Author response:

We thank the editors and reviewers for their generally positive and thoughtful feedback on this work. Below are provisional responses to some of the concerns raised:

Reviewer 1:

At a total scan duration of 2 minutes, the ASL sequence utilized in this cohort is much shorter than that of a typical ASL sequence (closer to 5 minutes as mentioned by the authors). However, this implementation also included multiple (n=5) PLDs. As currently described, it is unclear how any repetitions were acquired at each PLD and whether these were acquired efficiently (i.e., with a Look-Locker readout) or whether individual repetitions within this acquisition were dedicated to a single PLD. If the latter, the number of repetitions per PLD (and consequently signal-to-noise-ratio, SNR) is likely to be very low. Have the authors performed any analyses to determine whether the signal in individual subjects generally lies above the noise threshold? This is particularly relevant for white matter, which is the focus of several findings discussed in the study.

We agree that this was a short acquisition compared to most ASL protocols, necessitated by the strict time-keeping requirements for running such a large study. We apologise if this was not clear in the original manuscript, but due to this time constraint and the use of a segmented readout (which was not Look-Locker) there was only time available for a single average at each PLD. This does mean that the perfusion weighted images at each PLD are relatively noisy, although the image quality with this sequence was still reasonable, as demonstrated in Figure 1, with perfusion weighted images visibly above the noise floor. In addition, as has been demonstrated theoretically and experimentally in recent work (Woods et al., 2023, 2020), even though the SNR of each individual PLD image might be low in multi-PLD acquisitions, this is effectively recovered during the model fitting process, giving it comparable or greater accuracy than a protocol which collects many averages at a single (long) PLD. As also noted by the reviewers, this approach has the further benefit of allowing ATT estimation, which has proven to provide useful and complementary information to CBF. Finally, the fact that many of the findings in this study pass strict statistical thresholds for significance, despite the many multiple comparisons performed, and that the spatial patterns of these relationships are consistent with expectations, even in the white matter (e.g. Figure 6B), give us confidence that the perfusion estimation is robust. However, we will consider adding some additional metrics around SNR or fitting uncertainty in a revised manuscript, as well as clarifying details of the acquisition.

Hematocrit is one of the variables regressed out in order to reduce the effect of potential confounding factors on the image-derived phenotypes. The effect of this, however, may be more complex than accounting for other factors (such as age and sex). The authors acknowledge that hematocrit influences ASL signal through its effect on longitudinal blood relaxation rates. However, it is unclear how the authors handled the fact that the longitudinal relaxation of blood (T1Blood) is explicitly needed in the kinetic model for deriving CBF from the ASL data. In addition, while it may reduce false positives related to the relationships between dietary factors and hematocrit, it could also mask the effects of anemia present in the cohort. The concern, therefore, is two-fold: (1) Were individual hematocrit values used to compute T1Blood values? (2) What effect would the deconfounding process have on this?

We agree this is an important point to clarify. In this work we decided not to use the haematocrit to directly estimate the T1 of blood for each participant a) because this would result in slight differences in the model fitting for each subject, which could introduce bias (e.g. the kinetic model used assumes instantaneous exchange between blood water and tissue, so changing the T1 of blood for each subject could make us more sensitive to inaccuracies in this assumption); and b) because typically the haematocrit measures were quite some time (often years) prior to the imaging session, leading to an imperfect correction. We therefore took the pragmatic approach to simply regress each subject’s average haematocrit reading out of the IDP and voxelwise data to prevent it contributing to apparent correlations caused by indirect effects on blood T1. However, we agree with the reviewer that this certainly would mask the effects of anaemia in this cohort, so for researchers interested in this condition a different approach should be taken. We will update the revised manuscript to try to clarify these points.

The authors leverage an observed inverse association between white matter hyperintensity volume and CBF as evidence that white matter perfusion can be sensitively measured using the imaging protocol utilized in this cohort. The relationship between white matter hyperintensities and perfusion, however, is not yet fully understood, and there is disagreement regarding whether this structural imaging marker necessarily represents impaired perfusion. Therefore, it may not be appropriate to use this finding as support for validation of the methodology.

We appreciate the reviewer’s point that there is still debate about the relationship between white matter hyperintensities and perfusion. We therefore agree that this observed relationship therefore does not validate the methodology in the sense that it is an expected finding, but it does demonstrate that the data quality is sufficient to show significant correlations between white matter hyperintensity volume and perfusion, even in white matter regions, which would not be the case if the signal there were dominated by noise. Similarly, the clear spatial pattern of perfusion changes in the white matter that correlate with DTI measures in the same regions also suggests there is sensitivity to white matter perfusion. However, we will update the wording in the revised manuscript to try to clarify this point.

Reviewer 2:

This study primarily serves to illustrate the efficacy and potential of ASL MRI as an imaging parameter in the UK Biobank study, but some of the preliminary observations will be hypothesis-generating for future analyses in larger sample sizes. However, a weakness of the manuscript is that some of the reported observations are difficult to follow. In particular, the associations between ASL and resting fMRI illustrated in Figure 7 and described in the accompanying Results text are difficult to understand. It could also be clearer whether the spatial maps showing ASL correlates of other image-derived phenotypes in Figure 6B are global correlations or confined to specific regions of interest. Finally, while addressing partial volume effects in gray matter regions by covarying for cortical thickness is a reasonable approach, the Methods section seems to imply that a global mean cortical thickness is used, which could be problematic given that cortical thickness changes may be localized.

We apologise if any of the presented information was unclear and will try to improve this in our revised manuscript. To clarify, the spatial maps associated with other (non-ASL) IDPs were generated by calculating the correlation between the ASL CBF or ATT in every voxel in standard space with the non-ASL IDP of interest, not the values of the other imaging modality in the same voxel. No region-based masking was used for this comparison. This allowed us to examine whether the correlation with this non-ASL IDP was only within the same brain region or if the correlations extended to other regions too.

We also agree that the associations between ASL and resting fMRI are not easy to interpret. We therefore tried to be clear in the manuscript that these were preliminary findings that may be of interest to others, but clearly further study is required to explore this complex relationship further. However, we will try to clarify how the results are presented in the revised manuscript.

In relation to partial volume effects, we did indeed use only a global measure of cortical thickness in the deconfounding and we acknowledged that this could be improved in the discussion: [Partial volume effects were] “mitigated here by the inclusion of cortical thickness in the deconfounding process, although a region-specific correction approach that is aware of the through-slice blurring (Boscolo Galazzo et al., 2014) is desirable in future iterations of the ASL analysis pipeline.” As suggested here, although this is a coarse correction, we did not feel that a more comprehensive partial volume correction approach could be used without properly accounting for the through-slice blurring effects from the 3D-GRASE acquisition (that will vary across different brain regions), which is not currently available, although this is an area we are actively working on for future versions of the image analysis pipeline. We again will try to clarify this point further in the revised manuscript.

