprocal gaze events occurred with a frequency of approximately 1 in 500 social gaze events (comprising less than 0.2% of our social dataset). Given this extreme scarcity, we felt that any statistical comparisons across sex or familiarity would be underpowered and potentially misleading, leading to our decision to focus on partner and joint gaze states.

We speculate that the rarity of reciprocal gaze is primarily due to our task-free experimental setup. Unlike directed cooperation tasks where animals must look at each other to coordinate actions for a reward (e.g., Miss & Burkart, 2018), our study focused on task-free interactions. In a free-moving context without a common goal, marmosets may prioritize monitoring the environment or the partner’s actions (joint or partner gaze) over direct, sustained mutual eye contact, which can sometimes be perceived as a confrontational or high-arousal signal in primate social hierarchies.

b) Does a transition from a marmoset viewing their partner, to a joint gaze, count as a gaze-following event? It appears the authors are reluctant to use that terminology. What are the potential concerns in that terminology? Is there a concern that both animals orient to the same object that is salient to them without it being due to their gaze?

A transition from a partner-directed gaze to a joint gaze is indeed a gaze-following event. We distinguish these events from a transition between partner-directed gazes (e.g., male-to-female to female-to-male). In these "reciprocation" cases, once the second animal looked at the first, the first animal shifted their gaze away. Because the two gaze cones did not intersect on a common object at the end of the transition, I classified such events as a social exchange of attention rather than a coordinated gaze-following event.

Reviewer #2 (Recommendations for the authors):

I do have a few questions/points for clarification:

(1) While your approach appears to be able to track head orientation when the face is occluded or turned away from the primary cameras, how was the accuracy of this validated? Since you have multiple cameras, it should be possible to make the estimate using the occluded cameras and then validate using the non-occluded ones.

We appreciate the reviewer's comment regarding the validation of our tracking during partial occlusions.

We wish to clarify that our system does not utilize "primary" vs "auxiliary" cameras. Rather, any two or more cameras that capture facial features with high confidence are used to triangulate the points into 3D space. Thus, the "primary" cameras are dynamically determined frame-by-frame based on the animal's orientation.

To validate the accuracy of our 3D reconstruction during occlusions, we utilized a "projection-validation" approach. As demonstrated in Figure 2B (left panel), when the face is turned away from a specific camera, leaving only the back of the head visible, we used the facial features triangulated from the other non-occluded cameras and projected them onto the image plane of the occluded camera. The fact that these projected points aligned precisely with the expected (but hidden) anatomical landmarks confirms the global accuracy of our 3D model.

We previously benchmarked this approach using a three-camera system where we triangulated coordinates via two cameras and successfully projected them onto the third camera's image plane with high accuracy. This ensures that even when a camera is "blind" to the face, the 3D position estimated by the rest of the array remains robust.

(2) Marmosets, like other non-human primates, also look at other body postures for their social communication, though admittedly marmosets are far more likely to look others in the face than larger primates. The tail-raised genital displays come to mind. While the paper primarily focuses on shared vs deviant gaze, and I believe tracks not only the angle of viewing towards the target but also the distance from the face (please clarify if I am wrong), it would also be useful to know how often marmosets are looking at each other beyond just the face. This is particularly interesting if the gaze towards the partner varies depending on whether that partner was generally oriented towards the gazer, or not. For the joint gaze, were there conditions in which the two were looking at the same target, but had body postures that were not oriented toward one another (i.e. looking at a distant target beyond one of the animals, like looking over someone else's shoulder)?

We thank the reviewer for highlighting the importance of body postures and non-facial social signals (e.g., genital displays) in marmoset communication.

At the inception of this project, we explored tracking multiple body parts. However, due to the marmoset's dense fur and the lack of distinct skeletal markers under naturalistic lighting, human annotators and early automated tools struggled to achieve the precision required for high-resolution 3D kinematics. While recent advances in whole-body tracking now make these questions approachable, we chose to focus on the face normal vector because it provided the most robust and high-confidence signal for social orientation in our current dataset.

Regarding the "looking over the shoulder" scenario, we utilized a hierarchical classification system to prevent wrong categorization. Intersection with the partner’s face always took priority. If one animal’s gaze cone contained the other’s face, the state was classified as "Partner Gaze", even if the two gaze cones happened to intersect at a distant point in space. This ensures that "Joint Gaze" specifically captures instances where both animals ignore one another’s face regions to focus on a shared external target.

We agree that the relationship between body posture and head gaze is a fascinating area for future research. In our current setup, while "Joint Gaze" requires the head-gaze cones to intersect, the animals' bodies could indeed be oriented in different directions (e.g., looking at a distant target behind the partner). We have added a note to the Discussion acknowledging that incorporating whole-body gestures would further deepen the understanding of marmoset social ethology.

(3) In the introduction, (line 70), you raise the question of ecological relevance, using rhesus in laboratory settings. This could use a little more expansion/explanation of the limitations of current/past approaches.

We thank the reviewer for the suggestion to expand upon the ecological limitations of traditional laboratory paradigms.

We have substantially revised the Introduction (Lines 70–82) to provide a more detailed critique of past approaches. Specifically, we now highlight how traditional head-fixed or screen-based paradigms decouple eye movements from natural head-body dynamics and lack the reciprocal, multi-agent complexity found in real-world social environments (e.g., Land, 2006; Shepherd, 2010). By contrasting these constraints with the spatially and socially embedded nature of marmoset interactions, we clarify why a more naturalistic, quantitative approach is necessary to understand the true dynamics of social gaze. These additions provide a stronger theoretical foundation for our move toward a free-moving experimental model.

eLife Assessment

This important study provides a nuanced analysis of the impact of cues on cost/benefit decision-making deficits in male rats that could have translational relevance to many addictive disorders. The main findings are that cues paired with rewarded outcomes increase the proportion of risky outcomes, whereas risky choice is reduced when cues are paired with reward loss. The experimental data are compelling, whereas the computational analysis based on the optimisation of different Q-learning models is solid. The findings will be of interest to behavioural neuroscientists and clinicians with an interest in risk, decision making, and gambling disorders.

Reviewer #2 (Public review):

Summary:

The manuscript by Hathaway et al. describes a set of elegant behavioral experiments designed to understand which aspects of cue-reward contingencies drive risky choice behavior. The authors developed several clever variants of the well established rodent gambling task (also developed by this group) to understand how audiovisual cues alter learning, choice behavior, and risk. Computational and sophisticated statistical approaches were used to provide evidence that: 1) audiovisual cues drive risky choice if they are paired with rewards and decrease risk if only paired with loss, 2) pairing cues with rewards reduces learning from punishment, and 3) differences in risk taking seem to be present early on in training.

Strengths:

The paper is well written, the experiments well designed, and the results are highly interesting particularly for understanding how cues can motivate and invigorate normal and abnormal behavior.

Comments on revisions:

The authors have done an exceptional job at addressing my initial concerns and questions regarding the evidence to support their claims. I have no additional suggestions or concerns.

Reviewer #3 (Public review):

Summary:

In this work, Hathaway and colleagues aim to understand how audiovisual cues at time of outcome promote selection of risky choices. A real life illustration of this effect is used in electronic gambling machines which signal a win with flashing lights and jingles, encouraging the player to keep betting. More specifically, the author asks whether the cue has to be paired exclusively to wins, or whether it can be paired to both outcomes, or exclusively loss outcomes, or occur randomly. To tackle this question, they employ a version of the Iowa Gambling Task adapted to rats, and test the effect of different rules of cue-outcome associations on the probability of selecting the riskier options; they then test the effect of prior reward devaluation on the task; finally, the optimise computational models on the early phases of the experiment to investigate potential mechanisms underlying the behavioural differences.

Strengths:

The experimental approach is very thorough, in particular the choice of the different task variants cover a wide range of different potential hypotheses. Using this approach, they find that, although rats prefer the optimal choices, there is a shift towards selecting riskier options in the variants of the task where the cue is paired to win outcomes. They analyse this population average shift by showing that there is a concurrent increase in the number of risk-taking individuals in these tasks. They also make the novel discovery that pairing cues with loss outcomes instead reduces the tendency for risky decisions.

The computational strategy is appropriate and in keeping with the accepted state of the art: defining a set of candidate models, optimising them, comparing them, simulating the best ones to ensure they replicate the main experimental results, then analysing parameter estimates in the different tasks to speculate abut potential mechanisms.

Weaknesses:

While the overall computational approach is excellent, there is a missed opportunity in the computational modelling section due to the choice of models which is dependent on a preceding study by Langdon et al. (2019). Loss trials come at a double cost: firstly the lost opportunity of not having selected a winning option which is reflected straightforwardly in Q-learning by the fact that r=0, secondly a waiting period which will affect the overall reward rate. The authors combine these costs by converting the time penalty into "reward currency" using three different functions which make up the three different tested models. This means the question when comparing models is not something along the lines of "are individuals in the paired win-cue tasks more sensitive to risk? or less sensitive to time? etc." but rather "what is the best way of converting time into Q-value currency to fit the data?". Instead, the authors could have contrasted other models which explicitly track time as a separate variable (see for example "Impulsivity and risk-seeking as Bayesian inference under dopaminergic control" (Mikhael & Gershman 2021)) or give actions an extra risk bonus (as in "Nicotinic receptors in the VTA promote uncertainty seeking" (Naude et al 2016)) to better disentangle the mechanisms at play.

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer 1 (Public review):

When do behavioral differences emerge between the task variants? Based on the results and discussion, the cues increase the salience of either the wins or the losses, biasing behavior in favor of either risky or optimal choice. If this is the case, one might expect the cues to expedite learning, particularly in the standard and loss condition. Providing an analysis of the acquisition of the tasks may provide insight into how the cues are "teaching" decision-making and might explain how biases are formed and cemented.

While considerable differences in decision making emerge in early sessions of training, we do not observe any evidence that cuing outcomes expedites the development of stable choice patterns. Indeed, since the outcomes are cued across all four options, there is no categorical difference in salience between optimal and risky choices. Thus, our interpretation is that cuing wins and/or losses alters the integration of this feedback into choice preference, rather than the rate of the development of choice preference. To quantitatively address this point, we have included the following analysis:

“To quantitatively examine choice variability during training, we binned sessions 1-5 and 6-10 and analyzed variability in choice patterns across task variants. Analysis of the first five sessions of training revealed a significant shift in decision score across sessions (F(3, 502) = 31.23, p <.0001), which differed between task variants (session x task: F(16, 502) = 2.13, p = .007). Conversely, while significant differences in overall score were observed between task variants in sessions 6-10 (task: F(5, 156) = 6.81, p <.0001), there was no significant variability across sessions (session: F(3, 481) = 2.06, p = .10, task x session: F(15, 481) = 0.78, p = .71). This indicates that the variability in choice preference (and presumably, learning about outcomes) is maximized in the first five sessions, and there are no obvious differences in the rate of development of stable choice patterns between task variants.”

Does the learning period used for the modeling impact the interpretation of the behavioral results? The authors indicate that computational modeling was done on the first five sessions and used these data to predict preferences at baseline. Based on these results, punishment learning predicts choice preference. However, these animals are not naïve to the contingencies because of the forced choice training prior to the task, which may impact behavior in these early sessions. Though punishment learning may initially predict risk preference, other parameters later in training may also predict behavior at baseline.

The first five sessions were chosen based on a previously developed method used in Langdon et al. (2019). When choosing the number of sessions to include, there is a balance between including more data points to improve estimation of parameters while also targeting the timeframe of maximal learning. As training continues, the impact of outcomes on subsequent choice should decrease, and the learning rate would trend towards zero. This can be observed in the reduction in inter-session choice variability as training progresses, as demonstrated in the analyses above. Once learning has ceased, presumably other cognitive processes may dictate choice (for example, habitual stimulus-response associations), which would not be appropriately captured by reinforcement learning models. It would be a separate research question to determine the point at which parameters no longer become predictive, requiring a larger dataset to thoroughly assess. We acknowledge that we did not provide sufficient justification for the learning period used for the modeling. In conjunction with the analysis of early sessions outlined above, we have added the following to the text:

“We investigated differences in the acquisition of each task variant by fitting several reinforcement learning (RL) models to early sessions. Our modeling approach closely follows methods outlined in Langdon et al. (2019), in which a much larger dataset (>100 rats per task) was used to develop the RL models applied here. Due to the comparatively small n per group in the current study, we limited our model selection to those previously validated in Langdon et al. (2019), with minor extensions. As in previous work, models were fit to valid choices from the first five sessions. As training continues, the impact of outcomes on subsequent choice should decline, and parameter values may evolve over time (e.g., decreasing learning rate). To target the period of learning during which outcomes have maximal influence over choice, and parameters likely have fixed values, we limited our analyses to the first five sessions.”

The authors also present simulated data from the models for sessions 18-20, but according to the statistical analysis section, sessions 35-40 were used for analysis (and presumably presented in Figure 1). If the simulation is carried out in sessions 35-40, do the models fit the data?

Based on our experience, choice patterns are well instantiated by session 20, and training only continues to 30+ sessions to achieve stability in other task variables (e.g., latencies, premature responding, etc.). That being said, the discrepancy between session numbers is confusing, so we’ve extended the simulations to match the same session numbers that were analyzed in the experimental data.

Finally, though the n's are small, it would be interesting to see how the devaluation impacts computational metrics. These additional analyses may help to explain the nuanced effects of the cues in the task variants. 

Unfortunately, as the devaluation experiment is only one session, there are insufficient data to run the same models. Furthermore, changes in choice are subtle and not uniform across rats, making it difficult to reliably model this effect at the individual level. A separate experiment could investigate the specific cognitive processes underlying the devaluation effect.

Reviewer #1 (Recommendations for the authors):

The authors do not present individual data points for behavior. Including these data points would improve the interpretability of the results. Adding significant notations to the bar graphs would also help the reader. Although the stats are provided and significant comparisons highlighted, it isn't easy to go between the table and the figure to detect significant outcomes. If done, the statistics tables could be moved to the supplement. Including estimates of effect size for main findings in the main text would also benefit the reader. 

We thank the reviewer for their feedback on our approach to the figures and significance reporting – we have updated the relevant figures to include individual data points. Furthermore, we’ve added significance notations for task variants that are significantly different from the uncued or standard cued tasks on the figures. We’ve also moved some statistics tables to the supplement, as suggested. 

The authors allude to other metrics of the task (trials, omissions, etc.) but do not present these data anywhere. Including supplementary figures including individual data points and statistical analyses in the supplement is strongly encouraged.

A supplementary figure visualizing these metrics (choice latency, trials completed, and omissions) has been added, with individual data points included. Statistical analyses are reported in the main text – no significant effect in the ANOVAs were observed for any of these metrics, so post hoc analyses were not performed. 

Figure 4 is confusing. Presenting the WAIC values for each model rather than compared to the nonlinear model would be easier to understand. It is also unclear if statistical tests were used to assess differences in model fit as no test information is provided.

Figure 4 has been updated to increase clarity and address feedback from another reviewer. Raw WAIC values are not ideal for visualization, as the task variants have differing amounts of data and thus would be difficult to include on the same Y-axis. Instead, we present each model’s difference in WAIC relative to a basic model with no timeout penalty transform, so that all three models are visible, and the direction of model improvement is clearly indicated. Statistical tests of WAIC differences are not standard, as the numerical differences themselves indicate a better fit.

The authors do not provide a data availability statement.

We thank the reviewer for calling our attention to this oversight. A data availability statement has been added. 

Reviewer 2 (Public review):

Additional support and evidence are needed for the claims made by the authors. Some of the statements are inconsistent with the data and/or analyses or are only weakly supportive of the claims.

We appreciate the reviewer’s overarching concern that some claims in the original manuscript were insufficiently supported by the data or analyses. To address this, we have provided further rationale for the devaluation experiment and clarified our interpretation of those results, expanded the computational modeling analyses, and revised figures and wording to improve clarity. Below, we respond to the reviewer’s specific comments in detail.

Reviewer #2 (Recommendations for the authors):

Different variants of an RL model were used to understand how loss outcomes impacted choice behavior across the gambling task variants. Did the authors try different variants for rewarded outcomes? I wonder whether the loss specific RL effects are constrained to that domain or perhaps emerged because choice behavior to losses was better estimated with the different RL variants. For example, rewarded outcomes across the different choices may not scale linearly (e.g., 1, 2, 3, 4) so including a model in which Rtr is scaled by a free parameter might improve the fit for win choices.

We agree that asymmetries in model flexibility could, in principle, contribute to the observed effects. While we are somewhat limited in our ability to develop and validate further models due to the small size of the datasets compared to the high degree of choice variability between rats, we have explored the possibility as far as the data allow by fitting a model that includes a scaling parameter for rewards in addition to punishments:

“While we restricted our model selection to those previously validated on larger datasets, the specificity of the main finding to the punishment learning rate may be due to the greater flexibility afforded to loss scaling, rather than a true asymmetry in learning. To test this hypothesis, we fit a model featuring a scaling parameter for rewards, in addition to scaled costs:

where mRew is a linear scaling parameter for reward size. A separate scaling parameter was used for timeout penalty duration (i.e., same as scaled cost model). Group-level parameter estimates (Figure S3) reflected similar differences in the punishment learning rate and reward learning rate as the scaled cost model (Figure S4). Furthermore, all 95% HDIs for the mRew scaling parameter included 1, indicating that at least at the group level, scaling of reward size across the P1-P4 options closely follows the actual number of earned sucrose pellets. Thus, we find no evidence that our results can be simply attributed to the increased parameterization of losing outcomes.”

Additionally, I would like to see evidence that these alternative models provide a better fit compared to a standard delta-rule updating for unrewarded choices.

Each model is now compared directly to a standard delta-rule update model in the WAIC figure to demonstrate that the current models are a better fit for the data.

Could the authors provide some visualization of how variation in the r, m, or b parameters impact choices and/or patterns of choices?

We have added a figure to the supplementary section to visualize how different values for the r, m, and b parameters could alter the size of updates to Q-values on each trial across the four different options, thereby impacting subsequent choice. 

It was challenging to understand the impact of the reported effects and interpretation of the authors at various points in the manuscript. For example, the authors state that "only rats trained on tasks without win-paired cues exhibited shifts in risk preference following reinforcer devaluation". Figure 3 however seems to indicate that rats trained on the reverse-cued task show shifts in risk preference. 

We agree the original wording did not fully capture the nuance apparent in the figure. While not significantly different from baseline, rats in the reverse-cued experiment could have indeed updated their choice patterns and we were underpowered to detect the effect. We have updated the results section to include this point, and to more specifically outline that win-paired cues that scale with reward size lead to insensitivity to reinforcer devaluation:

“This indicates that pairing audiovisual cues with reward induces some degree of inflexibility in risk-preferring rats. Importantly, pairing cues with losses alone does not elicit rigidity in choice. Thus, in keeping with the observed effect on overall choice patterns, pairing cues with wins has a unique impact on sensitivity to reinforcer devaluation. Although not statistically significant, visual inspection of the reverse-cued task suggests that some choice flexibility may be present, and the study may be underpowered to detect this effect. Nonetheless, win-paired cues that scale with reward size reduce flexibility in choice patterns following reinforcer devaluation.”

It was not clear to me why the authors did a devaluation test and what was expected. Adding details regarding the motivation for specific analyses and/or experiments would improve understanding of these exciting results.

Further explanation has been added to the results section for the devaluation test to clarify the rationale and expected results:

“We next tested whether pairing salient audiovisual cues with outcomes on the rGT impacts flexibility in decision making when outcome values are updated. Reinforcer devaluation, in which subjects are sated on the sugar pellet reinforcer prior to task performance (presumably devaluing the outcome), is a common test of flexibility of decision making (Adams & Dickinson, 1981). We have previously employed this method to demonstrate that rats trained on the standard-cued task are insensitive to reinforcer devaluation (i.e., choice patterns do not shift despite devaluation of the sugar pellet reward; Hathaway et al., 2021).”

Some rats in the rGT become risk takers and some do not, but whether this is an innate phenomenon or emerges with training is not known. The authors report some correlations between the RL parameters and subsequent risk scores but this may be an artifact because the risk scores and many of the parameters differ between the experimental groups. Restricting these analyses to the rats in the standard procedure (or even conducting it in other rats that have been run in the rGT standard task) would alleviate this concern. The authors should also expand upon this result in the discussion. (if it holds up) and provide graphs of this relationship in the manuscript.

In a previous paper on which these analyses were based (Langdon et al., 2019), analyses of the relationship between RL parameter estimates and final decision score were conducted separately for rats trained on either the uncued or standard cued task, as the reviewer has suggested here. Those analyses showed that parameters controlling the learning from negative outcomes were specifically related to final score in both tasks. While we don’t have the appropriate n per group to split the analyses by task variant in the current study, we have highlighted these previous findings in the results section to address this concern:

“In Langdon et al. (2019), analyses were conducted to test whether parameters controlling sensitivity to punishment predicted final decision score at the end of training in the uncued and standard cued task variants. These analyses showed that across both task variants, there was evidence of reduced punishment sensitivity (i.e., lower m parameter or punishment learning rate) in risky versus optimal rats. We conducted similar analyses here to examine whether parameter estimates covary with decision score at end of training. To accomplish this, we fit simple linear regression models for each parameter and assessed whether the slopes were significantly different from zero.”

I don't see a b parameter in the nonlinear cost model, but is presented in Figure 6 and also in the "Parameters predicting risk preference on the rGT". The authors either need to update the formula or clarify what the b parameter quantifies in the nonlinear model.

We thank the reviewer for pointing out this oversight; the equation has been updated to include the b parameter.

The risk score is very confusing as high numbers or % indicate less risk and lower (more negative numbers) indicate greater risk. I've had to reread the text multiple times to remind myself of this, so I anticipate the same will be true for other readers. Perhaps the authors can add a visual guide to their y-axis indicating more positive numbers are less risky choices.

We acknowledge that this measure can be confusing – the calculation of this score is standard for the Iowa Gambling Task conducted in humans, on which the rGT is based, and was therefore adopted here. We’ve changed the name from “risk score” to “decision score”, along with including a visual guide to the y-axis in Figure 2, to address this point.

Negative learning rate is confusing as it almost implies that the learning was a negative value, rather than being a learning rate for negative outcomes. Please revise in the figures and in the text.

We have updated the text and figures where appropriate from “negative learning rate” to “punishment learning rate”. We have also changed the text from “positive learning rate” to “reward learning rate” to match this terminology.

Reviewer 3 (Public review):

There is a very problematic statistical stratagem that involves categorising individuals as either risky or optimal based on their choice probabilities. As a measurement or outcome, this is fine, as previously highlighted in the results, but this label is then used as a factor in different ANOVAs to analyse the very same choice probabilities, which then constitutes a circular argument (individuals categorised as risky because they make more risky choices, make more risky choices...).

Risk status was included as a factor to test whether the effects of the cue paradigms differed between risky versus optimal rats (i.e., interaction effects), not as an independent predictor of choice preference. We focus on results showing a significant task x risk status interaction, and conducted follow-up analyses separately within each group, at which point risk status was no longer included as a factor. We do not interpret main effects or choice x status interactions, which would indeed be circular for the reason noted by the reviewer.

A second experiment was done to study the effect of devaluation on risky choices in the different tasks. The results, which are not very clear to understand from Figure 3, would suggest that reward devaluation affects choices in tasks where the win-cue pairing is not present. The authors interpret this result by saying that pairing wins with cues makes the individuals insensitive to reward devaluation. Counter this, if an individual is prone to making risky choices in a given task, this points to an already distorted sense of value as the most rewarding strategy is to make optimal non-risky choices.

We have included significance notations in Figure 3 and included further detail in the text to improve clarity of the findings for the devaluation test. The reviewer raises an interesting point that risk-preferring rats have a distorted sense of value, since they do not follow the optimal strategy. However, we believe that this is at least partially separable from insensitivity to devaluation, since risk-preferring rats trained on tasks that don’t feature win-paired cues still exhibit flexibility in choice. We have added the following point to the discussion to address this:

“While risk-preferring rats exhibit some degree of distortion in reward valuation, as they do not follow the most rewarding strategy (i.e., selecting optimal options), we believe this to be at least partially separable from choice inflexibility, as risk-preferring rats on tasks that don’t feature win-paired cues remain sensitive to devaluation.”

While the overall computational approach is excellent, I believe that the choice of computational models is poor. Loss trials come at a double cost, something the authors might want to elaborate more upon, firstly the lost opportunity of not having selected a winning option which is reflected in Q-learning by the fact that r=0, and secondly a waiting period which will affect the overall reward rate. The authors choose to combine these costs by attempting to convert the time penalty into "reward currency" using three different functions that make up the three different tested models. This is a bit of a wasted opportunity as the question when comparing models is not something like "are individuals in the paired win-cue tasks more sensitive to risk? or less sensitive to time? etc" but "what is the best way of converting time into Q-value currency to fit the data?" Instead, the authors could have contrasted other models that explicitly track time as a separate variable (see for example "Impulsivity and risk-seeking as Bayesian inference under dopaminergic control" (Mikhael & Gershman 2021)) or give actions an extra risk bonus (as in "Nicotinic receptors in the VTA promote uncertainty seeking" (Naude et al 2016)).

We thank the reviewer for their thoughtful suggestions and agree that alternative modeling frameworks that explicitly track time or incorporate uncertainty bonuses would be highly informative for understanding the mechanisms underlying risky choice. However, the models employed here are drawn from previous work that required >100 rats per group for model development and validation. Due to the high degree of variability in decision making within the groups and the relatively small number of rats, this dataset is not well suited for substantial model innovation. Indeed, the most complex model from previous work had to be simplified to achieve model convergence. Testing models that greatly diverge from the previously validated RL models would make it difficult to determine whether poor model fit reflects a misspecified model or insufficient data.

We’d also like to note that the driving question for this study is to investigate the impact of different cue variants on choice patterns – untangling the relationship between timing, uncertainty, and risky choice is an important and interesting question, but beyond the scope of the present work. 

To address this limitation, we have expanded our justification of model choice in the results section to emphasize that we are applying previously developed models, with minor extensions:

“We investigated differences in the acquisition of each task variant by fitting several reinforcement learning (RL) models to early sessions. Our modeling approach closely follows methods outlined in Langdon et al. (2019), in which a much larger dataset (>100 rats per task) was used to develop the RL models applied here. Due to the comparatively small n per group in the current study, we limited our model selection to those previously validated in Langdon et al. (2019), with minor extensions.”

Another weakness of the computational section is the fact, that despite simulations having been made, figure 5 only shows the simulated risk scores and not the different choice probabilities which would be a much more interesting metric by which to judge model validity. 

We have expanded Figure 5 to show the simulated choice of each option.

In the last section, the authors ask whether the parameter estimates (obtained from optimisation on the early sessions) could be used to predict risk preference. While this is an interesting question to address, the authors give very little explanation as to how they establish any predictive relationship. A figure and more detailed explanation would have been warranted to support their claims.

We have expanded this section to provide clearer detail on the methods used to conduct this analysis and added a figure. To address a point raised by another reviewer, the statistical approach has been revised to more closely align with that used in Langdon et al. (2019), and the results have been updated appropriately:

“We next tested whether any of the subject-level parameter estimates in the nonlinear or scaled + offset model could reliably predict risk preference scores at the end of training. In Langdon et al. (2019), analyses were conducted to test whether parameters controlling sensitivity to punishment predicted final decision score at the end of training in the uncued and standard cued task variants. These analyses showed that across both task variants, there was evidence of reduced punishment sensitivity (i.e., lower m parameter or punishment learning rate) in risky versus optimal rats. We conducted similar analyses here to examine whether parameter estimates covary with decision score at end of training. To accomplish this, we fit simple linear regression models for each parameter and assessed whether the slopes were significantly different from zero.”

Why were the simulated risk scores calculated for sessions 18-20 and not 35-39 as in the experimental data, and why were the models optimised only on the first sessions?

These points were addressed in response to reviewer #1:

Based on our experience, choice patterns are well instantiated by session 20, and training only continues to 30+ sessions to achieve stability in other task variables (e.g., latencies, premature responding, etc.). That being said, the discrepancy between session numbers is confusing, so we’ve extended the simulations to match the same session numbers that were analyzed in the experimental data.

The first five sessions were chosen based on a previously developed method used in Langdon et al. (2019). When choosing the number of sessions to include, there is a balance between including more data points to improve estimation of parameters while also targeting the timeframe of maximal learning. As training continues, the impact of outcomes on subsequent choice should decrease, and the learning rate would trend towards zero. This can be observed in the reduction in inter-session choice variability as training progresses, as demonstrated in the analyses above. Once learning has ceased, presumably other cognitive processes may dictate choice (for example, habitual stimulus-response associations), which would not be appropriately captured by reinforcement learning models. It would be a separate research question to determine the point at which parameters no longer become predictive, requiring a larger dataset to thoroughly assess. We acknowledge that we did not provide sufficient justification for the learning period used for the modeling. In conjunction with the analysis of early sessions outlined above, we have added the following to the text:

“We investigated differences in the acquisition of each task variant by fitting several reinforcement learning (RL) models to early sessions. Our modeling approach closely follows methods outlined in Langdon et al. (2019), in which a much larger dataset (>100 rats per task) was used to develop the RL models applied here. Due to the comparatively small n per group in the current study, we limited our model selection to those previously validated in Langdon et al. (2019), with minor extensions. As in previous work, models were fit to valid choices from the first five sessions. As training continues, the impact of outcomes on subsequent choice should decline, and parameter values may evolve over time (e.g., decreasing learning rate). To target the period of learning during which outcomes have maximal influence over choice, and parameters likely have fixed values, we limited our analyses to the first five sessions.”

Concerning the figures, could you consider replacing or including with the bar plots, the full distribution of individual dots, or a violin plot, something to better capture the distribution of the data. This would be particularly beneficial for Figure 2B the risk score which, without a distribution suggests all individuals are optimal, something which in the text claim is not the case. 

Individual data points have been added to the relevant figures.

Is this not a case of compositional data where ANOVA is definitely not an appropriate method (compositional data consist in reporting proportions of different elements in a whole, eg this rock is 60% silicate, 20% man-made cement, etc.) because of violation of normality and mostly dependence between measurements (the sum must be 100% as in your case where knowing the proportions of P1, P2 and P3, I automatically deduce P4). I leave to you the care of finding a potential alternative. In any case, I also had difficulties understanding the varying degrees of freedom of the different reported F statistics which worry me that this has not been done properly.

This is a fair criticism, as choice proportions across P1-P4 are not fully independent. While alternative approaches do exist, there is no widely adopted or straightforward method that has been validated for this task. Accordingly, ANOVA remains the standard analytical approach for this task, as it facilitates comparison with previous work and is readily understood by readers. As mentioned in the methods, an arcsine transformation was applied to the proportional data to mitigate issues associated with bounded measures (i.e., summing to 100%). We thank the reviewer for drawing our attention to the discrepancies in the degrees of freedom – these have now been corrected.

eLife Assessment

This study provides a useful analysis of the changes in chromatin organization and gene expression that occur during the differentiation of two cell types (anterior endoderm and prechordal plate) from a common progenitor in zebrafish, together with investigations into the molecular factors involved. Although the findings are consistent with previous work, the evidence presented appears to be incomplete and would benefit from more rigorous quantification of live imaging and Cre-Lox experiments, a stronger rationale and controls for experiments manipulating chromatin remodeling factors, and a strong justification for the explant model especially given differences between explant and whole embryo data. This work may be of interest to zebrafish developmental biologists investigating the mechanisms underlying specification.

Reviewer #1 (Public review):

Summary:

During vertebrate gastrulation, mesendoderm cells are initially specified by morphogens (e.g. Nodal) and segregate into endoderm and mesoderm in part based on Nodal concentrations. Using zebrafish genetics, live imaging, and single-cell multi-omics, the manuscript by Cheng et al presents evidence to support a claim that anterior endoderm progenitors derive primarily from prechordal plate progenitors, with transcriptional regulators goosecoid (gsc) and ripply1 playing key roles in this cell fate determination. Such a finding would represent a significant advance in our understanding of how anterior endoderm is specified in vertebrate embryos.

Strengths:

Live imaging based tracking of PP and endo reporters (Fig 2) are well executed and convincing, though a larger number of individual cell tracks will be needed. In the first round of review, only a single cell track (n=1) was quantified. Now, more tracks have been collected but these data are still not clearly reported in a way that warrants their evaluation.

Weaknesses:

(1) While the authors have made an effort to include a gsc:CRE lineage tracing component to their study, the experimental data now presented (Figure S4E and reviewer figures) could be much stronger and more thorough. In the new panel, authors show a single microscopy image containing both red and green fluorescent cells. The green signal, which seems to mark the PP, is presumably derived from Tg(gsc:EGFP). The red mCherry signal is presumably derived from the combined effects of a Tg(gsc:CRE) and Tg(sox17-lox-STOP-lox-mCherry), i.e., labeling the progeny of gsc+ progenitors which expressed CRE and underwent recombination to create a productive endoderm-specific Tg(sox17:mCherry) reporter. The result appears to be promising and in line with the authors' predictions. However, this result should be strengthened by performing the experiment in stable transgenic lines (not just freshly injected F0 embryos) and should be properly quantified. The authors state in the legend that "the experiment was performed on at least 3 independent replicates", but offer no further detail, explanations, or quantifications. This issue is reminiscent of concerns from the previous round of review, where live tracking data derived from examining just a single (n=1) cell were presented. These standards might be adequate for generating preliminary insights, but fall far below what we would have previously expected from an Elife publication.

(2) I found the authors' rebuttal to my concerns about URD-trajectory derived insights and gsc/sox17 expression timing confusing. The authors claim that they get different results regarding gsc expression prevalence in the hypothetical PP/endoderm progenitor cluster when comparing scRNAseq data from embryos vs explants. Then they seem to use this difference to justify the use of the explants over the embryos - presumably because the explants enriched for the behavior that they wanted to see? They conclude that "directly using embryonic data to dissect the mechanism of fate separation between PP and anterior endoderm might not yield highly accurate results." I strongly disagree with this. I would argue that the whole-embryo dataset is likely doing a better job of cleanly separating these trajectories from each other.

(3) My concern about the use of n=1 cell for live tracking has been partially but not fully addressed. The authors should plot data point from each individual cell in the revised Figure 2D, instead of just saying "multiple cells" they should report the total number of cells that are actually included now (n=?), and should provide representative movies for a few additional examples.

At present the authors' data, as presented, still only partially support their aims and conclusions.

Reviewer #2 (Public review):

Summary:

During vertebrate gastrulation, the mesoderm and endoderm arise from a common population of precursor cells and are specified by similar signaling events, raising questions as to how these two germ layers are distinguished. Here, Cheng and colleagues use zebrafish gastrulation as a model for mesoderm and endoderm segregation. By reanalyzing published single cell sequencing data, they identify a common progenitor population for anterior endoderm and the mesodermal prechordal plate (PP). They find that expression levels of PP genes gsc and ripply are among the earliest differences between these populations, and that their increased expression suppresses the expression of endoderm markers. Further analysis of chromatin accessibility and Ripply CUT-and-TAG is consistent with direct repression of endoderm by this PP marker. This study demonstrates roles for Gsc and Ripply in suppressing anterior endoderm fate, but this role for Gsc was already known and the effect of Ripply is limited to a small population of anterior endoderm.

Strengths:

Integrated single cell ATAC- and RNA-seq convincingly demonstrate changes in chromatin accessibility that may underlie segregation of mesoderm and endoderm lineages, including gsc and ripply. Identification of Ripply-occupied genomic regions augments this analysis. The genetic mutants for both genes provide strong evidence for their function anterior mesendoderm development, although these phenotypes are subtle.

Weaknesses:

The use of zebrafish embryonic explants for cell fate trajectory analysis (rather than intact embryos) is not justified. Much of the work is focused on the role of Nodal in the mesoderm/endoderm fate decision, but the results largely confirm previous studies and again provide few new insights. The authors similarly confirm previous findings that FGF signaling likely plays a larger role in this fate decision, but these results are largely overlooked by the authors.

Reviewer #3 (Public review):

Summary of work:

Cheng, Liu, Dong, et al. demonstrate that anterior endoderm cells can arise from prechordal plate progenitors, which is suggested by pseudotime reanalysis of published scRNAseq data, pseudotime analysis of new scRNAseq data generated from Nodal-stimulated explants, live imaging from sox17:DsRed and gsc:eGFP transgenics, fluorescent in situ hybridization, and a Cre/Lox system. Early fate mapping studies already suggested that progenitors at the dorsal margin give rise to both of these cell types (Warga), and live imaging from the Heisenberg lab (Sako 2016, Barone 2017) convincingly showed this previously. However, the data presented for this point are very nice and further cement this result. Though better demonstrated by previous work (Alexander 1999, Gritsman 1999, Gritsman 2000, Sako 2016, Rogers 2017, others), the manuscript presents confirmatory data that high Nodal signaling is required for both cell types. The manuscript generates new single-cell RNAseq data from Nodal-stimulated explants with increased (lft1 KO) or decreased (ndr1 KD) Nodal signaling and multi-omic ATAC+scRNAseq data from wild-type 6 hpf embryos, which can be used as a resource, though few new conclusions are drawn from it in this manuscript. Lastly, the manuscript presents suggests that SWI/SNF remodelers and Ripply1 may be involved in the anterior endoderm - prechordal plate decision, but these data are less convincing. The SWI/SNF remodeler experiments are unconvincing because the demonstration that these factors are differentially expressed or active between the two cell types are weak. The Ripply1 gain-of function experiments are unconvincing because they are based on incredibly high overexpression of ripply1 (500 pg or 1000 pg) that generates a phenotype that is not in line with previously demonstrated overexpression studies (with phenotypes from 10-20x lower expression). Similarly, the cut-and-tag data seems low quality and is based on high overexpression, so may not support direct binding of ripply1 to these loci.