References

Woods JG, Achten E, Asllani I, Bolar DS, Dai W, Detre J, Fan AP, Fernández-Seara M, Golay X, Günther M, Guo J, Hernandez-Garcia L, Ho M-L, Juttukonda MR, Lu H, MacIntosh BJ, Madhuranthakam AJ, Mutsaerts HJ, Okell TW, Parkes LM, Pinter N, Pinto J, Qin Q, Smits M, Suzuki Y, Thomas DL, Van Osch MJP, Wang DJ, Warnert EAH, Zaharchuk G, Zelaya F, Zhao M, Chappell MA. 2023. Recommendations for Quantitative Cerebral Perfusion MRI using Multi-Timepoint Arterial Spin Labeling: Acquisition, Quantification, and Clinical Applications (preprint). Open Science Framework. doi:10.31219/osf.io/4tskr

Woods JG, Chappell MA, Okell TW. 2020. Designing and comparing optimized pseudo-continuous Arterial Spin Labeling protocols for measurement of cerebral blood flow. NeuroImage 223:117246. doi:10.1016/j.neuroimage.2020.117246

Reviewer #2 (Public review):

Summary:

This is an interesting study exploring methods for reconstructing visual stimuli from neural activity in the mouse visual cortex. Specifically, it uses a competition dataset (published in the Dynamic Sensorium benchmark study) and a recent winning model architecture (DNEM, dynamic neural encoding model) to recover visual information stored in ensembles of mouse visual cortex.

Strengths:

This is a great start for a project addressing visual reconstruction. It is based on physiological data obtained at a single-cell resolution, the stimulus movies were reasonably naturalistic and representative of the real world, the study did not ignore important correlates such as eye position and pupil diameter, and of course, the reconstruction quality exceeded anything achieved by previous studies. There appear to be no major technical flaws in the study, and some potential confounds were addressed upon revision. The study is an enjoyable read.

Weaknesses:

The study is technically competent and benchmark-focused, but without significant conceptual or theoretical advances. The inclusion of neuronal data broadens the study's appeal, but the work does not explore potential principles of neural coding, which limits its relevance for neuroscience and may create some disappointment to some neuroscientists. The authors are transparent that their goal was methodological rather than explanatory, but this raises the question of why neuronal data were necessary at all, as more significant reconstruction improvements might be achievable using noise-less artificial video encoders alone (network-to-network decoding approaches have been done well by teams such as Han, Poggio, and Cheung, 2023, ICML). Yet, even within the methodological domain, the study does not articulate clear principles or heuristics that could guide future progress. The finding that more neurons improve reconstruction aligns with well-established results in the literature that show that higher neuronal numbers improve decoding in general (for example, Hung, Kreiman, Poggio, and DiCarlo, 2005) and thus may not constitute a novel insight.

Specific issues:

(1) The study showed that it could achieve high-quality video reconstructions from mouse visual cortex activity using a neural encoding model (DNEM), recovering 10-second video sequences and approaching a two-fold improvement in pixel-by-pixel correlation over attempts. As a reader, I was left with the question: okay, does this mean that we should all switch to DNEM for our investigations of mouse visual cortex? What makes this encoding model special? It is introduced as "a winning model of the Sensorium 2023 competition which achieved a score of 0.301...single trial correlation between predicted and ground truth neuronal activity," but as someone who does not follow this competition (most eLife readers are not likely to do so, either), I do not know how to gauge my response. Is this impressive? What is the best theoretical score, given noise and other limitations? Is the model inspired by the mouse brain in terms of mechanisms or architecture, or was it optimized to win the competition by overfitting it to the nuances of the data set? Of course, I know that as a reader, I am invited to read the references, but the study would stand better on its own, if it clarified how its findings depended on this model.

The revision helpfully added context to the Methods about the range of scores achieved by other models, but this information remains absent from the Abstract and other important sections. For instance, the Abstract states, "We achieve a pixel-level correlation of 0.57 between the ground truth movie and the reconstructions from single-trial neural responses," yet this point estimate (presented without confidence intervals or comparisons to controls) lacks meaning for readers who are not told how it compares to prior work or what level of performance would be considered strong. Without such context, the manuscript undercuts potentially meaningful achievements.

(2) Along those lines, the authors conclude that "the number of neurons in the dataset and the use of model ensembling are critical for high-quality reconstructions." If true, these principles should generalize across network architectures. I wondered whether the same dependencies would hold for other network types, as this could reveal more general insights. The authors replied that such extensions are expected (since prior work has shown similar effects for static images) but argued that testing this explicitly would require "substantial additional work," be "impractical," and likely not produce "surprising results." While practical difficulty alone is not a sufficient reason to leave an idea untested, I agree that the idea that "more neurons would help" would be unsurprising. The question then becomes: given that this is a conclusion already in the field, what new principle or understanding has been gained in this study?

(3) One major claim was that the quality of the reconstructions depended on the number of neurons in the dataset. There were approximately 8000 neurons recorded per mouse. The correlation difference between the reconstruction achieved by 1000 neurons and 8000 neurons was ~0.2. Is that a lot or a little? One might hypothesize that 7000 additional neurons could contribute more information, but perhaps, those neurons were redundant if their receptive fields are too close together or if they had the same orientation or spatiotemporal tuning. How correlated were these neurons in response to a given movie? Why did so many neurons offer such a limited increase in correlation? Originally, this question was meant to prompt deeper analysis of the neural data, but the authors did not engage with it, suggesting a limited understanding of the neuronal aspects of the dataset.

(4) We appreciated the experiments testing the capacity of the reconstruction process, by using synthetic stimuli created under a Gaussian process in a noise-free way. But this originally further raised questions: what is the theoretical capability for reconstruction of this processing pipeline, as a whole? Is 0.563 the best that one could achieve given the noisiness and/or neuron count of the Sensorium project? What if the team applied the pipeline to reconstruct the activity of a given artificial neural network's layer (e.g., some ResNet convolutional layer), using hidden units as proxies for neuronal calcium activity? In the revision, this concern was addressed nicely in the review in Supplementary Figure 3C. Also, one appreciates that as a follow up, the team produced error maps (New Figure 6) that highlight where in the frames the reconstruction are likely to fail. But the maps went unanalyzed further, and I am not sure if there was a systematic trend in the errors.