During revision, the authors addressed some comments, including eliminating references to "lineage" when referring to pseudotime trajectories, eliminating conclusions drawn from locations of cells on UMAP plots, and reducing use of the term "cooperative" which may have been confusing in this context, as well as increasing the number of embryos analyzed for some experiments. The authors also point out that whole-embryo transcriptional trajectories typically do not associate endodermal cells with prechordal plate cells, despite classical evidence that they are related. This is most likely because endodermal cells arise from several different previous transcriptional states in different regions of the embryonic margin and are, as the authors point out, difficult to computationally sort into dorsal, lateral, and ventral populations. Thus, there is value in generating data to more specifically look at the relationship between dorsal mesodermal and endodermal populations. However, the decision to use an artificial Nodal-treated explant system, rather than isolating the relevant population from whole embryos (such as by dissection prior to dissociation) remains a weakness of the manuscript, since it is unclear whether endodermal specification has been altered in this system (there seem to be few endodermal cells produced and the system involves manipulating one of the signals under study in this work). Concerns about the rigor of experiments concerning ripply1 and SWI/SNF experiments remains. While the authors improved peak calling in their ripply1 cut-and-tag, it is still based on massive overexpression of ripply1 that may drive binding outside of its endogenous loci.

In the end, this study provides some additional details in the cell fate decision between the prechordal plate and anterior endoderm and generates new data that may be useful for reanalysis by other experts in the field. However, this work does not make clear how Nodal signaling, FGF signaling, and elements of the gene regulatory network (including gsc, possibly ripply1, and other factors) interact to make the decision. I suggest that this manuscript is of interest to Nodal signaling or zebrafish germ layer patterning afficionados, but may not be of interest to a broad audience. While it provides new datasets and observations, it does not weave these into a convincing story that advances our understanding of the specification of these cell types.

Author response:

The following is the authors’ response to the original reviews

eLife Assessment

This study provides a useful analysis of the changes in chromatin organization and gene expression that occur during the differentiation of two cell types (anterior endoderm and prechordal plate) from a common progenitor in zebrafish. Although the findings are consistent with previous work, the evidence presented in the study appears to be incomplete and would benefit from more rigorous interpretation of single-cell data, more in-depth lineage tracing, overexpression experiments with physiological levels of Ripply, and a clearer justification for using an explant system. With these modifications, this paper will be of interest to zebrafish developmental biologists investigating mechanisms underlying differentiation.

We sincerely thank the editor and the reviewers for their valuable time and efforts. Their insightful comments were greatly appreciated and have been largely addressed in the revised manuscript. We are confident that these revisions have enhanced the overall quality and clarity of our paper.

Reviewer #1 (Public review):

Summary:

During vertebrate gastrulation, mesendoderm cells are initially specified by morphogens (e.g. Nodal) and segregate into endoderm and mesoderm in part based on Nodal concentrations. Using zebrafish genetics, live imaging, and single-cell multi-omics, the manuscript by Cheng et al presents evidence to support a claim that anterior endoderm progenitors derive primarily from prechordal plate progenitors, with transcriptional regulators goosecoid (Gsc) and ripply1 playing key roles in this cell fate determination. Such a finding would represent a significant advance in our understanding of how anterior endoderm is specified in vertebrate embryos.

We would like to thank reviewer #1 for his/her comments and positive feedbacks about our manuscript.

Strengths:

Live imaging-based tracking of PP and endo reporters (Figure 2) is well executed and convincing, though a larger number of individual cell tracks will be needed. Currently, only a single cell track (n=1) is provided.

We thank the reviewer for the positive comments and the valuable suggestion. As the reviewer suggested, we re-performed live imaging analyses on the embryos of Tg(gsc:EGFP;sox17:DsRed). We tracked dozens of cells during their transformation from gsc-positive to sox17-positive. Furthermore, we performed quantification of the RFP/GFP signal intensity ratio in these cells over the course of development (Please see the revised Figure 2D and MovieS4).

Weaknesses:

(1) The central claim of the paper - that the anterior endoderm progenitors arise directly from prechordal plate progenitors - is not adequately supported by the evidence presented. This is a claim about cell lineage, which the authors are attempting to support with data from single-cell profiling and genetic manipulations in embryos and explants. The construction of gene expression (pseudo-time) trajectories, while a modern and powerful approach for hypothesis generation, should not be used as a substitute for bona fide lineage tracing methods. If the authors' central hypothesis is correct, a CRE-based lineage tracing experiment (e.g. driving CRE using a PP marker such as Gsc) should be able to label PP progenitor cells that ultimately contribute to anterior endoderm-derived tissues. Such an experiment would also allow the authors to quantify the relative contribution of PP (vs non-PP) cells to the anterior endoderm, which is not possible to estimate from the indirect data currently provided. Note: while the present version of the manuscript does describe a sox17:CRE lineage tracing experiment, this actually goes in the opposite direction that would be informative (sox:17:CRE-marked descendants will be a mixture of PP-derived and non-PP derived cells, and the Gsc-based reporter does not allow for long-term tracking the fates of these cells).

We sincerely thank the reviewer for the professional comments and the constructive suggestions. As the reviewer indicated, utilizing the single-cell transcriptomic trajectory analyses on zebrafish embryos and Nodal-injected explants system, along with the live imaging analyses on Tg(gsc:EGFP;sox17:DsRed) embryos, we revealed that anterior endoderm progenitors arise from prechordal plate progenitors. To further verify this observation, we conducted two sets of lineage-tracing assays. Initial evidence came from the results of co-injecting sox17:Cre and gsc:loxp-STOP-loxp-mcherry plasmids. We observed RFP-positive cells at 8 hpf, demonstrating the presence of cells that had expressed both genes. To explicitly follow the proposed lineage, we then implemented a reciprocal strategy, as suggested by the reviewer, that constructed and co-injected sox17:loxp-STOP-loxp-mcherry and gsc:Cre plasmids. The appearance of RFP-positive cells in the anterior dorsal region at 8 hpf provides direct evidence for a transition from gsc-positive to sox17-positive identity. These results are now included in the revised manuscript (Please see Author response image 1 and Figure S4E). However, in accordance with the reviewer's caution, we acknowledge that this does not prove this is the sole origin of anterior endoderm. Consequently, we have revised the text to clarify that our findings demonstrate that anterior endoderm can be specified from prechordal plate progenitors, without claiming that it is the only source.

Author response image 1.

Characterization of anterior endoderm lineage by Cre-Lox recombination system.

(2) The authors' descriptions of gene expression patterns in the single-cell trajectory analyses do not always match the data. For example, it is stated that goosecoid expression marks progenitor cells that exist prior to a PP vs endo fate bifurcation (e.g. lines 124-130). Yet, in Figure 1C it appears that in fact goosecoid expression largely does not precede (but actually follows) the split and is predominantly expressed in cells that have already been specified into the PP branch. Likewise, most of the cells in the endo branch (or prior) appear to never express Gsc. While these trends do indeed appear to be more muddled in the explant data (Figure 1H), it still seems quite far-fetched to claim that Gsc expression is a hallmark of endoderm-PP progenitors.

We thank the reviewer for pointing out this issue. Our initial analysis proposed that the precursors of the prechordal plate (PP) and anterior endoderm (endo) more closely resemble a PP cell fate, as their progenitor populations highly express PP marker genes, such as gsc. The gsc gene is widely recognized as a PP marker[1]. The reviewer pointed out that in our analysis, these precursor cells do not initially exhibit high gsc expression; rather, gsc expression gradually increases as PP fate is specified.

The reason for this observation is as follows: First, for the in vivo data, we used the URD algorithm to trace back all possible progenitor cells for both the PP and anterior endo trajectory. As mentioned in the manuscript, the PP and anterior endo are relatively distant in the trajectory tree of the zebrafish embryonic data. Consequently, this approach likely included other, confounding progenitor cells that do not express gsc (like ventral epiblast, Author response image 2). However, we further investigated the expression of gsc and sox17 along these two trajectories. The conclusion remains that gsc expression is indeed higher than sox17 in the progenitor cells common to both trajectories (Author response image 2). Combined with the live imaging analysis presented in this study, which shows that gsc expression increases progressively in the PP, this supports the notion that the progenitor cells for both PP and anterior endoderm initially bias towards a PP cell fate.

On the other hand, in our previously published work using the Nodal-injected explant system, which specifically induces anterior endo and PP, the cellular trajectory analysis also revealed that the specifications of PP and anterior endo follow very similar paths. Therefore, we proceeded to analyze the Nodal explant data. Similarly, when using URD to trace the differentiation trajectories of PP and anterior endo cells, a small number of other progenitor cells were also captured. This explains why a minority of cells do not express gsc—these are likely ventral epiblast cells (Author response image 2). However, based on the Nodal explant data, gsc is specifically highly expressed in the progenitor cells of the PP and anterior endo. Its expression remains high in the PP trajectory but gradually decreases in the endoderm trajectory (Figure 1H).

Author response image 2.

(A) The expression of ventral epiblast markers in PP and anterior Endo URD trajectory. (B) The expression of gsc, sox32 and sox17 in the progenitors of PP and anterior endo in embryos and Nodal explants.

(3) The study seems to refer to "endoderm" and "anterior endoderm" somewhat interchangeably, and this is potentially problematic. Most single-cell-based analyses appearing in the study rely on global endoderm markers (sox17, sox32) which are expressed in endodermal precursors along the entire ventrolateral margin. Some of these cells are adjacent to the prechordal plate on the dorsal side of the gastrula, but many (most in fact) are quite some distance away. The microscopy-based evidence presented in Figure 2 and elsewhere, however, focuses on a small number of sox17-expressing cells that are directly adjacent to, or intermingled with, the prechordal plate. It, therefore, seems problematic for the authors to generalize potential overlaps with the PP lineage to the entire endoderm, which includes cells in ventral locations. It would be helpful if the authors could search for additional markers that might stratify and/or mark the anterior endoderm and perform their trajectory analysis specifically on these cells.

We thank the reviewer for these comments and suggestions. We fully agree with the reviewer's point that the expression of sox32 and sox17 cannot be used to distinguish dorsal endoderm from ventral-lateral endoderm cells. However, during the gastrulation stage, all endodermal cells express sox32 and sox17, and there are currently no specific marker genes available to distinguish between them.

After gastrulation ends, the dorsal endoderm (i.e., the anterior endoderm) begins to express pharyngeal endoderm marker genes, such as pax1b. Therefore, in the analysis of embryonic data in vivo, when studying the segregation of the anterior endoderm and PP trajectory, we specifically used the pharyngeal endoderm as the subject to trace its developmental trajectory.

In the case of Nodal explants, Nodal specifically induces the fate of the dorsal mesendoderm, which includes both the PP and pharyngeal endoderm (anterior endoderm). Precisely for this reason, we consider the Nodal explant system as a highly suitable model for investigating the mechanisms underlying the cell fate separation between anterior endoderm and PP. Thus, in the Nodal explant data, we included all endodermal cells for downstream analysis.

To avoid any potential confusion for readers, we have revised the term "endoderm" in the manuscript to "anterior endoderm" as suggested by the reviewer.

(4) It is not clear that the use of the nodal explant system is allowing for rigorous assessment of endoderm specification. Why are the numbers of endoderm cells so vanishingly few in the nodal explant experiments (Figure 1H, 3H), especially when compared to the embryo itself (e.g. Figures 1C-D)? It seems difficult to perform a rigorous analysis of endoderm specification using this particular model which seems inherently more biased towards PP vs. endoderm than the embryo itself. Why not simply perform nodal pathway manipulations in embryos?

We sincerely thank the reviewer for raising this important question. In our study of the fate separation between the PP and anterior endoderm, we initially analyzed zebrafish embryonic data. However, when reconstructing the transcriptional lineage tree using URD, we observed that these two cell trajectories were positioned relatively far apart on the tree. Yet, existing studies have shown that the anterior endoderm and PP are not only spatially adjacent but also both originate from mesendodermal progenitor cells[2-4], and they share transcriptional similarities[5]. Therefore, as the reviewer pointed out, when tracing all progenitor cells of these two trajectories using the URD algorithm, it is easy to include other cell types, such as ventral epiblast cells (Author response image 2). For this reason, we concluded that directly using embryonic data to dissect the mechanism of fate separation between PP and anterior endoderm might not yield highly accurate results.

In contrast, our group’s previous work, published in Cell Reports, demonstrated that the Nodal-induced explant system specifically enriches dorsal mesendodermal cells, including anterior endoderm, PP, and notochord[5]. Thus, we considered the Nodal explant system to be a highly suitable model for investigating the mechanism of fate separation between PP and anterior endoderm. Ultimately, by analyzing both in vivo embryonic data and Nodal explant data, we consistently found that the anterior endoderm likely originates from PP progenitor cells—a conclusion further validated by live imaging experiments.

Regarding the reviewer’s concern about the relatively low number of endodermal cells in the Nodal explant system, we speculate that this is because the explants predominantly induce anterior endoderm. Since endodermal cells constitute only a small proportion of cells during gastrulation, and anterior endoderm represents an even smaller subset, the absolute number is naturally limited. Nevertheless, the anterior endodermal cells captured in our Nodal explants were sufficient to support our analysis of the fate separation mechanism between anterior endoderm and PP. Finally, to further strengthen the findings from scRNA-seq analyses, we subsequently performed live imaging validation experiments using both zebrafish embryos and the explant system.

(5) The authors should not claim that proximity in UMAP space is an indication of transcriptional similarity (lines 207-208), especially for well-separated clusters. This is a serious misrepresentation of the proper usage of the UMAP algorithm. The authors make a similar claim later on (lines 272-274).

We would like to extend our gratitude to the reviewer for their insightful comments. We have revised the descriptions regarding UMAP throughout the manuscript as suggested (Please see the main text in revised manuscript).

Reviewer # 1 (Recommendations For The Authors):

- Pseudotime trajectories constructed from single-cell snapshots are not true "lineage" measurements. Authors should refrain from referring to such data as lineage data (e.g. lines 99, 100, 103, 109, 112, 127, etc). Such models should be referred to as "trajectories", "hypothetical lineages", or something else.

We are grateful to the reviewer for this comment. Following their recommendation, we have revised the terminology from "transcriptional lineage tree" to "trajectory" across the entire manuscript (Please see main text in revised manuscript).

- The live imaging data presented in Figure 2 (and supplemental figures) are compelling and do seem to show that some cells can switch between PP and endo states. However, the number of cells reported is still too low to be able to ascertain whether or not this is just a rare/edge-case phenomenon. Tracks for just a single cell are reported in Figure 2C-D. This is insufficient. Tracks for many more cells should be collected and reported alongside this current sole (n=1) example. The choice of time window for these live imaging experiments should also be better explained. These live imaging experiments are being performed at or after 6hpf, but authors claim in the text that "... the segregation between PP and Endo has already occurred by 6hpf." (lines 126-127). Why not perform these live imaging experiments earlier, when the initial fate decision between PP and endo is supposedly occurring?

We sincerely appreciate the reviewer’s insightful questions and constructive feedback. In response, we have made several important revisions. First, the reviewer noted that our original manuscript tracked only a single cell and suggested increasing the number of tracked cells. Following this recommendation, we repeated the live-imaging experiments and expanded the number of tracked endodermal cells (Please see the revised Movie S4 and Figure 2D). The experimental conditions were kept identical to the previous setup, and these cells consistently exhibited a gradual transition from a gsc+ fate to a sox17+ endodermal fate. In addition, the reviewer recommended performing live imaging at an earlier time point (Movie S5). Accordingly, we conducted additional experiments initiating live imaging at around 5.7 hours and observed the onset of a sox17 expression in gsc+ cells at approximately 6 hpf, which is consistent with our single-cell transcriptomic analysis.

- The sections devoted to lengthy descriptions of GO terms (lines 131-146, 239-254) and receptor-ligand predictions (lines 170-185) are largely speculative. Consider streamlining.

Thanks for the reviewer's comment. We have streamlined the content related to the GO analysis as suggested (Please see Lines 128-132, 157-167, 221-225).

- The use of a "Nodal Activity Score" (lines 212-226) is clever but might actually be less informative than showing contributions from individual nodal target genes. The combining of counts data from 29 predicted nodal targets means that the contribution (or lack of contribution) from each gene becomes masked. The authors should include supplementary dot plots that break down the score across all 29 genes, allowing the reader to assess overall contributions and/or sub-clusters of gene co-expression patterns, if present.

Thank you very much for the reviewer's positive feedback on our use of the "Nodal Activity Score" and the valuable suggestions provided. Following the recommendation, we analyzed the expression of the 29 Nodal direct targets used in our study across the WT, ndr1 knockdown (kd), and lft1 knockout (ko) groups. We found that the known axial mesoderm genes, such as chrd, tbxta, noto, and gsc, contributed significantly to the Nodal score. The newly conducted analysis has been included in the Supplementary Information (Please see Figure S7L).

- The differential expression trends being reported for srcap (line 251) do not appear to be significant. Are details and P-values for these DEG tests reported somewhere in the manuscript?

We thank the reviewer for raising this question. Based on the reviewer's comment, we performed statistical tests (Wilcoxon test) to compare the expression of srcap in PP and Endo. Our analysis revealed that while srcap expression is slightly higher in PP than in Endo, this difference is not statistically significant. The specific p-value and fold change have been indicated in the revised figure (Please see Figure 4J and S7H). Based on this analysis, we revised our description to state that srcap expression is slightly higher in the PP compared to in the anterior endoderm.

- Following the drug experiments with the drug AU15330 (lines 254-263), authors have only reported #s of endodermal cells, which seem to have increased, which the authors suggest indicates a fate switch from PP to endo. However, the authors have not reported whether the numbers of PP cells decreased or stayed the same in these embryos. This would be helpful information to include, as it is very difficult to discern quantitative trends from the images presented in Fig 4H and 4L.

Thank the reviewer for his/her comments and suggestions. Following the reviewer's suggestions, we performed Imaris analysis on the HCR staining results from the DMSO (control), 1μM AU15330-treated, and 5μM AU15330-treated groups. Our analysis focused on the number of frzb-positive cells (PP), and the comparison revealed that treatment with AU15330 significantly reduces the PP cell number. These findings have been incorporated into the revised manuscript and supplementary information (Please see Figures S7J and S7K).

Reviewer #2 (Public review):

Summary:

During vertebrate gastrulation, the mesoderm and endoderm arise from a common population of precursor cells and are specified by similar signaling events, raising questions as to how these two germ layers are distinguished. Here, Cheng and colleagues use zebrafish gastrulation as a model for mesoderm and endoderm segregation. By reanalyzing published single-cell sequencing data, they identify a common progenitor population for the anterior endoderm and the mesodermal prechordal plate (PP). They find that expression levels of PP genes Gsc and ripply are among the earliest differences between these populations and that their increased expression suppresses the expression of endoderm markers. Further analysis of chromatin accessibility and Ripply cut-and-tag is consistent with direct repression of endoderm by this PP marker. This study demonstrates the roles of Gsc and Ripply in suppressing anterior endoderm fate, but this role for Gsc was already known and the effect of Ripply is limited to a small population of anterior endoderm. The manuscript also focuses extensively on the function of Nodal in specifying and patterning the mesoderm and endoderm, a role that is already well known and to which the current analysis adds little new insight.

We would like to thank the reviewer #2 for the constructive comments and positive feedback regarding our manuscript.

Strengths:

Integrated single-cell ATAC- and RNA-seq convincingly demonstrate changes in chromatin accessibility that may underlie the segregation of mesoderm and endoderm lineages, including Gsc and ripply. Identification of Ripply-occupied genomic regions augments this analysis. The genetic mutants for both genes provide strong evidence for their function in anterior mesendoderm development, although these phenotypes are subtle.

We thank the reviewer for recognizing our work, and we greatly appreciate the constructive suggestions from the reviewer.

Weaknesses:

The use of zebrafish embryonic explants for cell fate trajectory analysis (rather than intact embryos) is not justified. In both transcriptomic comparisons between the two fate trajectories of interest and Ripply cut-and-tag analysis, the authors rely too heavily on gene ontology which adds little to our functional understanding. Much of the work is focused on the role of Nodal in the mesoderm/endoderm fate decision, but the results largely confirm previous studies and again provide few new insights. Some experiments were designed to test the relationship between the mesoderm and endoderm lineages and the role of epigenetic regulators therein, but these experiments were not properly controlled and therefore difficult to interpret.

We sincerely thank the reviewer for the comments. As we previously answered, in our study of the fate differentiation between the PP and the anterior endoderm, we initially analyzed zebrafish embryonic data. However, when we used URD to reconstruct the transcriptional trajectory tree, we found that these two cell trajectories were distantly located on the tree. Existing studies have shown that the anterior endoderm and the PP are not only spatially adjacent but also both originate from mesendodermal progenitor cells and share transcriptional similarities[2-4]. Therefore, when tracing all progenitor cells of these two trajectories using the URD algorithm, it is easy to include other cell types, such as ventral mesendodermal cells (Please see Author response image 2A). Based on this, we believe that directly using embryonic data to decipher the mechanism of fate differentiation between the PP and the anterior endoderm may not yield sufficiently precise results. In contrast, our group’s previous study published in Cell Reports demonstrated that the Nodal-induced explant system can specifically enrich dorsal mesendodermal cells, including the anterior endoderm, PP, and notochord[5]. Thus, we consider the Nodal explant system as an ideal model for studying the fate differentiation mechanism between the PP and the anterior endoderm. Ultimately, through comprehensive analysis of in vivo embryonic data and Nodal explant data, we consistently found that the anterior endoderm likely originates from PP progenitor cells—a conclusion further validated by live imaging experiments.

Regarding the GO analysis, we have streamlined it as suggested by the reviewers. In the revised manuscript, we analyzed the expression of specific genes contributing to key GO functions. Additionally, in the revised version, we conducted more live imaging experiments and quantitative cell assays. We designed gRNA for srcap using the CRISPR CAS13 system to knock down srcap, which further corroborated the morpholino knockdown results, showing consistency with the morpholino data. We also performed Western blot validation of the SWI/SNF complex's response to the drug AU15330, confirming the drug's effectiveness. We hope these additional experiments adequately address the reviewers' concerns.

Reviewer #2 (Recommendations For The Authors):

(1) In the introduction, the authors state that mesendoderm segregates into mesoderm and endoderm in a Nodal-concentration dependent manner. While it is true that higher Nodal signaling levels are required for endoderm specification, A) this is also true for some mesoderm populations, and B) Work from Caroline Hill's lab has shown that Nodal activity alone is not determinative of endoderm fate. Although the authors cite this work, it is conclusions are not reflected in this over-simplified explanation of mesendoderm development. The authors also state that it is not clear when PP and endoderm can be distinguished transcriptionally, but this was also addressed in Economou et al, 2022, which found that they can be distinguished at 60% epiboly but not 50% epiboly.

We sincerely thank the reviewer for raising this question and reminding us of the conclusions drawn from that excellent study. As the reviewer pointed out, Economou et al. demonstrated that Nodal signaling alone is insufficient to determine the cell fate segregation of mesendoderm[6]. However, their study primarily focused on the fate segregation of the ventral-lateral mesendoderm lineage. In contrast, we believe that the mechanisms underlying dorsal mesendoderm specification may differ.

First, it is well-studied that in zebrafish embryos, the most dorsal mesendoderm is initially specified by the activity of the dorsal organizer. Notably, the Nodal signaling ligands ndr1 and ndr2 begin to be expressed in the dorsal organizer as early as the sphere stage[7]. In our study, through single-cell transcriptomic trajectory analysis and live imaging analysis, we observed that the cell fate segregation of the dorsal mesendoderm can be traced back to the shield stage.

Second, the regulatory mechanisms governing dorsal mesendoderm fate differentiation may differ from those of the ventral-lateral mesendoderm. For instance, the gsc gene is exclusively expressed in the dorsal mesendoderm and is absent in the ventral-lateral mesendoderm. Given that gsc is a critical master gene, its overexpression in the ventral side can induce a complete secondary body axis. Similarly, ripply1, identified in our study, is also expressed early and specifically in the dorsal mesendoderm. Overexpression of ripply1 in the ventral side similarly induces a secondary body axis, albeit with the absence of the forebrain[5]. In this study, we found that gsc and ripply1 as the repressor, collectively inhibited dorsal (anterior) endoderm specified from PP progenitors.

In summary, our study focuses on the regulatory mechanisms of fate segregation in the dorsal (anterior) mesendoderm, which differs from the mechanisms of ventral-lateral mesendoderm lineage segregation reported by Economou et al. We believe that this distinction represents a key novelty of our work.

(2) As noted in the manuscript, Warga and Nusslein-Volhard determined long ago that PP and anterior endoderm share a common precursor. It is surprising that this close relationship is not apparent from the lineage trees in whole embryos but is apparent in lineage trees from explants. The authors speculate that the resolution of the whole embryo dataset is insufficient to detect this branch point and propose explants as the solution, but it is not clear why the explant dataset is higher resolution and/or more appropriate to address this question.

We sincerely thank the reviewer for their thoughtful comments. As we mentioned previously, our investigation of fate differentiation between the PP and the anterior endoderm initially involved the analysis of zebrafish embryonic data. However, when we used URD to reconstruct the transcriptional trajectory tree, we observed that these two cell trajectories were located far apart. Previous elegant studies, as the reviewer mentioned, have shown that the anterior endoderm and the PP are not only spatially adjacent but also both originate from mesendodermal progenitor cells and share transcriptional similarities[2,3,8]. Consequently, when tracing all progenitor cells of these two trajectories using the URD algorithm, other cell types—such as ventral mesendodermal cells—are easily included. Based on this, we believe that directly using embryonic data to elucidate the mechanism of fate differentiation between the PP and the anterior endoderm may lack sufficient precision.

In contrast, our group’s previous study published in Cell Reports demonstrated that the Nodal-induced explant system specifically enriches dorsal mesendodermal cells, including the anterior endoderm, PP, and notochord[5]. Therefore, we consider the Nodal explant system as an ideal model for studying the mechanism underlying fate differentiation between the PP and the anterior endoderm. Through comprehensive analyses of both in vivo embryonic and Nodal explant data, we consistently found that the anterior endoderm likely originates from PP progenitor cells—a conclusion further supported by live imaging experiments.

(3) Much of the analysis of DEGs between the lineages of interest is focused on GO term enrichment. But this logic is circular. The endoderm lineage is defined as such because it expresses endoderm-enriched genes, therefore the finding that the endoderm lineage is enriched for endoderm-related GO terms adds no new insights.

We thank the reviewer for these comments. As the reviewers suggested, in the revised manuscript, we indicated specific genes associated with key GO terms (Please see Figure 4B). Additionally, we have streamlined the content related to the GO analysis as suggested.

(4) The authors describe the experiment in Figure S4 as key evidence that Gsc+ cells can give rise to endoderm, but no controls are presented. Only a few cells are shown that express mCherry upon injection of sox17:cre constructs. Is mCherry also expressed in the occasional cell injected with Gsc:lox-stop-lox-mCherry in the absence of cre? Although they report 3 independent replicates, it appears that only 2 individual embryos express mCherry. This very small number is not convincing, especially in the absence of appropriate controls.

We thank the reviewer for raising this question. Following the reviewer's suggestion, we injected gsc:loxp-stop-loxp-mCherry into zebrafish embryos at the 1-cell stage as a control. After performing at least three independent replicates and analyzing no fewer than 100 embryos, we did not observe any mCherry-positive cells. Additionally, we co-injected gsc:loxp-stop-loxp-mCherry with sox17:cre and increased the sample size. Furthermore, we constructed plasmids of sox17:loxp-stop-loxp-mCherry and gsc:cre, and upon injection at the 1-cell stage, we observed RFP-positive cells at 8 hpf (Please see Author response image 1 and Figure S4E). Together with our live imaging data, these experiments collectively demonstrate that anterior endodermal cells can originate from PP progenitors.

(5) The authors spend a lot of effort demonstrating that PP and anterior endoderm are Nodal dependent. First, these data (especially Figures 3E and 3I) are not very convincing, as the differences shown are very small or not apparent. Second, this is already well-known and adds nothing to our understanding of mesoderm-endoderm segregation.

We sincerely thank the reviewer for their insightful questions. First, the reviewer mentioned that in the initial version of our manuscript, the effects of ndr1 knockdown and lefty1 knockout on Nodal signaling and cell fate—particularly prechordal plate (PP) and anterior endoderm (endo)—in Nodal-induced explants were not very pronounced. We recognize that the negative feedback mechanism between Nodal and Lefty signaling may explain why Nodal acts as a morphogen, regulating pattern formation through a Turing-like model[9]. Therefore, knocking down a Nodal ligand gene, such as ndr1 in this study, or knocking out a Nodal inhibitor, such as lft1, may only have a subtle impact on Nodal signaling[10].

Accordingly, in this study, we performed extensive pSmad2 immunofluorescence analysis and observed that although the overall intensity of Nodal activity did not change dramatically, there was a statistically significant difference. Importantly, this subtle variation in Nodal signaling strength is precisely what we intended to capture, since PP and anterior endoderm are highly sensitive to Nodal signaling[11], and even minor differences may bias their fate segregation.

This leads directly to the reviewer’s second concern. While numerous studies suggest that the strength of Nodal signaling influences mesendodermal fate—with high Nodal promoting endoderm and lower concentrations inducing mesoderm—most of these studies focus on ventral-lateral mesendoderm development[4,6,10]. In contrast, the mechanisms underlying dorsal mesendoderm fate specification differ, which is a key innovation of our study.

Previous work by Bernard Thisse and colleagues demonstrated that even a slight reduction in Nodal signaling, achieved by overexpressing a Nodal inhibitor, is sufficient to cause defects in the specification of PP and endoderm[11]. This indicates that PP and endoderm require the highest levels of Nodal signaling for proper specification. Moreover, the most dorsal mesendoderm, PP and anterior endoderm are not only spatially adjacent but also share similar transcriptional states, making the regulation of their fate separation particularly challenging to study.

The Dr. C.P. lab made important contributions to this issue, showing that the duration of Nodal exposure is critical for segregating PP and anterior endoderm fates: prolonged Nodal signaling promotes expression of the transcriptional repressor Gsc, which directly suppresses the key endodermal transcription factor Sox17, thereby inhibiting anterior endoderm specification[3]. They also found that tight junctions among PP cells facilitate Nodal signal propagation[8]. However, their studies revealed that Gsc mutants do not exhibit endodermal phenotypes, suggesting that additional factors or mechanisms regulate PP versus anterior endoderm fate separation[3].

In our study, we first observed that subtle differences in Nodal concentration may bias the fate choice between PP and anterior endoderm. Given that ndr1 knockdown and lft1 knockout mildly reduce or enhance Nodal signaling, respectively, we reasoned that using these two perturbations in a Nodal-induced explant system combined with single-cell RNA sequencing could generate transcriptomic profiles under slightly reduced and enhanced Nodal signaling. This approach may help identify key decision points and transcriptional differences during PP and anterior endoderm segregation, ultimately uncovering the molecular mechanisms downstream of Nodal that govern their fate separation.

(6) The authors claim that scrap expression differs between the 2 lineages of interest, but this is not apparent from Figure 4J-K. Experiments testing the role of SWI/SNF and scrap also require additional controls. Can scrap MO phenotypes be rescued by scrap RNA? Is there validation that SWI/SNF components are degraded upon treatment with AU15330?

We are very grateful for the reviewers' questions. Using single-cell data from zebrafish embryos and Nodal explants, we compared the expression of srcap in the PP and anterior Endo cell populations. We found that srcap expression showed a slight increase in PP compared to anterior Endo, but the difference was not statistically significant (Please see Figure 4J and S7H). Therefore, we modified our description in the revised manuscript. However, we speculate that this slight difference might influence the distinct cell fate specification between PP and anterior endo. In the original version of the manuscript, we reported that either treatment with AU15330, an inhibitor of the SWI/SNF complex, or injection of morpholino targeting srcap—a key component of the SWI/SNF complex—enhanced anterior endo fate while reducing PP cell specification. During this round of revision, we initially attempted to follow the reviewer’s suggestion to co-inject srcap mRNA along with srcap morpholino to rescue the phenotype. However, we found that the length of srcap mRNA exceeds 10,000 bp, and despite multiple attempts, we were unable to successfully obtain the srcap mRNA. Therefore, we were unable to perform the rescue experiment and instead adopted an alternative approach to validate the function of srcap. We aimed to use anthor knockdown approach (CRISPR/Cas system) to determine whether a phenotype similar to that observed with morpholino knockdown could be achieved. Using the CRISPR/Cas13 system, we designed gRNA targeting srcap, knocked down srcap, and examined the cell specification of PP and anterior endo. We found that, consistent with our previous results, knocking down srcap obviously reduced PP cell fate while increasing anterior endo cell fate (Author response image 3). Additionally, the reviewer raised the question of whether the SWI/SNF complex is degraded after AU15330 treatment. Following the reviewer’s suggestion, we attempted to perform Western blot analysis on BRG1, one of the components of the SWI/SNF complex. However, despite multiple attempts, we were unable to achieve successful detection of the BRG1 protein by the antibody in zebrafish. Several studies have reported that knockdown or knockout of brg1 leads to defects in neural crest cell specification in zebrafish[12,13]. Therefore, alternatively, we treated zebrafish embryos at the one-cell stage with 0 μM (DMSO), 1 μM, and 5 μM AU15330, and examined the expression of sox10 and pigment development around 48 h. We found that treatment with 1 μM AU15330 reduced sox10 expression and pigment production, though not significantly, whereas treatment with 5 μM AU15330 significantly disrupted neural crest cell development. Thus, this experiment demonstrates that AU15330 is functional in zebrafish. (Author response image 3).

Author response image 3.

(A) Characterization of anterior endoderm and PP cells following CRISPR-Cas13d-mediated srcap knockdown. (B) Validation of srcap mRNA expression by RT‑qPCR following CRISPR‑Cas13d knockdown. (C) RT‑qPCR shows the expression of sox10 after treatment with increasing concentrations of AU15300. (D) Morphology of zebrafish embryos at 48 hpf after treatment with increasing concentrations of AU15300.

(7) The authors conclude from their chromatin accessibility analysis that variations in Nodal signaling are responsible for expression levels of PP and endoderm genes, but they do not consider the alternative explanation that FGF signaling is playing this role. Such a function for FGF was established by Caroline Hill's lab, and the authors also show in Figure S5G that FGF signaling in enriched between these cell populations.

Thank you very much for raising this issue. As the reviewer pointed out, Caroline Hill's lab has conducted elegant work demonstrating that FGF signaling plays a crucial role in the separation of ventral-lateral mesendoderm cell fates[4,6]. In contrast, our study primarily focuses on studying the mechanisms underlying the separation of dorsal mesendoderm cell fates. However, our research also reveals that FGF signaling significantly regulates the fate separation of the dorsal mesendoderm, as inhibiting FGF signaling suppresses PP cell specification while promoting anterior Endo fate. In our previously published work, we found that Nodal signaling can directly activate the expression of FGF ligand genes[5]. Therefore, we hypothesize that Nodal signaling, acting as a master regulator, activates various downstream target genes—including FGF—and how FGF signaling regulates the cell fate separation of the dorsal mesendoderm warrants further investigation in our further studies.

(8) When interpreting the results of their Ripply cut-and-run experiment, the authors again rely heavily on GO term analysis and claim that this supports a role for Ripply as a transcriptional repressor. GO term enrichment does not equal functional analysis. It would be more convincing to intersect DEGs between WT and ripply-/- embryos with Ripply-enriched loci.

Thanks for raising this important issue and the constructive suggestion. In response to the reviewer's valid concern regarding the GO term analyses from our CUT&Tag data, we implemented a more stringent filtering strategy. We identified peaks enriched in the treatment group and applied differential analysis, selecting genes with a log₂FoldChange > 3, padj < 0.05, and baseMean > 30 as high-confidence Ripply1 binding targets. A GO enrichment analysis of these genes revealed significant terms related to muscle development, consistent with Ripply1's established role in somite development, thereby validating our approach. We supplemented the related gene list in the revised manuscript. Moreover, within this refined analysis, we found that sox32 met our binding threshold, while sox17 did not. Furthermore, as suggested, we examined mespbb—a known Ripply1-repressed gene—which was present, and gsc, a Nodal target used as a negative control, which was absent. This confirms the specificity of our analysis (Figure 6 and Figure S11). Consequently, our revised analyses support a model in which Ripply1 directly binds the sox32 promoter. Given that Sox32 is a known upstream regulator of sox17, this binding provides a plausible direct mechanism for the observed regulation of sox17 expression. We have updated the figures and text accordingly. We attempted to generate ripply1^-/- mutants but found that homozygous loss results in embryonic lethality.

(9) The way N's are reported is unconventional. N= number of embryos used in the experiment, n= number of embryos imaged. If an embryo was not imaged or analyzed in any way, it cannot be considered among the embryos in an experiment. If only 4 embryos were imaged, the N for that experiment is 4 regardless of how many embryos were stained. Authors should also report not only the number of embryos examined but also the number of independent trials performed for all experiments.

Thank you very much for the reviewer's suggestion. As suggested, we have revised the description regarding the number of embryos and experimental replicates in the figure legends.

(10) The authors should avoid the use of red-green color schemes in figures to ensure accessibility for color-blind readers.

Thanks for the suggestions. We have updated the figures in our revised manuscript and adjusted the color schemes to avoid red-green combinations.

Reviewer #3 (Public Review):

Summary:

Cheng, Liu, Dong, et al. demonstrate that anterior endoderm cells can arise from prechordal plate progenitors, which is suggested by pseudo time reanalysis of published scRNAseq data, pseudo time analysis of new scRNAseq data generated from Nodal-stimulated explants, live imaging from sox17:DsRed and Gsc:eGFP transgenics, fluorescent in situ hybridization, and a Cre/Lox system. Early fate mapping studies already suggested that progenitors at the dorsal margin give rise to both of these cell types (Warga) and live imaging from the Heisenberg lab (Sako 2016, Barone 2017) also pretty convincingly showed this. However, the data presented for this point are very nice, and the additional experiments in this manuscript, however, further cement this result. Though better demonstrated by previous work (Alexander 1999, Gritsman 1999, Gritsman 2000, Sako 2016, Rogers 2017, others), the manuscript suggests that high Nodal signaling is required for both cell types, and shows preliminary data that suggests that FGF signaling may also be important in their segregation. The manuscript also presents new single-cell RNAseq data from Nodal-stimulated explants with increased (lft1 KO) or decreased (ndr1 KD) Nodal signaling and multi-omic ATAC+scRNAseq data from wild-type 6 hpf embryos but draws relatively few conclusions from these data. Lastly, the manuscript presents data that SWI/SNF remodelers and Ripply1 may be involved in the anterior endoderm - prechordal plate decision, but these data are less convincing. The SWI/SNF remodeler experiments are unconvincing because the demonstration that these factors are differentially expressed or active between the two cell types is weak. The Ripply1 gain-of-function experiments are unconvincing because they are based on incredibly high overexpression of ripply1 (500 pg or 1000 pg) that generates a phenotype that is not in line with previously demonstrated overexpression studies (with phenotypes from 10-20x lower expression). Similarly, the cut-and-tag data seems low quality and like it doesn't support direct binding of ripply1 to these loci.

In the end, this study provides new details that are likely important in the cell fate decision between the prechordal plate and anterior endoderm; however, it is unclear how Nodal signaling, FGF signaling, and elements of the gene regulatory network (including Gsc, possibly ripply1, and other factors) interact to make the decision. I suggest that this manuscript is of most interest to Nodal signaling or zebrafish germ layer patterning afficionados. While it provides new datasets and observations, it does not weave these into a convincing story to provide a major advance in our understanding of the specification of these cell types.

We sincerely thank the reviewer for their thorough and thoughtful assessment of our work. The reviewer acknowledged several strengths of our study, such as the use of multiple technical approaches to demonstrate that anterior endoderm differentiates from PP progenitor cells, and recognized the value of the newly added single-cell omics data. The reviewer also raised some concerns regarding the initial version of our work, including the SWI/SNF remodeler experiments and the Ripply1 gain-of-function experiment. In the revised manuscript, we have supplemented these parts with additional control experiments to better support our conclusions. We hope that our updated manuscript adequately addresses the points raised by the reviewer.

Major issues:

(1) UMAPs: There are several instances in the manuscript where UMAPs are used incorrectly as support for statements about how transcriptionally similar two populations are. UMAP is a stochastic, non-linear projection for visualization - distances in UMAP cannot be used to determine how transcriptionally similar or dissimilar two groups are. In order to make conclusions about how transcriptionally similar two populations are requires performing calculations either in the gene expression space, or in a linear dimensional reduction space (e.g. PCA, keeping in mind that this will only consider the subset of genes used as input into the PCA). Please correct or remove these instances, which include (but are not limited to):

p.4 107-110

p.4 112

p.8 207-208

p.10 273-275

We would like to thank the reviewer for raising this question. The descriptions of UMAP have been revised throughout the manuscript in accordance with the reviewer's suggestion (Please see the main text in the revised manuscript).

(2) Nodal and lefty manipulations: The section "Nodal-Lefty regulatory loop is needed for PP and anterior Endo fate specification" and Figure 3 do not draw any significant conclusions. This section presents a LIANA analysis to determine the signals that might be important between prechordal plate and endoderm, but despite the fact that it suggests that BMP, Nodal, FGF, and Wnt signaling might be important, the manuscript just concludes that Nodal signaling is important. Perhaps this is because the conclusion that Nodal signaling is required for the specification of these cell types has been demonstrated in zebrafish in several other studies with more convincing experiments (Alexander 1999, Gritsman 1999, Gritsman 2000, Rogers 2017, Sako 2016). While FGF has recently been demonstrated to be a key player in the stochastic decision to adopt endodermal fate in lateral endoderm (Economou 2022), the idea that FGF signaling may be a key player in the differentiation of these two cell types has strangely been relegated to the discussion and supplement. Lastly, the manuscript does not make clear the advantage of performing experiments to explore the PP-Endo decision in Nodal-stimulated explants compared to data from intact embryos. What would be learned from this and not from an embryo? Since Nodal signaling stimulates the expression of Wnts and FGFs, these data do not test Nodal signaling independent of the other pathways. It is unclear why this artificial system that has some disadvantages is used since the manuscript does not make clear any advantages that it might have had.

We sincerely thank the reviewers for their valuable comments. As mentioned in our manuscript, although a substantial number of studies have reported on the mechanisms governing the segregation of mesendoderm fate in zebrafish embryos—including the Dr. Hill laboratory’s work cited by the reviewers, which demonstrated the involvement of FGF signaling in the ventral mesendoderm fate specification—research on the regulatory mechanisms underlying anterior mesendoderm differentiation remains relatively limited. This is largely due to the challenges posed by the close physical proximity and similar transcriptional states of anterior mesendoderm cells, as well as their shared dependence on high levels of Nodal signaling for specification.

Several studies from the Dr. C.P. Heisenberg’s laboratory have attempted to elucidate the fate segregation between anterior mesendoderm cells, namely the prechordal plate (PP) and anterior endoderm (endo) cells. They found that PP cells are tightly connected, facilitating the propagation of Nodal signaling[8]. Prolonged exposure to Nodal activates the expression of Gsc, which acts as a transcriptional repressor to inhibit sox17 expression, thereby suppressing endodermal fate[3]. However, they also noted that Gsc mutants do not exhibit endoderm developmental defects, suggesting the involvement of additional factors in this process.

The reviewer inquired about our rationale for using the Nodal-injected explant system. In our investigation of the fate separation between the PP and the anterior endo, we initially analyzed zebrafish embryonic data. Using URD to reconstruct the transcriptional lineage tree, we found that these two cell types were positioned distantly from each other. However, existing literature indicates that the anterior endoderm and PP are not only spatially adjacent but also derive from common mesendodermal progenitors and exhibit transcriptional similarities[2,8]. As the reviewer noted, when tracing all progenitor cells of these two lineages using URD, it is easy to inadvertently include other cell types—such as ventral epiblast cells—which may compromise the accuracy of the analysis. We therefore concluded that directly using embryonic data to dissect the mechanism of fate separation between PP and anterior endoderm might not yield highly precise results.

By contrast, our group’s earlier study published in Cell Reports demonstrated that the Nodal-induced explant system specifically enriches dorsal mesendodermal cells, including anterior endo, PP, and notochord[5]. This makes the Nodal explant system a highly suitable model for studying the fate separation between PP and anterior endo. Ultimately, by analysing in vivo embryonic data and Nodal explant data, we consistently found that the anterior endoderm likely originates from PP progenitors—a conclusion further supported by live imaging experiments.

As we answered above, we first used the analyses of single-cell RNA sequencing and live imaging to demonstrate that anterior endoderm can originate from PP progenitor cells. Understanding the mechanism underlying the fate segregation between these two cell populations became a key focus of our research. We began by applying cell communication analysis to our single-cell data to identify signaling pathways that may be involved. This analysis specifically highlighted the Nodal-Lefty signaling pathway. Since Lefty acts as an inhibitor of Nodal signaling, we hypothesized that differences in Nodal signaling strength might regulate the fate of these two cell populations. By overexpressing different concentrations of Nodal mRNA and examining the fates of PP and anterior Endo cells, we confirmed this hypothesis.

Thus, we propose that even subtle differences in Nodal signaling levels may influence anterior mesendoderm fate decisions. To test this, we generated systems with slightly reduced Nodal signaling (via ndr1 knockdown) and slightly elevated Nodal signaling (via lft1 knockout). Using these models, we precisely captured the critical stage of fate segregation between PP and anterior endo cells and identified a novel transcriptional repressor, Ripply1, which works in concert with Gsc to suppress anterior endoderm differentiation.

(3) ripply1 mRNA injection phenotype inconsistent with previous literature: The phenotype presented in this manuscript from overexpressing ripply1 mRNA (Fig S11) is inconsistent with previous observations. This study shows a much more dramatic phenotype, suggesting that the overexpression may be to a non-physiological level that makes it difficult to interpret the gain-of-function experiments. For instance, Kawamura et al 2005 perform this experiment but do not trigger loss of head and eye structures or loss of tail structures. Similarly, Kawamura et al 2008 repeat the experiment, triggering a mildly more dramatic shortening of the tail and complete removal of the notochord, but again no disturbance of head structures as displayed here. These previous studies injected 25 - 100 pg of ripply1 mRNA with dramatic phenotypes, whereas this study uses 500 - 1000 pg. The phenotype is so much more dramatic than previously presented that it suggests that the level of ripply1 overexpression is sufficiently high that it may no longer be regulating only its endogenous targets, making the results drawn from ripply1 overexpression difficult to trust.

We sincerely thank the reviewer for raising this question. First, we apologize for not providing a detailed description of the amount of HA-ripply1 mRNA injected in our previous manuscript. We injected 500 pg of HA-ripply1 mRNA at the 1-cell stage and allowed the embryos to develop until 6 hpf for the CUT&Tag experiment. In the supplementary materials, we included a bright-field image of an 18 hpf-embryo injected with HA-ripply1 mRNA, which morphologically exhibited severe developmental abnormalities. The reviewer pointed out that the amount of ripply1 mRNA we injected might be excessive, potentially leading to non-specific gain-of-function effects. The injection dose of 500 pg was determined based on conclusions from our previous study. In that study, injecting 24 pg of ripply1 mRNA into one cell of zebrafish embryos at the 16–32 cell stage was sufficient to induce a secondary axis lacking the forebrain[5]. From this, we estimated that an injection concentration of approximately 500–1000 pg would be appropriate at the 1-cell stage, so that after several rounds of cell division, each cell gained 20-30 pg mRNA at 32 cell stage. Additionally, we conducted supplementary experiments injecting 100 pg, 250 pg, and 500 pg of ripply1 mRNA, and observed 500 pg of ripply1 mRNA led to a dramatic suppression of endoderm formation (Author response image 4).

Finally, our study focuses on the mechanism of cell fate segregation in the anterior mesendoderm, primarily during gastrulation. The embryos injected with ripply1 mRNA underwent normal gastrulation, and our CUT&Tag experiment was performed at 6 hpf. Therefore, we believe that the amount of ripply1 mRNA injected in this study is appropriate for addressing our research question.

Author response image 4.

Different concentrations of ripply1 mRNA were injected into zebrafish embryos at the one-cell stage, with RFP fluorescence labeling sox17-positive cells.

(4) Ripply1 binding to sox17 and sox32 regulatory regions not convincing: The Cut and Tag data presented in Fig 6J-K does not seem to be high quality and does not seem to provide strong support that Ripply 1 binds to the regulatory regions of these genes. The signal-to-noise ratio is very poor, and the 'binding' near sox17 that is identified seems to be even coverage over a 14 kb region, which is not consistent with site-specific recruitment of this factor, and the 'peaks' highlighted with yellow boxes do not appear to be peaks at all. To me, it seems this probably represents either: (1) overtagmentation of these samples or (2) an overexpression artifact from injection of too high concentration of ripply1-HA mRNA. In general, Cut and Tag is only recommended for histone modifications, and Cut and Run would be recommended for transcriptional regulators like these (see Epicypher's literature). Given this and the previous point about Ripply1 overexpression, I am not convinced that Ripply1 regulates endodermal genes. The existing data could be made somewhat more convincing by showing the tracks for other genes as positive and negative controls, given that Ripply1 has known muscle targets (how does its binding look at those targets in comparison) and there should be a number of Nodal target genes that Ripply1 does not bind to that could be used as negative controls. Overall this experiment doesn't seem to be of high enough quality to drive the conclusion that Ripply1 directly binds near sox17 and sox32 and from the data presented in the manuscript looks as if it failed technically.

We sincerely thank the reviewer for raising this question. We apologize that the binding regions of sox17 marked in our previous analysis were incorrect, and we have made the corresponding revisions in the latest version of the manuscript.

The reviewer noted that our CUT&Tag data contain considerable noise. To address this, we further refined our data processing: we annotated all peaks enriched in the treatment group and performed differential analysis, selecting genes with log₂FoldChange > 3, padj < 0.5, and baseMean > 30 as candidate targets of Ripply1 binding. Subsequent GO enrichment analysis of these genes revealed significant enrichment of muscle development-related GO terms, which is consistent with previously reported roles of Ripply1 in regulating somite development. Therefore, we believe our filtering method effectively removes a large number of noise peaks and their associated genes.

Under these screening criteria, we found that sox32 meets the threshold, while sox17 does not. In addition, following the reviewer’s suggestion, we examined mespbb—a known gene repressed by Ripply1—and gsc, a Nodal target gene, as a negative control.

Based on these new analyses, we have revised our figures and text accordingly. Our data now support the possibility that Ripply1 may directly bind to the promoter region of sox32. Since sox32 acts as a direct upstream regulator of sox17, this binding could influence sox17 expression (Figure 6 and Figure S11).

Finally, we would like to note that studies have reported Ripply1 as a transcriptional repressor, which may function by recruiting other co-factors, such as Groucho, to form a complex[14,15]. This might explain why our CUT&Tag data detected Ripply1 binding to a broad set of genes.

(5) "Cooperatively Gsc and ripply1 regulate": I suggest avoiding the term "cooperative," when describing the relationship between Ripply1 and Gsc regulation of PP and anterior endoderm - it evokes the concept of cooperative gene regulation, which implies that these factors interact with each biochemically in order to bind to the DNA. This is not supported by the data in this manuscript, and is especially confusing since Ripply1 is thought to require cooperative binding with a T-box family transcription factor to direct its binding to the DNA.

We sincerely thank the reviewer for raising this important issue. The reviewer pointed out that the term "Cooperatively" may not be entirely appropriate in the context of our study. In accordance with the reviewer's suggestion, we have replaced "Cooperatively" with "Collectively" in the relevant sections.

(6) SWI/SNF: The differential expression of srcap doesn't seem very remarkable. The dot plots in the supplement S7H don't help - they seem to show no expression at all in the endoderm, which is clearly a distortion of the data, since from the violin plots it's obviously expressed and the dot-size scale only ranges from ~30-38%. Please add to the figure information about fold-change and p-value for the differential expression. Publicly available scRNAseq databases show scrap is expressed throughout the entire early embryo, suggesting that it would be surprising for it to have differential activity in these two cell types and thereby contribute to their separate specification during development. It seems equally possible that this just mildly influences the level of Nodal or FGF signaling, which would create this effect.

Thank the Reviewer for this question. As suggested, we performed Wilcoxon tests to compare srcap expression between PP and Endo populations. The analysis shows that while srcap expression is moderately elevated in PP compared to in Endo, this difference is not statistically significant. The corresponding p-value and fold change have now been included in the revised figure (Please see Figure 4J and S7H). Although the transcriptional level of srcap shows no significant difference between PP and anterior endoderm, our subsequent experiments—using AU15330 (an inhibitor of the SWI/SNF complex) and injecting morpholino targeting srcap, a key component of the SWI/SNF complex—demonstrated that its inhibition indeed promotes anterior endoderm fate while reducing PP cell specification. Therefore, we propose that subtle differences in the SWI/SNF complex may regulate the fate specification of PP and anterior endoderm through two mechanisms. First, as mentioned in our study, these chromatin remodelers modulate the expression of master regulators such as Gsc and Ripply1, thereby influencing cell fate decisions. Second, as noted by the reviewer, these chromatin remodelers may affect the interpretation of Nodal signaling, ultimately contributing to the divergence between PP and anterior endoderm fates.

The multiome data seems like a valuable data set for researchers interested in this stage of zebrafish development. However, the presentation of the data doesn't make many conclusions, aside from identifying an element adjacent to ripply1 whose chromatin is open in prechordal plate cells and not endodermal cells and showing that there are a number of loci with differential accessibility between these cell types. That seems fairly expected since both cell types have several differentially expressed transcriptional regulators (for instance, ripply1 has previously been demonstrated in multiple studies to be specific to the prechordal plate during blastula stages). The manuscript implies that SWI/SNF remodeling by Srcap is responsible for the chromatin accessibility differences between these cell types, but that has not actually been tested. It seems more likely that the differences in chromatin accessibility observed are a result of transcription factors binding downstream of Nodal signaling.

We thank the reviewer for recognizing the value of our newly generated data. Through integrative analysis of single-cell data from wild-type, ndr1 kd, and lft1 ko groups of Nodal-injected explants at 6 hours post-fertilization (hpf), we identified a critical branching point in the fate segregation of the prechordal plate (PP) and anterior endoderm (Endo), where chromatin remodelers may play a significant role. Based on this finding, we performed single-cell RNA and ATAC sequencing on zebrafish embryos at 6 hpf. Analysis of this multi-omics dataset revealed that transcriptional repressors such as Gsc, Ripply1, and Osr1 exhibit differences in both transcriptional and chromatin accessibility levels between the PP and anterior Endo. Subsequent overexpression and loss-of-function experiments further demonstrated that Gsc and Ripply1 collaboratively suppress endodermal gene expression, thereby inhibiting endodermal cell fate. Previous studies have reported that for the activation of certain Nodal downstream target genes, the pSMAD2 protein of the Nodal signaling pathway recruits chromatin remodelers to facilitate chromatin opening and promote further transcription of target genes[16]. Therefore, our data provide chromatin accessibility profiles for Gsc and Ripply1, offering a valuable resource for future investigations into their pSMAD2 binding sites.

Minor issues:

Figure 2 E-F: It's not clear which cells from E are quantitated in F. For instance, the dorsal forerunner cells are likely to behave very differently from other endodermal progenitors in this assay. It would be helpful to indicate which cells are analyzed in Fig F with an outline or other indicator of some kind. Or - if both DFCs and endodermal cells are included in F, to perhaps use different colors for their points to help indicate if their fluorescence changes differently.

Thank you for the reviewer's suggestion. In the revised version of the figure, we have outlined the regions of the analyzed cells.

Fig 3 J: Should the reference be Dubrulle et al 2015, rather than Julien et al?

Thanks, we have corrected.

References:

Alexander, J. & Stainier, D. Y. A molecular pathway leading to endoderm formation in zebrafish. Current biology : CB 9, 1147-1157 (1999).

Barone, V. et al. An Effective Feedback Loop between Cell-Cell Contact Duration and Morphogen Signaling Determines Cell Fate. Dev. Cell 43, 198-211.e12 (2017).

Economou, A. D., Guglielmi, L., East, P. & Hill, C. S. Nodal signaling establishes a competency window for stochastic cell fate switching. Dev. Cell 57, 2604-2622.e5 (2022).

Gritsman, K. et al. The EGF-CFC protein one-eyed pinhead is essential for nodal signaling. Cell 97, 121-132 (1999).

Gritsman, K., Talbot, W. S. & Schier, A. F. Nodal signaling patterns the organizer. Development (Cambridge, England) 127, 921-932 (2000).

Kawamura, A. et al. Groucho-associated transcriptional repressor ripply1 is required for proper transition from the presomitic mesoderm to somites. Developmental cell 9, 735-744 (2005).

Kawamura, A., Koshida, S. & Takada, S. Activator-to-repressor conversion of T-box transcription factors by the Ripply family of Groucho/TLE-associated mediators. Molecular and cellular biology 28, 3236-3244 (2008).

Sako, K. et al. Optogenetic Control of Nodal Signaling Reveals a Temporal Pattern of Nodal Signaling Regulating Cell Fate Specification during Gastrulation. Cell Rep. 16, 866-877 (2016).

Rogers, K. W. et al. Nodal patterning without Lefty inhibitory feedback is functional but fragile. eLife 6, e28785 (2017).

Warga, R. M. & Nüsslein-Volhard, C. Origin and development of the zebrafish endoderm. Development 126, 827-838 (1999).

References:

(1) Steinbeisser, H., and De Robertis, E.M. (1993). Xenopus goosecoid: a gene expressed in the prechordal plate that has dorsalizing activity. C R Acad Sci III 316, 959-971.

(2) Warga, R.M., and Nusslein-Volhard, C. (1999). Origin and development of the zebrafish endoderm. Development (Cambridge, England) 126, 827-838. 10.1242/dev.126.4.827.

(3) Sako, K., Pradhan, S.J., Barone, V., Inglés-Prieto, Á., Müller, P., Ruprecht, V., Čapek, D., Galande, S., Janovjak, H., and Heisenberg, C.P. (2016). Optogenetic Control of Nodal Signaling Reveals a Temporal Pattern of Nodal Signaling Regulating Cell Fate Specification during Gastrulation. Cell reports 16, 866-877. 10.1016/j.celrep.2016.06.036.

(4) van Boxtel, A.L., Economou, A.D., Heliot, C., and Hill, C.S. (2018). Long-Range Signaling Activation and Local Inhibition Separate the Mesoderm and Endoderm Lineages. Developmental cell 44, 179-191.e175. 10.1016/j.devcel.2017.11.021.

(5) Cheng, T., Xing, Y.Y., Liu, C., Li, Y.F., Huang, Y., Liu, X., Zhang, Y.J., Zhao, G.Q., Dong, Y., Fu, X.X., et al. (2023). Nodal coordinates the anterior-posterior patterning of germ layers and induces head formation in zebrafish explants. Cell reports 42, 112351. 10.1016/j.celrep.2023.112351.

(6) Economou, A.D., Guglielmi, L., East, P., and Hill, C.S. (2022). Nodal signaling establishes a competency window for stochastic cell fate switching. Developmental cell 57, 2604-2622 e2605. 10.1016/j.devcel.2022.11.008.

(7) Schier, A.F., and Talbot, W.S. (2005). Molecular genetics of axis formation in zebrafish. Annual review of genetics 39, 561-613. 10.1146/annurev.genet.37.110801.143752.

(8) Barone, V., Lang, M., Krens, S.F.G., Pradhan, S.J., Shamipour, S., Sako, K., Sikora, M., Guet, C.C., and Heisenberg, C.P. (2017). An Effective Feedback Loop between Cell-Cell Contact Duration and Morphogen Signaling Determines Cell Fate. Developmental cell 43, 198-211.e112. 10.1016/j.devcel.2017.09.014.

(9) Muller, P., Rogers, K.W., Jordan, B.M., Lee, J.S., Robson, D., Ramanathan, S., and Schier, A.F. (2012). Differential diffusivity of Nodal and Lefty underlies a reaction-diffusion patterning system. Science (New York, N.Y.) 336, 721-724. 10.1126/science.1221920.

(10) Rogers, K.W., Lord, N.D., Gagnon, J.A., Pauli, A., Zimmerman, S., Aksel, D.C., Reyon, D., Tsai, S.Q., Joung, J.K., and Schier, A.F. (2017). Nodal patterning without Lefty inhibitory feedback is functional but fragile. eLife 6. 10.7554/eLife.28785.

(11) Thisse, B., Wright, C.V., and Thisse, C. (2000). Activin- and Nodal-related factors control antero-posterior patterning of the zebrafish embryo. Nature 403, 425-428. 10.1038/35000200.

(12) Eroglu, B., Wang, G., Tu, N., Sun, X., and Mivechi, N.F. (2006). Critical role of Brg1 member of the SWI/SNF chromatin remodeling complex during neurogenesis and neural crest induction in zebrafish. Developmental dynamics : an official publication of the American Association of Anatomists 235, 2722-2735. 10.1002/dvdy.20911.

(13) Hensley, M.R., Emran, F., Bonilla, S., Zhang, L., Zhong, W., Grosu, P., Dowling, J.E., and Leung, Y.F. (2011). Cellular expression of Smarca4 (Brg1)-regulated genes in zebrafish retinas. BMC developmental biology 11, 45. 10.1186/1471-213X-11-45.

(14) Kawamura, A., Koshida, S., Hijikata, H., Ohbayashi, A., Kondoh, H., and Takada, S. (2005). Groucho-associated transcriptional repressor ripply1 is required for proper transition from the presomitic mesoderm to somites. Developmental cell 9, 735-744. 10.1016/j.devcel.2005.09.021.

(15) Kawamura, A., Koshida, S., and Takada, S. (2008). Activator-to-repressor conversion of T-box transcription factors by the Ripply family of Groucho/TLE-associated mediators. Mol Cell Biol 28, 3236-3244. 10.1128/MCB.01754-07.

(16) Ross, S., Cheung, E., Petrakis, T.G., Howell, M., Kraus, W.L., and Hill, C.S. (2006). Smads orchestrate specific histone modifications and chromatin remodeling to activate transcription. EMBO J 25, 4490-4502. 10.1038/sj.emboj.7601332.

eLife Assessment

This important study investigates how the Cdv division proteins of Metallosphaera sedula assemble on and interact with curved membranes in vitro, advancing our understanding of this reduced ESCRT-like machinery. The data provide support for sequential protein recruitment and curvature-dependent enrichment at membrane necks, based on well-controlled reconstitution assays and quantitative analysis. The work establishes a convincing experimental framework for dissecting Cdv-mediated membrane remodeling. The study will be of broad interest to evolutionary and synthetic biologists as well as membrane biophysicists.

Reviewer #1 (Public review):

Summary:

The authors aimed to elucidate the recruitment order and assembly of the Cdv proteins during Sulfolobus acidocaldarius archaeal cell division using a bottom-up reconstitution approach. They employed liposome-binding assays, EM, and fluorescence microscopy with in vitro reconstitution in dumbbell-shaped liposomes to explore how CdvA, CdvB, and the homologues of ESCRT-III proteins (CdvB, CdvB1, and CdvB2) interact to form membrane remodeling complexes.<br /> The study sought to reconstitute the Cdv machinery by first analyzing their assembly as two sub-complexes: CdvA:CdvB and CdvB1:CdvB2ΔC. The authors report that CdvA binds lipid membranes only in the presence of CdvB and localizes preferentially to membrane necks. Similarly, the findings on CdvB1:CdvB2ΔC indicate that truncation of CdvB2 facilitates filament formation and enhances curvature sensitivity in interaction with CdvB1. Finally, the authors reconstitute a quaternary CdvA:CdvB:CdvB1:CdvB2 complex and demonstrate its enrichment at membrane necks. The mechanistic details of how these complexes drive membrane remodeling, particularly through subcomplex removal by the proteasome and/or CdvC, remain insufficiently addressed, and the study therefore mainly provides an experimental framework for future mechanistic investigation.

Strengths:

The study of machinery assembly and its involvement in membrane remodeling, particularly using bottom-up reconstituted in vitro systems, presents significant challenges. This is particularly true for systems like the ESCRT-III complex, which localizes uniquely at the lumen of membrane necks prior to scission. The use of dumbbell-shaped liposomes in this study provides a promising experimental model to investigate ESCRT-III and ESCRT-III-like protein activity at membrane necks.<br /> The authors present intriguing evidence regarding the sequential recruitment of ESCRT-III proteins in crenarchaea-a close relative of eukaryotes.

Weaknesses:

The findings of this study suggest that the hierarchical recruitment characteristic of eukaryotic systems may predate eukaryogenesis, which represents a significant and exciting contribution. However, the broader implications of these findings for membrane remodeling mechanisms remain largely unexplored. Nevertheless, this study provides a valuable experimental framework to address these questions in the future.

Reviewer #2 (Public review):

Summary:

The Crenarchaeal Cdv division system represents a reduced form of the universal and ubiquitous ESCRT membrane reverse-topology scission machinery, and therefore a prime candidate for synthetic and reconstitution studies. The work here represents a convincing extension of previous work in the field, clarifying the order of recruitment of Cdv proteins to curved membranes.

Strengths:

The use of a recently developed approach to produce dumbbell-shaped liposomes (De Franceschi et al. 2022), which allowed the authors to assess recruitment of various Cdv assemblies to curved membranes or membrane necks; reconstitution of a quaternary Cdv complex at a membrane neck.

Weaknesses:

The initial manuscript was a bit light on quantitative detail, across the various figures - addressing this would make the paper much stronger. The authors could also include in the discussion a short paragraph on implications for our understanding of ESCRT function in other contexts and/or in archaeal evolution - for the interests of a broad audience. These issues have been addressed in the authors' revision.

Reviewer #3 (Public review):

In this revised report, De Franceschi et al. purify components of the Cdv machinery in archaeon M. sedula and probe their interactions with membrane and with one-another in vitro using two main assays - liposome flotation and fluorescent imaging of encapsulated proteins. This has the potential to add to the field by showing how the order of protein recruitment seen in cells is related to the differential capacity of individual proteins to bind membranes when alone or when combined.

Using the floatation assay, they demonstrate that CdvA, CdvB, and CdvB1 bind liposomes. CdvB2 lacking its C-terminus is not efficiently recruited to membranes unless CdvAB or CdvB1 are present. The authors then employ a clever liposome assay that generates chained spherical liposomes connected by thin membrane necks, which allows them to accurately control the buffer composition inside and outside of the liposome. With this, they show that all four proteins accumulate in necks of dumbbell-shaped liposomes that mimic the shape of constricting necks in cell division, possibly indicating a sensing of catenoid membrane geometry. Taken altogether, these data lead them to propose that Cdv proteins are sequentially recruited to the membrane as has also been suggested by in vivo studies of ESCRT-III dependent cell division in crenarchaea.

In their revision, the authors have addressed the vast majority of our previous concerns. The paper is much improved as a result. The Figures are improved and the authors have added appropriate controls and additional experiments, strengthening their conclusions.

There are still some discrepancies between these results and what is know about Sulfolobus division. Since the initial submission, other work has shown that in S. acidocaldarius, CdvA is the first component to assemble a ring (in absence of CdvB , doi.org/10.1073/pnas.2513939122) and that CdvB2 is able to bind membranes in vitro (doi.org/10.1073/pnas.2525941123). This might reflect differences between Sulfolobus and Metallosphaera, but probably should be discussed.

Author response:

The following is the authors’ response to the original reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

The authors aimed to elucidate the recruitment order and assembly of the Cdv proteins during Sulfolobus acidocaldarius archaeal cell division using a bottom-up reconstitution approach. They employed liposome-binding assays, EM, and fluorescence microscopy with in vitro reconstitution in dumbbellshaped liposomes to explore how CdvA, CdvB, and the homologues of ESCRT-III proteins (CdvB, CdvB1, and CdvB2) interact to form membrane remodeling complexes.

The study sought to reconstitute the Cdv machinery by first analyzing their assembly as two subcomplexes: CdvA:CdvB and CdvB1:CdvB2ΔC. The authors report that CdvA binds lipid membranes only in the presence of CdvB and localizes preferentially to membrane necks. Similarly, the findings on CdvB1:CdvB2ΔC indicate that truncation of CdvB2 facilitates filament formation and enhances curvature sensitivity in interaction with CdvB1. Finally, while the authors reconstitute a quaternary CdvA:CdvB:CdvB1:CdvB2 complex and demonstrate its enrichment at membrane necks, the mechanistic details of how these complexes drive membrane remodeling by subcomplexes removal by the proteasome and/or CdvC remain speculative.

Although the work highlights intriguing similarities with eukaryotic ESCRT-III systems and explores unique archaeal adaptations, the conclusions drawn would benefit from stronger experimental validation and a more comprehensive mechanistic framework.

Strengths:

The study of machinery assembly and its involvement in membrane remodeling, particularly using bottom-up reconstituted in vitro systems, presents significant challenges. This is particularly true for systems like the ESCRT-III complex, which localizes uniquely at the lumen of membrane necks prior to scission. The use of dumbbell-shaped liposomes in this study provides a promising experimental model to investigate ESCRT-III and ESCRT-III-like protein activity at membrane necks.

The authors present intriguing evidence regarding the sequential recruitment of ESCRT-III proteins in crenarchaea-a close relative of eukaryotes. This finding suggests that the hierarchical recruitment characteristic of eukaryotic systems may predate eukaryogenesis, which is a significant and exciting contribution. However, the broader implications of these findings for membrane remodeling mechanisms remain speculative, and the study would benefit from stronger experimental validation and expanded contextualization within the field.

We thank the Referee for his/her appreciation of our work.

Weaknesses:

This manuscript presents several methodological inconsistencies and lacks key controls to validate its claims. Additionally, there is insufficient information about the number of experimental repetitions, statistical analyses, and a broader discussion of the major findings in the context of open questions in the field.

We have now added more controls, information about repetitions, and discussion.

Reviewer #2 (Public review):

Summary:

The Crenarchaeal Cdv division system represents a reduced form of the universal and ubiquitous ESCRT membrane reverse-topology scission machinery, and therefore a prime candidate for synthetic and reconstitution studies. The work here represents a solid extension of previous work in the field, clarifying the order of recruitment of Cdv proteins to curved membranes.

Strengths:

The use of a recently developed approach to produce dumbbell-shaped liposomes (De Franceschi et al. 2022), which allowed the authors to assess recruitment of various Cdv assemblies to curved membranes or membrane necks; reconstitution of a quaternary Cdv complex at a membrane neck.

We thank the Referee for his/her appreciation of the work.

Weaknesses:

The manuscript is a bit light on quantitative detail, across the various figures, and several key controls are missing (CdvA, B alone to better interpret the co-polymerisation phenotypes and establish the true order of recruitment, for example) - addressing this would make the paper much stronger. The authors could also include in the discussion a short paragraph on implications for our understanding of ESCRT function in other contexts and/or in archaeal evolution, as well as a brief exploration of the possible reasons for the discrepancy between the foci observed in their liposome assays and the large rings observed in cells - to better serve the interests of a broad audience.

We have now added more controls, information about repetitions, and discussion.

Reviewer #3 (Public review):

Summary:

In this report, De Franceschi et al. purify components of the Cdv machinery in archaeon M. sedula and probe their interactions with membrane and with one-another in vitro using two main assays - liposome flotation and fluorescent imaging of encapsulated proteins. This has the potential to add to the field by showing how the order of protein recruitment seen in cells is related to the differential capacity of individual proteins to bind membranes when alone or when combined.

Strengths:

Using the floatation assay, they demonstrate that CdvA and CdvB bind liposomes when combined. While CdvB1 also binds liposomes under these conditions, in the floatation assay, CdvB2 lacking its C-terminus is not efficiently recruited to membranes unless CdvAB or CdvB1 are present. The authors then employ a clever liposome assay that generates chained spherical liposomes connected by thin membrane necks, which allows them to accurately control the buffer composition inside and outside of the liposome. With this, they show that all four proteins accumulate in necks of dumbbell-shaped liposomes that mimic the shape of constricting necks in cell division. Taken altogether, these data lead them to propose that Cdv proteins are sequentially recruited to the membrane as has also been suggested by in vivo studies of ESCRT-III dependent cell division in crenarchaea.

We thank the Referee for his/her appreciation of the work.

Weaknesses:

These experiments provide a good starting point for the in vitro study the interaction of Cdv system components with the membrane and their consecutive recruitment. However, several experimental controls are missing that complicate their ability to draw strong conclusions. Moreover, some results are inconsistent across the two main assays which make the findings difficult to interpret:

(1) Missing controls.

Various protein mixtures are assessed for their membrane-binding properties in different ways. However, it is difficult to interpret the effect of any specific protein combination, when the same experiment is not presented in a way that includes separate tests for all individual components. In this sense, the paper lacks important controls. For example, Fig 1C is missing the CdvB-only control. The authors remark that CdvB did not polymerise (data not shown) but do not comment on whether it binds membrane in their assays. In the introduction, Samson et al., 2011 is cited as a reference to show that CdvB does not bind membrane. However, here the authors are working with protein from a different organism in a different buffer, using a different membrane composition and a different assay. Given that so many variables are changing, it would be good to present how M. sedula CdvB behaves under these conditions.

We thank the referee for raising this point. We have now added these data in Figure 1C. Indeed it turns out that CdvB from M. sedula exhibits clear membrane binding on its own in a flotation assay.

Similarly, there is no data showing how CdvB alone or CdvA alone behave in the dumbbell liposome assay.

Without these controls, it's impossible to say whether CdvA recruits CdvB or the other way around. The manuscript would be much stronger if such data could be added.

We have now added these data in Figure 1E, 1F and 1G. Overall, we can confirm that CdvA binds the membrane better in the presence of CdvB (although both proteins can bind the membrane on their own). Both proteins appear to recognize the curved region of the membrane neck.

(2) Some of the discrepancies in the data generated using different assays are not discussed.

The authors show that CdvB2∆C binds membrane and localizes to membrane necks in the dumbbell liposome assay, but no membrane binding is detected in the flotation assay. The discrepancy between these results further highlights the need for CdvB-only and CdvA-only controls.

We have now added these controls in Figure 1. In addition, we would like to clarify that the flotation assay and the SMS dumbbell assay serve different purposes and are not directly comparable in quantitative terms. In the flotation assay, all the protein present as input is eventually recovered and visualized. Thus, quantitative information on the proportion of the fraction of the total protein bound to lipids can be inferred from this assay. The SMS assay, in contrast, provides a very different kind of information. Because of the particular protocol required to generate dumbbells (De Franceschi, 2022), the total amount of protein in the inner buffer in dumbbells is not accurately defined, because protein that is not correctly reconstituted (e.g. which aggregates while still in the droplet phase) will interfere with vesicle generation, with the result that dumbbell with such aggregates is generally not formed in the first place. This renders it impossible to draw any quantitative conclusions about the proportion of the sample bound to lipids. The SMS is therefore not directly comparable to the flotation assay, and it is rather complementary to it. Indeed, the purpose of the SMS is to provide information about curvature selectivity of the protein.