(5) I was encouraged by Figure 4, which shows how the reconstructions succeeded or failed across different spatial frequencies. The authors note that "the reconstruction process failed at high spatial frequencies," yet it also appears to struggle with low spatial frequencies, as the reconstructed images did not produce smooth surfaces (e.g., see the top rows of Figures 4A and 4B). In regions where one would expect a single continuous gradient, the reconstructions instead display specular, high-frequency noise. This issue is difficult to overlook and might deserve further discussion.

Author response:

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

Public Reviews: 

Reviewer #2 (Public review): 

Summary: 

This is an interesting study exploring methods for reconstructing visual stimuli from neural activity in the mouse visual cortex. Specifically, it uses a competition dataset (published in the Dynamic Sensorium benchmark study) and a recent winning model architecture (DNEM, dynamic neural encoding model) to recover visual information stored in ensembles of mouse visual cortex. 

Strengths: 

This is a great start for a project addressing visual reconstruction. It is based on physiological data obtained at a single-cell resolution, the stimulus movies were reasonably naturalistic and representative of the real world, the study did not ignore important correlates such as eye position and pupil diameter, and of course, the reconstruction quality exceeded anything achieved by previous studies. There appear to be no major technical flaws in the study, and some potential confounds were addressed upon revision. The study is an enjoyable read. 

Weaknesses: 

The study is technically competent and benchmark-focused, but without significant conceptual or theoretical advances. The inclusion of neuronal data broadens the study's appeal, but the work does not explore potential principles of neural coding, which limits its relevance for neuroscience and may create some disappointment to some neuroscientists. The authors are transparent that their goal was methodological rather than explanatory, but this raises the question of why neuronal data were necessary at all, as more significant reconstruction improvements might be achievable using noise-less artificial video encoders alone (network-to-network decoding approaches have been done well by teams such as Han, Poggio, and Cheung, 2023, ICML). Yet, even within the methodological domain, the study does not articulate clear principles or heuristics that could guide future progress. The finding that more neurons improve reconstruction aligns with well-established results in the literature that show that higher neuronal numbers improve decoding in general (for example, Hung, Kreiman, Poggio, and DiCarlo, 2005) and thus may not constitute a novel insight. 

We thank the reviewer for this second round of comments and hope we were able to address the remaining points below. 

Indeed, using surrogate noiseless data is interesting and useful when developing such methods, or to demonstrate that they work in principle. But in order to evaluate if they really work in practice, we need to use real neuronal data. While we did not try movie reconstruction from layers within artificial neural networks as surrogate data, in Supplementary Figure 3C we provide the performance of our method using simulated/predicted neuronal responses from the dynamic neural encoding model alongside real neuronal responses.

Specific issues: 

(1)The study showed that it could achieve high-quality video reconstructions from mouse visual cortex activity using a neural encoding model (DNEM), recovering 10-second video sequences and approaching a two-fold improvement in pixel-by-pixel correlation over attempts. As a reader, I was left with the question: okay, does this mean that we should all switch to DNEM for our investigations of mouse visual cortex? What makes this encoding model special? It is introduced as "a winning model of the Sensorium 2023 competition which achieved a score of 0.301...single trial correlation between predicted and ground truth neuronal activity," but as someone who does not follow this competition (most eLife readers are not likely to do so, either), I do not know how to gauge my response. Is this impressive? What is the best theoretical score, given noise and other limitations? Is the model inspired by the mouse brain in terms of mechanisms or architecture, or was it optimized to win the competition by overfitting it to the nuances of the data set? Of course, I know that as a reader, I am invited to read the references, but the study would stand better on its own, if it clarified how its findings depended on this model. 

The revision helpfully added context to the Methods about the range of scores achieved by other models, but this information remains absent from the Abstract and other important sections. For instance, the Abstract states, "We achieve a pixel-level correlation of 0.57 between the ground truth movie and the reconstructions from single-trial neural responses," yet this point estimate (presented without confidence intervals or comparisons to controls) lacks meaning for readers who are not told how it compares to prior work or what level of performance would be considered strong. Without such context, the manuscript undercuts potentially meaningful achievements. 

We appreciate that the additional information about the performance of the SOTA DNEM to predict neural responses could be made more visible in the paper and will therefore move it from the methods to the results section instead: 

Line 348 “This model achieved an average single-trial correlation between predicted and ground truth neural activity of 0.291 during the competition, this was later improved to 0.301. The competition benchmark models achieved 0.106, 0.164 and 0.197 single-trial correlation, while the third and second place models achieved 0.243 and 0.265. Across the models, a variety of architectural components were used, including 2D and 3D convolutional layers, recurrent layers, and transformers, to name just a few.” will be moved to the results.

With regard to the lack of context for the performance of our reconstruction in the abstract, we may have overcorrected in the previous revision round and have tried to find a compromise which gives more context to the pixel-level correlation value: 

Abstract: “We achieve a pixel-level correlation of 0.57 (95% CI [0.54, 0.60]) between ground-truth movies and single-trial reconstructions. Previous reconstructions based on awake mouse V1 neuronal responses to static images achieved a pixel-level correlation of 0.238 over a similar retinotopic area.”

(2) Along those lines, the authors conclude that "the number of neurons in the dataset and the use of model ensembling are critical for high-quality reconstructions." If true, these principles should generalize across network architectures. I wondered whether the same dependencies would hold for other network types, as this could reveal more general insights. The authors replied that such extensions are expected (since prior work has shown similar effects for static images) but argued that testing this explicitly would require "substantial additional work," be "impractical," and likely not produce "surprising results." While practical difficulty alone is not a sufficient reason to leave an idea untested, I agree that the idea that "more neurons would help" would be unsurprising. The question then becomes: given that this is a conclusion already in the field, what new principle or understanding has been gained in this study? 

As mentioned in our previous round of revisions, we chose not to pursue the comparison of reconstructions using different model architectures in this manuscript because we did not think it would add significant insights to the paper given the amount of work it would require, and we are glad the reviewer agrees. 

While the fact that more neurons result in better reconstructions is unsurprising, how quickly performance drops off will depend on the robustness of the method, and on the dimensionality of the decoding/reconstruction task (decoding grating orientation likely requires fewer neurons than gray scale image reconstruction, which in turn likely requires fewer neurons than full color movie reconstruction). How dependent input optimization based image/movie reconstruction is on population size has not been shown, so we felt it was useful for readers to know how well movie reconstruction works with our method when recording from smaller numbers of neurons. 