(3) Validation of the liposome assay.

The experimental setup to create dumbbell-shaped liposomes seems great and is a clever novel approach pioneered by the team. Not only can the authors manipulate liposome shape, they also state that this allows them to accurately control the species present on the inside and outside of the liposome. Interpreting the results of the liposome assay, however, depends on the geometry being correct. To make this clearer, it would seem important to include controls to prove that all the protein imaged at membrane necks lie on the inside of liposomes. In the images in SFig3 there appears to be protein outside of the liposome. It would also be helpful to present data to show test whether the necks are open, as suggested in the paper, by using FRAP or some other related technique.

We thank the Referee for his/her appreciation. The proteins are encapsulated inside the liposomes, not outside of them. While Figure S3 might give the appearance that there is some protein outside, this is actually just an imaging artifact. Author response image 1 (below) explains this: When the membrane and protein channel are shown separately, it is clear that the protein cluster that appeared to be ‘outside’ actually colocalizes with an extra small dumbbell lobe (yellow arrowhead). The protein appeared to be outside of it because (1) the protein fluorescent signal is stronger than the signal from the membrane, and (2) there is a certain time delay in the acquisition of the two channels (0.5-1 second), thus the membrane may have slightly shifted out of focus when the fluorescence was being acquired. We are confident that the protein is inside in these dumbbells because the procedure for preparing the dumbbells requires extensive emulsification by pipetting, which requires ≈ 1 minute. This time is more than sufficient for proteins with high affinity for the membrane, like ESCRT and Cdv, to bind the membrane. For an example of how fast binding under confinement can be, please see movie 2 from this paper: De Franceschi N, Alqabandi M, Miguet N, Caillat C, Mangenot S, Weissenhorn W, Bassereau P. The ESCRT protein CHMP2B acts as a diffusion barrier on reconstituted membrane necks. J Cell Sci. 2018 Aug 3;132(4):jcs217968.

Moreover, in many instances, we observed that the protein is inside because, by increasing the gain in the images post-acquisition, a clear protein signal appear in the lumen (see Author response image 2).

Author response image 1.

Separate channels showing colocalization of protein and lipids (adapted from Figure S3). The zoom-in shows separate channels, highlighting that the CdvB2 cluster that seems to be ‘outside the dumbbell’ actually colocalizes with the small terminal lobe of the dumbbell, indicating that the protein is encapsulated within that lobe.

Author response image 2.

Residual protein present inside lumen of dumbbells as visualized by increasing the brightness post-acquisition.

We are not sure what the referee means by “test whether the necks are open, as suggested in the paper”. We are confident that the lobes of dumbbells originated from a single floppy vesicle, and were therefore mutually connected with an open neck (at least at the onset of the experiment). We have performed extensive FRAP assays on dumbbells in previous papers (De Franceschi et al., ACS nano 2022 and De Franceschi et al., Nature Nanotech 2024) which unequivocally proved that these chains of dumbbells are connected with open necks. We now also performed a few FRAP assay with reconstituted Cdv proteins, which confirmed this point. We have added a movie of such an experiment to the manuscript (Movie 1).

Investigating whether the necks are open or closed after Cdv reconstitution is indeed a very relevant question, that could be rephrased as “verify whether Cdv proteins or their combination can induce membrane scission”. This is however beyond the scope of this manuscript, as the current work merely addressed the question of hierarchical recruitment of Cdv proteins at the membrane. We plan to examine this in future work.

(4) Quantification of results from the liposome assay.

The paper would be strengthened by the inclusion of more quantitative data relating to the liposome assay. Firstly, only a single field of view is shown for each condition. Because of this, the reader cannot know whether this is a representative image, or an outlier? Can the authors do some quantification of the data to demonstrate this? The line scan profiles in the supplemental figures would be an example of this, but again in these Figures only a single image is analyzed.

The images that we showed are indeed representative. The dumbbells that are generated by the SMS approach contain an “internal control”: in each dumbbell, the protein has the option of localizing at the neck or localizing elsewhere in the region of flat membrane. We see consistently that Cdv proteins have a strong preference for localizing at the neck.

We would recommend that the authors present quantitative data to show the extent of co-localization at the necks in each case. They also need a metric to report instances in which protein is not seen at the neck, e.g. CdvB2 but not CdvB1 in Fig2I, which rules out a simple curvature preference for CdvB2 as stated in line 182.

While the request for better quantitation is reasonable, this would require carrying out very significant new experiments at the microscope, which is rendered near-impossible since both first authors left the lab on to new positions.

Secondly, the authors state that they see CdvB2∆C recruited to the membrane by CdvB1 (lines 184-187, Fig 2I). However, this simple conclusion is not borne out in the data. Inspecting the CdvB2∆C panels of Fig 2I, Fig3C, and Fig3D, CdvB2∆C signal can be seen at positions which don't colocalize with other proteins. The authors also observe CdvB2∆C localizing to membrane necks by itself (Fig 2E). Therefore, while CdvB1 and CdvB2∆C colocalize in the flotation assay, there is no strong evidence for CdvB2∆C recruitment by CdvB1 in dumbbells. This is further underscored by the observation that in the presented data, all Cdv proteins always appear to localize at dumbbell necks, irrespective of what other components are present inside the liposome. Although one nice control is presented (ZipA), this suggests that more work is required to be sure that the proteins are behaving properly in this assay. For example, if membrane binding surfaces of Cdv proteins are mutated, does this lead to the accumulation of proteins in the bulk of the liposome as expected?

In the particular example of Figure 2I, it indeed appears that there are some clusters of CdvB2ΔC that do not contain CdvB1 (we indicated them in Author response image 3 by red arrowheads), while the yellow arrowheads indicate clusters that contain both proteins. It can be clearly seen that the clusters that do contain both proteins (yellow arrows) are localized at necks, while those that only contain CdvB2ΔC (red arrows) are not localized at necks. This is no coincidence. The clusters indicated by the red arrow do contain CdvB1. However, these clusters rapidly diffuse on the membrane plane because they are not fixed at the neck: therefore, they constantly shift in and out of focus. Because there is a time delay in the acquisition of each channel (between 0.5 and 1 second), these cluster were in focus when the CdvB2ΔC signal was being acquired, but sifted out of focus when the CdvB1 signal was being acquired. This implies that the clusters indicated by the yellow arrowheads are stably localized at necks, which is precisely the point we wished to make with this experiment: because Cdv proteins have an affinity for curved geometry, they preferentially and stably localize at necks. Why don’t all the clusters localize at necks then? We estimate that the simple answer is that, in this particular case, there are more clusters than there are necks, so some of the clusters must necessarily localize somewhere else.

Author response image 3.

Current Figure 2H, where clusters that are double-positive for both CdvB1 and CdvB2ΔC are indicated by yellow arrowheads, while cluster that apparently only contain CdvB2ΔC are indicated by red arrowheads. It is observed that all the double-positive clusters are localized at necks.

(5) Rings.

The authors should comment on why they never observe large Cdv rings in their experiments. In crenarchaeal cell division, CdvA and CdvB have been observed to form large rings in the middle of the 1 micron cell, before constriction. Only in the later stages of division are the ESCRTs localized to the constricting neck, at a time when CdvA is no longer present in the ring. Therefore, if the in vitro assay used by the authors really recapitulated the biology, one would expect to see large CdvAB rings in Figs 1EF. This is ignored in the model. In the proposed model of ring assembly (line 252), CdvAB ring formation is mentioned, but authors do not discuss the fact that they do not observe CdvAB rings - only foci at membrane necks. The discussion section would benefit from the authors commenting on this.

The referee is correct: it is intriguing that we don’t see micron-sized rings for CdvA and CdvB. We do note that our EM data (Fig.S1) show that CdvA in its own can form rings of about 100-200nm diameter, well below the diffraction limit, that could well correspond to the foci that we optically resolve in Figure 1. We now added a brief comment on this to the manuscript on lines 256-264.

(6) Stoichiometry

It is not clear why 100% of the visible CdvA and 100% of the the visible CdvB are shifted to the lipid fraction in 1C. Perhaps this is a matter of quantification. Can the authors comment on the stoichiometry here?

We agree that this was unclear. Since that particular gel was stained by coumassie, the quantitative signals might be unreliable, and hence we have repeated this experiment using fluorescently labelled proteins, which show indeed a less extreme distribution. This was also done to make the data more uniform, as requested by the referees.

(7) Significance of quantification of MBP-tagged filaments.

Authors use tagging and removal of MBP as a convenient, controllable system to trigger polymerisation of various Cdv proteins. However, it is unclear what is the value and significance of reporting the width and length of the short linear filaments that are formed by the MBP-tagged proteins. Presumably they are artefactual assemblies generated by the presence of the tag?

Providing a measure of the changes induced by MBP removal, in fact, validates that this actually has an effect. But perhaps this places too much emphasis on the short filaments. We now opted for a compromise, removing the quantification of the width and length of short filaments formed by MBPtagged protein from the text, but keeping the supplementary figure showing their distribution as compared to the other filaments (Figure S2E, SF).

Similar Figure 2C doesn't seem a useful addition to the paper.

We removed panel 2C, and now merely report these values in the text.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

I would suggest the authors perform a deeper discussion about their findings, such as what are the evolutionary implications, how they think lipids from these archaea may affect the recruitment process,...

Because there is no exact homology between Archaea Cdv proteins and Eukaryotic ESCRT-III proteins, we do not feel our work brings new evolutionary implications beyond what we already state in the manuscript. We also dis not perform experiments using Archaea lipids, thus we would rather not speculate on how they may potentially affect the recruitment of Cdv proteins.

In general, the manuscript lacks information regarding some scale bars, number of experimental repetitions (n or N), statistical analysis when needed, information about protein concentrations used in their assays.

We have now added this information in the manuscript.

Below, I provide a list of comments that I think the authors should address to improve the manuscript:

(1) Line 113-114: The authors test protein-membrane interactions using flotation assays with positively curved SUV membranes but encapsulate proteins in dumbbell-shaped liposomes with negative curvature at the connecting necks. Might the use of membranes with opposite curvatures affect the recruitment process? Since the proteins are fluorescently labeled, I suggest testing recruitment using flat giant unilamellar vesicles or supported lipid bilayers (with zero curvature) to validate their findings.

We thank the referee for this suggestion. Please do note that we are not claiming in our paper that Cdv proteins recognize negative curvature. We merely observe that they localize at necks. The neck of a dumbbell exhibits the so-called “catenoid” geometry, which is characterized by having both positive and negative curvature.

Experimentally, on the SUVs, we now realize there was a mistake in the method section: In the flotation assay we in fact used multilamellar vesicles, not SUVs, precisely for the reason mentioned by the referee. We apologize for the oversight and have now corrected this in the methods. Multilamellar vesicles are not characterized by a strong positive curvature as SUVs do, but we do agree that they likely don’t have negative curvature there either. Because of the heterogeneous nature of the multilamellar vesicles, they provide a binding assay that was rather independent of the curvature. Complementary to the flotation assay, the SMS approach was employed to reveal the curvature preference of proteins.

Finally, we performed the experiment on large GUVs suggested by the referee using CdvB as an example, but this turned out to be inconclusive because the protein forms clusters: these clusters may be creating local curvature at the nanometer scale, which cannot be resolved by optical microscopy (Author response image 4). This is quite typical for proteins that recognize curvature (cf. for instance: De Franceschi N, Alqabandi M, Miguet N, Caillat C, Mangenot S, Weissenhorn W, Bassereau P. The ESCRT protein CHMP2B acts as a diffusion barrier on reconstituted membrane necks. J Cell Sci. 2018 Aug 3;132(4):jcs217968.)

Author response image 4.

Fluorescently labelled CdvB bound to giant unilamellar vesicle. The protein was added in the outer buffer. CdvB forms distinct clusters, which may generate a local region of high membrane curvature.

(2) Line 138-139: How is His-ZipA binding the membrane? Wouldn't Ni²⁺-NTA lipids be required? If not, how is the binding achieved?

Indeed, NTA-lipids were present. This is now stated both in the legend and in the methods.

(3) In the encapsulated protein assays, why does the luminal fluorescence intensity of the encapsulated protein sometimes appear similar to the bulk fluorescence signal? Since only a small fraction of the protein assembles at membrane necks, shouldn't the luminal pool of unbound protein show higher fluorescence intensity inside the liposomes?

We thank the referee for raising this point and giving us the opportunity to explain this. The reason is that Cdv proteins have a very high affinity for the neck, and when they cluster at the neck the fluorescence intensity of the cluster is many times higher than the background fluorescence. Because we were interested in imaging the clusters and avoiding overexposing them, we adjusted the imaging conditions accordingly, with the result that the fluorescence from both the lumen and the bulk is at very low level.

By choosing different imaging conditions, however, it can be actually seen that the signal inside the lumen is clearly higher than the bulk: this can be seen for instance in Author response image 2, where the brightness has been properly adjusted.

(4) Line 184-185: In Fig. 2I, some CdvB2ΔC puncta seem independent of CdvB1 and are not localized at membrane necks. How many such puncta exist? For example, in the provided micrograph, 2 out of 5 clusters are independent of CdvB1. This proportion is significant. Could the authors quantify the prevalence of these structures and discuss why they form?

We thank the referee for giving us the opportunity to explain this apparent discrepancy. We’ll like to stress the fact that CdvB2ΔC and CdvB1 form an obligate heterodimer: in all our experiments, without exception, we find that they form a strong complex when we mix the two proteins. This is true both in dumbbells and in flotation assays.

In the particular example of Figure 2I, it indeed appears that there are some clusters of CdvB2ΔC that do not contain CdvB1 (we indicated them in Author response image 3 by red arrowheads), while the yellow arrowheads indicate clusters that contain both proteins. It can be clearly seen that the clusters that do contain both proteins (yellow arrows) are localized at necks, while those that only contain CdvB2ΔC (red arrows) are not localized at necks. This is no coincidence. The clusters indicated by the red arrow do contain CdvB1. However, these clusters rapidly diffuse on the membrane plane because they are not fixed at the neck: therefore, they constantly shift in and out of focus. Because there is a time delay in the acquisition of each channel (between 0.5 and 1 second), these cluster were in focus when the CdvB2ΔC signal was being acquired, but sifted out of focus when the CdvB1 signal was being acquired. This implies that the clusters indicated by the yellow arrowheads are stably localized at necks, which is precisely the point we wished to make with this experiment: because Cdv proteins have affinity for curved geometry, they preferentially and stably localize at necks. Why don’t all the clusters localize at necks then?

(5) Figure 1E and 1F: Why do lipids accumulate and colocalize with the proteins? How can the authors confirm lumen connectivity between vesicles? Performing FRAP assays could validate protein localization and enrichment at the lumen of the membrane necks.

At first sight, indeed some lipid enrichment seems to be observed at the neck between lobes of dumbbells.

This is, however, an imaging artifact due to the fact that the neck is diffraction limited. As shown in the Author response image 5, we are acquiring the membrane signal from both lobes at the neck region, and therefore the signal is roughly double, hence the apparent lipid enrichment.

Author response image 5.

Schematic illustrating that the neck between two lobes is smaller than the diffraction limit of optical microscopy (the size of a typical pixel is indicated by the green square). Because of this technical limitation, the fluorescence intensity of the membrane at the neck is twice that of a single membrane.

The referee is correct in pointing out that these images do not prove that the lobes are connected, and that FRAP assays is the only way to prove this point. However, in previous papers we have confirmed extensively that in chains of dumbbells the lobes are connected:

- De Franceschi N, Pezeshkian W, Fragasso A, Bruininks BMH, Tsai S, Marrink SJ, Dekker C. Synthetic Membrane Shaper for Controlled Liposome Deformation. ACS Nano. 2022 Nov 28;17(2):966–78. doi: 10.1021/acsnano.2c06125.

- De Franceschi N, Barth R, Meindlhumer S, Fragasso A, Dekker C. Dynamin A as a one-component division machinery for synthetic cells. Nat Nanotechnol. 2024 Jan;19(1):70-76. doi: 10.1038/s41565023-01510-3.

Random sticking of liposomes would also generate clusters of vesicles, not linear chains. We now provide also a Movie (Movie 1) supporting this point.

Investigating whether the necks are open or closed after Cdv reconstitution is indeed a very relevant question, that could be rephrased as “verify whether Cdv proteins or their combination can induce membrane scission”. This is however beyond the scope of this manuscript, as the current work merely addressed the question of hierarchical recruitment of Cdv proteins at the membrane. We plan to examine this in future work.

(6) Why didn't the authors use the same lipid composition, particularly the same proportion of negatively charged lipids, on the SUVs of the flotation assays and on the dumbbell-shaped liposomes?

In flotation assays, it is typical to use a relatively large proportion of negatively charged lipids, to promote protein binding. This is because the aim is to maximize membrane coverage by the protein. The SMS procedure to generate dumbbell-shaped GUVs is completely different, however. Rather than covering the membrane with protein, the idea is to reduce the amount of protein to a minimum, so that any curvature preference can be best visualized. This is e.g. routinely done in tube pulling experiments, for the same reason (See for instance Prévost C, Zhao H, Manzi J, Lemichez E, Lappalainen P, Callan-Jones A, Bassereau P. IRSp53 senses negative membrane curvature and phase separates along membrane tubules. Nat Commun. 2015 Oct 15;6:8529. doi: 10.1038/ncomms9529).

(7) Line 117-119: The suggestion that polymer formation between CdvA and CdvB facilitates membrane recruitment is intriguing. However, fluorescence microscopy experiments could better elucidate whether there is sequential recruitment of CdvB followed by CdvA, or if these proteins form a heteropolymer composite for membrane binding. Can CdvB bind membranes independently, or does this require synergy between CdvA and CdvB.

We thank the referee for prompting us to perform this experiment. As we now show in Figure 1C, CdvB indeed is able to bind the membrane independently of CdvA. Whether this happens sequentially or simultaneously is an interesting question, but one that is impossible to address with either the SMS or the flotation assay, because in both cases we can only observe the endpoint of the recruitment.

We would also like to clarify one specific experimental detail. Perhaps unsurprisingly, the results from the flotation assay are dependent on the way the assay is performed. In particular, we observed that the same protein can exhibit a different binding profile depending on whether it is being loaded either at the top or at the bottom of the gradient. This can be seen in Author response image 6. This is counterintuitive, since once the equilibrium is reached, the result should only depend on the density of the sample. We performed an overnight centrifugation (> 16 hours) on a short tube (< 3 cm tall), thus equilibrium is being reached (which is corroborated by the fact that CdvB1 and CdvB2 can float to the top of the gradient within this timespan, as shown in Figure 2C, 2E, 2G). We ascribe the difference between top and bottom loading to the fact that, when the sample is loaded at the bottom, it has to be mixed with a concentrated sucrose solution, while in the case of loading from the top, this is not done.

In literature, both loading from top and from bottom have been used:

- Lata S, Schoehn G, Jain A, Pires R, Piehler J, Gottlinger HG, Weissenhorn W. Helical structures of ESCRTIII are disassembled by VPS4. Science. 2008 Sep 5;321(5894):1354-7. doi: 10.1126/science.1161070

- Moriscot C, Gribaldo S, Jault JM, Krupovic M, Arnaud J, Jamin M, Schoehn G, Forterre P, Weissenhorn W, Renesto P. Crenarchaeal CdvA forms double-helical filaments containing DNA and interacts with ESCRT-III-like CdvB. PLoS One. 2011;6(7):e21921. doi: 10.1371/journal.pone.0021921.

- Senju Y, Lappalainen P, Zhao H. Liposome Co-sedimentation and Co-flotation Assays to Study LipidProtein Interactions. Methods Mol Biol. 2021;2251:195-204. doi: 10.1007/978-1-0716-1142-5_14. In performing the flotation assay for CdvB1 and CdvB2ΔC, or when using all 4 proteins together, we loaded the sample at the bottom, and we could detect reproducible binding to liposomes (Figures 2D, 2F, 2H, 3A). However, CdvB does not bind the membrane when loaded at the bottom. Thus, for the experiments shown in figure 1C, we loaded the proteins at the top. This experimental setup allowed us to highlight that CdvB indeed induce a stronger interaction between CdvA and the membrane.

Author response image 6.

CdvB binding to multilamellar vesicles in a flotation assay. In the left panel, the sample was loaded at the top of the sucrose gradient; in the right panel it was loaded at the bottom.

(8) Line 165-173: The authors claim that filament curvature differs between CdvB2ΔC alone and the CdvB1:CdvB2ΔC complex. Are these differences statistically significant? What is the sample size (N)? Furthermore, how do the authors confirm interactions between these proteins in the absence of membranes based solely on EM micrographs?

We can confirm that the filaments are composed by both proteins, because the filaments have different curvature when both proteins are present. However, as requested by referee 3, point (7), we removed the quantification of curvature from panel 2C. We report the N number in the text.

(9) Line 121-123: Are the authors referring to positive or negative membrane curvatures? The cited literature suggests ESCRT-III proteins either lack curvature preferences (e.g., Snf7, CHMP4B) or prefer high positive curvature (e.g., late ESCRT-III subunits). This is confusing since the authors later test recruitment to negatively curved necks.

We do not claim that Cdv proteins prefer positive or negative curvature, because the necks present in dumbbells have a catenoid geometry, which include both positive and negative curvature. We have now clarified this in the discussion.

(10) Since the conclusions rely on the oligomeric state of the proteins, providing SEC-MALS spectra to show the protein oligomeric state right after the purification would strengthen the claims.

While such SEC-MALDI experiments may be interesting, practical implementation of this is not possible since both first authors left the lab on to new positions.

(11) Line 157-160: Suppl. Fig. 2 shows only a single EM micrograph of a small filament. Could the authors provide lower magnification images showing more filaments?

As requested by Referee 3, point (7), we have toned down the importance of these short filaments.

Also, why are the sample sizes for filament length (N=161) and width (N=129) different?

Protein filaments formed by Cdv tend to stick to each other side by side, so that for some filaments the width could not be accurately assessed, and accordingly those were removed from the analysis.

(12) The introduction states that CdvA binds membranes while CdvB does not. However, the results suggest CdvB facilitates membrane binding, helping CdvA attach. This discrepancy needs further explanation.

We thank the referee for raising this point. We have now performed additional experiments (both SMS assay and flotation assays) showing that indeed CdvB from M. sedula is (unlike CdvB from Sulfolobus) able to bind the membrane on its own (Figure 1C, 1F).

Reviewer #2 (Recommendations for the authors):

Best practice would be to show single fluorescence channels in grayscale or inverted grayscale, retaining pseudocolouring only for the merged multichannel image.

We decided to retain and standardize the colors, both for gels and for microscopy images, in order to have the same color-code for each protein. We believe this improves readability, and this was also a request from Referee 3. Thus, throughout the manuscript, CdvA is in grayscale, CdvB in yellow, CdvB1 in green, CdvB2ΔC in cyan and the membrane in magenta.

It would be great to include a quantification of liposome curvature vs focal intensity of the various Cdv components - across figures.

Quantification of liposome curvature at the neck can be done (De Franceschi et al., Nature Nanotech. 2024). However, in practice, this requires transferring of the sample post-preparation into a new chamber in order to increase the signal-to-noise ratio of the encapsulated dye, a procedure that drastically reduces the yield of dumbbells. The very sizeable amount of work required to obtain reliable measurements, especially considering all the proteins and protein combinations used in this study, indicates that this represents a project in itself, which goes well beyond the scope of this manuscript.

Reviewer #3 (Recommendations for the authors):

(1) We would encourage the authors to consider including the length of the scale bar next to the scale bar in each image and not in the figure description. This would greatly aid in clarity and interpretation of figures.

We have now written the length of the scale bar in the figures.

(2) In a similar vein, could the authors consider labeling panels throughout the manuscript, writing that sample is being presented? This goes mainly for the negative stain and the dumbbell fluorescence images, as having to continuously consult the figure legend again hinders clarity.

We have now labelled the EM images as requested by the referee.

(3) Lines 254-256: would the statement hold not only for CdvB2∆C, but for all imaged proteins? They all seem to localize to membrane necks, presumably favoring membrane binding to a specific membrane topology.

We agree with the referee, and changed the phrasing accordingly.

(4) CdvB2∆C construct - presumably this was a truncation of helix 5 of the ESCRT-III domain? Figure 1A shows that the ESCRT-III domain spans residues 34-170 and therefore implies that all five ESCRT-III helices (which make up the ESCRT-III domain) are present in the C-terminal truncation. Could the authors clarify?

Indeed, the truncation was done at residue 170.

(5) Results of the liposome flotation assays are presented inconsistently across the three figures (Figs 1C, 2DFH, and 3A). This makes it more difficult than it needs to be to interpret and compare results. Could the authors consider presenting the three gels in a more similar, standardized way across the three figures?

To improve readability, we now standardized the colors, both for gels and for microscopy images, in order to have the same color-code for each protein. Thus, throughout the manuscript, CdvA is in grayscale, CdvB in yellow, CdvB1 in green, CdvB2ΔC in cyan and the membrane in magenta.

(6) From the data presented in Fig 1EF, it cannot be concluded whether CdvB and CdvA colocalize, as only one protein is labelled. Is there a technical reason for this?

We have now repeated the same experiment by having both proteins labelled, confirming that there is co-localization at the neck (Figure 1G).

(7) Fig 2C: is the difference between the two samples significant

As requested by Referee 3, we have removed Figure 2C.

(8) Fig 2I is missing a 'merged' panel.

We have now added the merged panel.

(9) The fluorescence intensity plots in Supp Figs 1C and 3C would be easier to interpret if the lipid and protein signal would be plotted on the same plot (say, with normalized fluorescence intensity)

It is not immediately obvious to us what the signal should be normalized to. What we wished to convey with these plots was that the intensity of proteins spikes at the neck region. In an attempt to improve clarity, we have now aligned the plots vertically, and highlighted the position of the neck.

(10) CdvA should have a capital "A" in Figure 3A, panel 3.

We have now corrected this.

(11) The discussion doesn't comment on the need to truncate CdvB2.

This is explained in the result session.

eLife Assessment

This study systematically characterizes the activity patterns of a lateral supramammillary nucleus (SuM)-medial septum (MS)-hippocampus circuit across sleep-wake cycles and its role in memory consolidation. This work is fundamental because it identifies a previously unrecognized brain hub that helps coordinate how different types of memory are supported during a specific sleep state, advancing our understanding of how sleep contributes to memory organization. The work is well-designed, and the data are solid, supporting clear and significant conclusions; however, some mechanistic details and causal relationships would benefit from further clarification or additional experiments. The paper provides new insights into how distinct memory modalities could be processed by parallel, sleep-active subcortical-hippocampal circuits, which would be of general interest to a broad neuroscience audience.

Reviewer #1 (Public review):

In this manuscript, the authors aim to define how rapid eye movement sleep supports memory consolidation by identifying the brain circuits that are selectively engaged during this sleep state. They focus on a pathway linking a hypothalamic region involved in sleep regulation to the medial septum and onward to a hippocampal subregion that is critical for social memory. By combining recordings of neural activity with sleep-state-specific circuit manipulations, the study seeks to explain how information is routed during sleep to support distinct types of memory.

A major strength of the work is the use of state-of-the-art circuit-based approaches to link sleep dynamics to defined long-range connections and behavioral outcomes. The authors show that neurons in the lateral supramammillary region projecting to the medial septum are selectively active during rapid eye movement sleep, and that silencing this pathway during this sleep state disrupts consolidation of both social and contextual fear memories. Further dissection of downstream circuitry reveals that inhibition of the medial septum-to-hippocampal CA2 pathway during rapid eye movement sleep selectively impairs social memory. These results provide support for functional specialization within parallel pathways and suggest that this circuit acts as a hub for routing memory-related information during sleep.

While the evidence supporting a role for this circuit in sleep-dependent memory consolidation is compelling, several important mechanistic details remain unresolved. The chemical signaling used by the neurons connecting the hypothalamus to the medial septum is not clearly defined, leaving open whether these cells release excitatory signals, inhibitory signals, or a combination of both. In addition, the medial septum contains multiple neuronal populations with distinct downstream targets, and the specific cell types receiving input from this pathway are not clearly identified. Similarly, the nature of the signals delivered from the medial septum to the hippocampus remains unclear, making it difficult to link circuit anatomy to the observed behavioral specificity. Finally, because different circuit segments are manipulated independently, the causal relationship between upstream and downstream pathways remains suggestive rather than definitive and should be discussed explicitly as a limitation or addressed experimentally.

Overall, the authors largely achieve their aims by identifying a rapid eye movement sleep-specialized circuit that contributes to memory consolidation in a modality-specific manner. The findings are likely to have a meaningful impact on the field by advancing understanding of how sleep organizes memory through parallel neural pathways and by providing a useful framework for future studies of sleep-dependent brain state regulation. With additional clarification of circuit mechanisms or a clearer discussion of current limitations, the study would offer even greater value to the neuroscience community.

Reviewer #2 (Public review):

Summary:

This study systematically characterizes the activity patterns of a lateral supramammillary nucleus (SuM)-medial septum (MS)-hippocampus circuit across sleep-wake cycles and its role in memory consolidation. The authors demonstrate that the lateral SuM-MS projection is specifically active during REM sleep, and that REM-selective inhibition of this circuit, and of its downstream MS-CA2 pathway, impairs the consolidation of social memory. The work is well-designed, and the data are robust in supporting clear and significant conclusions. It provides important new insights into how distinct memory modalities could be processed by parallel, sleep-active subcortical-hippocampal circuits. The manuscript is of high quality overall, with some points to address as detailed below.

Strengths:

(1) Novel finding:<br /> The identification of a REM-specialized subpopulation within the lateral SuM-MS pathway and its specific role in social memory consolidation via the lateral SuM-MS-CA2 projection is a significant advance. It effectively complements the previously described direct SuM-CA2 pathway and supports a model of the SuM as a "REM-hub" routing information through dedicated downstream targets.

(2) Technical rigor:<br /> The combination of retrograde tracing, in vivo calcium imaging, single-unit identification via optrode recording, and temporally precise (REM-sleep-specific) optogenetic manipulation provides strong correlative and causal evidence.

(3) Appropriate controls:<br /> Behavioral experiments include crucial controls for optogenetic inhibition (GtACR1 group, NREM/Wake inhibition control, mCherry control), effectively ruling out nonspecific effects of light or timing.

Weaknesses:

(1) Figure titles/descriptions:<br /> For clarity, the authors should consider specifying the recording method in the figure titles or legends. For instance, Figure 2: "Bulk Ca2+ activity (fiber photometry) of lateral SuM-MS projecting neurons..." and Figure 3: "Single-unit activity patterns (optrode recordings) of lateral SuM-MS projecting neurons...".

(2) Statistical details:<br /> The use of "LSD post-hoc comparison" following ANOVA is noted. LSD is sensitive but can increase Type I error risk with multiple comparisons. Please justify its use or consider employing a more conservative post-hoc test (e.g., Tukey's or Bonferroni) for key comparisons like the social preference index in Figure 4h to bolster robustness.

(3) Data presentation:<br /> When reporting statistical results in figure legends (e.g., Figures 2d, 3i-k), please provide the specific test statistic values (e.g., F, χ²) and exact P values where possible, rather than only significance asterisks.

(4) Deepening mechanistic insight:<br /> The study excellently demonstrates "what" the circuit does. The discussion could be strengthened by further exploring "how" it might work. The finding that SuM-MS inhibition does not affect CA1 theta power is interesting and distinguishes it from other MS/hippocampal pathways. The suggestion of a theta-independent mechanism is plausible. Could the authors hypothesize more specifically? For example, might this circuit modulate reactivation events in the local CA2 network, neurochemical milieu (e.g., acetylcholine), or synapse-specific plasticity during REM sleep to facilitate social memory consolidation?

(5) Implications of regional heterogeneity:<br /> The functional divergence between lateral (90% REM-active) and medial SuM-MS neurons is intriguing. A brief discussion on the potential anatomical basis (differential inputs/outputs) and functional significance (e.g., integration of specific affective or arousal signals) of this subdivision would be valuable.

eLife Assessment

This paper describes an important advance in a 2D in vitro neural culture system to generate mature, functional, diverse, and geometrically consistent cultures, in a 384-well format with defined dimensions and the absence of the necrotic core, which persists for up to 300 days. The well-based format and conserved geometry make it a promising tool for arrayed screening studies. The evidence is compelling and provides a method for generating consistent 3D cortical layer-like organization.

Reviewer #3 (Public review):

Summary:

Kroeg et al. introduced a novel method for generating 3D cortical layer-like organization in hiPSC-derived models, achieving remarkably consistent topography within compact dimensions. Their approach involves seeding frontal cortex-patterned iPSC-derived neural progenitor cells into 384-well plates, which triggers the spontaneous assembly of adherent cortical organoids comprising diverse neuronal subtypes, astrocytes, and oligodendrocyte-lineage cells.

Strengths:

Compared with existing brain organoid models, these adherent cortical organoids exhibit enhanced reproducibility and improved cell viability during prolonged culture, thereby providing versatile opportunities for high-throughput drug discovery, neurotoxicological screening, and investigation of brain disorder pathophysiology. Overall, this study addresses an important and timely need for advancing current brain organoid systems.

Weaknesses:

Highlighting the consistency of differentiation across different cell lines and standardizing functional outputs are crucial to emphasize the broad future potential of this new organoid system for large-scale pharmacological screening. The authors provided a substantial amount of new data during the revision process to support the reproducibility of neuronal activity. The next step would be to leverage this platform for functional screening of chemical and genetic perturbations to identify new drug candidates.

Comments on revisions:

Most of my previous concerns were adequately addressed through the revision.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

Kroeg et al. describe a novel method for 2D culture human induced pluripotent stem cells (hiPSCs) to form cortical tissue in a multiwell format. The method claims to offer a significant advancement over existing developmental models. Their approach allows them to generate cultures with precise, reproducible dimensions and structure with a single rosette; consistent geometry; incorporating multiple neuronal and glial cell types (cellular diversity); avoiding the necrotic core (often seen in free-floating models due to limited nutrient and oxygen diffusion). The researchers demonstrate the method's capacity for long-term culture, exceeding ten months, and show the formation of mature dendritic spines and considerable neuronal activity. The method aims to tackle multiple key problems of in vitro neural cultures: reproducibility, diversity, topological consistency, and electrophysiological activity. The authors suggest their potential in high-throughput screening and neurotoxicological studies.

Strengths:

The main advances in the paper seem to be: The culture developed by the authors appears to have optimal conditions for neural differentiation, lineage diversification, and long-term culture beyond 300 days. These seem to me as a major strength of the paper and an important contribution to the field. The authors present solid evidence about the high cell type diversity present in their cultures. It is a major point and therefore it could be better compared to the state of the art. I commend the authors for using three different IPS lines, this is a very important part of their proof. The staining and imaging quality of the manuscript is of excellent quality.

We thank the reviewer for the positive comments on the potential of our novel platform to address key problems of in vitro neural culture, highlighting the longevity and reproducibility of the method across multiple cell lines.

Weaknesses:

(1) The title is misleading: The presented cultures appear not to be organoids, but 2D neural cultures, with an insufficiently described intermediate EB stage. For nomenclature, see: doi: 10.1038/s41586-022-05219-6. Should the tissue develop considerable 3D depth, it would suffer from the same limited nutrient supply as 3D models - as the authors point out in their introduction.

We appreciate the opportunity to clarify this point. We respectfully disagree that the cultures do not meet the consensus definition of an organoid. In fact, a direct quote from the seminal nomenclature paper referenced by the reviewer states: “We define organoids as in vitro-generated cellular systems that emerge by self-organization, include multiple cell types, and exhibit some cytoarchitectural and functional features reminiscent of an organ or organ region. Organoids can be generated as 3D cultures or by a combination of 3D and 2D approaches (also known as 2.5D) that can develop and mature over long periods of time (months to years).” (Pasca et al, 2022 doi10.1038/s41586-022-05219-6). Therefore, while many organoid types indeed have a more spherical or globular 3D shape, the term organoid also applies to semi-3D or nonglobular adherent organoids, such as renal (Czerniecki et al 2018, doi.org/10.1016/j.stem.2018.04.022) and gastrointestinal organoids (Kakni et al 2022, doi.org/10.1016/j.tibtech.2022.01.006). Accordingly, the adherent cortical organoids described in the manuscript exhibit self-organization to single radial structures consisting of multiple cell layers in the z-axis, reaching ~200um thickness (therefore remaining within the limits for sufficient nutrient supply), with consistent cytoarchitectural topology and electrophysiological activity, and therefore meet the consensus definition of an organoid.

(2) The method therefore should be compared to state-of-the-art (well-based or not) 2D cultures, which seems to be somewhat overlooked in the paper, therefore making it hard to assess what the advance is that is presented by this work.

It was not our intention to benchmark this model quantitatively against other culture systems. Rather, we have attempted to characterize the opportunities and limitations of this approach, with a qualitative contrast to other culture methods. Compared to stateof-the-art 2D neural network cultures, adherent cortical organoids provide distinct advantages in:

(1) Higher order self-organized structure formation, including segregation of deeper and upper cortical layers.

(2) Longevity: adherent cortical organoids can be successfully kept in culture for at least 1 year, whereas 2D cultures typically deteriorate after 8-12 weeks.

(3) Maturity, including the formation of dendritic mushroom spines and robust electrophysiological activity.