(3) One major claim was that the quality of the reconstructions depended on the number of neurons in the dataset. There were approximately 8000 neurons recorded per mouse. The correlation difference between the reconstruction achieved by 1000 neurons and 8000 neurons was ~0.2. Is that a lot or a little? One might hypothesize that 7000 additional neurons could contribute more information, but perhaps, those neurons were redundant if their receptive fields are too close together or if they had the same orientation or spatiotemporal tuning. How correlated were these neurons in response to a given movie? Why did so many neurons offer such a limited increase in correlation? Originally, this question was meant to prompt deeper analysis of the neural data, but the authors did not engage with it, suggesting a limited understanding of the neuronal aspects of the dataset. 

We apologize that we did not engage with this comment enough in the previous round. We assumed that the question arose because there was a misunderstanding about figure 5: 1000 not 1 neuron is sufficient to reconstruct the movies to a pixel-level correlation of 0.344. Of course, the fact that increasing the number of neurons from 1000 to 8000 only increased the reconstruction performance from 0.344 to 0.569 (65% increase in correlation) is still worth discussing. To illustrate this drop in performance qualitatively, we show 3 example frames from movie reconstructions using 1000-8000 neurons in Author response image 1.

Author response image 1.

3 example frames from reconstructions using different numbers of neurons. 

As the reviewer points out, the diminishing returns of additional neurons to reconstruction performance is at least partly because there is redundancy in how a population of neurons represents visual stimuli. In supplementary figure S2, we inferred the on-off receptive fields of the neurons and show that visual space is oversampled in terms of the receptive field positions in panel C. However, the exact slope/shape of the performance vs population size curve we show in Figure 5 will also depend on the maximum performance of our reconstruction method, which is limited in spatial resolution (Figure 4 & Supplementary Figure S5). It is possible that future reconstruction approaches will require fewer neurons than ours, so we interpret this curve rather as a description of the reconstruction method itself than a feature of the underlying neuronal code. For that reason, we chose caution and refrained from making any claims about neuronal coding principles based on this plot. 

(4) We appreciated the experiments testing the capacity of the reconstruction process, by using synthetic stimuli created under a Gaussian process in a noise-free way. But this originally further raised questions: what is the theoretical capability for reconstruction of this processing pipeline, as a whole? Is 0.563 the best that one could achieve given the noisiness and/or neuron count of the Sensorium project? What if the team applied the pipeline to reconstruct the activity of a given artificial neural network's layer (e.g., some ResNet convolutional layer), using hidden units as proxies for neuronal calcium activity? In the revision, this concern was addressed nicely in the review in Supplementary Figure 3C. Also, one appreciates that as a follow up, the team produced error maps (New Figure 6) that highlight where in the frames the reconstruction are likely to fail. But the maps went unanalyzed further, and I am not sure if there was a systematic trend in the errors. 

We are happy to hear that we were able to answer the reviewers’ question of what the maximum theoretical performance of our reconstruction process is in figure 3C. Regarding systematic trends in the error maps, we also did not observe any clear systematic trends. If anything, we noticed that some moving edges were shifted, but we do not think we can quantify this effect with this particular dataset. 

(5) I was encouraged by Figure 4, which shows how the reconstructions succeeded or failed across different spatial frequencies. The authors note that "the reconstruction process failed at high spatial frequencies," yet it also appears to struggle with low spatial frequencies, as the reconstructed images did not produce smooth surfaces (e.g., see the top rows of Figures 4A and 4B). In regions where one would expect a single continuous gradient, the reconstructions instead display specular, high-frequency noise. This issue is difficult to overlook and might deserve further discussion. 

Thank you for pointing this out, this is indeed true. The reconstructions do have high frequency noise. We mention this briefly in line 102 “Finally, we applied a 3D Gaussian filter with sigma 0.5 pixels to remove the remaining static noise (Figure S3) and applied the evaluation mask.” In revisiting this sentence, we think it is more appropriate to replace “remove” with “reduce”. This noise is more visible in the Gaussian noise stimuli (Figure 4) because we did not apply the 3D Gaussian filter to these reconstructions, in case it interfered with the estimates of the reconstruction resolution limits. 

Given that the Gaussian noise and drifting grating stimuli reconstructions were from predicted activity (“noise-free”), this high-frequency noise is not biological in origin and must therefore come from errors in our reconstruction process. This kind of high-frequency noise has previously been observed in feature visualization (optimizing input to maximize the activity of a specific node within a neural network to visualize what that node encodes; Olah, et al., "Feature Visualization", https://distill.pub/2017/feature-visualization/, 2017). It is caused by a kind of overfitting, whereby a solution to the optimization is found that is not “realistic”. Ways of combating this kind of noise include gradient smoothing, image smoothing, and image transformations during optimization, but these methods can restrict the resolution of the features that are recovered. Since we were more interested in determining the maximum resolution of stimuli that can be reconstructed in Figure 4 and Supplementary Figures 5-6, we chose not to apply these methods.

Reviewer #3 (Public review): 

Summary: 

This paper presents a method for reconstructing input videos shown to a mouse from the simultaneously recorded visual cortex activity (two-photon calcium imaging data). The publicly available experimental dataset is taken from a recent brain-encoding challenge, and the (publicly available) neural network model that serves to reconstruct the videos is the winning model from that challenge (by distinct authors). The present study applies gradient-based input optimization by backpropagating the brain-encoding error through this selected model (a method that has been proposed in the past, with other datasets). The main contribution of the paper is, therefore, the choice of applying this existing method to this specific dataset with this specific neural network model. The quantitative results appear to go beyond previous attempts at video input reconstruction (although measured with distinct datasets). The conclusions have potential practical interest for the field of brain decoding, and theoretical interest for possible future uses in functional brain exploration. 

Strengths: 

The authors use a validated optimization method on a recent large-scale dataset, with a state-of-the-art brain encoding model. The use of an ensemble of 7 distinct model instances (trained on distinct subsets of the dataset, with distinct random initializations) significantly improves the reconstructions. The exploration of the relation between reconstruction quality and number of recorded neurons will be useful to those planning future experiments. 

Weaknesses: 

The main contribution is methodological, and the methodology combines pre-existing components without any new original component. 

We thank the reviewer for their balanced assessment of our manuscript.

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

Public Reviews:

Reviewer #1 (Public review): 

Summary: 

This paper presents a method for reconstructing videos from mouse visual cortex neuronal activity using a state-of-the-art dynamic neural encoding model. The authors achieve high-quality reconstructions of 10-second movies at 30 Hz from two-photon calcium imaging data, reporting a 2-fold increase in pixel-by-pixel correlation compared to previous methods. They identify key factors for successful reconstruction including the number of recorded neurons and model ensembling techniques. 

Strengths: 

(1) A comprehensive technical approach combining state-of-the-art neural encoding models with gradient-based optimization for video reconstruction. 

(2) Thorough evaluation of reconstruction quality across different spatial and temporal frequencies using both natural videos and synthetic stimuli. 

(3) Detailed analysis of factors affecting reconstruction quality, including population size and model ensembling effects. 

(4) Clear methodology presentation with well-documented algorithms and reproducible code. 

(5) Potential applications for investigating visual processing phenomena like predictive coding and perceptual learning. 

We thank the reviewer for taking the time to provide this valuable feedback. We would like to add that in our eyes one additional main contribution is the step of going from reconstruction of static images to dynamic videos. We trust that in the revised manuscript, we have now made the point more explicit that static image reconstruction relies on temporally averaged responses, which negates the necessity of having to account for temporal dynamics altogether. 

Weaknesses: 

The main metric of success (pixel correlation) may not be the most meaningful measure of reconstruction quality: 

High correlation may not capture perceptually relevant features.

Different stimuli producing similar neural responses could have low pixel correlations The paper doesn't fully justify why high pixel correlation is a valuable goal 

This is a very relevant point. In retrospect, perhaps we did not justify this enough. Sensory reconstruction typically aims to reconstruct sensory input based on brain activity as faithfully as possible. A brain-to-image decoder might therefore be trained to produce images as close to the original input as possible. The loss function to train the decoder would therefore be image similarity on the pixel level. In that case, evaluating reconstruction performance based on pixel correlation is somewhat circular. 

However, when reconstructing videos, we optimize the input video in terms of its perceptual similarity to the original video and only then evaluate pixel-level similarity. The perceptual similarity metric we optimize for is the estimate of how the neurons in mouse V1 respond to that video. We then evaluate the similarity of this perceptually optimized video to the original input video with pixel-level correlation. In other words, we optimize for perceptual similarity and then evaluate pixel similarity. If our method optimized pixel-level similarity, then we would agree that perceptual similarity is a more relevant evaluation metric. We do not think it was clear in our original submission that our optimization loss function is a perceptual loss function, and have now made this clearer in Figure 1C-D and have clarified this in the results section, line 70:

“In effect, we optimized the input video to be perceptually similar with respect to the recorded neurons.”

And in line 110: 

“Because our optimization of the movies was based on a perceptual loss function, we were interested in how closely these movies matched the originals on the pixel level.”

We chose to use pixel correlation to measure pixel-level similarity for several reasons. 1) It has been used in the past to evaluate reconstruction performance (Yoshida et al., 2020), 2) It is contrast and luminance insensitive, 3) correlation is a common metric so most readers will have an intuitive understanding of how it relates to the data. 

To further highlight why pixel similarity might be interesting to visualize, we have included additional analysis in Figure 6 illustrating pixel-level differences between reconstructions from experimentally recorded activity and predicted activity. 

We expect that the type of perceptual similarity the reviewer is alluding to is pretrained neural network image embedding similarity (Zhang et al., 2018: https://doi.org/10.48550/arXiv.1801.03924). While these metrics seem to match human perceptual similarity, it is unclear if they reflect mouse vision. We did try to compare the embedding similarity from pretrained networks such as VGG16, but got results suggesting the reconstructed frames were no more similar to the ground truth than random frames, which is obviously not true. This might be because the ground truth videos were too different in resolution from the training data of these networks and because these metrics are typically very sensitive to decreases in resolution. 

The best alternative approach to evaluate mouse perceptual similarity would be to show the reconstructed videos to the same animals while recording the same neurons and to compare these neural activation patterns to those evoked by the original ground truth videos. This has been done for static images in the past: Cobos et al., bioRxiv 2022, found that static image reconstructions generated using gradient descent evoked more similar trial-averaged (40 trials) responses to those evoked by ground truth images compared to other reconstruction methods. Unfortunately, we are currently not able to perform these in vivo experiments, which is why we used publicly available data for the current paper. We plan to use this method in the future. But this method is also not flawless as it assumes that the average response to an image is the best reflection of how that image is represented, which may not be the case for an individual trial.

As far as we are aware, there is currently no method that, given a particular activity pattern in response to an image/video, can produce an image/video that induces a neural activity pattern that is closer to the original neural response than simply showing the same image/video again. Hypothetically, such a stimulus exists because of various visual processing phenomena we mention in our discussion (e.g., predictive coding and selective attention), which suggest that the image that is represented by a population of neurons likely differs from the original sensory input. In other words, what the brain represents is an interpretation of reality not a pure reflection. Experimentally verifying this is difficult, as these variations might be present on a single trial level. The first step towards establishing a method that captures the visual representation of a population of neurons is sensory reconstruction, where the aim is to get as close as possible to the original sensory input. We think pixel-level correlation is a stringent and interpretable metric for this purpose, particularly when optimizing for perceptual similarity rather than image similarity directly.

Comparison to previous work (Yoshida et al.) has methodological concerns: Direct comparison of correlation values across different datasets may be misleading; Large differences in the number of recorded neurons (10x more in the current study); Different stimulus types (dynamic vs static) make comparison difficult; No implementation of previous methods on the current dataset or vice versa. 

Yes, we absolutely agree that direct comparison to previous static image reconstruction methods is problematic. We primarily do so because we think it is standard practice to give related baselines. We agree that direct comparison of the performance of video reconstruction methods to image reconstruction methods is not really possible. It does not make sense to train and apply a dynamic model on a static image data set where neural activity is time-averaged, as the temporal kernels could not be learned. Conversely, for a static model, which expects a single image as input and predicts time averaged responses, it does not make sense to feed it a series of temporally correlated movie frames and to simply concatenate the resulting activity perdition. The static model would need to be substantially augmented to incorporate temporal dynamics, which in turn would make it a new method. This puts us in the awkward position of being expected to compare our video reconstruction performance to previous image reconstruction methods without a fair way of doing so. We have now added these caveats in line 119:

“However, we would like to stress that directly comparing static image reconstruction methods with movie reconstruction approaches is fundamentally problematic, as they rely on different data types both during training and evaluation (temporally averaged vs continuous neural activity, images flashed at fixed intervals vs continuous movies).”

We have also toned down the language, emphasising the comparison to previous image reconstruction performance in the abstract, results, and conclusion. 

Abstract: We removed “We achieve a ~2-fold increase in pixel-by-pixel correlation compared to previous state-of-the-art reconstructions of static images from mouse V1, while also capturing temporal dynamics.” and replaced with “We achieve a pixel-level correction of 0.57 between the ground truth movie and the reconstructions from single-trial neural responses.”

Discussion: we removed “In conclusion, we reconstruct videos presented to mice based on the activity of neurons in the mouse visual cortex, with a ~2-fold improvement in pixel-by-pixel correlation compared to previous static image reconstruction methods.” and replaced with “In conclusion, we reconstruct videos presented to mice based on single-trial activity of neurons in the mouse visual cortex.”