(4) Cell type diversity including a more physiological ratio of inhibitory and excitatory neurons (10% GAD67+/NeuN+ neurons in adherent cortical organoids, vs 1% in 2D neural networks), and the emergence of oligodendrocyte lineage cells.

On the other hand, limitations of adherent cortical organoids compared to 2D neural network cultures include:

(1) Culture times for organoids are much longer than for 2D cultures and the method can therefore be more laborious and more expensive.

(2) Whole cell patch clamping is not easily feasible in adherent cortical organoids because of the restrictive geometry of 384-well plates.

(3) Reproducibility is prominently claimed throughout the manuscript. However, it is challenging to assess this claim based on the data presented, which mostly contain single frames of unquantified, high-resolution images. There are almost no systematic quantifications presented. The ones present (Figure S1D, Figure 4) show very large variability. However, the authors show sets of images across wells (Figure S1B, Figure S3) which hint that in some important aspects, the culture seems reproducible and robust.

We made considerable efforts to establish quantitative metrics to assess reproducibility. We applied a quantitative scoring system of single radial structures at different time points for multiple batches of all three lines as indicated in Figure S1C. This figure represents a comprehensive dataset in which each dot represents the average of a different batch of organoids containing 10-40 organoids per batch. To emphasize this, we have adapted the graph to better reflect the breadth of the dataset. Additional quantifications are given in Figure S2 for progenitor and layer markers for Line 1 and in Figure 2 for interneurons across all three lines, showing relatively low variability. That being said, we acknowledge the reviewer’s concerns and have modified the text to reduce the emphasis of this point, pending more extensive data addressing reproducibility across an even broader range of parameters.

(4) What is in the middle? All images show markers in cells present around the center. The center however seems to be a dense lump of cells based on DAPI staining. What is the identity of these cells? Do these cells persist throughout the protocol? Do they divide? Until when? Addressing this prominent cell population is currently lacking.

A more comprehensive characterization of the cells in the center remains a significant challenge due to the high cell density hindering antibody penetration. However, dyebased staining methods such as DAPI and the LIVE/DEAD panel confirm a predominance of intact nuclei with very minimal cell death. The limited available data suggest that a substantial proportion of the cells in the center are proliferative neural progenitors, indicated by immunolabeling for SOX2 (Figure 2A,D;Figure S4C). Furthermore, we are currently optimizing the conditions to perform single cell / nuclear RNA sequencing to further characterize the cellular composition of the organoids.

(5) This manuscript proposes a new method of 2D neural culture. However, the description and representation of the method are currently insufficient. (a) The results section would benefit from a clear and concise, but step-by-step overview of the protocol. The current description refers to an earlier paper and appears to skip over some key steps. This section would benefit from being completely rewritten. This is not a replacement for a clear methods section, but a section that allows readers to clearly interpret results presented later.

We have revised the manuscript to include a more detailed step-by-step overview of the protocol.

(b) Along the same lines, the graphical abstract should be much more detailed. It should contain the time frames and the media used at the different stages of the protocol, seeding numbers, etc.

As suggested, we have adapted the graphical abstract to include more detail.

Reviewer #2 (Public review):

Summary:

In this manuscript, van der Kroeg et al have developed a method for creating 3D cortical organoids using iPSC-derived neural progenitor cells in 384-well plates, thus scaling down the neural organoids to adherent culture and a smaller format that is amenable to high throughput cultivation. These adherent cortical organoids, measuring 3 x 3 x 0.2 mm, self-organize over eight weeks and include multiple neuronal subtypes, astrocytes, and oligodendrocyte lineage cells.

Strengths:

(1) The organoids can be cultured for up to 10 months, exhibiting mature dendritic spines, axonal myelination, and robust neuronal activity.

(2) Unlike free-floating organoids, these do not develop necrotic cores, making them ideal for high-throughput drug discovery, neurotoxicological screening, and brain disorder studies.

(3) The method addresses the technical challenge of achieving higher-order neural complexity with reduced heterogeneity and the issue of necrosis in larger organoids. The method presents a technical advance in organoid culture.

(4) The method has been demonstrated with multiple cell lines which is a strength.

(5) The manuscript provides high-quality immunostaining for multiple markers.

We appreciate the reviewer’s acknowledgement of the strengths of this novel platform as a technical advance in organoid culture that reduces heterogeneity and shows potential for higher throughput experiments.

Weaknesses:

(1) Direct head-to-head comparison with standard organoid culture seems to be missing and may be valuable for benchmarking, ie what can be done with the new method that cannot be done with standard culture and vice versa, ie what are the aspects in which new method could be inferior to the standard.

In our opinion, it would be extremely difficult to directly compare methods. Most notably, whole brain organoids grow to large and irregular globular shapes, while adherent cortical organoids have a more standardized shape confined by the geometry of a 384well. Moreover, it was not our intention to benchmark this model quantitatively against other culture systems. Rather, we have attempted to characterize the opportunities and limitations of this approach, with a qualitative contrast to other culture methods, as addressed in response to comment 2 of Reviewer 1 above.

(2) It would be important to further benchmark the throughput, ie what is the success rate in filling and successfully growing the organoids in the entire 384 well plate?

Figure S1 shows the success rate of organoid formation and stability of the organoid structures over time. In addition, we have added the number of wells that were filled per plate.

(3) For each NPC line an optimal seeding density was estimated based on the proliferation rate of that NPC line and via visual observation after 6 weeks of culture. It would be important to delineate this protocol in more robust terms, in order to enable reproducibility with different cell lines and amongst the labs.

Figure S1 provides the relationship between proliferation rate and seeding density, allowing estimation of seeding densities based on the proliferation rate of the NPCs. However, we appreciate the reviewers' feedback and have modified the methods to provide more detail.

Reviewer #3 (Public review):

Summary:

Kroeg et al. have introduced a novel method to produce 3D cortical layer formation in hiPSC-derived models, revealing a remarkably consistent topography within compact dimensions. This technique involves seeding frontal cortex-patterned iPSC-derived neural progenitor cells in 384-well plates, triggering the spontaneous assembly of adherent cortical organoids consisting of various neuronal subtypes, astrocytes, and oligodendrocyte lineage cells.

Strengths:

Compared to existing brain organoid models, these adherent cortical organoids demonstrate enhanced reproducibility and cell viability during prolonged culture, thereby providing versatile opportunities for high-throughput drug discovery, neurotoxicological screening, and the investigation of brain disorder pathophysiology. This is an important and timely issue that needs to be addressed to improve the current brain organoid systems.

We thank the reviewer for highlighting the strengths of our novel platform. We appreciate that all three reviewers agree that the adherent cortical organoids presented in this manuscript reliably demonstrate increased reproducibility and longevity. They also commend its potential for higher throughput drug discovery and neurotoxicological/phenotype screening purposes.

Weaknesses:

While the authors have provided significant data supporting this claim, several aspects necessitate further characterization and clarification. Mainly, highlighting the consistency of differentiation across different cell lines and standardizing functional outputs are crucial elements to emphasize the future broad potential of this new organoid system for large-scale pharmacological screening.

We appreciate the feedback and have added more detail on consistency and standardization of functional outputs.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Minor points

(1) As the preprint is officially part of the eLife review, I have to remark that the preprint which is made available on bioarxiv, suffers from some serious compatibility or format problem: one cannot highlight sentences as in a regular PDF and when trying to copypaste sentences from it jumbled characters are copied to the clipboard.

The updated version of the paper on bioRxiv should not suffer from these compatibility issues.

(2) Since the paper is presenting a new method it should briefly describe how each step, including the hiPSC culture was done, the reference to an earlier publication in this case is not sufficient, and this practice is generally best to avoid for methods papers.

Each step in the culturing process has now been described in the methods.

(3) The EB stage is insufficiently described. The "2D - 3D - 2D" transitions should be clearly explained.

The methods section has been rewritten and expanded to include these processes in more detail.

(4) Is there one FACS sorting in the protocol, or multiple (additional at IPS culture)? What markers each? What is the motivation for sorting and purifying the neural progenitors? Was the culture impure? What was purity? What cell types are expected after sorting, and what is removed?

Only one FACS sorting step is performed at the NPC stage. This was added as an improvement to our original neural network protocol (Günhanlar et al 2018) to ensure consistency over different hiPSC source cell lines that can yield variable amounts of frontal cortical patterned NPCs. Positive sorting for neural lineage markers CD184 and CD24, and negative sorting for mesenchymal/neural crest CD217 and CD44 glial progenitor markers, according to Yuan et al 2011, ensures frontal-patterned cortical NPCs as confirmed for all batches by immunohistochemistry for SOX2, Nestin and FOXG1. We have added new text to the Methods section to clarify this more explicitly.

(5) Seeding protocol and parameters are insufficiently described, and from what I read they are poorly defined: "Specifically, the optimal seeding density was determined by visual inspection of the organoids between 28 to 42 days after seeding a range of cell densities in the 384-well plate wells." For a new method, precise, actionable instructions are needed. I may have overlooked those elsewhere, in this case, please clarify these sections.

The Methods section was rewritten and expanded to describe the methodology in greater detail with more actionable instructions.

(6) The timeline in Figure 1 is not clearly delineated; I found it hard to understand which figure corresponds to which stage (e.g. facs sorting is not mentioned in the first part of the results but it is part of Figure 1A, neural rosette formation can happen both before and after facs sorting, simply referring to rosettes is not clear). Later parts of the manuscript 
> clearly introduce the terms sorting and seeding in the context of this method, and how ages (days) refer to these time points.

Figure 1 was adapted to clarify the generation of Neural Progenitor Cells (NPCs) and subsequent seeding of NPCs to generate Adherent Cortical Organoids (ACOs).

(7) The authors define: "cortical organized defined as a single radial structure." This is not a commonly used definition of organoids, for nomenclature, please see: doi: 10.1038/s41586-022-05219-6 (Pasca et al 2022).

To clarify, the statement is not meant to reflect a definition of organoids in general, but rather the scoring of proper structure formation for Figure S1C. For discussion on nomenclature, see our response to point 1 of Reviewer 1 in the public review. We changed the wording to be more accurate.

(8) In Figure S1d, the authors write: "the fraction of structurally intact cultures decreased to 50%", but I'm looking at that graph there seems to be no notable decrease, but huge variability. The authors should quantify claims of decrease by linear regression and an R square. Variation within and the cross-cell lines seem to be large. Also, it is unclear if dots are corresponding to the same wells/plates, in other words: is this a longitudinal experiment? What is the overall success rate? How is success determined? Are there clear criteria? to the same wells/plates, in other words: is this a longitudinal experiment? What is the overall success rate? How is success determined? Are there clear criteria?

We agree with the reviewer that the claim on fraction of intact cultures decreasing over time to 50% is an overinterpretation due the large variability. We changed the wording in the manuscript to: While some later batches show moderately reduced success rates compared with the earliest batches, properly formed single-structure organoids were still obtained at 40–90% success across all examined time points (Figure S1C), indicating that long-term culture is feasible albeit with variable efficiency. The data are not longitudinal as each dot represents an endpoint of a different batch of organoids, totaling 18 independent batches across the three lines. We have clarified this in the figure legend. Success was defined at the well level as the presence of a single, continuous radial structure occupying the well, without obvious fragmentation or fusion events, as assessed by LIVE/DEAD that also confirmed viability. Wells were scored as successful only when the radial structure showed predominantly live signal with no large necrotic areas. Wells containing multiple radial structures, fused aggregates, or predominantly dead tissue were scored as unsuccessful.

(9) Figure s1c: the numbering to this panel should be swapped, because it is referenced after other panels in the text. The reference is confusing: "Plotting the interaction between proliferation and the amount of NPCs required to be seeded for the successful generation of adherent cortical organoids" - success is not present in this graph at all? How is that measured?

Figures S1C and S1D have been adapted to clarify the measure of ‘successful organoid formation’.

(a) The description of this plot is confusing: "The doubling time of the NPCs explains more than half the variation (r2 = 0.67) of the required seeding density." What else is there? I thought that this was the formula the authors suggested to determine seeding density, but it seems not. Or is "manual inspection" the determinant, and that seems to correlate with this metric?

Even though the rate of proliferation, measured as doubling time, is the main determinant of the seeding density, it is not the only determinant of the seeding density. For instance, intrinsic differences in differentiation potential could also play a role. Therefore, NPC lines with similar doubling times might still have slightly different optimal seeding densities. We have added clarification of this conclusion to the Results section.

(b) Seeding density is a key parameter in many in vitro differentiation and culture protocols. This importance however does not mean that this density is attributable to differences in cell proliferation rate. Alternatively, the amount of cells determines the amount of secreted molecules and cell-to-cell contacts.

Here, when we refer to the cell density, we specifically refer to the cell density needed to generate the ACO. We show that the most important contributor to the variation in ACO formation is the proliferation, measured here as the doubling time. We agree that there are other factors involved such as the secreted molecules, cell-to-cell contacts as well as the ability of a given NPC line to differentiate into a post-mitotic cell.

(c) Is it mentioned which cell line this experiment corresponds to?

The data in Figure S1D is from the 3 reported cell lines, as well as 2 clones from a fourth IPS cell line. This is detailed in the Methods section of the proliferation assay.

(d) Without a more detailed explanation, seeding density and doubling time could be independent variables.

These two variables are highly correlated as shown in Figure S1D, but it is true that there can be other variables that account for the observed variance, as discussed above in Point 9b.

(e) In this figure the success rate is not visible at all so I have no idea how the autors arrive at a conclusion about success rate.

We have adapted the figure legend to reflect which cell lines the dots in Fig. S1D represent. NPC lines can have substantial variation in proliferation rates. The figure reflects data of NPCs of 5 clones of 4 different hiPSC lines (as indicated in the Methods) with different proliferation rates. Also, the ACO success rate (operationally defined uniformly to the data shown in Fig. S1C) was also included.

(10) Figure 2: Clean spatial segregation seems to be a strength of the system and therefore I would recommend putting more of the relevant microscopy images to the main figure, which are now currently in Figure S4.

We have adapted Figure 2 accordingly, and included additional representative cortical layering images in Figure S4.

(11) The variability in interneuron content seems to be significant, as currently presented in the figure. However, this may be due to a special organization. It would first quantify in consecutive rings around the centers whether interneurons have a tendency to be enriched towards the center or the edge of the culture. Maybe this explains the variability that is currently present in Figure s5b.

We agree that spatial organization of interneurons could, in principle, contribute to variability. In our analysis, however, images were acquired from positions selected by a random sampling grid across the entire culture, rather than from specific central or peripheral regions. Each field contained on average 130.6 ± 16.1 NeuN+ nuclei, which provided a relatively large sampling volume per position. If interneurons were strongly enriched at the center or edge, we would expect systematic differences in interneuron fraction between fields assigned to central versus peripheral grid positions. We did not observe such a pattern in our dataset, suggesting that spatial organization is not the main driver of the observed variability.

(12) Because in previous figures it seems like there is considerable variability across individual cultures and images here are coming from separate cultures, please use different shapes of the points coming from different cultures/wells, to see if maybe there is a culture-to-culture difference that explains the variability present in the figure.

We have added different symbols per organoid for the interneuron quantifications and moved this quantification to main Figure 2.

(13) I believe it is currently the standard error of the mean which is displayed in the figure, which is not an appropriate representation for variability, or the reproducibility across individual data points. SEM quantifies the reproducibility of the mean, not the reproducibility of the individual data points, which matters here. Mean refers to the mean of this quantification experiment and therefore it's not a biological entity. A box plot showing the interquartile range besides the individual data points would be an accurate representation of the spread of the data.

We agree and have adapted the data, now in Figure 5, accordingly.

(14) Again, in general, the main figures should contain much more of the quantification, as opposed to just raw images.

Quantifications have been added in Figure 2 for the GAD67/NeuN for all cell lines as well as a time course quantification of GAD67/NeuN for 1 of the cell lines. In Figure 4, we have added excitatory and inhibitory synaptic quantifications.

(15) Figure 2F-I the location of the center of the rosette should be marked with a star so that the conclusion about the direction of processes can be established.

The suggested addition of a marker at the center of each rosette was evaluated but not implemented, because it reduced rather than improved figure clarity.

(16) Figure 3 b and c:

High magnification images of single cells, can't show changes in cell type morphology, and one cannot conclude that these cells are present in significant numbers across time. Zoomed-out images or quantification would be necessary for such a claim. The authors already have such images as presented in the next panels, so quantification without new experiments.
> I am uncertain about the T3 supplement here - do these images correspond to the same conditions?

(a) It is unclear to me why different markers are used in the different panels, namely why NG2 is not used in any of the other images.

NG2 was used at early developmental time points to show the presence of Oligodendrocyte Precursor Cells (OPCs). At later time points, the focus switched to MBP staining to indicate more mature oligodendrocyte lineage cells. Although NG2 and MBP are not in the same panels, the staining was performed for both antibodies at the same developmental time point (Day 119) as seen in Figure 3C and 3D.

(b) Color coding in Figure 3G is ambiguous; the use of two blues should be avoided, and the Sub-sub panels should be individually labeled for the color code.

We agree, and have now used different colors.

(c) It is unclear if the presence of the t3 molecule is part of the standard procedure or if it was a side experiment to enhance the survival of oligodendrocytes. Are there no oligodendrocytes without? How does T3 affect other cell types, and the general health and differentiation of the cultures?

Indeed, T3 is essential for oligodendrocyte formation. We did not observe obvious effects on the general health or differentiation potential of the cultures.

(d) Is the 2ng/ml t3 from day one to the final day?

Indeed, in the organoids cultured to study oligodendrocyte formation, T3 was added from Day 1. These details have now been clarified in the Methods and Results sections.

(17) Figure 4:

(a) Microscopy in this figure is high quality and very convincing about neural maturity.

(b) The term "cluster" should be avoided. Unclear what it means here, but my best guess is "cells in a frame of view." Cluster is used with a different meaning in electrophysiology.

This was adapted to ‘neurons in a field of view (FOV)’.

(c) Panel J: I assume each row corresponds to a single cell? Could this be clarified? Are these selected cells from each frame, or all active cells are represented?

Indeed, each row corresponds to a single cell, showing all active cells in the frame. This is now clarified in the legend.

(d) How many Wells do these data correspond to, and in which line it was measured?

As reported in the legend for Figure 5, these data correspond to 2 wells at Day 61 to which we have now added calcium imaging data from 3 wells from a different batch at Day 100. We have included in the legend that these recordings were from Line 1.

(e) Panels G to I, again, the use of standard error of the mean is inappropriate and misleading: looking at the error bar one must conclude that there is minimal variation, which is the exact opposite of the conclusions, when one would look at the variability of the raw data points.

As suggested, the graphs have been adapted as boxplots with interquartile ranges to highlight the distribution of data points.

(f) It is unclear how many neurons and how many total actively firing neurons are present in the videos analyzed

All neurons that were active in the field of view and showed at least one calcium event during the ~10 minute recording were included in the analysis. Using this method, we cannot comment on the proportion of neurons that were active from the total amount of neurons present, since the AAV virus we used does not transduce all neurons.

(g) This figure shows the strength of the method in achieving neural maturity and function. There seems to be that there is considerable activity in the neuronal cultures analyzed. To conclude how reliably the method leads to such mature cultures one would need to measure at least a dozen wells (even if with some simpler and low-resolution method). Concluding reproducibility from one or two hand-picked examples is not possible.

We agree with the reviewer that the number of wells used for calcium imaging analysis was limited. We are currently working on more advanced methods to increase the throughput of this analysis. However, we’ve now added another timepoint to the calcium imaging data in Figure 5 from an independent batch of 3 adherent cortical organoids, which demonstrates continued robust activity at Day 100, as well as Day 61.

Methods:

(1) Stem cell culture. The artist described that line 3 is grown on MEFs. Is this true for the other two lines, furthermore were they cultured in identical conditions?

Line 2 and 3 were not grown on MEFs. We specifically chose different sources of NPCs to reflect the robust nature of the differentiation protocol. We have recently also adapted the protocol from Line 3 NPCs to confirm that the protocol also works starting from hiPSCs grown in feeder-free conditions in StemFlex medium, by adapting NPC differentiation according to our recent publication in Frontiers in Cellular Neuroscience (Eigenhuis et al 2023).

(2) "NPCs were differentiated to adherent cortical organoids between passages 3 and 7 after sorting." Please clarify this sentence. I assume it refers to the first facs sorting of the protocol, but a section is not sufficiently detailed.

We have adapted the methods to clarify that the FACS purification step occurs at the NPC stage.

(3) I didn't fully understand: It seems to be that there are two steps of fact sorting involved, one after passage 3 and one after week 4. This should be represented in the graphical abstract of Figure 1.

As outlined above, there is only 1 FACS sorting step at NPC stage. We have adapted this in the Methods and in the graphical abstract.

(4) Neural differentiation: The authors write that optimal seeding density was determined by visual inspection of the organoids - this is.

We have clarified the Methods section to better explain the process of optimizing the seeding density for each NPC line to generate the ACOs.

(5) What does the following sentence mean: "Cells were refreshed every 2-3 days." Does it mean in replacement of the complete media? How much Media was added to the Wells?

This is a very good point that we have now clarified in the Methods, as full replenishment of media is neither feasible, nor desirable. From the total volume of 110 µl per well, 80 µl is taken out and replaced with 85 µl to compensate for evaporation.

(6) Calcium imaging: can the authors explain the decision to move the cultures one day before imaging into brainphys neural differentiation medium? In 3D organoid protocols, brainphys is gradually introduced to avoid culture shock (very different composition), and used for multiple months to enhance neural differentiation. For recording electrophysiological activity, artificial CSF is the most common choice.

Indeed, for whole cell recordings of 2D neural networks as performed in Günhanlar et al 2018, we used gradual transition to aCSF. For the current ACOs, we found that using BrainPhys from the start of organoid differentiation prevents structure formation, probably because of increased speed of maturation disrupting proliferation and organization of radial glia differentiation. However, by changing the media to BrainPhys just one day before recording (reflecting a gradual change as not all medium is fully replenished and easier than switching to aCSF during recording), we saw greatly improved neuronal activity.

(7) Statistical analysis : As I pointed out before, the standard error of the mean is not an appropriate metric to represent the variability of the data. It is meant to represent the variability of the estimated average. The following thought experiment should make it clear: I measured the expression of a gene in my system. 50 times I measured 0 and 50 times I measured 100. The average is 50, but of course it is a very bad representation of the data because no such data points exist with that value. Yet the standard error of the mean would be plus minus 5.

We have revised Figures 5C–5D to boxplots displaying the interquartile range with all individual data points overlaid, which more accurately represents the variability in the dataset.

Discussion

(1) The discussion focuses on human cortical development, however, the methods presented by the authors entail dissociation and replating through multiple stages not part of brain development. I see the approach as more valuable as a possibly reliable method that generates both diverse and mature neural cultures.

We have revised the Discussion to avoid explicitly invoking an in vitro recapitulation of human cortical development. Nevertheless, given that the NPCs from which the organoids originate exhibit frontal cortical identity, coupled with the timely emergence of cortical neuronal markers and rudimentary cortical layering, we are increasingly confident that the development of these cultures most likely mirrors that of the frontal cortex. To further substantiate this hypothesis, single-cell RNA sequencing experiments will be conducted in the future to provide additional insights.

(2) One of the major claims of the authors is that the method is very reproducible. However, there is almost no data on reproducibility throughout the paper. Mostly single, high magnification images are presented, which therefore represent a small region of a single well of a single batch of a single cell line. Based on the data presented it is not possible to evaluate the reproducibility of the method.

We agree that the original version did not sufficiently document reproducibility. To address this, we have refined and expanded our presentation of reproducibility data. The previous success-rate panel (original Figure S1D) has been moved and adapted as the new Figure S1C. In this updated version, each dot still represents the endpoint success rate of an independent batch, but dot size now scales with batch size (10–40 organoids), and the legend specifies the total numbers of organoids analyzed per line (line 1: n=248; line 2: n=70; line 3: n=70). Together with the distribution of success rates between ~40– 90% across multiple time points and three iPSC lines, this more detailed representation allows readers to directly assess the robustness of line-to-line and batch-to-batch performance. In addition, new time course quantifications of interneuron proportion (Figure 2G,H), synaptic marker densities (Figure 4H, I), and late-stage calcium imaging (Figure 5C,D,E) further demonstrate that key structural and functional read-outs show overlapping ranges across lines and independent differentiations, reinforcing that the method yields reproducible core phenotypes despite some biological variability.

(3) The data presented is very promising, and it suggests that the authors derived optimal conditions for neural differentiation and neural culture diversification. I am confident that the authors can show that reproducibility, at least in a practical sense (e.g. in wells that form a culture) is high.

Overall, this is a very promising and exciting work, that I am looking forward to reading in a mature manuscript.

Reviewer #2 (Recommendations for the authors):

(1) Direct head-to-head comparison with standard organoid culture seems to be missing and may be valuable for benchmarking, ie what can be done with the new method that cannot be done with standard culture and vice versa, ie what are the aspects in which new method could be inferior to the standard.

We have now more clearly elaborated the differences with other methods. As addressed in our response to point 2 of Reviewer 1 in the public reviews, there are several limitations and advantages to the adherent cortical organoids model listed as follows:

Advantages of adherent cortical organoids:

(1) Higher order self-organized structure formation, including segregation of deeper and upper cortical layers.

(2) Longevity: adherent cortical organoids can be successfully kept in culture for at least 1 year, whereas 2D cultures typically deteriorate after 8-12 weeks.

(3) Maturity, including the formation of dendritic mushroom spines and robust electrophysiological activity.

(4) Cell type diversity including a more physiological ratio of inhibitory and excitatory neurons (10% GAD67+/NeuN+ neurons in adherent cortical organoids, vs 1% in 2D neural networks), and the emergence of oligodendrocyte lineage cells.

On the other hand, limitations of adherent cortical organoids compared to 2D neural network cultures include:

(1) Culture times for organoids are much longer than for 2D cultures and the method can therefore be more laborious and more expensive.

(2) Whole cell patch clamping is not easily feasible in adherent cortical organoids because of the restrictive geometry of 384-well plates.

(2) It would be important to further benchmark the throughput, ie what is the success rate in filling and successfully growing the organoids in the entire 384 well plate?

We have addressed this question in the current version of Fig. S1C, in which multiple batches of organoids of all three lines were scored for their success rate. The graph reflects the proportion of properly formed organoids of +/- 400 seeded wells scored at different timepoints, in which each timepoint is a different batch. As mentioned in the response to Reviewer 1, we have also added data on the number of organoids seeded per line in the figure legend.

(3) For each NPC line an optimal seeding density was estimated based on the proliferation rate of that NPC line and via visual observation after 6 weeks of culture. It would be important to delineate this protocol in more robust terms, in order to enable reproducibility with different cell lines and amongst the labs.

As outlined in the response to Reviewer 1, we have clarified the Methods and Discussion sections on seeding density and proliferation rate.

Reviewer #3 (Recommendations for the authors):

Kroeg et al. have introduced a novel method to produce 3D cortical layer formation in hiPSC-derived models, revealing a remarkably consistent topography within compact dimensions. This technique involves seeding frontal cortex-patterned iPSC-derived neural progenitor cells in 384-well plates, triggering the spontaneous assembly of adherent cortical organoids consisting of various neuronal subtypes, astrocytes, and oligodendrocyte lineage cells. Compared to existing brain organoid models, these adherent cortical organoids demonstrate enhanced reproducibility and cell viability during prolonged culture, thereby providing versatile opportunities for high-throughput drug discovery, neurotoxicological screening, and the investigation of brain disorder pathophysiology. This is an important and timely issue that needs to be addressed to improve the current brain organoid systems. While the authors have provided significant data supporting this claim, several aspects necessitate further characterization and clarification. Particularly, highlighting the consistency of differentiation across different cell lines and standardizing functional outputs are crucial elements to emphasize the future broad potential of this new organoid system for large-scale pharmacological screening.

(1) Considering the emergence of astrocyte markers (GFAP, S100b) and upper layer neuron marker (CUX1) around Day 60, the overall differentiation speed is significantly faster compared to other forebrain organoid protocols. Are these accelerated sequences of neurodevelopment consistent across different hiPSC lines?

As shown in Fig. S5, astrocytes are present around Day 60 for all three lines. For comparison with other organoid protocols, an important consideration is that the timeline for these organoids starts at NPC plating, while for other protocols timing often starts from the hiPSC stage. We have clarified the timeline in the graphical abstract in Figure 1A and in the Methods.

(2) The calcium imaging results in Figure 4G were recorded at a single time point, Day 61, a relatively early time window compared to other forebrain organoid protocols (more than 100 days, PMID: 31257131; PMID: 36120104). Are the neurons in adherent cortical organoids functionally mature enough around Day 61? How consistent is this functional activity across different cell lines and independent differentiation batches?

As discussed above in Point 1, it is important to consider that the specified timeline starts from NPC plating. In analogy to 2D neural networks, robust neuronal activity can be observed after ~8 weeks in culture. In addition, we have now added calcium imaging data for an additional batch of organoids at Day 100 in Figure 5, which exhibit comparable levels of neuronal activity as observed on Day 61.

(3) Along the same line, Various cell types, such as oligodendrocytes and astrocytes, are believed to influence neuronal maturation. Therefore, longitudinal studies until the late stage are necessary to observe changes in electrophysiological activity based on the degree of neuronal maturation (at least two more later time points, such as 100 days and 150 days).

As described in the previous points, we have now included a Day 100 time point in the calcium imaging data, in addition to the recordings at Day 61 (Figure 5C-E).

(4) The authors assert that heterogeneity among organoids has been diminished using the human adherent cortical organoids protocol. However, there is inadequate quantitative data to prove the consistency of neuronal activities between different wells. Therefore, experiments quantifying the degree of heterogeneity between organoids, such as through methods like calcium imaging, are necessary to determine if neuron activity occurs consistently across each organoid well.

We agree with the review and have added several quantitative experiments: a) we’ve added another timepoint to the calcium imaging data in Figure 5 from an independent batch of 3 adherent cortical organoids, which demonstrates continued robust activity at day 100, as well as day 61; b) we added synapse quantification in Figure 4, and c) interneuron quantification in Figure 2. We are currently also pursuing high throughput measures of activity to assess the longitudinal activity of ACOs in a larger number of wells. This way we can more definitively quantify the time-dependent variance in organoid activity.

(5) Is this platform applicable to other functional measurements for neuronal activity, such as the MEA system? When observing the morphology of neurons formed in organoids, they appear to extend axons and dendrites in a consistent direction, suggesting a radial structure that demonstrates high reproducibility across wells. A culture system where neurons are arranged with such consistency in directionality could be highly beneficial for experiments utilizing the MEA system to assess parameters such as the speed of electrical activity transmission and stimulus-response. Therefore, there seems to be a need for a more detailed explanation of the utility of the structural characteristics of the culture system.

The ACO platform is indeed suitable for MEA recordings. We are in the process of engineering the required geometry using HD-MEA systems through specialized inserts to generate ACOs on MEA systems.

(6) In Figure 2E-I, authors suggest morphological diversity of GFAP+/S100b+ astrocyte, but the imaging data presented in Figure F-I is only based on GFAP immunoreactivity.

Since GFAP is also expressed in radial glial cells at this stage (Figure 2I), many fibrous astrocytes and interlaminar astrocytes are likely radial glial neural progenitor cells instead of astrocytes. It appears necessary to perform additional staining using astrocyte markers such as S100B or outer radial glia markers such as HOPX to demonstrate that the figure depicts subtype-specific morphologies of astrocytes.

In Figure 2M, we stained for GFAP and PAX6 to mark radial glia that look different than the astrocyte morphologies we describe in Figure 2J-L. We see a large overlap in GFAP and S100B staining in Figure 2I, in which most GFAP+ cells are double positive for S100B (yellow) that is more consistent with astrocyte maturation than radial glia. Furthermore, we have not seen PAX6 staining outside the dense edges of the center of the ACO.

(7) In Figure 4D, the axon appears to exhibit directionality. Additional explanation regarding the organization of the axon is necessary. Further research utilizing sparse staining to examine the morphology of single neurons seems warranted.

The polarized directionality of the axons is something we indeed have also noticed. We are looking into options to further investigate this intriguing property of the ACOs.

(8) Figure 1E-F only showed cell viability in the early stages around Day 40-50. To demonstrate the superior long-term viability of ACO culture, it appears necessary to illustrate the ratio of dead cells to live cells over the course of a time course.

Figure S1B shows LIVE/DEAD staining for ACOs of all three lines, revealing minimal DEAD staining at Day 56. A longitudinal time course experiment was not performed, however the line- and batch-specific quantifications over developmental timepoints in Figure S1C provide an indication of the robust long-term viability of the ACOs.

eLife Assessment

This study presents a valuable finding on the neural representation of time from two distinct egocentric and allocentric reference frames. The evidence is solid and largely supports the hypothesis, with one caveat that the task differences could impact the observed effects. The work will be of interest to cognitive neuroscientists working on the perception and memory of time.

Reviewer #2 (Public review):

Summary:

Xu et al. used fMRI to examine the neural correlates associated with retrieving temporal information from an external compared to internal perspective ('mental time watching' vs. 'mental time travel'). Participants first learned a fictional religious ritual composed of 15 sequential events of varying durations. They were then scanned while they either (1) judged whether a target event happened in the same part of the day as a reference event (external condition); or (2) imagined themselves carrying out the reference event and judged whether the target event occurred in the past or will occur in the future (internal condition). Behavioural data suggested that the perspective manipulation was successful: RT was positively correlated with sequential distance in the external perspective task, while a negative correlation was observed between RT and sequential distance for the internal perspective task. Neurally, the two tasks activated different regions, with the external task associated with greater activity in the supplementary motor area and supramarginal gyrus, and the internal condition with greater activity in default mode network regions. Of particular interest, only a cluster in the posterior parietal cortex demonstrated a significant interaction between perspective and sequential distance, with increased activity in this region for longer sequential distances in the external task but increased activity for shorter sequential distances in the internal task. Only a main effect of sequential distance was observed in the hippocampus head, with activity being positively correlated with sequential distance in both tasks. No regions exhibited a significant interaction between perspective and duration, although there was a main effect of duration in the hippocampus body with greater activity for longer durations, which appeared to be driven by the internal perspective condition. On the basis of these findings, the authors suggest that the hippocampus may represent event sequences allocentrically, whereas the posterior parietal cortex may process event sequences egocentrically.

Strengths:

The topic of egocentric vs. allocentric processing has been relatively under-investigated with respect to time, having traditionally been studied in the domain of space. As such, the current study is timely and has the potential to be important for our understanding of how time is represented in the brain in the service of memory. The study is well thought out and the behavioural paradigm is, in my opinion, a creative approach to tackling the authors' research question. A particular strength is the implementation of an imagination phase for the participants while learning the fictional religious ritual. This moves the paradigm beyond semantic/schema learning and is probably the best approach besides asking the participants to arduously enact and learn the different events with their exact timings in person. Importantly, the behavioural data point towards successful manipulation of internal vs. external perspective in participants, which is critical for the interpretation of the fMRI data. The use of syllable length as a sanity check for RT analyses as well as neuroimaging analyses is also much appreciated.

Suggestions:

The authors have satisfactorily addressed my last remaining suggestion.

Author response:

The following is the authors’ response to the previous reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

In this fMRI study, the authors wished to assess neural mechanisms supporting flexible temporal construals. For this, human participants learned a story consisting of fifteen events. During fMRI, events were shown to them, and participants were instructed to consider the event from "an internal" or from "an external" perspective. The authors found distinct patterns of brain activity in the posterior parietal cortex (PPC) and anterior hippocampus for the internal and the external viewpoint. Specifically, activation in the posterior parietal cortex positively correlated with distance during the external-perspective task, but negatively during the internal-perspective task. The anterior hippocampus positively correlated with distance in both perspectives. The authors conclude that allocentric sequences are stored in the hippocampus, whereas egocentric sequences are supported by the parietal cortex.

We thank the reviewer for the accurate summary of our study.

Strengths:

The research topic is fascinating, and very few labs in the world are asking the question of how time is represented in the human brain. Working hypotheses have been recently formulated, and the work tackles them from the perspective of construals theory.

We appreciate the reviewer's positive and encouraging comments.

Weaknesses:

Although the work uses two distinct psychological tasks, the authors do not elaborate on the cognitive operationalization the tasks entail, nor the implication of the task design for the observed neural activation.

We thank the reviewer for bringing this issue to our attention. In the revised manuscript, we have added a paragraph to the Discussion acknowledging this potential limitation of the study. Please see our response below.

Reviewer #1 (Recommendations for the authors):

Overall, I thank the authors for providing clear responses and much-needed detail on their original work, which enables a better understanding of their perspectives. I still have some detailed questions about the reported work, which I provide below. It could help clarify the work for a more general audience and its replicability by the community.

We thank the reviewer for their positive evaluation of our previous revisions.

Main general concern:

I have one remaining core concern, which I distill as being a very different take on the usefulness of task design with neuroimaging. This concern follows from the authors' response to my original comment, which suggested possible confounds in fMRI data analysis and interpretation, as differences in task design and behavioral outcomes were not incorporated in the analytical approach.

The authors confirmed that "there is a substantial difference between the two tasks" but argue that these differences are not relevant seing that "the primary goal of this study was not to directly compare these tasks to isolate a specific cognitive component " However, the authors do perform such contrasts in their analysis (e.g. p. 10: "We first directly contrasted the activity level between external- and internal-perspective tasks in the time window of...") and build inferences on brain activation from them (e.g., p. 10: "Compared with the internal-perspective task, the externalperspective task specifically activated the...").