We have also removed the performance table and have instead added supplementary figure 3 with in-depth comparison across different versions of our reconstruction method (variations of masking, ensembling, contrast & luminance matching, and Gaussian blurring). 

Limited exploration of how the reconstruction method could provide insights into neural coding principles beyond demonstrating technical capability. 

The aim of this paper was not to reveal principles of neural coding. Instead, we aimed to achieve the best possible performance of video reconstructions and to quantify the limitations. But to highlight its potential we have added two examples of how sensory reconstruction has been applied in human vision research in line 321: 

“Although fMRI-based reconstruction techniques are starting to be used to investigate visual phenomena in humans (such as illusions [Cheng et al., 2023] and mental imagery [Shen et al., 2019; Koide-Majima et al., 2024; Kalantari et al., 2025]), visual processing phenomena are likely difficult to investigate using existing fMRI-based reconstruction approaches, due to the low spatial and temporal resolution of the data.”

We have also added a demonstration of how this method could be used to investigate which parts of a reconstruction from a single trial response differs from the model's prediction (Figure  6). We do this by calculating pixel-level differences between reconstructions from the recorded neural activity and reconstructions from the expected neural activity (predicted activity by the neural encoding model). Although difficult to interpret, this pixel-by-pixel error map could represent trial-by-trial deviations of the neural code from pure sensory representation. But at this point we cannot know whether these errors are nothing more than errors in the reconstruction process. To derive meaningful interpretations of these maps would require a substantial amount of additional work and in vivo experiments and so is outside the scope of this paper, but we include this additional analysis now to highlight a) why pixel-level similarity might be interesting to quantify and visualize and b) to demonstrate how video reconstruction could be used to provide insights into neural coding, namely as a tool to identify how sensory representations differ from a pure reflection of the visual input.  

The claim that "stimulus reconstruction promises a more generalizable approach" (line 180) is not well supported with concrete examples or evidence. 

What we mean by generalizable is the ability to apply reconstruction to novel stimuli, which is not possible for stimulus classification. We now explain this better in the paragraph in line 211: 

“Stimulus identification, i.e. identifying the most likely stimulus from a constrained set, has been a popular approach for quantifying whether a population of neurons encodes the identity of a particular stimulus [Földiák, 1993, Kay et al., 2008]. This approach has, for instance, been used to decode frame identity within a movie [Deitch et al., 2021, Xia et al., 2021, Schneider et al., 2023, Chen et al.,2024]. Some of these approaches have also been used to reorder the frames of the ground truth movie [Schneider et al., 2023] based on the decoded frame identity. Importantly, stimulus identification methods are distinct from stimulus reconstruction where the aim is to recreate what the sensory content of a neuronal code is in a way that generalizes to new sensory stimuli [Rakhimberdina et al., 2021]. This is inherently a more demanding task because the range of possible solutions is much larger. Although stimulus identification is a valuable tool for understanding the information content of a population code, stimulus reconstruction could provide a more generalizable approach, because it can be applied to novel stimuli.”

All the stimuli we reconstructed were not in the training set of the model, i.e., novel. We have also downed down the claim: we have replaced “promises” with “could provide”. 

The paper would benefit from addressing how the method handles cases where different stimuli produce similar neural responses, particularly for high-speed moving stimuli where phase differences might be lost in calcium imaging temporal resolution. 

Thank you for this suggestion, we think this is a great question. Calcium dynamics are slow and some of the high temporal frequency information could indeed be lost, particularly phase information. In other words, when the stimulus has high temporal frequency information, it is harder to decode spatial information because of the slow calcium dynamics. Ideally, we would look at this effect using the drifting grating stimuli; however, this is problematic because we rely on predicted activity from the SOTA DNEM, and due to the dilation of the first convolution, the periodic grating stimulus causes aliasing. At 15Hz, when the temporal frequency of the stimulus is half the movie frame rate, the model is actually being given two static images, and so the predicted activity is the interleaved activity evoked by two static images. We therefore do not think using the grating stimuli is a good idea. But we have used the Gaussian stimuli as it is not periodic, and is therefore less of a problem. 

We have now also reconstructed phase-inverted Gaussian noise stimuli and plotted the video correlation between the reconstructions from activity evoked by phase-inverted stimuli. On the one hand, we find that even for the fastest changing stimuli, the correlation between the reconstructions from phase inverted stimuli are negative, meaning phase information is not lost at high temporal frequencies. On the other hand, for the highest spatial frequency stimuli, the correlation is negative. So, the predicted neural activity (and therefore the reconstructions) are phase-insensitive when the spatial frequency is higher than the reconstruction resolution limit we identified (spatial length constant of 1 pixel, or 3.38 degrees). Beyond this limit, the DNEM predicts activity in response to phase-inverted stimuli, which, when used for reconstruction, results in movies which are more similar to each other than the stimulus that actually evokes them. 

However, not all information is lost at these high spatial frequencies. If we plot the Shannon entropy in the spatial domain or the motion energy in the temporal domain, we find that even when the reconstructions fail to capture the stimulus at a pixel-specific level (spatial length constant of 1 pixel, or 3.38 degrees), they do capture the general spatial and temporal qualities of the videos. 

We have added these additional analyses to Figure 4 and Supplementary Figure 5.

Reviewer #2 (Public review): 

This is an interesting study exploring methods for reconstructing visual stimuli from neural activity in the mouse visual cortex. Specifically, it uses a competition dataset (published in the Dynamic Sensorium benchmark study) and a recent winning model architecture (DNEM, dynamic neural encoding model) to recover visual information stored in ensembles of the mouse visual cortex. 

This is a great project - the physiological data were measured at a single-cell resolution, the movies were reasonably naturalistic and representative of the real world, the study did not ignore important correlates such as eye position and pupil diameter, and of course, the reconstruction quality exceeded anything achieved by previous studies. Overall, it is great that teams are working towards exploring image reconstruction. Arguably, reconstruction may serve as an endgame method for examining the information content within neuronal ensembles - an alternative to training interminable numbers of supervised classifiers, as has been done in other studies. Put differently, if a reconstruction recovers a lot of visual features (maybe most of them), then it tells us a lot about what the visual brain is trying to do: to keep as much information as possible about the natural world in which its internal motor circuits may act consequently. 

While we enjoyed reading the manuscript, we admit that the overall advance was in the range of those that one finds in a great machine learning conference proceedings paper. More specifically, we found no major technical flaws in the study, only a few potential major confounds (which should be addressable with new analyses), and the manuscript did not make claims that were not supported by its findings, yet the specific conceptual advance and significance seemed modest. Below, we will go through some of the claims, and ask about their potential significance. 