To clarify, my original concern was not about comparing neural activity in response to the two tasks but about the brain activity generated by two distinct tasks, which aim to reveal fundamentally distinct neural processes. The authors' response raises several concerns about the theoretical, methodological and empirical foundation of the work that are beyond the scope of a single empirical study and too long to detail here. Cognitive neuroscience relies on tasks to infer neural processes; this is the fertile and essential ground for using behavior in neuroscience to get to a mechanistic understanding of brain functions (e.g., Krakauer et al., 2017). In short, task design is fundamental because it shapes what neural processes are being investigated. Any inferences about brain activity recorded while a participant performs a task result from manipulated variables that should be under the control of the experimenter. Acknowledging that two tasks are distinct is acknowledging that different (neural) processes may govern their resolution. My initial remark was meant to highlight that, from basic signal detection theory, a same/different task and a temporal order task may not yield the same kind of basic biases and decision-making processes; these are far below and more basic than the posited sophisticated representations herein (construals, perspective taking).

In short, the general approach is far coarser than the level of interpretational granularity being pushed forward in the paper would suggest.

We greatly appreciate the reviewer’s comments and agree that this is a very fair point. We acknowledge that the two tasks differ in their underlying decision-making processes. In the revised manuscript, we have added a paragraph at the end of the Discussion to explicitly acknowledge this limitation and to outline possible avenues for future research (Page 23).

“One limitation of the present study is that the external- and internal-perspective tasks differed not only in the type of perspective-taking they were intended to elicit, but also in their underlying decision-making processes. The external-perspective task explicitly required participants to compare two events with respect to external temporal landmarks and judge whether they occurred in the same or different parts of the day (i.e., a same/different judgment), whereas the internalperspective task explicitly required participants to project themselves into a reference event and judge whether the target event occurred in the future or the past relative to that reference (i.e., a temporal-order judgment). This task design ensured that participants adopted two distinct perspectives on the event series, but at the expense of coherence in the cognitive operations required to make the two types of judgments. One alternative approach would be to more closely align the response demands of the two tasks by drawing on McTaggart’s (1908) A-series and Bseries distinction: in the external-perspective task, participants could judge whether the target event occurred before or after the reference event (i.e., a before/after judgment), whereas in the internal-perspective task they could judge whether the target event occurred in the past or future relative to the reference event (i.e., a past/future judgment). Although such a design would improve coherence in the underlying decision-making processes (i.e., both are temporal-order judgments), it would reduce experimental control over the perspective-taking manipulation. For example, before/after judgments could still be made from an internal perspective. Future studies are therefore needed to determine whether findings obtained from these two task designs converge.”

Additional clarifications:

Intro/theory

In this revised MS, the authors provided some clarifications of their theoretical perspective in the introduction. From my standpoint, the motivation remains insufficiently precise for a scientific report. Some theoretical aspects, such as construals or perspective taking remain evasive in relation to ego and allocentric representations. A couple of paragraphs dedicated to explaining what the authors mean precisely when using these terms would greatly help to situate the validity of the working hypothesis. In the absence of clear definitions, it remains difficult to evaluate what is being tested. For instance, what do the authors mean by "time construal"? How is a time construal the same or not as a "temporal distance" or a "temporal sequence"? This would greatly help the readership.

Additionally, some assertions are not clearly identified or fairly attributed. For instance, the assertion that EST provides a means to spatialize time is the authors' point of view or interpretation of this work, not an original proposition of the theory. Another example is McTaggart's metaphysics on time series (in the ontology of time in physics) "echoed" in linguistics; it has effectively been proposed and popularized by L. Boroditskty. The prospective and retrospective views of time should not be attributed to Tsao et al but to Hicks or Block in the 70's, who studied the psychology of time in humans.

We sincerely thank the reviewer for this criticism, which prompted us to clarify the relevant concepts in our manuscript. In the revised version, we made the following three main changes to the Introduction.

In the second paragraph of the Introduction (page 3), we clarify that event segmentation theory is independent of, but related to, the spatial construal of time hypothesis. We also clarify what we mean by time construals and explain that the two temporal components—duration and sequence—can be represented within such time construals, rather than constituting time construals themselves. These revisions were intended to prevent potential misunderstandings for the reader. In addition, we incorporated Boroditsky’s contributions relevant to this framework:

“One solution, which might be unique to humans, is to conceptualize time in terms of space (i.e., the spatial construal of time; e.g., Clark, 1973; Traugott, 1978; Lakoff & Johnson, 1980). Within this framework, time is usually first segmented into events—the basic temporal entities that observers conceive as having a beginning and an end (Zacks & Tversky, 2001). These temporal entities are then ordered in space, such that events occurring at different times can be maintained in working memory, allowing them to be flexibly accessed from different perspectives and easily referenced during communication (e.g., Casasanto & Boroditsky, 2008; Núñez & Cooperrider, 2013; Bender & Beller, 2014; Abrahamse et al., 2014; Figure 1A). The two core temporal components—duration and sequence—can be readily represented in such time construals.”

In the third paragraph of the Introduction (pages 3-4), we acknowledge the contributions of earlier behavioral studies on prospective and retrospective timing by citing the work suggested by the reviewer (Block & Zakay, 1997), which indicates that two distinct cognitive systems underlie timing processes. These behavioral findings converge with the conclusions of more recent neuroimaging studies:

“Unlike prospective timing tracking the continuous passage of time, durations in time construals are event-based (Sinha & Gärdenfors, 2014): the interval boundaries are constituted by events, and the event durations reflect their span (Figure 1A). Accumulating evidence suggests that distinct cognitive systems underlie these two types of duration (e.g., Block & Zakay, 1997). The motor and attentional system—particularly the supplementary motor area—has been associated with prospective timing (e.g., Protopapa et al., 2019; Nani et al., 2019; De Kock et al., 2021; Robbe, 2023), whereas the episodic memory system—particularly the hippocampus—is considered to support the representation of duration embedded within an event sequence (e.g., Barnett et al., 2014; Thavabalasingam et al., 2018; see also the comprehensive review by Lee et al., 2020).”

Block, R. A., & Zakay, D. (1997). Prospective and retrospective duration judgments: A meta-analytic review. Psychonomic Bulletin & Review, 4(2), 184-197.

In the fifth paragraph of the Introduction (page 5), we added a sentence to clarify the relationship between allocentric and egocentric reference frames and perspective taking:

“However, the neural mechanisms that enable the brain to generate distinct construals of an event sequence remain largely unknown. Valuable insights may be drawn from research in the spatial domain, which posits the existence of stable allocentric representations that are independent of viewpoint, from which variable egocentric representations corresponding to different perspectives can be generated.”

Methods:

While more detail is provided in the Methods, some additional detail would be helpful to enable the replication of this work. For instance,

- The table reports a sequence of phrases with assigned durations. Are the event phrases actual sentences given to participants? If so, how were participants made aware of the duration of the events, seeing that these sentence parts do not provide time information?

We apologize that we did not make this clear. The full text used during the reading phase of learning has already been provided in Figure 1—source data 1, which includes the information about event durations. In the revised manuscript, we now explicitly refer to this information in the Methods section (page 38): In the reading phase, participants read a narrative describing the whole ritual on a computer screen twice (Figure 1—source data 1).

- One of my original questions was about the narrative. In the Methods section, the authors state that participants read a text. Providing the full text would be helpful, also as a sanity check for sequentiality.

As clarified in the previous response, the texts are provided in Figure 1—source data 1, which illustrates the texts for both even- and odd-numbered participants.

- In the imagination phase, the authors introduce proportionality between imagination and experience (p. 37). What scale was used? What motivated it?

We thank the reviewer for bringing this issue to our attention. In this study, participants did not directly experience the events; instead, they learned the event information through narrative reading or imagination to ensure experimental control and efficiency. As clarified in the Methods section, the ratio between imagination duration and actual event duration was 30 seconds to 1 hour. In the revised manuscript, we have further explained our motivation for this design choice (page 39):

Here, we let participants learn the event information through narrative reading or imagination. Compared to learning through actual experience, this approach prioritizes experimental control and efficiency. The timing of the events is compressed, akin to the process of retrospectively recalling our experiences, in which we mentally traverse events without requiring the actual time they originally took. However, future studies may be needed to investigate whether the encoding of events from first- and second-hand experience differs.

Results:

- p. 10: the interpretation of the data on chunking and boundary effects should be properly referenced to e.g. Davachi's published work.

We thank the reviewer for highlighting Davachi’s important work on event boundaries. We have appropriately cited these studies in the revised manuscript (page 10), as reflected in the following passage: This pattern can be interpreted as a categorical effect: sequential distances within the same part of the day were perceived as shorter (i.e., a chunking effect), whereas distances spanning different parts of the day were perceived as longer (i.e., a boundary effect). Similar boundary- or chunking-related effects on event cognition have been reported in previous studies (e.g., Ezzyat & Davachi, 2011; DuBrow & Davachi, 2013; Radvansky & Zacks, 2017).

Ezzyat, Y., & Davachi, L. (2011). What constitutes an episode in episodic memory?. Psychological Science, 22(2), 243-252.

DuBrow, S., & Davachi, L. (2013). The influence of context boundaries on memory for the sequential order of events. Journal of Experimental Psychology: General, 142(4), 1277.

Radvansky, G. A., & Zacks, J. M. (2017). Event boundaries in memory and cognition. Current Opinion in Behavioral Sciences, 17, 133-140.

Reviewer #2 (Public review):

Summary:

Xu et al. used fMRI to examine the neural correlates associated with retrieving temporal information from an external compared to internal perspective ('mental time watching' vs. 'mental time travel'). Participants first learned a fictional religious ritual composed of 15 sequential events of varying durations. They were then scanned while they either (1) judged whether a target event happened in the same part of the day as a reference event (external condition); or (2) imagined themselves carrying out the reference event and judged whether the target event occurred in the past or will occur in the future (internal condition). Behavioural data suggested that the perspective manipulation was successful: RT was positively correlated with sequential distance in the external perspective task, while a negative correlation was observed between RT and sequential distance for the internal perspective task. Neurally, the two tasks activated different regions, with the external task associated with greater activity in the supplementary motor area and supramarginal gyrus, and the internal condition with greater activity in default mode network regions. Of particular interest, only a cluster in the posterior parietal cortex demonstrated a significant interaction between perspective and sequential distance, with increased activity in this region for longer sequential distances in the external task but increased activity for shorter sequential distances in the internal task. Only a main effect of sequential distance was observed in the hippocampus head, with activity being positively correlated with sequential distance in both tasks. No regions exhibited a significant interaction between perspective and duration, although there was a main effect of duration in the hippocampus body with greater activity for longer durations, which appeared to be driven by the internal perspective condition. On the basis of these findings, the authors suggest that the hippocampus may represent event sequences allocentrically, whereas the posterior parietal cortex may process event sequences egocentrically.

We sincerely appreciate the reviewers for providing an accurate, comprehensive, and objective summary of our study.

Strengths:

The topic of egocentric vs. allocentric processing has been relatively under-investigated with respect to time, having traditionally been studied in the domain of space. As such, the current study is timely and has the potential to be important for our understanding of how time is represented in the brain in the service of memory. The study is well thought out and the behavioural paradigm is, in my opinion, a creative approach to tackling the authors' research question. A particular strength is the implementation of an imagination phase for the participants while learning the fictional religious ritual. This moves the paradigm beyond semantic/schema learning and is probably the best approach besides asking the participants to arduously enact and learn the different events with their exact timings in person. Importantly, the behavioural data point towards successful manipulation of internal vs. external perspective in participants, which is critical for the interpretation of the fMRI data. The use of syllable length as a sanity check for RT analyses as well as neuroimaging analyses is also much appreciated.

We thank the reviewer for the positive and encouraging comments.

Suggestions:

The authors have done a commendable job addressing my previous comments. In particular, the additional analyses elucidating the potential contribution of boundary effects to the behavioural data, the impact of incorporating RT into the fMRI GLMs, and the differential contributions of RT and sequential distance to neural activity (i.e., in PPC) are valuable and strengthen the authors' interpretation of their findings.

My one remaining suggestion pertains to the potential contribution of boundary effects. While the new analyses suggest that the RT findings are driven by sequential distance and duration independent of a boundary effect (i.e., Same vs. Different factor), I'm wondering whether the same applies to the neural findings? In other words, have the authors run a GLM in which the Same vs. Different factor is incorporated alongside distance and duration?

We thank the reviewer for their positive evaluation of our previous revisions and are pleased that the additional analyses adequately address the boundary effects in the behavioral data and the RT effects in the neural data.

With respect to boundary effects in the neural data, we followed the reviewer’s suggestion and constructed a more complex GLM that incorporated the Same/Different part of the day as an additional regressors modulating the target events. Importantly, the same PPC region continued to show an interaction effect between Task Type and Sequential Distance. We have added this important control analysis in our revised manuscript (Pages 13–14):

“To further assess whether the observed PPC reactivation can be attributed to boundary or chunking effects introduced by the Parts of the Day, as well as other behavioral outputs, we performed an additional control analysis. Using a more complex first-level model, we included two extra regressors modulating the target events in both internal- and external-perspective tasks, alongside Sequential Distance and Duration: (1) Same/Different parts of the day (coded as 1/−1) and (2) Future/Past (coded as 1/−1). Even with these additional controls, the same PPC region remained the strongest area across the entire brain, showing an interaction effect between Task Type and Sequential Distance, although the cluster size was slightly reduced (voxel-level p < 0.001; clusterlevel FWE-corrected p = 0.054).”

eLife Assessment

The study provides important mechanistic insight into the transcriptional control of γδT17 development, elegantly demonstrating how HEB and Id3 act sequentially and cooperatively to regulate γδT17 cell specification and maturation. The study provides compelling evidence that advances the understanding of E-Id protein dynamics in thymic T cell specification. The work is comprehensive, technically rigorous, and conceptually clear, and will be of interest to immunologists, developmental biologists, and those studying the molecular underpinnings of physiological outcomes.

Reviewer #1 (Public review):

The authors use Flow cytometry and scRNA seq to identify and characterize the defect in gdT17 cell development from HEB f/f, Vav-icre (HEB cKO) and Id3 germline-deficient mice. HEB cKO mice showed defects in the gdT17 program at an early stage, and failed to properly upregulate expression of Id3 along with other genes downstream of TCR signaling. Id3KO mice showed a later defect in maturation. The results together indicate HEB and Id3 act sequentially during gdT17 development. The authors further showed that HEB and TCR signaling synergize to upregulate Id3 expression in the Scid-adh DN3-like T cell line. Analysis of previously published Chip-seq data revealed binding of HEB (and Egr2) at overlapping regulatory regions near Id3 in DN3 cells.The study provides insight into mechanisms by which HEB and Id3 act to mediate gdT17 specification and maturation. The work is well performed and clearly presented.

Comments on revisions:

The authors have answered all of my questions. I am strongly supportive of the revised work.

Reviewer #2 (Public review):

Summary:

The manuscript by Selvaratnam et al. defines how the transcription factor HEB integrates with TCR signaling to regulate Id3 expression in the context of gdT17 maturation in the fetal thymus. Using conditional HEB ablation driven by Vav Cre, flow cytometry, scRNA-seq, and reanalysis of ChIP-seq data the authors, provide evidence for a sequential model in which HEB and TCR-induced Egr2 cooperatively upregulate Id3, enabling gdT17 maturation and limiting diversion to the ab lineages. The work provides an important mechanistic insight into how the E/ID-protein axis coordinates gd T cell specification and effector maturation.

Strengths include:

(1) The proposed model that HEB primes, TCR induces, and Id3 stabilizes gdT17 cells in embryonal development is elegant and consistent with the findings.

(2) The choice of animal models and the study of a precise developmental window.

(3) The cross-validation of flow, scRNA-seq, and ChIP-seq reanalyses strengthens the conclusions.

(4) The study clarifies the dual role of Id3, first as an HEB-dependent maturation factor for gdT17 cells, and as a suppressor of diversion to the ab lineages.

Comments on revisions:

In this revised version of their manuscript the authors have effectively addressed all of my previous concerns. In its current form the study represents a significant advancement in our understanding of how the transcription factor HEB integrates with TCR signaling to regulate Id3 expression in the context of gdT17 maturation in the fetal thymus. In this revised version of their manuscript the authors have effectively addressed all of my previous concerns. In its current form the study represents a significant advancement in our understanding of how the transcription factor HEB integrates with TCR signaling to regulate Id3 expression in the context of gdT17 maturation in the fetal thymus.

Reviewer #3 (Public review):

Summary:

The authors of this manuscript have addressed a key concept in T cell development: how early thymus gd T cells subsets are specified and the elements that govern gd T17 versus other gd T cell subset or ab T cell subsets are specified. They show that the transcriptional regulator HEB/Tcf12 plays a critical role in specifying the gd T17 lineage and, intriguingly that it up regulates the inhibitor Id3 which is later required for further gd T17 maturation.

Strengths:

The conclusions drawn by the authors are amply supported by a detailed analysis of various stages of T cell maturation in WT and KO mouse strains at the single cell level both phenotypically, by flow cytometry for various diagnostic surface markers, and transcriptionally, by single cell sequencing. Their conclusions are balanced and well supported by the data and citations of previous literature.

Weaknesses:

I actually found this work to be quite comprehensive.

Comments on revisions:

Nothing to add here. The authors were very thorough in their original submission, and all minor issues identified have been addressed to my satisfaction.

Author response:

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

We thank the reviewers for their enthusiasm and insightful suggestions. Our responses to specific concerns and questions are detailed below.

Public Reviews:

Reviewer #1 (Public review):

The authors use Flow cytometry and scRNA seq to identify and characterize the defect in gdT17 cell development from HEB f/f, Vav-icre (HEB cKO), and Id3 germline-deficient mice. HEB cKO mice showed defects in the gdT17 program at an early stage, and failed to properly upregulate expression of Id3 along with other genes downstream of TCR signaling. Id3KO mice showed a later defect in maturation. The results together indicate HEB and Id3 act sequentially during gdT17 development. The authors further showed that HEB and TCR signaling synergize to upregulate Id3 expression in the Scid-adh DN3-like T cell line. Analysis of previously published Chi-seq data revealed binding of HEB (and Egr2) at overlapping regulatory regions near Id3 in DN3 cells.

The study provides insight into mechanisms by which HEB and Id3 act to mediate gdT17 specification and maturation. The work is well performed and clearly presented. We only have minor comments.

Reviewer #2 (Public review):

Summary:

The manuscript by Selvaratnam et al. defines how the transcription factor HEB integrates with TCR signaling to regulate Id3 expression in the context of gdT17 maturation in the fetal thymus. Using conditional HEB ablation driven by Vav Cre, flow cytometry, scRNA-seq, and reanalysis of ChIP-seq data the authors, provide evidence for a sequential model in which HEB and TCR-induced Egr2 cooperatively upregulate Id3, enabling gdT17 maturation and limiting diversion to the ab lineages. The work provides an important mechanistic insight into how the E/ID-protein axis coordinates gd T cell specification and effector maturation.

Strengths include:

(1) The proposed model that HEB primes, TCR induces, and Id3 stabilizes gdT17 cells in embryonal development is elegant and consistent with the findings.

(2) The choice of animal models and the study of a precise developmental window.

(3) The cross-validation of flow, scRNA-seq, and ChIP-seq reanalyses strengthens the conclusions.

(4) The study clarifies the dual role of Id3, first as an HEB-dependent maturation factor for gdT17 cells, and as a suppressor of diversion to the ab lineages.

Weaknesses:

(1) The ChIP-seq reanalysis indicates overlapping HEB, E2A, and Egr2 peaks ~60 kb upstream of Id3. Given that the Egr2 data are not generated using the same thymocyte subsets, some form of validation should be considered for the co-binding of HEB and Egr2, potentially ChIP-qPCR in sorted gdT17 progenitors.

We agree that this is a valid concern and continue to work on confirming the mechanism from several other angles. Validating HEB/E2A and Egr2 co-binding in gdT17 cell progenitors by ChIP-qPCR would/will be a very precise and definitive experiment, but it will be very challenging to perform, in part due to the low numbers of gdT17 precursors in the fetal thymus (note the y-axis scales in Fig. 1F, J). As a complementary approach, we have analyzed additional ChIP-seq data for HEB/E2A binding in Rag2^-/- DN3 cells retrovirally transduced with the KN6 gdTCR cultured with stroma expressing the weak KN6 ligand T10 for 4 days. This analysis revealed that the binding of HEB/E2A on those sites persisted after weak gdTCR signaling, strengthening the likelihood that concurrent binding of HEB/E2A and Egr2 occurs during this developmental transition. We noted that HEB/E2A binding was slightly dampened in Rag2^-/- DN3 + gdTCR cells relative to Rag2^-/- DN3 cells, consistent with the induction of Id3 and subsequent Id3-mediated disruption of E protein binding. We also located HEB/E2A and Egr binding sites in close proximity in the two regions that shared peaks between HEB/E2A and Egr2 analyses (HE1 and HE2), in line with the potential participation of these two transcription factors in an enhanceosome binding complex.

Furthermore, we examined the chromatin landscape of the Id3 locus by sorting WT DN3 and DN4 cells, as well as Rag2^-/- DN3 cells to provide a genuine pre-selection context, and performing ATAC-seq (Figure 7–suppl 7A). Given the known ability of E2A and HEB to induce chromatin remodeling, we also examined accessibility in DN3 and DN4 cells from HEB cKO mice. Alignment of ATAC-seq and ChIP-seq peaks in the Id3 locus revealed accessibility of HE1 and HE2 in Rag2^-/-, WT DN3, and WT DN4 cells. However, accessibility of HE1 and HE2 was dampened in HEB cKO cells, especially at the DN3 stage, suggesting that HEB may be involved in remodeling the Id3 locus, resulting in a poised state that enables TCR-dependent transcription factors to induce Id3 proportionally to TCR signal strength. These data are now presented as a new “Figure 7 – figure supplement 1” with corresponding Results, Discussion, and Methods updates.

Our next story will be focused on a finer dissection of the Id3 cis-regulatory elements and their combinatorial regulation by HEB/E2A and other transcription factors, and how they relate to specific signaling pathways. For this study, we will modify the language regarding Egr2 to reflect the open questions that still remain to be addressed.

(2) E2A expression is not affected in HEB-deficient cells, raising the question of partial compensation, a point that should be specifically discussed.

This confounding factor is always an issue with E proteins. We have now added a section to the discussion that highlights previous literature and relates it to our findings.

(3) All experiments are done at E18, when fetal gdT17 development predominates. The discussion could address whether these mechanisms extend to neonatal or adult gdT17 subsets.

In our 2017 paper (PMID 29222418) we showed that HEB cKO mice have defects in the production of functional gdT17 cells in fetal and neonatal thymus and in the adult periphery (in lungs and spleen). While the adult thymus does not support the development of fully functional innate gd T cells, it does contain gdTCR+ cells that have activated the Sox-Maf-Rorc network (Yang 2023, PMID 37815917). It will be very interesting to assess the impact of HEB loss on these cells, and we are actively pursuing this goal. For now, we will add a paragraph to the discussion addressing what we know from previous work and what is yet to be learned.

Reviewer #3 (Public review):

Summary:

The authors of this manuscript have addressed a key concept in T cell development: how early thymus gd T cell subsets are specified and the elements that govern gd T17 versus other gd T cell subsets or ab T cell subsets are specified. They show that the transcriptional regulator HEB/Tcf12 plays a critical role in specifying the gd T17 lineage and, intriguingly, that it upregulates the inhibitor Id3, which is later required for further gd T17 maturation.

Strengths:

The conclusions drawn by the authors are amply supported by a detailed analysis of various stages of T cell maturation in WT and KO mouse strains at the single cell level, both phenotypically, by flow cytometry for various diagnostic surface markers, and transcriptionally, by single cell sequencing. Their conclusions are balanced and well supported by the data and citations of previous literature.

Weaknesses:

I actually found this work to be quite comprehensive. I have a few suggestions for additional analyses the authors could explore that are unrelated to the predominant conclusions of the manuscript, but I failed to find major flaws in the current work.

I note that HEB is expressed in many hematopoietic lineages from the earliest progenitors and throughout T cell development. It is also noteworthy that abortive gamma and delta TCR rearrangements have been observed in early NK cells and ILCs, suggesting that, particularly in early thymic development, specification of these lineages may have lower fidelity. It might prove interesting to see whether their single-cell sequencing or flow data reveal changes in the frequency of these other T-cell-related lineages. Is it possible that HEB is playing a role not only in the fidelity of gdT17 cell specification, but also perhaps in the separation of T cells from NK cells and ILCs or the frequency of DN1, DN2, and DN3 cells? Perhaps their single-cell sequencing data or flow analyses could examine the frequency of these cells? That minor caveat aside, I find this to be an extremely exciting body of work.

Excellent question, and the underlying answer is yes, loss of HEB renders the cells more open to divergence to non-T lineages, even at the DN3 stage. Although our datasets did not reveal those cells, we have examined this question previously. In our 2011 paper (Braunstein, 2011, PMID 21189289) where we identified “DN1-like” cells arising from HEB-/- DN3 cells in OP9-DL1 co-cultures. These cells responded to IL-15 and IL-7 by differentiating into cytotoxic NK-like cells. We did not detect TCRb rearrangements but did not look for gdTCR rearrangements. Subsequently, multiple papers from other labs showed that ILC2 were greatly expanded in the thymus using Id-overexpression transgenic mice and HEB/E2A-double deficient mice (Miyazaki, 2023, PMID 28514688; Miyazaki, 2025, PMID 39904558; Berrett, 2019, PMID 31852728; Qian, 2019, PMID 30898894; Peng, 2020, PMID:32817168). The ILCs in these mice had TCRg rearrangements, consistent with a shared origin with WT thymic-derived ILCs. In unpublished data from our lab, we found an increase in the numbers of ILC2 but not ILC3 in HEB cKO fetal thymic organ cultures. We did not follow up on this work any further since the topic was being heavily pursued in other labs, but remain very interested in this branchpoint, and will mention the literature in the discussion.

Joint recommendations for the authors:

(1) Experimental validation (for mechanistic clarity)

The ChIP-seq reanalysis indicates overlapping HEB, E2A, and Egr2 peaks ~60 kb upstream of Id3. Given that the Egr2 data are not generated using the same thymocyte subsets, some form of validation should be considered for the co-binding of HEB and Egr2, potentially ChIP-qPCR in sorted gdT17 progenitors to substantiate the proposed cooperative mechanism.

See above; new experiments with ATAC-seq and additional ChIP-seq analysis.

(2) Figures

Potential inconsistencies in Figure 1H: In the legend to Figure 1H, Vg1-Vg5- cells are considered Vg6+ cells. Flow plots show reduced A Vg1-Vg5- population in HEBc ko mice, but the accompanying bar plot shows increased frequency of Vg6+ cells.

Vg6 cells are actually considered to be Vg4-Vg5-Vg1- cells (not Vg4- Vg1- cells, which is important in the fetal context). The flow plot shows the percentage of Vg6 cells out of the Vg1-Vg4- population, whereas the bar plot shows the percentage of Vg6 cells out of all gdTCR+ cells. The ratio of Vg6 to Vg5 cells decreases within the Vg1-Vg4- population, whereas the overall percentages and numbers of Vg6 cells in all gd T cells is increased in HEB cKO mice. We have now more clearly explained this in the text and the figure legend.

Clarify which cells produce IL-17A in Figure 1L.

This plot is gated on all gd T cells stimulated with PMA/ionomycin; this has been added to the results and figure legend.

In Supplementary Figure 2, legend, do the authors mean that TRGV4 was depleted? The authors write TRDV4. Please check.

Thank you for catching this mistake, we have corrected it.

In Figure 7, the Author showed Id3 mRNA expression. Can the expression of Id2 be included?

That is a really interesting question, and we will follow up on it in future studies.

If Id1 or Id4 are relevant for any of these studies, can their expression be shown in Supplementary Figure 3A? If these are minimally expressed or not expressed, this could be mentioned.

Id1 and Id4 were not detectable in our studies, this is now stated in the results section describing expression of E proteins and Id proteins.

(3) Discussion

Discuss possible redundancy between HEB and E2A, as E2A expression appears unaffected in HEB-deficient cells.

See above

Address whether the mechanisms identified at E18 (embryonic stage) also apply to neonatal or adult γδT17 subsets.

See above

Expand on how HEB function may relate to other hematopoietic or early lymphoid lineages (NK/ILC, DN1-DN3 stages), based on reviewer curiosity.

See above

(4) Methods and terminology

Define the terms γδTe1 and γδTe2 (e.g., early effector subsets).

This has been defined more clearly in several sections of the text.

Add details to the scRNA-seq methods section (average number of cells analyzed and sequencing depth per cell).

These details have been added.

eLife Assessment

This important study examines the role of TNF in modulating energy metabolism during parasite infection. The authors perform an elegant set of studies, however the evidence supporting the major claims of the manuscript is incomplete, particularly in highlighting a direct role for GLUT1 in monocytes. This work integrates an interesting set of observations that will be of interest to the Plasmodium and pathogenesis communities with an expanded set of experiments.

Reviewer #2 (Public review):

Summary:

The premise of the manuscript by Matteucci et al. is interesting and elaborates a mechanism via which TNFa regulates monocyte activation and metabolism to promote murine survival during Plasmodium infection. The authors show that TNF signaling (via an unknown mechanism) induces nitrite synthesis, which (via yet an unknown mechanism), and stabilizes the transcription factor HIF1a. Furthermore, that HIF1a (via an unknown mechanism) increases GLUT1 expression and increases glycolysis in monocytes. The authors demonstrate that this metabolic rewiring towards increased glycolysis in a subset of monocytes is necessary for monocyte activation including cytokine secretion, and parasite control.

Strengths:

The authors provide elegant in vivo experiments to characterize metabolic consequences of Plasmodium infection, and isolate cell populations whose metabolic state is regulated downstream of TNFa. Furthermore, the authors tie together several interesting observations to propose an interesting model regarding

Weaknesses:

The main conclusion of this work - that "Reprogramming of host energy metabolism mediated by the TNF-iNOS-HIF1a axis plays a key role in host resistance to Plasmodium infection" is unsubstantiated. The authors show that TNFa induces GLUT1 in monocytes, but never show a direct role for GLUT1 or glucose uptake in monocytes in host resistance to infection (nor the hypoglycemia phenotype they describe).

Comments on revisions:

The demonstration that the established TNF-iNOS-HIF-1α-glycolysis axis operates in vivo during P. chabaudi infection is valuable and relevant. However, it constitutes contextual validation and must be carefully described as such. This distinction, i.e., "what has already been shown vs. what is new" is not consistently reflected in the framing of the manuscript raising overstatement concerns. This is particularly evident in the abstract and other conclusive statements, where mechanistic novelty is implied, even when the underlying pathways/mechanisms are already known. To improve the manuscript, all sentences that refer to already established findings should be accurately described as such.

For example, the abstract states: "Here, we show that TNF signaling hampers physical activity, food intake, and energy expenditure while enhancing glucose uptake by the liver and spleen as well as controlling parasitemia in P. chabaudi-infected mice." In this sentence, the effects of TNF signaling on physical activity, food intake, energy expenditure, glucose metabolism and control of parasitemia are unequivocally established and therefore do not, in themselves, constitute new findings. Feeding behavior, not cell-intrinsic metabolism, may drive glycemic differences

The authors propose that TNF signaling leads to GLUT1 upregulation (in inflammatory monocytes, MO-DCs, and within the liver and spleen) during Plasmodium infection, and that this results in increased glucose uptake contributing to systemic hypoglycemia. While this is an intriguing hypothesis, we urge the authors to consider an alternative explanation that, at present, is not adequately ruled out. Given that glycemia serves as a central functional readout in the manuscript, this distinction is essential to clarify.

The observed regulation of glycemia is likely not a direct consequence of increased glucose uptake by immune cells or by tissues but may instead reflect broader differences in disease severity across genotypes. The iNOS KO, TNFR KO, and HIF-19775ΔαLyz2 mice likely experience a dampened inflammatory response, which would blunt infection-induced anorexia and help preserve overall metabolic homeostasis. This alternate interpretation is supported by the authors' metabolic cage data showing increased physical activity in TNFR KO mice and the elevated food intake shown in Figure 2B.

Since anorexia and energy expenditure are tightly coupled to the inflammatory milieu, it is plausible that these behavioral and systemic differences-not monocyte nor tissue GLUT1 expression per se-are the primary contributors to the observed glycemic patterns. To support their current interpretation, the authors should perform a pair-feeding experiment in which (at least) TNFR KO mice are restricted to the same food intake as infected WT controls. This would help disentangle whether differences in glycemia truly reflect immune-driven metabolic rewiring or are secondary to differences in caloric intake.

The contribution of monocyte-specific glucose metabolism to host resistance remains unresolved.

We appreciate the authors' effort to address the mechanistic role of glycolysis in host resistance using in vivo 2-deoxyglucose (2DG) treatment. However, I would like to point out that while this experiment is informative, it does not fully resolve the specific concern raised regarding the cell-intrinsic role of TNF-induced glycolysis in monocytes. 2DG acts systemically, inhibiting glycolysis across a wide range of cell types-including hepatocytes, endothelial cells, lymphocytes, and myeloid populations. Therefore, the observed increase in parasitemia following 2DG treatment may reflect the broad importance of glycolysis for host defense, or alternatively, may result from elevated circulating glucose levels induced by 2DG (PMID: 35841892), which could enhance parasite growth by increasing nutrient availability. Therefore, this experiment does not allow for a specific conclusion about the requirement for TNF-driven metabolic reprogramming in monocytes.

Author response:

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

We now performed new experiments that were included in the manuscript. Our new results show that that monocyte-derived dendritic cells primed in vivo during P. chabaudi infection, or in vitro with TNF express high levels or GLUT-1 (Figures 4M, 5D, 6L). Furthermore, our new data show that mice treated with 2-DG (na inhibitor of glycolysis) are more susceptible to infection (Figures 6N, O). In addition, new results of glucose uptake by muscle and adipose tissues were added to the manuscript. Finally, figure legends were revised, densitometric analysis performed, and other issues addressed in the text.

Please see below a point-by-point reply to the Reviewers’ comments.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

The manuscript by Kely C. Matteucci et al. titled "Reprogramming of host energy metabolism mediated by the TNF-iNOS-HIF-1α axis plays a key role in host resistance to Plasmodium infection" describes that TNF induces HIF-1α stabilization that increases GLUT1 expression as well as glycolytic metabolism in monocytic and splenic CD11b+ cells in P. chabaudi infected mice. Also, TNF signaling plays a crucial role in host energy metabolism, controlling parasitemia, and regulating the clinical symptoms in experimental malaria.

This paper involves an incredible amount of work, and the authors have done an exciting study addressing the TNF-iNOS-HIF-1α axis as a critical role in host immune defense during Plasmodium infection.

Reviewer #2 (Public Review):

Summary:

The premise of the manuscript by Matteucci et al. is interesting and elaborates on a mechanism via which TNFa regulates monocyte activation and metabolism to promote murine survival during Plasmodium infection. The authors show that TNF signaling (via an unknown mechanism) induces nitrite synthesis, which (via yet an unknown mechanism), and stabilizes the transcription factor HIF1a. Furthermore, HIF1a (via an unknown mechanism) increases GLUT1 expression and increases glycolysis in monocytes. The authors demonstrate that this metabolic rewiring towards increased glycolysis in a subset of monocytes is necessary for monocyte activation including cytokine secretion, and parasite control.

Strengths:

The authors provide elegant in vivo experiments to characterize metabolic consequences of Plasmodium infection, and isolate cell populations whose metabolic state is regulated downstream of TNFa. Furthermore, the authors tie together several interesting observations to propose an interesting model.

Weaknesses:

The main conclusion of this work - that "Reprogramming of host energy metabolism mediated by the TNF-iNOS-HIF1a axis plays a key role in host resistance to Plasmodium infection" is unsubstantiated. The authors show that TNFa induces GLUT1 in monocytes, but never show a direct role for GLUT1 or glucose uptake in monocytes in host resistance to infection (nor the hypoglycemia phenotype they describe).

We kindly disagree with the Reviewer. There is a series of experiments showing that TNFR KO (Figures 1, 2, 4), HIF1a KO (Figure 5) and iNOS KO (Figure 6) mice have partially impaired inflammatory response and control of parasitemia (Figures Figures 1E, 5G and 6B).

To further address the issue raised by the reviewer, we performed two sets of experiments. First, we show, in vitro, the impact of TNF stimulation on GLUT1 expression and glucose uptake (Figure 4M, 5D, 6L). Our results show that GLUT1 is increased after 18 hours with TNF (100 ng/mL) stimulation in MODCs from WT mice but not from iNOS KO, HIF1a KO e TNFR KO mice. Similar results were obtained with monocytic cells derived from infected mice (Figure 4L, 5C, 6K). The results support the discussion by demonstrating that TNF stimulation influences GLUT1 expression in monocytic cells. This aligns with the proposed mechanism that TNF signaling regulates HIF-1α stabilization and glycolytic metabolism via RNI. The absence of GLUT1 upregulation and glucose uptake in TNFR KO, iNOS KO and HIF-1α KO mice further reinforces the role of RNI in promoting HIF-1α stabilization, as suggested in the discussion.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

Major points

All Figure legends are not precise about the data express means {plus minus} standard errors of the means (SEM) or SD. Figure 1D shows no SD in the data from the uninfected group. It strongly suggests precise and improving all figure legends, giving more details in terms of including an explanation of all symbols, non-standard abbreviations, error bars (standard deviation or standard error), experimental and biological replicates, and the number of animals, and representative of the independent experiments.

We apologize for the lack of details in the Figure legends. As requested, we are now indicating whether we used SEM or STDV, number of mice per group, number of replicate experiments. We also clarified the groups that are being compared, and the statistical significance indicated by the symbols. We also standardized symbols as asterisk only, and number of asterisk indicating the significance.

Figure 1. The figure legend has no information about the organ for which TNF mRNA was measured (Figure 1D). Also, regarding the TNF data, Figure 1 C e 1D shows that the circulating levels of TNF and the expression of TNF mRNA in the liver peaked at the same time point, and after 6h, there is no difference between infected and uninfected mice. It would be expected that the TNF mRNA expression would be detected earlier than the protein, assuming that the primary source of TNF is from the liver. Is there another organ that could mainly source blood TNF levels? Did the authors have a chance to measure the blood TNF levels during infection (0-8dpi), besides the measurement at different times only on day 8?

We included in the legend of Figure 1D that mRNA was extracted from liver.

Liver and spleen are the main reservoir of infected erythrocytes and the main source of cytokines during the infection with the erythrocytic stage of malaria. The results presented in Figures 1C and 1D are from in vivo experiments, not a controlled cellular experiment in vitro. So, we can not conclude about exact time and synchronous production of TNF mRNA and protein. We have published earlier that during P. chabaudi infection, the peaks of TNF mRNA expression and the levels of circulating TNF protein occur between midnight and 6 am (Hirako at al., 2018). Hence the results are consistent in the results described here. In addition, this earlier study also shows that the same pattern of TNF at days 6 and 8 post-infection are similar. Furthermore, in another studies, we reported that the peak of TNF production occurs between days 6 and 10 post P. chabaudi infection (Franklin et al, PNAS, 2009; Franklin et al, Microbes and Infection, 2007). This is now clarified in the text (page 05, line 132):

“As previously demonstrated, the circulating levels of TNF and expression of TNF mRNA in the liver peaked at 6 am (end of dark cycle) at 8 dpi (Figure 1C and 1D), and has been reported to peak between days 6 and 10 post-infection, with a consistent pattern observed on days 6 and 8.”

Figure 2. "We observed that in naïve animals, all of these parameters were similar in TNFR^-/- and C57BL/6 mice (Figures 2A-D, top panels, and Figures 2E-H)." Interestingly, the respiratory exchange rate of TNFR^-/- uninfected mice seems higher in TNFR^-/- uninfected mice than in naïve uninfected mice, and this pattern seems to be more pronounced in TNFR^-/- uninfected mice. Is there any suggestion that could explain the change in respiratory exchange rate behavior without infection in those animals?

At the moment, we have not investigated the basis of this difference between uninfected WT and TNFR KO mice, which goes beyond the scope of this research. This is indeed an interesting observation that should be pursued in the future by our group and elsewhere. We mentioned this difference, when describing the results (page 06, lines 155):

“We observed that in naïve animals, all of these parameters were similar in TNFR^-/- and C57BL/6 mice (Figures 2A-D, top panels and Figures 2E-H), with a slightly higher respiratory exchange rate in uninfected TNFR^-/- mice. In contrast, all the evaluated parameters were decreased in infected C57BL/6 mice compared to their naïve counterparts during the light and dark cycles. When we analyzed only infected mice, the alterations in all parameters were milder in TNFR^-/- compared to C57BL/6 mice (Figures 2A-D bottom panels and 2E-H).”

Figure 3. To give an idea of the main population of non-parenchymal cells, it will be helpful to clarify briefly how non-parenchymal cells from the liver of infected or uninfected mice were isolated.

We described in detail at Material and Methods (Page 19, Lines 566.)

Figure 3, B, C, D, G and Figure 4K and Figure 5 A and B - Semi-quantitative data through the densitometric analysis of western blots should be included in all figures.

Thank you for the suggestion. We now included the densitometric analysis for all Western blot results in Supplementary figure.

Figure 4. The author describes, "We observed that except for Hexokinase-3, the expression of mRNAs of glycolytic enzymes (Hexokinase-1, PFKP, and PKM) was increased in C57BL/6 but not TNFR-/- 8dpi." Sometimes, it is hard to understand which groups have been compared to some data. Be precise in describing the statistical analysis between the groups. It seems that those genes were increased in "infected C57BL/6 in comparison to uninfected mice, but not TNFR-/- 8-dpi. Moreover, even though the authors include statistic symbols "ι, ιι, ιιι" in other legends, there is no explanation about statistic symbols in the legend of Figure 4.

As mentioned above, we improved the descriptions of all figures in the legend, and when necessary in the main text describing the results.

Figure 5. The authors describe, "We found that GLUT1 protein and glycolysis (ECAR) was impaired, respectively, in monocytic cells and splenic CD11b+ cells from infected, as compared to uninfected HIF-1aΔLyz2 mice (Figures 5C-5E)." The GLUT-1 expression was inhibited in both cells compared to HIF-1afl/fl mice but not even close to impaired GLUT-1 expression. There is still a robust amount of GLUT-1 expression, and significantly higher when compared to cells from uninfected mice.

We tuned our statement to partially impaired, indicating that other host or parasite components maybe be also influencing GLUT-1 expression. In fact, we have recently published that IFNγ has also an important role in regulating GLUT1 expression in MO-DCs and this reference is mentioned in the text (page 10, line 291):

“We found that glycolysis (ECAR) and GLUT1 expression were impaired, though partially, in monocytic and splenic CD11b+ cells from infected HIF-1aΔLyz2 mice (Figures 5C-5E) compared to infected WT mice. The level of GLUT1 expression that is still maintained is likely due to other host or parasite factors, such as IFN-γ (Ramalho 2024).”

Figure 6. It is essential to have more information about the number of replicates in Figure 6A. However, there are just two dots replicates in the condition CD11b+ splenic cells from C57BL/6 stimulated with or without LPS (purple bars). It is essential to be precise regarding the number of experimental and biological replicates in each experiment and the statistical analysis that has been applied, including this group. Furthermore, the author concludes, "...these data demonstrated that RNI induces HIF-1α expression...." This conclusion needs a more careful description since no data supports that monocytic cells or splenic CD11b+ cells from iNOS-/- infected mice decrease stabilization of HIF-1αm using blotting, as shown in Figure 5 A.

As mentioned above the number of replicates for each experiment was included in the figure legends.

Minor Points.

Figure 3. "Hepatocytes have an important role in glucose uptake from the circulation, and they do this primarily through GLUT2 (38), whose mRNA expression was downregulated (Figure 3A) and protein expression unchanged in response to Pc infection (Figure 4K)." I suggest moving the Figure 4K to Figure 3 to make it easy to follow the data description.

We thank the reviewer for the suggestion. However, we chose to keep Figure 4K in Figure 4, as this panel includes data from TNF receptor deficient mice, and the analysis of TNF knockout models is first introduced and discussed in Figure 4. For clarity and consistency, we therefore maintained this panel within Figure 4.

Line 433. Replace iNOS for iNOS-/- mice.

iNOS is now replaced for iNOS-/- mice.

Reviewer #2 (Recommendations For The Authors):

The premise of the manuscript by Matteucci et al. is interesting and elaborates on a mechanism via which TNFa regulates monocyte activation and metabolism to promote murine survival during Plasmodium infection. The authors show that TNF signaling (via an unknown mechanism) induces nitrite synthesis, which (via yet an unknown mechanism), and stabilizes the transcription factor HIF1a. Furthermore, HIF1a (via an unknown mechanism) increases GLUT1 expression and increases glycolysis in monocytes. The authors demonstrate that this metabolic rewiring towards increased glycolysis in a subset of monocytes is necessary for monocyte activation including cytokine secretion, and parasite control.

The main goal of this work is to study the interplay of TNF/HIF1a/iNOs in the pathogenesis in an experimental model of malaria. To dissect the molecular mechanism by which TNF induces reactive nitrogen species and regulates HIFa expression is beyond the scope of our research. Nevertheless, there is a vast literature addressing these issues. We now include in the discussion a paragraph describing the main conclusion of these studies published previously (page 12, line 363):

"Previous studies have shown that TNF induces the production of RNI through the upregulation of iNOS via the NF-κB pathway (63, 64). TNF-mediated iNOS expression is critical for NO production, which in turn stabilizes HIF-1α by inhibiting prolyl hydroxylases (PHDs) even under normoxic conditions (58, 59). HIF-1α then upregulates the expression of glycolytic genes, including GLUT1 (22, 62).”

Major comments

Issues concerning novelty

Some of the reported observations are not novel. TNFa and TNFa signaling has been demonstrated to contribute to the release of certain cytokines, and to contribute to the control parasitemia (PMID: 10225939). TNFa has been shown to increase glucose uptake in tissues (PMID: 2589544). There is a textbook about the role of INOS during the pathogenesis of malaria, including its association with parasite control (https://link.springer.com/chapter/10.1007/0-306-46816-6_15). Furthermore, other mechanisms controlling glycemia during Plasmodium infection have been shown (PMID: 35841892). The authors should adequately discuss other papers which have reported some of their findings.

Thanks for the comments on previously existing literature. We are well aware of some of this earlier literature. Some of these earlier findings are mentioned in our manuscript. We emphasized these fundamental findings in the discussion, as requested (page 12, line 368):

“TNF has been described as a critical mediator in malaria, driving cytokine release and parasitemia control (PMID: 10225939). It also enhances glucose uptake in tissues, aligning with our findings of increased glycolysis in monocytes (PMID: 2589544). The role of iNOS in malaria is well documented. IFN-γ and TNF induced the production of NO, which inhibits parasite growth but can cause tissue damage and organ dysfunction, especially in severe malaria (Mordmüller et al., 2002). Recent studies also highlight the complexity of glycemia regulation during Plasmodium infection describing its role in modulating parasite virulence and transmission (PMID:35841892). These studies demonstrate the critical function of TNF and iNOS in immune responses against Plasmodium, aligning with our findings of this axis and metabolic rewiring that are essential for monocyte activation and outcome of Pc infection.”

The authors claim that "Reprogramming of host energy metabolism mediated by the TNF-iNOS-HIF1a axis plays a key role in host resistance to Plasmodium infection," and contributes significantly to their effector functions (particularly parasite clearing), and the systemic drop in glycemia observed during Pc infection. Although the authors show that TNFa does result in altered metabolism and increased GLUT1 levels in a subpopulation of monocytes, the evidence that TNFa-induced glylcolysis plays a key role in host resistance is correlative at best.

This is an important question. We did show that TNFR KO have higher parasitemia. But TNF is pleiotropic cytokine and has multiple roles on innate and acquired immunity. The experiment we have performed and helps to address this issue is the in vivo treatment with 2DG. We found that treatment with this inhibitor of glycolysis results in a increase of parasitemia. These results are now included in Figure 6.

When considering that the majority of monocytic populations are reduced in frequency and only a small subset (i.e., Monocyte-derived DCs) increase in frequency (Fig 3K) during Pc infection, this makes it very difficult to demonstrate that a cell population whose overall frequency reduces contributes significantly to the drop in glycemia during Pc infection. The authors should therefore include experiments that demonstrate that the inhibition of glycolysis induced by TNFa in monocytes is protective and/or contributes to a decrease in extracellular glucose. The authors could assess the impact of the loss of function of GLUT1 on activated monocytes and monocyte-derived DCs on glycemia upon TNFa stimulation.

We agree. We focused on monocytes and the derived inflammatory monocytes and MO-DCs. In fact, the frequency of monocytes, considering the inflammatory monocytes and MO-DCs, is increased both in spleen and liver. One interesting result is that the HIF1a Lysm KO mice has impaired metabolism, attenuated hypoglycemia and increased parasitemia (Figure 5). Nevertheless, we agree that our current data thus not proof that the glycemia is due to the consumption of glucose by the activated monocytes, and that these are the only cells with increased glucose consumption. This is now added to the discussion (page 13, line 395):

"Although the frequency of MO-DCs increases during infection, other cell populations may also contribute to glucose consumption. Further experiments, including the assessment of GLUT1 function in these populations, are needed to clarify their contribution to glucose consumption during infection."

Furthermore, in the current state of the manuscript, it is unclear how activated monocyte populations uptake glucose. The authors claim that glucose uptake by activated monocytes is GLUT1-dependent, however, glucose transport via GLUT1 is insulin-dependent. Since Plasmodium infection is associated with insulin resistance, and almost unquantifiable levels of insulin (PMID: 35841892), and TNFa itself induces insulin resistance (PMCID: PMC43887), it is unclear how the activated monocyte population uptakes glucose. If the authors consider TNFa to be sufficient for GLUT1 induction, in vitro experiments (TNFa+monocytes) could bolster this claim (and support that GLUT1 is induced in an insulin-independent mechanism.

There is significant evidences indicating that in contrast to GLUT4, induction of GLUT1 in mice is independent of insulin (PMID: 9801136). In our case, seems to be induced by the cytokines TNF and IFN𝛾(this study and Ramalho et al., 2024). We now performed experiments exposing monocytes to TNF and evaluating GLUT1 expression. The results indicate that monocytes exposed to TNF (100 ng/mL) for 18 hours from WT mice exhibited a significant increase in GLUT1 expression. This increase was comparable to the increased-GLUT1 phenotype observed in infected animals. The results of this experiment were included in the manuscript.

A text was included to the discussion to clarify the issue of insulin dependence of GLUT1 expression (page 13, line 388):

“GLUT1 expression is recognized as independent of insulin, in contrast to GLUT4 (PMID: 9801136). In our model, this regulation appears to be driven by pro-inflammatory cytokines, particularly TNF. Supporting this, our results show that in vitro stimulation with TNF, significantly increases GLUT1 expression in monocytes, accordingly to the ex vivo phenotype observed in infected animals.”

Alternative hypothesis which might explain their phenotypes

Figure 2 A-H: The metabolic effects of the genetic manipulations including INOS KO, TNFR KO, and HIF-1α∆Lyz2 could be explained by lesser disease morbidity owed to a reduction of inflammatory response during infection. Under this condition, the development of anorexia will not be as profound in the knock-outs compared with wild-type littermate controls, since anorexia of infection is tightly linked to the magnitude of inflammatory response. Accordingly, infected knock-out animals can keep eating, which presumably impacts glycemia, maintenance of core body temperature, and overall energetics of infected mice. The authors should exclude this possibility.

We consider this possibility and the discussion now elaborates about this alternative hypothesis. We believe, that these two mechanisms are not mutually exclusive (page 16, line 474):

“Although restored physical activity, food consumption and energy expenditure in knockout mice may contribute to the observed systemic metabolic parameters by altering energy balance, these effects are not mutually exclusive with the TNF-driven, cell-intrinsic metabolic mechanisms described here.”

Minor comments

The authors showed increased parasitemia upon TNFR and HIF1a depletion in the LyZ2 compartment. The same was observed upon organismal INOS depletion. This raises the question of whether the TNFHIF-INOS signaling axis is adaptive or maladaptive during Pcc infection. The authors should show host survival in mice lacking TNFR and HIF1a in the LyZ2 compartment, and in mice lacking INOS (presumably, they have these data).

Despite the fact the various knockout mice have increased parasitemia and signs of disease, they all survive the infection. This is now included in the Figure legends.

Are the higher tissue glucose levels specific to the liver and the spleen or this is a more general event? Have the authors looked at other organs?

We now added the results of glucose uptake in the muscle and adipose tissues in figure 2. The fact that the glucose uptake is not increased in muscle and adipose tissue, further suggest that the increased glucose uptake in this model is insulin independent.

Figure 1F: All core body temperatures are within the physiological range, i.e., >36 degrees C. This makes it unclear why the authors regarded this as hypothermia. The authors should present experiments demonstrating the development of hypothermia in Figure 1F, as they claim this.

Temperature changes in mouse kept in animal house have been an issue discussed in the field. It is clear, however, that early in the morning (end of active period) mice have torpor. Lower temperature and physical activity.

In Figure 4, since the authors already suggested that extra-hepatic cells, and not the liver parenchyma, contribute to glucose uptake, the authors should clarify why they analyzed the whole liver in Figure 4, and not extra-hepatic cells. Furthermore, the authors should quantify the hepatic monocytic population in non-infected versus infected wild-type animals.

The reason we used whole liver, is that the number of non-parenchymal cells obtained from liver is limited for Western blot analysis. We thought that was important to show that expression of GLUT1 was decreased in the liver of TNFR KO mice. Nevertheless, the level of TNFR expression in different cell types in the liver was shown by flow cytometry. In addition, we performed the WB with cells extracted from the spleen, where lymphoid and myeloid cells are more abundant.

Line 87: Phagocytizing parasitized what?

This has been corrected in the manuscript.

Line 111 Define RNI before being used.

Is there a gender disparity in the TNFR KO phenotype? If yes, the authors should comment about this in their discussion.

This has been defined and addressed in the manuscript

Line 192: Did the authors mean 3B??

In 3M, please plot monocytes from uninfected animals.

The plot of uninfected animals are now included in Figure 3M

Line 390 Remove the extra dash in HIF1a.

Extra dash has been removed.

Line 397 Define RA

RA is now defined.

eLife Assessment

This study presents a valuable finding on how the GAP DLC1, a deactivator of the small GTPase RhoA, regulates RhoA activity globally as well as at Focal Adhesions. Using a new acute optogenetic system coupled to a RhoA activity biosensor, the authors present convincing evidence that DLC1 amplifies local Rho activity at Focal Adhesions. Thanks to modeling, they show that DLC1 is needed for a negative feedback loop that engage more RhoA deactivators upon RhoA activation, highlighting the complex regulation of RhoGTPases in space and time.

Reviewer #1 (Public review):

Summary:

The manuscript of Heydasch et al. addresses the spatiotemporal regulation of Rho GTPase signaling in living cells and its coupling to the mechanical state of the cell. They focus on a GAP of RhoA, the Rho specific GAP Deleted in Liver Cancer 1 (DLC1). They first show that removing DLC1 either by a CRISPR KO or by downregulation using siRNA leads to an increased contractility and globally elevated RhoA activity, as revealed by a FRET biosensor. This result was expected, since DLC1 is deactivating RhoA its absence should lead to increasing amounts of active RhoA. To go beyond global and steady levels of RhoA activity, the authors developed an acute optogenetic system to study transient RhoA activity dynamics in different genetic and subcellular contexts. In WT cells, they found that pulses of activation lead to an increased RhoA activity at focal adhesions (FA) compared to plasma membrane (PM), which suggests that FAs contain less RhoA GAPs, more RhoA, or that FAs involve positive feedbacks implying others GEFs for example. In DLC1 KO cells, they found that the RhoA response upon pulses of optogenetic activation was increased (higher peak) both at FA and PM, which could be expected since less GAP should increase the amount of active RhoA. But surprisingly, they observed also a higher rate of RhoA deactivation in DLC1 KO cells, which is counterintuitive: less GAP should result in a slower rate of deactivation. Less GAP should also lead to a lower rate of observed RhoA activation (smaller koff) and delayed peak. Using a modeling approach and control experiments (to monitor the optogenetic intrinsic dynamics), the authors propose that there is a negative feedback in WT cell between activated RhoA and the activity of its GAPs (other than DLC1). More active RhoA decreases GAP activity such that active RhoA relaxation to its basal state is relatively slow. This negative feedback would be absent in DCL1-deficient cells, explaining the relatively faster relaxation. This hypothesis is convincing given the data and the model, and it shows that there are compensatory mechanisms at play when DLC1 is knocked down. Further on, the authors study the dynamics of DLC1 on FAs depending on the mechanical state and nicely show a causal decrease of DLC1 enrichment at FA upon FA reinforcement, hereby probing a positive feedback where RhoA activation is further amplified as the force exerted at FA is increasing. Altogether, this work highlight the extremely fine regulation in space and time of RhoGTPases that is only revealed through acute perturbations, while at the cell scale and long time scale, complex compensatory mechanisms are at play rendering knock-down or overexpression experiments not always straightforward to interpret (in the present case, knock-down of a deactivator lead to an increase of deactivation rate through the induced absence of other activity dependent-deactivators).

Strengths:

- Experiments are precise and well done.

- Technically, the work brings original and interesting data. The use of transient optogenetic activation within focal adhesions together with a biosensor of activity is new and elegant.

- The link between DLC1 and global contractility/RhoA activity is clear and convincing.

- The surprising higher rate of RhoA deactivation in DLC1 KO cells is convincing, as well as the differences in the dynamics of RhoA between focal adhesions and plasma membrane.

- The model is very helpful to support the hypothesis of the negative feedback loop.

- The correlation between DLC1 enrichment and focal adhesion dynamics is very clear.

Weaknesses:

- The negative and positive feedback loops could have been dug more deeply molecularly (in particular discover what are the compensatory mechanisms at play), but this could be the purpose of future work.

Comments on revised version:

I thank the authors for the great improvement of their work and their detailed answers to my comments. The modeling work is great and really brings novelty to the story. It also helps a lot to have the data for the optoLARG recruitment. I suggest authors move to the Version of Record.

Author response:

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

We thank the reviewers for their careful reading of our manuscript and for the constructive and insightful feedback. In response, we performed several new experiments and analyses that significantly strengthen the study. First, we addressed the important question of optoLARG recruitment dynamics by generating a new cell line expressing optoLARG-mScarlet3 together with paxillin-miRFP, enabling us to directly quantify the dynamics of the optogenetic activator at focal adhesions and the plasma membrane. Second, we introduced a quantitative modeling framework to analyze RhoA activity dynamics during transient optogenetic stimulation. Using the measured optoLARG kinetics as input, we fitted activation and deactivation parameters for both WT and DLC1 KO cells, revealing a loss of negative feedback regulation in the KO condition. Together, these additions clarify the temporal relationships between optogenetic activation, RhoA signaling, and biosensor responses, and provide a more rigorous, mechanistic interpretation of our data. We rewrote large parts of the discussion section to reflect this new information.

Below, we provide detailed, point-by-point responses to all reviewer comments.

Recruitment dynamics optoLARG

Reviewer #1:

Public Review:

For the optogenetic experiments, it is not clear if we are looking at the actual RhoA dynamics of the activity or at the dynamics of the optogenetic tool itself.

Recommendations for the authors:

For the transient optogenetic activations at FA and PM, it would be great to have one data set where the optoLARG is fused to a fluorescent protein, for example, mCherry, while FAs would be marked with paxillin-miRFP (by transient transfection to avoid making a new stable cell line). The dynamics of the optogenetic activator should be the same (on and off rates), but it can be possible that the activator is retained at FA for example. Such an experiment would help the understanding of the differential observed dynamics, where several timescales are involved: the dynamics of the opto tool, the dynamics of RhoA itself, and the dynamics of the biosensor.

We agree with the reviewers, this is an essential control for this manuscript and the cell line will be useful in future studies. We developed a new construct containing with the recruitable SSpB domain tagged in red (optoLARG-mScarlet3) compatible with the iLid system, and paxilin-miRFP to locate the focal adhesions. From previous experiments we know that the anchor part of optoLARG system is distributed evenly across the cell membrane and is not affected by cytoskeletal structures like focal adhesions. As for the recruitable part of the optoLARG system, that translocates from the cytosol to the membrane upon blue light stimulation, we illuminated focal adhesion and non-focal adhesion regions, and quantified optoLARG dynamics. The same scripts were used for automated stimulation and analysis as were used for the rGBD recruitment experiments. We illustrate these results in the new Suppl. Fig S3. We found no significant difference in recruitment dynamics between focal adhesion/non-focal adhesion regions (Fig. S3B). We found the optoLARG dynamics fits well with inverse-exponential during recruitment under blue light stimulation, and exponential decay after blue light stimulation (disassociation phase), consistent with the expected iLID dynamics (Fig S3C). This experiment is described in detail at the end of the section "Optogenetic interrogation of the Rho GTPase flux in WT and DLC1 KO cells" (Lines 303-320). We then went on to use the optoLARG dynamics as input for the models describing RhoA activity dynamics (see next comment). This should help to untangle the measured RhoA dynamics from the dynamics of the optogenetic tool.

Quantitative analysis RhoA activity dynamics

Public Review:

There is no model to analyze transient RhoA responses, however, the quantitative nature of the data calls for it. Even a simple model with linear activation-deactivation kinetics fitted on the data would be of benefit for the conclusions on the observed rates and absolute amounts.

Recommendations for the authors:

[...] for the transient optogenetic experiments, it would be great to make a simple model, or at least to fit the curves with an on rate, an off rate, and a peak value. This will clarify the conclusions drawn for the experiments. For example, the authors claim that they observe an increased Rho activation rate in DLC1 KO cells (see sections "Optogenetic interrogation of the Rho GTPase flux in WT and DLC1 KO cells" and "Discussion") but the rate is not well-defined. One can have two curves with the same activation rate but one that peaks higher (larger multiplicative prefactor) and it would resemble the presented data. This being said, the higher deactivation rate in DLC1 KO cells is evident from the data.

We agree that a quantitative analysis and model would improve our understanding of the data. We fit the activation/deactivation kinetics and provide the values in the chapter "Optogenetic interrogation of the Rho GTPase flux in WT and DLC1 KO cells" (Lines 287-299). We then modeled the RhoA activity dynamics at focal adhesions and at the plasma membrane after transient optogenetic stimulation using a system of ODEs, using the new measurements of optoLARG kinetics as activation input. We find a close fit for the experimental data, with WT following classic Michaelis-Menten dynamics. Interestingly, when fitting the DLC1-KO data with the same model as for WT, the parameter modeling the negative feedback loop (active RhoA recruiting a GAP) is set to zero; in other words, the factor that deactivates RhoA is present at a constant concentration. We added an additional main Figure 5 describing the models and fits, and added a new Results section "Modeling indicates loss of negative RhoA autoregulation in DLC1-KO cells" (Lines 326-378), and also updated the Methods and Discussion section of the paper accordingly. We use the findings to more clearly ground the mathematical terms used to describe our results.

Error figure 6E

Recommendations for the authors:

The scheme presented in Figure 6E is not supported by the data and should be modified. In this scheme, the authors show a strongly delayed peak in control cells versus DCL1 KO cells, whereas in the data the peaks appear to be at similar time points. Similarly, the authors show a strongly decreased rate of activation, whereas the initial rates appear identical in the data.

The delayed peak we illustrated is an error, we thank the reviewers for catching it. The decreased rate of deactivation and activation, although exaggerated in the scheme, is however present in the data (and is now quantified, see answer above). We updated the figure accordingly (now Fig. 7E in the manuscript).

Clarification term "signaling flux"

Recommendations for the authors:

It would be nice to define more precisely several terms that are used throughout the manuscript. For example, could the authors define what they mean by "signaling flux"? Is it the temporal derivative of the Rho levels? Or the spatial derivative?

We agree that this was not clear in the previous version of the manuscript. We refer to "signaling flux" as the continuous cycle of RhoA activation by GEFs and inactivation by GAPs, processes that persist even when bulk RhoA activity appears steady, as introduced by Miller & Bement (2009). We now explicitly define "signaling flux" in the abstract (Lines 20-24).

See: Miller, Ann L., and William M. Bement. "Regulation of cytokinesis by Rho GTPase flux." Nature cell biology 11.1 (2009): 71-77. https://doi.org/10.1038/ncb1814

Recommendations for the authors:

Also (see above) it would be nice to define precisely what are the rates: the activation rate is in general the k_on of a reaction scheme, but it will differ from the observed rate given by a biosensor. For example, with a k_on and a k_off the observed rate toward the steady-state will be given by the sum of the activation and deactivation rates. In the manuscript, the authors do not make the distinction between the activation rate with the rate of increase of the biosensor which is confounding for the reader and for the interpretation of the data.

We update the results section to make this distinction more clear (Lines 288-300), and add a note explicitly highlighting the difference between biosensor signal dynamics and the underlying RhoA activation/deactivation rates (Lines 298-300). In addition, our newly introduced model helps disentangle the combined activation/deactivation rates into distinct GEF and GAP activity parameters.

Improvements to figure 3

Minor recommendation:

In Figures 3 B and D, the stars (statistical differences) are not visible. It would be good to make them bigger or move them above the graphs.

Thank you! We updated the graphics.

Other changes

Additional panel (Figure 5D) showing paxillin intensity does not change after weak optogenetic stimulation, to better illustrate the weak stimulation regime that does not trigger FA reinforcement (contrasting Figure 7). Additional small layout changes to Figure 5.

Addition of authors that contributed to the revisions

eLife Assessment

This study represents an important advance in our understanding of how certain inhibitors affect the behavior of voltage gated potassium channels. Robust molecular dynamics simulation and analysis methods lead to a new proposed inhibition mechanism with strength of support being mostly convincing, though computational evidence is limited for some conformations discussed. This study has considerable significance for the fields of ion channel physiology and pharmacology and could aid in development of selective inhibitors for protein targets.

Reviewer #1 (Public review):

Summary:

The authors were seeking to identify a molecular mechanism whereby the small molecule RY785 selectively inhibits Kv2.1 channels. Specifically, the authors sought to explain some of the functional differences that RY785 exhibits in experimental electrophysiology experiments as compared to other Kv inhibitors, namely the charged and non-specific inhibitor tetraethylammonium (TEA). The authors used a recently published cryo-EM Kv2.1 channel structure in the open activated state and performed a series of multi-microsecond-long all-atom molecular dynamics simulations to study Kv2.1 channel conduction under the applied membrane voltage with and without RY785 or TEA present. They observed that while TEA directly blocks K+ permeation by occluding ion permeation pathway, RY785 binds to multiple non-polar residues near the hydrophobic gate of the channel driving it to a semi-closed non-conductive state. They confirmed this mechanism using an additional set of simulations and used it to explain experimental electrophysiology data,

Strengths:

The total length of simulation time is impressive, totaling many tens of microseconds. The authors develop their own forcefield parameters for the RY785 molecule based on extensive QM based parameterization. The computed permeation rate of K+ ions through the channel observed under applied voltage conditions is in reasonable agreement with experimental estimates of the single channel conductance. The authors have performed extensive simulations with the apo channel as well as both TEA and RY785. The simulations with TEA reasonably demonstrate that TEA directly blocks K+ permeation by binding in the center of the Kv2.1 channel cavity, preventing K+ ions from reaching the SCav site. The authors conclude that RY785 likely stabilizes a partially closed conformation of the Kv2.1 channel and thereby inhibits K+ current. This conclusion is plausible given that RY785 makes stable contacts with multiple hydrophobic residues in the S6 helix, which they can also validate using a recently published closed-state Kv2.1 channel cryo-EM structure. This further provides a possible mechanism for the experimental observations that RY785 speeds up the deactivation kinetics of Kv2 channels from a previous experimental electrophysiology study.

Weaknesses:

The authors, however, did not directly observe this semi-closed channel conformation and in fact acknowledge that more direct simulation evidence would require extensive enhanced-sampling simulations beyond the scope of this study. They have not estimated the effect of RY785 binding on the protein-based hydrophobic pore constriction, which may further substantiate their proposed mechanism. And while the authors quantified K+ permeation, they have not made any estimates of the ligand binding affinities or rates, which could have been potentially compared to experiment and used to validate their models.

However, despite those relatively minor weaknesses, the conclusions of the study are convincing, and overall this is a solid study helping us to understand two distinct molecular mechanisms of the voltage-gated potassium channel Kv2.1 inhibition by TEA and RY785, respectively.

Reviewer #2 (Public review):

Summary

In this manuscript, Zhang et al. investigate the conduction and inhibition mechanisms of the Kv2.1 channel, with a particular focus on the distinct effects of TEA and RY785 on Kv2 potassium channels. Using microsecond-scale molecular dynamics simulations, the authors characterize K⁺ ion permeation and RY785-mediated inhibition within the central pore. Their results reveal an inhibition mechanism that differs from those described for other Kv channel inhibitors.

Strengths

The study identifies a distinctive inhibitory mode for RY785, which binds along the channel walls in the open-state structure while still permitting a reduced level of K⁺ conduction. In addition, the authors propose a long-range allosteric coupling between RY785 binding in the central pore and changes in the structural dynamics of Kv2.1. Overall, this is a well-organized and carefully executed study, employing robust simulation and analysis methodologies. The work provides novel mechanistic insights into voltage-gated potassium channel inhibition and may offer useful guidance for future structure-based drug design efforts.

Weaknesses:

The study needs to consider the possibility of multiple binding sites for PY785, particularly given its impact on voltage sensors and gating currents. Specifically, the potential for allosteric binding sites in the voltage-sensing domain (VSD) should be assessed, as some allosteric modulators with thiazole moieties are known to bind VSD domains in multiple voltage-gated sodium channels (Ahuja et al., 2015; Li et al., 2022; McCormack et al., 2013; Mulcahy et al., 2019). Increasing structural and functional evidence supports the existence of multiple ligand-binding modes in voltage-gated ion channels. For example, polyunsaturated fatty acids have been shown to bind to KCNQ1 at both the voltage sensor domain and the pore domain (https://doi.org/10.1085/jgp.202012850). Similarly, cannabidiol has been structurally resolved in Nav1.7 at two distinct sites, one in a fenestration and another near the IFM-binding pocket (https://doi.org/10.1038/s41467-023-39307-6). These advances illustrate that ligand effects cannot always be interpreted based solely on a single binding site identified previously.

Author response:

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

eLife Assessment

This study represents an important advance in our understanding of how certain inhibitors affect the behavior of voltage gated potassium channels. Robust molecular dynamics simulation and analysis methods lead to a new proposed inhibition mechanism with strength of support being mostly convincing, and incomplete in some aspects. This study has considerable significance for the fields of ion channel physiology and pharmacology and could aid in development of selective inhibitors for protein targets 

We are encouraged by this favorable assessment and thank editors and reviewers for their constructive feedback and recommendations. We trust that the revisions made to the manuscript will clarify the aspects that had been perceived to be incomplete.

Reviewer #1 (Public review):

Summary: 

This study seeks to identify a molecular mechanism whereby the small molecule RY785 selectively inhibits Kv2.1 channels. Specifically, it sought to explain some of the functional differences that RY785 exhibits in experimental electrophysiology experiments as compared to other Kv inhibitors, namely the charged and non-specific inhibitor tetraethylammonium (TEA). This study used a recently published cryo-EM Kv2.1 channel structure in the open activated state and performed a series of multi-microsecond-long all-atom molecular dynamics simulations to study Kv2.1 channel conduction under the applied membrane voltage with and without RY785 or TEA present. While TEA directly blocks K+ permeation by occluding ion permeation pathway, RY785 binds to multiple nonpolar residues near the hydrophobic gate of the channel driving it to a semi-closed non-conductive state. This mechanism was confirmed using an additional set of simulations and used to explain experimental electrophysiology data.

Strengths:

The total length of simulation time is impressive, totaling many tens of microseconds. The study develops forcefield parameters for the RY785 molecule based on extensive QM-based parameterization. The computed permeation rate of K+ ions through the channel observed under applied voltage conditions is in reasonable agreement with experimental estimates of the singlechannel conductance. The study performed extensive simulations with the apo channel as well as both TEA and RY785. The simulations with TEA reasonably demonstrate that TEA directly blocks K+ permeation by binding in the center of the Kv2.1 channel cavity, preventing K+ ions from reaching the SCav site. The conclusion is that RY785 likely stabilizes a partially closed conformation of the Kv2.1 channel and thereby inhibits the K+ current. This conclusion is plausible given that RY785 makes stable contact with multiple hydrophobic residues in the S6 helix. This further provides a possible mechanism for the experimental observations that RY785 speeds up the deactivation kinetics of Kv2 channels from a previous experimental electrophysiology study.

Weaknesses:

The study, however, did not produce this semi-closed channel conformation and acknowledges that more direct simulation evidence would require extensive enhanced-sampling simulations. The study has not estimated the effect of RY785 binding on the protein-based hydrophobic pore constriction, which may further substantiate their proposed mechanism. And while the study quantified K+ permeation, it does not make any estimates of the ligand binding affinities or rates, which could have been potentially compared to the experiment and used to validate the models. 

As stated in the original manuscript, we concur that the mechanism we propose remains hypothetical until further studies of the complete conformational cycle of the channel are conducted. The recently determined structure of a Kv2.1 channel in the closed state (Mandala and MacKinnon, PNAS 2025) presents an excellent opportunity to do so. Indeed, a cursory analysis of that structure shows that a Pro-Ile-Pro motif in helix S6 marks the position of the intracellular gate, where the pore domain constricts maximally (aside from the selectivity filter). As illustrated in Fig. 5, this motif is precisely where the benzimidazole and thiazole moieties of RY785 bind in our simulations. The mechanism we outline in Fig. 7 thus seems very plausible, in our view; that is RY785 occludes the K⁺ permeation pathway before the pore domain reaches the closed conformation, explaining the observed electrophysiological effects (see Discussion). The Discussion has been revised to note the recent discovery of the aforementioned structure, its implications for the mechanism we propose, and the opportunities for further research that are now open.

Reviewer #3 (Public review):

Summary:

In this manuscript, Zhang et al. investigate the conductivity and inhibition mechanisms of the Kv2.1 channel, focusing on the distinct effects of TEA and RY785 on Kv2 potassium channels. The study employs microsecond-scale molecular dynamics simulations to characterize K+ ion permeation and compound binding inhibition in the central pore. 

Strengths:

The findings reveal a unique inhibition mechanism for RY785, which binds to the channel walls in the open structure while allowing reduced K+ flow. The study also proposes a long-range allosteric coupling between RY785 binding in the central pore and its effects on voltage-sensing domain dynamics. Overall, this well-organized paper presents a high-quality study with robust simulation and analysis methods, offering novel insights into voltage-gated ion channel inhibition that could prove valuable for future drug design efforts.