We thank the reviewer for the positive feedback on our paper.

(1) The study showed that it could achieve high-quality video reconstructions from mouse visual cortex activity using a neural encoding model (DNEM), recovering 10-second video sequences and approaching a two-fold improvement in pixel-by-pixel correlation over attempts. As a reader, I am left with the question: okay, does this mean that we should all switch to DNEM for our investigations of the mouse visual cortex? What makes this encoding model special? It is introduced as "a winning model of the Sensorium 2023 competition which achieved a score of 0.301... single-trial correlation between predicted and ground truth neuronal activity," but as someone who does not follow this competition (most eLife readers are not likely to do so, either), I do not know how to gauge my response. Is this impressive? What is the best achievable score, in theory, given data noise? Is the model inspired by the mouse brain in terms of mechanisms or architecture, or was it optimized to win the competition by overfitting it to the nuances of the data set? Of course, I know that as a reader, I am invited to read the references, but the study would stand better on its own if clarified how its findings depended on this model. 

This is a very good point. We do not think that everyone should switch to using this particular DNEM to investigate the mouse visual cortex, but we think DNEMs and stimulus reconstruction in general has a lot of potential. We think static neural encoding models have already been demonstrated to be an extremely valuable tool to investigate visual coding (Walker et al., 2019; Yoshida et al., 2021; Willeke et al., bioRxiv 2023). DNEMs are less common, largely because they are very large and are technically more demanding to train and use. That makes static encoding models more practical for some applications, but they do not have temporal kernels and are therefore only used for static stimuli. They cannot, for instance, encode direction tuning, only orientation tuning. But both static and dynamic encoding models have advantages over stimulus classification methods which we outline in our discussion. Here we provide the first demonstration that previous achievements in static image reconstruction are transferable to movies.

It has been shown in the past for static neural encoding models that choosing a better-performing model produces reconstructed static images that are closer to the original image (Pierzchlewicz et al., 2023). The factors in choosing this particular DNEM were its capacity to predict neural activity (benchmarked against other models), it was open source, and the data it was designed for was also available. 

To give more context to the model used in the paper, we have included the following, line 348:

“This model achieved an average single-trial correlation between predicted and ground truth neural activity of 0.291 during the competition, this was later improved to 0.301. The competition benchmark models achieved 0.106, 0.164 and 0.197 single-trial correlation, while the third and second place models achieved 0.243 and 0.265. Across the models, a variety of architectural components were used, including 2D and 3D convolutional layers, recurrent layers, and transformers, to name just a few.” 

Concerning biologically inspired model design. The winning model contained 3 fully connected layers comprising the “Cortex” just before the final readout of neural activity, but we would consider this level of biological inspiration as minor. We do not think that the exact architecture of the model is particularly important, as the crucial aspect of such neural encoders is their ability to predict neural activity irrespective of how they achieve it. There has been a move towards creating foundation models of the brain (Wang et al., 2025) and the priority so far has been on predictive performance over mechanistic interpretability or similarity to biological structures and processes. 

Finally, we would like to note that we do not know what the maximum theoretical score for single-trial responses might be, and don't think there is a good way of estimating it in this context. 

(2) Along those lines, two major conclusions were that "critical for high-quality reconstructions are the number of neurons in the dataset and the use of model ensembling." If true, then these principles should be applicable to networks with different architectures. How well can they do with other network types? 

This is a good question. Our method critically relies on the accurate prediction of neural activity in response to new videos. It is therefore expected that a model that better predicts neural responses to stimuli will also be better at reconstructing those stimuli given population activity. This was previously shown for static images (Pierzchlewicz et al., 2023). It is also expected that whenever the neural activity is accurately predicted, the corresponding reconstructed frames will also be more similar to the ground truth frames. We have now demonstrated this relationship between prediction accuracy and reconstruction accuracy in supplementary figure 4.

Although it would be interesting to compare the movie reconstruction performance of many different models with different architectures and activity prediction performances, this would involve quite substantial additional work because movie reconstruction is very resource- and time-intensive. Finding optimal hyperparameters to make such a comparison fair and informative would therefore be impractical and likely not yield surprising results. 

We also think it is unlikely that ensembling would not improve reconstruction performance in other models because ensembling across model predictions is a common way of improving single-model performance in machine learning. Likewise, we think it is unlikely that the relationship between neural population size and reconstruction performance would differ substantially when using different models, because using more neurons means that a larger population of noisy neurons is “voting” on what the stimulus is. However, we would expect that if the model were worse at predicting neural activity, then more neurons are needed for an equivalent reconstruction performance. In general, we would recommend choosing the best possible DNEM available, in terms of neural activity prediction performance, when reconstructing movies using input optimization through gradient descent. 

(3) One major claim was that the quality of the reconstructions depended on the number of neurons in the dataset. There were approximately 8000 neurons recorded per mouse. The correlation difference between the reconstruction achieved by 1 neuron and 8000 neurons was ~0.2. Is that a lot or a little? One might hypothesize that ~7,999 additional neurons could contribute more information, but perhaps, those neurons were redundant if their receptive fields were too close together or if they had the same orientation or spatiotemporal tuning. How correlated were these neurons in response to a given movie? Why did so many neurons offer such a limited increase in correlation? 

In the population ablation experiments, we compared the performance using ~1000, ~2000, ~4000, ~8000 neurons, and found an attenuation of 39.5% in video correlation when dropping 87.5% of the neurons (~1000 neurons remaining), we did not try reconstruction using just 1 neuron. 

(4) On a related note, the authors address the confound of RF location and extent. The study resorted to the use of a mask on the image during reconstruction, applied during training and evaluation (Line 87). The mask depends on pixels that contribute to the accurate prediction of neuronal activity. The problem for me is that it reads as if the RF/mask estimate was obtained during the very same process of reconstruction optimization, which could be considered a form of double-dipping (see the "Dead salmon" article, https://doi.org/10.1016/S1053-8119(09)71202-9). This could inflate the reconstruction estimate. My concern would be ameliorated if the mask was obtained using a held-out set of movies or image presentations; further, the mask should shift with eye position, if it indeed corresponded to the "collective receptive field of the neural population." Ideally, the team would also provide the characteristics of these putative RFs, such as their weight and spatial distribution, and whether they matched the biological receptive fields of the neurons (if measured independently). 

We can reassure the reviewer that there is no double-dipping. We would like to clarify that the mask was trained only on videos from the training set of the DNEM and not the videos which were reconstructed. We have added the sentence, line 91: 

“None of the reconstructed movies were used in the optimization of this transparency mask.”