Weaknesses:

(1) The study neglects to consider the possibility of multiple binding sites for RY785, particularly given its impact on voltage sensors and gating currents. Specifically, there is potential for allosteric binding sites in the voltage-sensing domain (VSD), as some allosteric modulators with thiazole moieties are known to bind VSD domains in multiple voltage-gated sodium channels (Ahuja et al., 2015; Li et al., 2022; McCormack et al., 2013; Mulcahy et al., 2019).

As noted in the manuscript, we designed our simulations to explore the possibility that RY785 binds within the pore domain, because TEA and RY785 are competitive and TEA is known to bind within the pore. That RY785 did in fact spontaneously and reproducibly bind within the pore was however not a predetermined outcome; if the site of interaction for the inhibitor was elsewhere in the channel, the simulation would not have shown a stable associated state, which would have prompted us to examine other possible sites, including the voltage sensors. It was also not predetermined or foreseeable a priori that the mode of interaction we observed in simulation provides a straightforward rationale for the electrophysiological effects of RY785. Based on our results, therefore, we believe that RY785 binds within the pore of Kv2. As stated by the reviewer, other allosteric modulators are known to bind instead to the sensors; to our knowledge, however, there is no precedent of a small-molecule inhibitor that simultaneously acts on the sensors and the pore domain. We therefore believe that future studies should focus on corroborating or refuting the mechanism we propose, through additional experimental and computational work; if, contrary to our claim, RY785 is found not to bind to the pore domain, it would be logical to explore other possible sites of interaction, as the reviewer suggests. The Discussion has been modified to address this point.

(2) The study describes RY785 as a selective inhibitor of Kv2 channels and characterizes its binding residues through MD simulations. However, it is not clear whether the identified RY785-binding residues are indeed unique to Kv2 channels.

To clarify this question, we have included a multiple sequence alignment as Supplementary Figure 1; the revised manuscript refers to this figure in the Discussion section. The alignment reveals that the cluster of residues forming contacts with RY785 (Val409, Pro406, Ile405, Ile401, and Val398) is indeed specific to Kv2.1. Among Kv channels, Kv3.1 and Kv4.1 exhibit the greatest similarity to Kv2.1 at these positions, but they differ in a crucial substitution: Ile405 in Kv2.1 is replaced by Val. This replacement shortens the sidechain, undoubtedly reducing the magnitude of the hydrophobic interaction between inhibitor and channel (Val is approximately 6 kcal/mol, i.e. 1,000 times, more hydrophilic than Ile). Kv5.1 differs from Kv2.1 at two positions: Pro406 is replaced by His, and Val409 by Ile. The introduction of His abolishes the hydrophobic interaction at that position, and the need for hydration likely perturbs all adjacent contacts with RY785. Lastly, Kv6-Kv10 and Cav channels feature entirely different residues at these positions. Consistent with these findings, a recent study by the Sack lab (https://elifesciences.org/articles/99410) has demonstrated that Kv5, Kv6, Kv8, and Kv9 pore subunits confer resistance to RY785, while a high-throughput electrophysiological study carried out by Merck (Herrington et al., 2011) reported that RY785 shows no significant activity against Cav channels. The sequence alignment offers a simple interpretation for these experimental observations, namely that RY785 is recognized by Kv2 channels through the abovementioned hydrophobic cluster within the pore domain.

(3) The study does not clarify the details, rationale, and ramifications of a biasing potential to dihedral angles.

We refer the reviewer to published work, for example Stix et al, 2023 and Tan et al, 2022. We provide additional comments below.

(4) The observation that the Kv2.1 central pore remains partially permeable to K+ ions when RY785 is bound is intriguing, yet it was not revealed whether polar groups of RY785 always interact with K+ ions.

We detected no persistent specific interactions between RY785 and the permeant K+ ions.

Recommendations for the authors:

Reviewer #2 (Recommendations for the authors):

The manuscript describes atomistic molecular dynamics (MD) simulations of a voltage-gated potassium channel Kv2.1 using its cryo-EM structure in the open activated state and its inhibition by a classical non-specific cationic blocker tetraethylammonium (TEA) as well as a novel selective inhibitor RY785. Using multi-microsecond-long all-atom MD runs under the applied membrane voltage of 100 mV the authors were able to confirm that the channel structure represents an open conducting state with the computed single-channel conductance lower than experimental values, but still in the same order of magnitude range. They also determined that both TEA and RY785 bind in the channel pore between the cytoplasmic hydrophobic gate and narrow selectivity filter (SF) region near the extracellular side. However, while TEA directly blocks a knock-on K+ conduction by physically obstructing ion access to the SF, the mechanism of action of RY785 is different. It does not directly prevent K+ access to the SF but rather binds to multiple residues in the hydrophobic gate region, which effectively narrows a pore and drives the channel toward a semi-closed nonconductive conformation, which might be distinct from one with the deactivated voltage sensors and closed pore observed at hyperpolarized membrane potentials. However, additional studies beyond the scope of this work might be needed to fully establish this mechanism as suggested by the authors.

The manuscript is written very well and represents a significant advance in the field of ion channel research. I do not have any major issues, which need to be addressed. However, I have several suggestions.

For the apo-channel K+ conduction MD simulation under the applied voltage, the authors seem to observe mostly a direct or Coulomb knock-on mechanism across the SF with almost no water copermeation. This is in line with computational electrophysiology studies with dual membrane setup by B. de Groot and others but in disagreement with multiple previous studies by B. Roux and others also using applied electric field and CHARMM force fields as in the present study. I wonder why the outcomes are so different. Is it related to the Kv2.1 channel itself, a relatively small applied electric field used (corresponding to a membrane potential of 100 mV vs. 500-750 mV used in many previous simulations), ion force field (e.g., LJ parameters), or some other factors? Could weak dihedral restraints on the protein backbone and side chains contribute to this mechanism? I also wonder if the authors might have considered different initial SF ion configurations. Related to that, I wonder if the authors observed any SF distortions in their simulations including frequently observed backbone carbonyl flipping and/or dilation/contraction.

We are aware of these discrepancies between published simulation studies, but cannot offer a satisfactory explanation, beyond speculation. The reviewer is correct that the mechanism of ion permeation we observe is comparable to that reported by de Groot, as we noted in Tan et al, 2022 and Stix et al, 2023. Neither in this nor in those previous studies did we observe any persistent distortions of the selectivity filter – but that outcome was expected by construction. The weak biasing potentials acting on the mainchain dihedral angles allow for local fluctuations but not a persistent deformation, relative to the conductive form determined experimentally.

For MD simulations with the ligand present, I wonder if the authors can comment on the effect of the ligand especially RY785 on the pore size or more importantly size of the hydrophobic gate. The presence of the ligand itself would definitely result in a narrower pore, but I also wonder if this would also lead to a rearrangement of pore sidechain and/or backbone residues, which would lead to a narrower pore from a protein itself thus confirming the proposed mechanism of driving the channel towards a semi-closed state. It is easy to compute but I wonder if the presence of weak dihedral restraints may preclude this analysis.

Yes, while the simulation design used in this study allows for local fluctuations in the mainchain structure and nearly unrestricted sidechain dynamics, changes in either the secondary or tertiary structure of the channel are strongly disfavored. This approach is thus sufficient to examine ligand binding or ion flow in the microsecond timescale but not channel gating. In the revised version of the Discussion, we outline a roadmap for future computational studies of that gating process, on the basis of the open-channel structure we used and the recently determined structure of the closed state.

The authors state that RY785 does not block K+ ion, but it does significantly slow the rate of K+ ion access to the pore Scav site. Is this not a part of the mechanism for inhibition of the channel? The authors seem to focus on the primary mechanism of inhibition as the RY785 promoting channel closing, but would it not also reduce K+ current in the open state by slowing the rate of K+ entry into the cavity and selectivity filter? The authors should address this point in the text. I am also somewhat confused that in the MD simulations performed by the authors, there is still some K+ conduction with RY785 in the pore, which is not in 100% agreement with electrophysiology experiments. Does it mean that the channel in the simulations has not yet reached that semiclosed state or a reduced K+ conduction is not observed experimentally?

The salient experimental observation is RY785 abrogates K+ currents through Kv2 channels (Herrington et al, 2011; Marquis et al, 2022). In our view, that observation can be explained in one of two ways: either RY785 completely blocks the flow of K+ ions across the channel while the pore domain remains in the conductive, open state – like TEA does – or RY785 induces or facilitates the closing of the channel, thereby abrogating K+ flow. The fact that we observe K+ flow while RY785 is bound to the channel is therefore not in disagreement with the electrophysiological measurements, but it does rule out the first of those two possible interpretations of the existing experiments. As it happens, the second possible explanation, i.e. that RY785 facilitates the closing of the pore domain, also provides a rationale for another puzzling experimental observation, namely that RY785 shifts the voltage dependence of the currents produced by the voltage sensors as they reconfigure to open or close the intracellular gate.

Also, I wonder if the authors considered that since there are 4 potential equivalent sites in the pore (although, overlapping) more than one RY785 might be needed to prevent K+ conduction, even though the experimental Hill coefficient of ~1 does not indicate cooperativity.

Admittedly, our simulation design was based on the premise that only one RY785 molecule might be recognized within the pore. Based on the outcome of the simulations, we are confident that this assumption was valid, as the binding pose that we identified rules out multiple occupancy – which would be indeed consistent with a Hill coefficient of ~1.

I also wonder if the authors considered estimating ligand binding affinities and/or "on" rates from their simulations to have a more direct comparison with experiments and test the accuracy of their models. There are multiple enhanced sampling techniques allowing to do that, although it can be a study on its own.

We thank the reviewer for this suggestion, which we will consider for future studies.

The authors also discussed that they could not study Kv2.1 deactivation in a reasonable simulation time. Indeed it is very challenging but they should cite previous studies e.g. 2012 Jensen et al paper (PMID: 22499946) on this subject. There are structures of Kv channels with the deactivated voltagesensing domains (VSDs) available, e..g of EAG1 channel (PDB 8EP1), although they do not have a domain-swapped architecture. There are structural modeling approaches including AlphaFold, which can be potentially used to get a Kv2.1 structure with deactivated VSDs, and targeted MD, string method etc. can be used to study transition between different states with and without bound ligands.

As noted, a structure of a Kv2 channel with a closed pore has now been determined experimentally. In the revised Discussion, we comment on what this structure tells us about the mechanism of inhibition we propose, and how it could be leveraged in future studies.

The authors should be commended for doing a thorough QM-based force field parameterization of RY785. However, a validation of the developed force field parameters is lacking. In terms of QM validation, a gas-phase dipole moment can be compared in terms of direction and magnitude (it's normal to be overestimated to implicitly reflect solvent-induced polarization). If there are any experimental data available for this compound, they can be tested as well.

We agree with the reviewer that forcefield validation is important, but to our knowledge no experimental data exists for RY785 to compare with, such as hydration free energies. We did however compare the gas-phase dipole moment computed with QM and with the MM forcefield we developed based on atomic charges optimized to reproduce QM interactions with water. The MM model yields a gas-phase dipole moment of 3.94 D, which is 20% greater than the QM dipole moment, or 3.23 D. That deviation is within the typical range for electroneutral molecules (Vanommeslaeghe et al, 2010), and as the reviewer notes, reflects the solvent-induced polarization implicit in the derivation of atomic charges. As shown in Author response image 1, the orientation of the dipole moment calculated with MM (right, blue arrow) is also in good agreement with that predicted with QM (left)

Author response image 1.

(1) p. 3 "the last two helices in each subunit" -> "the last two transmembrane helices in each subunit".

Thanks. Corrected.

(2) p. 5 "and therefore do not cause large density variations e.g. 100-fold or greater.". I would be more specific here and indicate what are the actual variations in density or free energy encountered and how they are compared e.g. with thermal fluctuations (~kT).

Thanks. The exact variations in K+ density had been included in the original manuscript, in Fig. 2C, but we failed to refer to this figure at this point in the description of the results. The ion density is plotted in a log scale to facilitate conversion to free-energy units. Corrected.

(3) p. 6 Figure 1 caption "and along the perpendicular to the membrane" -> "perpendicular to the membrane normal"?. "The channel is an assembly of four distinct subunits (in colors);" -> "The channel is an assembly of four identical subunits (distinct by colors);". I would use the same protein coloring method in panels B and C as was used in panel A.

Thanks. Corrected as needed.

(4) p. 6 Figure 2 In panel B I would appreciate a representative complete ion permeation event trace. In panel C caption I would indicate corresponding sites "S0-S4, Scav" for each residue mentioned. I also would not use gray color for site names in the figure.

We appreciate the suggestion, but believe the figure is clear as is. Panel B is meant to focused on the mechanism of knock-on. Panel A includes numerous complete permeation events. 

(5) p. 7 Figure 3 caption. Please indicate which atoms of residues T373 and P406 were used to define SF and gate positions. Chemical structures of both TEA and RY785 would be useful. In panels C and F channel interacting residues (if any) would be helpful to show.

The revised caption clarifies that the positions of T373 and P406 are represented by their carbonalpha atoms. A close-up view of the structures of TEA and RY785 is included in the Supplementary Information section.

(6) p. 8. Figure 4 caption. Please indicate if N atoms ere used for density maps in panels B and C, and which value of the density was used to show meshes. In panel A please indicate what are the units of the density shown by color maps. 

The caption has been revised to clarify these questions.

(7) p. 9 "inside the protein" -> "inside the channel pore".

Thanks. Corrected.

(8) p. 10 "which lines the cavity" -> "which lines the water-filled cavity"

We appreciate the suggestion but believe the wording is clear as is.

(9) p.10 Fig. 5. It would be helpful to distinguish residues from different chains e.g. by different colors rather than using different colors for different residues. The S atom in RY785 is hard to recognize due to the yellow color used for C atoms. Figure 5B is very confusing. It is not clear what this plot represents. For instance, what does it mean that Pro405 has ~10 contacts in 20% of simulation snapshots? Does it mean 10 C..C/S interactions within 4.5 A? I am not sure what the value of this is. I think a bar or radar chart plot showing % of contacts with one, two, or more residues of each type would be more helpful. 

Thanks. The revised caption ought to clarify how to interpret the plot.

(10) p. 12 "Due to its 2-fold molecular symmetry". TEA has a tetrahedral point group or Td symmetry. It has several two-fold rotational axes though. 

Thanks. Corrected.

(11) p. 12 "it prevents K+ ions in the cytoplasmic space from destabilizing the K+ ions that reside in the selectivity filter" I am not sure if this statement is entirely accurate as there might be destabilization of a multi-ion SF configuration not ions per see.

We believe this statement is clear as is.

(12) p. 13 Fig. 7 caption "includes non-conductive or transiently inactivated states" - I am not sure what "transiently inactivated state" is as inactivation is a specific term used in ion channel research and it does not seem to be explicitly considered in this study.

A reference has been included in the caption for readers interested in the process of inactivation.

(13) p. 14 "the net charge of these constructs is thus zero". This would depend on the number of basic and acidic residues in the protein. 

Yes, it does – and as a result the construct we model has a net zero charge.

(14) p. 14 I wonder if the protein was constrained or heavily restrained during MARTINI membrane building and equilibration procedure. Otherwise, C-alpha mapping would be problematic and clashes with lipid membrane atoms might take place as well.

It was indeed. When a protein is simulated using the MARTINI coarse-grained forcefield, its fold must be preserved through a network of strong ‘virtual’ bonds between adjacent carbon-alpha atoms. This is standard practice so we do not believe it requires further explanation.

(15) p. 15 PME - please spell out and provide reference.

Corrected.

(16) p. 15 "with a smooth switching function" - is it a special or standard switching function? Also, was it used for energy or forces? 

The switching function brings both forces and energies to a value of zero at the cut-off value, smoothly. We refer the reviewer to the NAMD manual for further details.

(17) p. 15 '𝑘 = 1 𝑘B𝑇.' Please confirm that there is a factor of "1" there, which can be actually skipped if this is the case. 

The value of k = 1 KBT is correct.

(18) p. 15. Please cite PMID: 22001851 for the transmembrane electric field application technique.

Corrected.

(19) p. 15 "and CHARMM36m" -> "and CHARMM36m force field". 

Corrected.

(20) p. 16 "the four proteins subunits" -> "the four protein subunits". 

Corrected.

(21) p. 16. Please provide the reference for CGenFF. It's reference 49. 

Corrected.

Supporting Information (SI): CGenFF is misspelled in multiple figure captions in the SI. All potential energy scans indicate "angle", but some are bond angles while others are dihedral angles. Using subscripts for atom numbers is confusing and does not match the numbering scheme used in Fig. S1. So, please use the same style of numbering throughout, e.g. C46-C42-N43 (without subscripts). Please label the X and Y axes in Figsures S2-S19 and S21. In Figure S22 please perform a linear regression analysis and/or compute Pearson correlation coefficients and indicate trend lines. Table S1. It would be good to compute RMS or mean unsigned errors to get an idea about accuracy. Also, please indicate if reference QM values were scaled by 1.16 for energies or offset for distances. 

The Supplementary Information has been corrected. We thank the reviewer for their detailed feedback. 

Reviewer #3 (Recommendations for the authors):

(1) The study needs to consider the possibility of multiple binding sites for RY785, particularly given its impact on voltage sensors and gating currents. Specifically, the potential for allosteric binding sites in the voltage-sensing domain (VSD) should be assessed, as some allosteric modulators with thiazole moieties are known to bind VSD domains in multiple voltage-gated sodium channels (Ahuja et al., 2015; Li et al., 2022; McCormack et al., 2013; Mulcahy et al., 2019). Molecular docking and/or MD simulations could quickly test this hypothesis. If this hypothesis is not true, a comprehensive search can exclude such a possibility, which can also confirm the long-range allosteric coupling between RY785 binding in the central pore and voltage-sensing domain dynamics. 

Please see our response above.

(2) The authors describe RY785 as a selective inhibitor of Kv2 channels and characterize its binding residues through MD simulations. To support this claim, Figure 5 needs to include a multiple sequence alignment with other Kv channels. This would help demonstrate whether the identified RY785-binding residues are indeed unique to Kv2 channels.

Please see our response above.

(3) The study applies a biasing potential to 𝜙, 𝜓, and 𝜒1 dihedral angles. Please clarify:

(a) Is this potential solely to prevent selectivity filter collapse/degradation, as mentioned in a previous D. E. Shaw Research publication (Jensen et al., 2012)?

Yes, that is correct.

(b) If it applies to all amino acids, can this potential prevent other changes, such as in the voltagesensing domain?

Yes, that is correct.

(c) What specific "large-scale structural changes" does this potential preclude? 

For example, it would preclude the spontaneous degradation of the secondary or tertiary structure of the protein. We have revised the Methods section to make these points clearer. 

(d) Given that such biasing potentials on backbone dihedral angles can decrease conformational flexibility, and considering that Kv channel permeability/conductivity could be highly sensitive to filter flexibility, what insights can you provide about the impact of the force constant k on channel conductivity?

In previous studies based on an identical methodology (Stix et al, 2023; Tan et al, 2022), we have observed good agreement between calculated and experimental conductance values – at least as good as can be hoped for, when all approximations are considered. Based on the data presented in those studies, we have no reason to believe our methodology inhibits the permeability of the channel, which is logical as the local structural fluctuations required for K+ flow across the selectivity filter are not impaired, by definition. To the contrary, the fact that these weak biasing potentials make the conductive form of the filter the most favorable state in simulation enable a clear-cut analysis of conductance under plausible simulation conditions, both in terms applied voltage and K+ concentration. We refer the reviewer to the abovementioned studies for further details and a discussion of this subject.

(4) The observation that the Kv2.1 central pore remains partially permeable to K+ ions when RY785 is bound is intriguing. Given the compact nature of the central cavity when RY785 is bound, it would be valuable to investigate whether polar groups of RY785 (e.g., nitrogens from the amide, benzimidazole, and thiazole moieties) always interact with K+ ions. Characterizing these interactions could inform the design of similar compounds with differential modulation effects.

We examined this possibility and detected no convincing interaction patterns between RY785 and K+ ions – logically, inhibitor and ions are in close proximity while residing concurrently within the pore, but we detected no evidence of specific interactions.

Minor points:

It is strongly recommended that the refined force field parameters for RY785 be shared as a separate supplementary file in CHARMM force field format. This addition would be valuable for the scientific community, allowing other researchers to use or compare these parameters in future studies.

We agree entirely. Upon publication of the VOR for this article the forcefield parameters for RY785 will be made freely available for download at https://github.com/Faraldo-Gomez-Lab-atNIH/Download.

The study uses a KCl concentration of 300 mM, which exceeds typical intracellular K+ levels. While this may be intentional to enhance K+ permeation probability, a brief justification for this choice should be included in the Methods section.

Yes, what motivated this choice in this and in our previous studies of K+ channels was the expectation of a greater number of permeation events, for a given simulation length, and therefore greater confidence (i.e. statistical significance) in the observed ion conductance, or in the degree to which it might be inhibited by a blocker. It worth noting that 300 mM KCl, while atypical in the intracellular environment, is often used in electrophysiological studies. The Methods section has been amended to clarify this point.

eLife Assessment

This study presents a valuable contribution by introducing a model-based, Bayesian method for inferring action potentials from calcium imaging data that directly quantifies uncertainty in spike timing through posterior distributions. Using a Monte Carlo particle Gibbs sampling approach, the method achieves temporal resolution and accuracy comparable to existing techniques while offering the key added benefit of principled uncertainty estimates. The underlying methodology and characterization are solid, and the work will be of particular interest to theoretically oriented neuroscientists seeking rigorous new tools for data-driven parameter inference.

Reviewer #1 (Public review):

Summary:

In this study, Diana et al. present a Monte Carlo-based method to perform spike inference from calcium imaging data. A particular strength of their approach is that they can estimate not only averages but also uncertainties of the modeled process. The authors focus on the quantification of spike time uncertainties in simulated data and in data recorded with high sampling rate in cebellar slices with GCaMP8f, and they demonstrate the high temporal precision that can be achieved with their method to estimate spike timing.

Strengths:

- The author provide a solid ground work for sequential Monte Carlo-based spike inference, which extends previous work of Pnevmatikakis et al., Greenberg et al. and others.

- The integration of two states (silence vs. burst firing) seems to improve the performance of the model.

- The acquisition of a GCaMP8f dataset in cerebellum is useful and helps make the point that high spike time inference precision is possible under certain conditions.

Weaknesses:

- Although the algorithm is compared (in the revised manuscript) to other models to infer individual spikes (e.g., MLSpike), these comparisons could be more comprehensive. Future work that benchmarks this and other algorithms under varying conditions (e.g., noise levels, temporal resolution, calcium indicators) would help assess and confirm robustness and useability of this algorithm.

- The mathematical complexity underlying the method may pose challenges for experimentalist who may want to use the methods for their analyses. While this is not a weakness of the approach itself, this highlights the need for further validation and benchmarking in future work, to build user confidence.

Comments on revisions:

Thank you for addressing the final comments, and congrats on this study!

Reviewer #2 (Public review):

Summary:

Methods to infer action potentials from fluorescence-based measurements of intracellular calcium dynamics are important for optical measurements of activity across large populations of neurons. The variety of existing methods can be separated into two broad classes: a) model-independent approaches that are trained on ground truth datasets (e.g., deep networks), and b) approaches based on a model of the processes that link action potentials to calcium signals. Models usually contains parameters describing biophysical variables, such as rate constants of the calcium dynamics and features of the calcium indicator. The method presented here, PGBAR, is model-based and uses a Bayesian approach. A novelty of PGBAR is that static parameters and state variables are jointly estimated using particle Gibbs sampling, a sequential Monte Carlo technique that can efficiently sample the latent embedding space.

Strengths:

A main strength of PGBAR is that it provides probability distributions rather than point estimates of spike times. This is different from most other methods and may be an important feature in cases when estimates of uncertainty are desired. Another important feature of PGBAR is that it estimates not only the state variable representing spiking activity, but also other variables such as baseline fluctuations and stationary model variables, in a joint process. PGBAR can therefore provide more information than various other methods. The information in the github repository is well-organized. The authors demonstrate convincingly that PGBAR can resolve inter-spike intervals in the range of 5 ms using fluorescence data obtained with a very fast genetically encoded calcium indicator at very high sampling rates (line scans at >= 1 kHz).

Weaknesses:

The accuracy of spike train reconstructions is not higher than that of other model-based approaches, and lower than the accuracy of a model-independent approach based on a deep network in a regime of commonly used acquisition rates.

Comments on revisions:

I have no further comments on the manuscript.

Author response:

The following is the authors’ response to the previous reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

In this study, Diana et al. present a Monte Carlo-based method to perform spike inference from calcium imaging data. A particular strength of their approach is that they can estimate not only averages but also uncertainties of the modeled process. The authors focus on the quantification of spike time uncertainties in simulated data and in data recorded with high sampling rate in cebellar slices with GCaMP8f, and they demonstrate the high temporal precision that can be achieved with their method to estimate spike timing.

Strengths:

- The author provide a solid ground work for sequential Monte Carlo-based spike inference, which extends previous work of Pnevmatikakis et al., Greenberg et al. and others.

- The integration of two states (silence vs. burst firing) seems to improve the performance of the model.

- The acquisition of a GCaMP8f dataset in cerebellum is useful and helps make the point that high spike time inference precision is possible under certain conditions.

Weaknesses:

- Although the algorithm is compared (in the revised manuscript) to other models to infer individual spikes (e.g., MLSpike), these comparisons could be more comprehensive. Future work that benchmarks this and other algorithms under varying conditions (e.g., noise levels, temporal resolution, calcium indicators) would help assess and confirm robustness and useability of this algorithm.

The metrics used for comparison follow the field's benchmarking conventions (see the CASCADE paper, Rupprecht et al. 2021). Indeed, improved standardized methods would be ideal to develop, which is beyond the scope of this manuscript.

- The mathematical complexity underlying the method may pose challenges for experimentalist who may want to use the methods for their analyses. While this is not a weakness of the approach itself, this highlights the need for further validation and benchmarking in future work, to build user confidence.

We acknowledge the challenges of understanding the mathematics underlying our method, but such a study is necessary to ensure its accuracy and reliability. Indeed, we will strive to improve the technique's user-friendliness in future instantiations.

Reviewer #2 (Public review):

Summary:

Methods to infer action potentials from fluorescence-based measurements of intracellular calcium dynamics are important for optical measurements of activity across large populations of neurons. The variety of existing methods can be separated into two broad classes: a) model-independent approaches that are trained on ground truth datasets (e.g., deep networks), and b) approaches based on a model of the processes that link action potentials to calcium signals. Models usually contains parameters describing biophysical variables, such as rate constants of the calcium dynamics and features of the calcium indicator. The method presented here, PGBAR, is model-based and uses a Bayesian approach. A novelty of PGBAR is that static parameters and state variables are jointly estimated using particle Gibbs sampling, a sequential Monte Carlo technique that can efficiently sample the latent embedding space.

Strengths:

A main strength of PGBAR is that it provides probability distributions rather than point estimates of spike times. This is different from most other methods and may be an important feature in cases when estimates of uncertainty are desired. Another important feature of PGBAR is that it estimates not only the state variable representing spiking activity, but also other variables such as baseline fluctuations and stationary model variables, in a joint process. PGBAR can therefore provide more information than various other methods. The information in the github repository is well-organized.

Weaknesses:

On the other hand, the accuracy of spike train reconstructions is not higher than that of other model-based approaches, and clearly lower than the accuracy of a model-independent approach based on a deep network. The authors demonstrate convincingly that PGBAR can resolve inter-spike intervals in the range of 5 ms using fluorescence data obtained with a very fast genetically encoded calcium indicator at very high sampling rates (line scans at >= 1 kHz).

In the revision, Figure 9 shows that temporal accuracy is very similar between PGBAR and the supervised method, CASCADE, and that PGBAR has a lower false positive rate. These results support the effectiveness of unsupervised Monte Carlo sampling, even with a simple autoregressive model.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

I'd like to thank the authors for their revisions. Their comments have addressed all my concerns, and I thank them for the clarifications. I have no further comments, except a few minor notes that the authors may consider or not:

- The paragraph starting in line 367 is newly written and not yet as clear and mature as other parts of the manuscript. It is at several sentences roughly clear what it is about, but the precision of the wording is lacking. For example "distributions of the average time from ground-truth" seems a bit unclear, maybe "distributions of the average time of estimate spikes from ground-truth spikes" instead. Similarly, "the false detection rate, defined as the difference between detected and ground-truth spikes ..." could be rephrased using the difference between "numbers of spikes" instead of the difference between "spikes". But all of this is minor.

- In the new Figure 9A, the error bars for the MLSpike method seem to be absent. In the same figure legend, it should be "excess" instead of "excess".

We thank the reviewer for the feedback. We revised the wording of the new paragraph in response to the reviewer’s suggestions, restored the missing error bar in Figure 9, and corrected the figure legend.

Reviewer #2 (Recommendations for the authors):

Comparison to CASCADE: as far as I know there are no CASCADE models that have been trained on ground truth data in the regime of very fast (line scan) sampling, which is rarely used. A fair comparison of spike time estimates between PGBAR and CASCADE should take this into account. This can be done by training a new CASCADE model using the dataset of this paper. Given that performance of PGBAR and CASCADE is very similar already now (except for the false positive rate), a CASCADE model optimized for high sampling rate may be expected to catch up with (or even exceed) the performance of PGBAR. At a minimum, this possibility should be discussed.

While this may be true, retraining a CASCADE model on high-frequency ground-truth data is beyond the scope of this manuscript. Indeed, a retrained CASCADE model optimized for line-scan or GCaMP8f data could improve performance and potentially match or exceed PGBAR, particularly in reducing false positives.

Our aim, however, is not to benchmark supervised methods under their optimal retraining conditions, but to provide an unsupervised alternative that does not rely on labeled training data. In practice, retraining supervised models is constrained by the availability of suitable ground-truth datasets and by the uncertainty in how the method generalizes to acquisition regimes that differ substantially from the training set.

We have therefore added a sentence in the Discussion (at the end of the subsection Comparison with benchmark datasets):

[...] “While retraining supervised methods such as CASCADE on high-frequency or GCaMP8f ground-truth datasets could further improve its performance, limitations in dataset availability and generalization across acquisition regimes motivate complementary, training-free approaches such as PGBAR.”

As stated in the manuscript, future extensions, such as using nonlinear biophysical models as the generative model for Monte Carlo–based inference, may further improve spike estimation accuracy.

eLife Assessment

Using a transposon sequencing (TN-seq) approach, the authors identified key genetic determinants of drug tolerance in Mycobacterium abscessus. Given that M. abscessus is inherently resistant to multiple antibiotics, this valuable study makes a significant contribution by uncovering how antibiotic tolerance is linked to reactive oxygen species (ROS) in this non-tuberculous mycobacterial (NTM) species. The solid findings further strengthen the growing evidence that ROS play a central role in the mechanism of antibiotic action and tolerance in mycobacteria. However, the use of words persistence or tolerance should follow the consensus definition given in the Balaban 2019 Nat Rev Micro paper.

Reviewer #2 (Public review):

Summary:

The work set out to better understand the phenomenon of antibiotic persistence in mycobacteria. Three new observations are made using the pathogenic Mycobacterium abscessus as an experimental system: phenotypic tolerance involves suppression of ROS, protein synthesis inhibitors can be lethal for this bacterium, and levofloxacin lethality is unaffected by deletion of catalase, suggesting that this quinolone does not kill via ROS.

Strengths:

The ROS experiments are supported in three ways: measurement of ROS by a fluorescent probe, deletion of catalase increases lethality of selected antibiotics, and a hypoxia model suppresses antibiotic lethality. A variety of antibiotics are examined, and transposon mutagenesis identifies several genes involved in phenotypic tolerance, including one that encodes catalase. The methods are adequate for making these statements.

Overall impact:

Showing that ROS accumulation is suppressed during phenotypic tolerance, while expected, adds to the examples of the protective effects of low ROS levels. Moreover, the work, along with a few others, extends the idea of antibiotic involvement with ROS to mycobacteria. These observations help solidify the field. The work raises an important unanswered question: why are rifampicin and many protein synthesis inhibitors bacteriostatic with E. coli but bactericidal with pathogenic mycobacteria?

Comments on revisions:

I call attention to word choice, because it can indicate how familiar the authors are with the field. An issue that caught my attention was the use of the words persistence and tolerance, because they are not uniformly used in the generally accepted way (see Balaban 2019 Nat Rev Micro). In this consensus statement persistence refers specifically to a subpopulation and as such has survival kinetics that are distinct from those seen with tolerance, a phenomenon that refers to the entire population. I notice that the Balaban paper is not in the reference list. My suggestion is to take a look at the Balaban paper and then examine every use of the words tolerance and persistence in the manuscript to be sure that they fit the Balaban definition.

Author response:

The following is the authors’ response to the previous reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

Persistence is a phenomenon by which genetically susceptible cells are able to survive exposure to high concentrations of antibiotics. This is especially a major problem when treating infections caused by slow growing mycobacteria such as M. tuberculosis and M. abscessus. Studies on the mechanisms adopted by the persisting bacteria to survive and evade antibiotic killing can potentially lead to faster and more effective treatment strategies.

To address this, in this study, the authors have used a transposon mutagenesis based sequencing approach to identify the genetic determinants of antibiotic persistence in M. abscessus. To enrich for persisters they employed conditions, that have been reported previously to increase persister frequency - nutrient starvation, to facilitate genetic screening for this phenotype. M.abs transposon library was grown in nutrient rich or nutrient depleted conditions and exposed to TIG/LZD for 6 days, following which Tnseq was carried out to identify genes involved in spontaneous (nutrient rich) or starvationinduced conditions. About 60% of the persistence hits were required in both the conditions. Pathway analysis revealed enrichment for genes involved in detoxification of nitrosative, oxidative, DNA damage and proteostasis stress. The authors then decided to validate the findings by constructing deletions of 5 different targets (pafA, katG, recR, blaR, Mab_1456c) and tested the persistence phenotype of these strains. Rather surprisingly only 2 of the 5 hits (katG and pafA) exhibited a significant persistence defect when compared to wild type upon exposure to TIG/LZD and this was complemented using an integrative construct. The authors then investigated the specificity of delta-katG susceptibility against different antibiotic classes and demonstrated increased killing by rifabutin. The katG phenotype was shown to be mediated through the production of oxidative stress which was reverted when the bacterial cells were cultured under hypoxic conditions. Interestingly, when testing the role of katG in other clinical strains of Mab, the phenotype was observed only in one of the clinical strains demonstrating that there might be alternative anti-oxidative stress defense mechanisms operating in some clinical strains.

Strengths:

While the role of ROS in antibiotic mediated killing of mycobacterial cells have been studied to some extent, this paper presents some new findings with regards to genetic analysis of M. abscessus susceptibility, especially against clinically used antibiotics, which makes it useful. Also, the attempts to validate their observations in clinical isolates is appreciated.

Weaknesses:

Amongst the 5 shortlisted candidates from the screen, only 2 showed marginal phenotypes which limits the impact of the screening approach.

We appreciate the reviewer’s comments, but we note that 4 out of 5 genes displayed phenotypes concordant with findings of the Tn-Seq data, with katG and pafA, as well as MAB_1456c (during starvation only) and blaR (in rich media only) having decreased survival as shown in Figure 3A-D. We do agree that some of the phenotypes were more modest in a single-mutant context than in the pooled Tn-Seq screen. In addition, several mutants that had modest changes in survival also showed profound defects in resuming growth after removal of antibiotics, with the pafA mutants particularly impaired. (Figure 3 - figure supplement 1).

While the role of KatG mediated detoxification of ROS and involvement of ROS in antibiotic killing was well demonstrated, the lack of replication of this phenotype in some of the clinical isolates limits the significance of these findings.

While the role of katG varied among strains, the antibiotic-induced accumulation of ROS was seen in all three strains (Figure 6A). This suggests that in some strains other ROS-detoxification pathways are able to compensate for the loss of katG.

(Figure 2—figure supplements 1–3)

Figure 1—figure supplement 1.

Reviewer #2 (Public review):

Summary:

The work set out to better understand the phenomenon of antibiotic persistence in mycobacteria. Three new observations are made using the pathogenic Mycobacterium abscessus as an experimental system: phenotypic tolerance involves suppression of ROS, protein synthesis inhibitors can be lethal for this bacterium, and levofloxacin lethality is unaffected by deletion of catalase, suggesting that this quinolone does not kill via ROS.

Strengths:

The ROS experiments are supported in three ways: measurement of ROS by a fluorescent probe, deletion of catalase increases lethality of selected antibiotics, and a hypoxia model suppresses antibiotic lethality. A variety of antibiotics are examined, and transposon mutagenesis identifies several genes involved in phenotypic tolerance, including one that encodes catalase. The methods are adequate for making these statements.

Weaknesses:

The work can be improved by a more comprehensive treatment of prior work, especially comparison of E. coli work with mycobacterial studies.

Moreover, the work still has some technical issues to fix regarding description of the methods, supplementary material, and reference formating.

See detailed responses below.

Overall impact: Showing that ROS accumulation is suppressed during phenotypic tolerance, while expected, adds to the examples of the protective effects of low ROS levels. Moreover, the work, along with a few others, extends the idea of antibiotic involvement with ROS to mycobacteria. These are fieldsolidifying observations.

Comments on revisions:

The authors have moved this paper along nicely. I have a few general thoughts.

It would be helpful to have more references to specific figures and panels listed in the text to make reading easier.

Text modified to add more figure references.

(1) I would suggest adding a statement about the importance of the work. From my perspective, the work shows the general nature of many statements derived from work with E. coli. This is important. The abstract says this overall, but a final sentence in the abstract would make it clear to all readers.

We appreciate the suggestion and have added a line to the abstract.