Making the mask dependent on eye position would be difficult to implement with the current DNEM, where eye position is fed to the model as an additional channel. When using a model where the image is first transformed into retinotopic coordinates in an eye position-dependent manner (such as in Wang et al., 2025) the mask could be applied in retinotopic coordinates and therefore be dependent on eye position. 

Effectively, the alpha mask defines the relative level of influence each pixel contributes to neural activity prediction. We agree it is useful to compare the shape of the alpha mask with the location of traditional on-off receptive fields (RFs) to clarify what the alpha mask represents and characterise the neural population available for our reconstructions. We therefore presented the DNEM with on-off patches to map the receptive fields of single neurons in an in silico experiment (the experimentally derived RF are not available). As expected, there is a rough overlap between the alpha mask (Supplementary Figure 2D), the average population receptive field (Supplementary Figure 2B), and the location of receptive field peaks (Supplementary Figure 2C). In principle, all three could be used during training or evaluation for masking, but we think that defining a mask based on the general influence of images on neural activity, rather than just on off patch responses, is a more elegant solution.

One idea of how to go a step further would be to first set the alpha mask threshold during training based on the % loss of neural activity prediction performance that threshold induces (in our case alpha=0.5 corresponds to ~3% loss in correlation between predicted vs recorded neural responses, see Supplementary Figure 3D), and second base the evaluation mask on a pixel correlation threshold (see example pixel correlation map in Supplementary Figure 2E) instead to avoid evaluating areas of the image with low image reconstruction confidence. 

We referred to this figure in the result section, line 83:

“The transparency masks are aligned with but not identical to the On-Off receptive field distribution maps using sparse-noise (Figure S2).” 

We have also done additional analysis on the effect of masking during training and evaluation with different thresholds in Supplementary Figure 3.

(5) We appreciated the experiments testing the capacity of the reconstruction process, by using synthetic stimuli created under a Gaussian process in a noise-free way. But this further raised questions: what is the theoretical capability for the reconstruction of this processing pipeline, as a whole? Is 0.563 the best that one could achieve given the noisiness and/or neuron count of the Sensorium project? What if the team applied the pipeline to reconstruct the activity of a given artificial neural network's layer (e.g., some ResNet convolutional layer), using hidden units as proxies for neuronal calcium activity? 

That’s a very interesting point. It is very hard to know what the theoretical best reconstruction performance of the model would be. Reconstruction performance could be decreased due to neural variability, experimental noise, the temporal kernel of the calcium indicator and the imaging frame rate, information compression along the visual hierarchy, visual processing phenomena (such as predictive coding and selective attention), failure of the model to predict neural activity correctly, or failure of the reconstruction process to find the best possible image which explains the neural activity. We don't think we can disentangle the contribution of all these sources, but we can provide a theoretical maximum assuming that the model and the reconstruction process are optimal. To that end, we performed additional simulations and reconstructed the natural videos using the predicted activity of the neurons in response to the natural videos as the target (similar to the synthetic stimuli) and got a correlation of 0.766. So, the single trial performance of 0.569 is ~75% of this theoretical maximum. This difference can be interpreted as a combination of the losses due to neuronal variability, measurement noise, and actual deviations in the images represented by the brain compared to reality. 

We thank the reviewer for this suggestion, as it gave us the idea of looking at error maps (Figure 6), where the pixel-level deviation of the reconstructions from recorded vs predicted activity is overlaid on the ground truth movie.

(6) As the authors mentioned, this reconstruction method provided a more accurate way to investigate how neurons process visual information. However, this method consisted of two parts: one was the state-of-the-art (SOTA) dynamic neural encoding model (DNEM), which predicts neuronal activity from the input video, and the other part reconstructed the video to produce a response similar to the predicted neuronal activity. Therefore, the reconstructed video was related to neuronal activity through an intermediate model (i.e., SOTA DNEM). If one observes a failure in reconstructing certain visual features of the video (for example, high-spatial frequency details), the reader does not know whether this failure was due to a lack of information in the neural code itself or a failure of the neuronal model to capture this information from the neural code (assuming a perfect reconstruction process). Could the authors address this by outlining the limitations of the SOTA DNEM encoding model and disentangling failures in the reconstruction from failures in the encoding model? 

To test if a better neural prediction by the DNEM would result in better reconstructions, we ran additional simulations and now show that neural activity prediction performance correlates with reconstruction performance (Supplementary Figure 4B). This is consistent with Pierzchlewicz et al., (2023) who showed that static image reconstructions using better encoding models leads to better reconstruction performance. As also mentioned in the answer to the previous comment, untangling the relative contributions of reconstruction losses is hard, but we think that improvements to the DNEM performance are key. Two suggestions to improving the DNEM we used would be to translate the input image into retinotopic coordinates and shift this image relative to eye position before passing it to the first convolutional layer (as is done in Wang et al. 2025), to use movies which are not spatially down sampled as heavily, to not use a dilation of 2 in the temporal convolution of the first layer and to train on a larger dataset. 

(7) The authors mentioned that a key factor in achieving high-quality reconstructions was model assembling. However, this averaging acts as a form of smoothing, which reduces the reconstruction's acuity and may limit the high-frequency content of the videos (as mentioned in the manuscript). This averaging constrains the tool's capacity to assess how visual neurons process the low-frequency content of visual input. Perhaps the authors could elaborate on potential approaches to address this limitation, given the critical importance of high-frequency visual features for our visual perception. 

This is exactly what we also thought. To answer this point more specifically, we ran additional simulations where we also reconstruct the movies using gradient ensembling instead of reconstruction ensembling. Here, the gradients of the loss with respect to each pixel of the movie is calculated for each of the model instances and are averaged at every iteration of the reconstruction optimization. In essence, this means that one reconstruction solution is found, and the averaging across reconstructions, which could degrade high-frequency content, is skipped. The reconstructions from both methods look very similar, and the video correlation is, if anything, slightly worse (Supplemental Figure 3A&C). This indicates that our original ensembling approach did not limit reconstruction performance, but that both approaches can be used, depending on what is more convenient given hardware restrictions. 

Reviewer #3 (Public review): 

Summary: 

This paper presents a method for reconstructing input videos shown to a mouse from the simultaneously recorded visual cortex activity (two-photon calcium imaging data). The publicly available experimental dataset is taken from a recent brain-encoding challenge, and the (publicly available) neural network model that serves to reconstruct the videos is the winning model from that challenge (by distinct authors). The present study applies gradient-based input optimization by backpropagating the brain-encoding error through this selected model (a method that has been proposed in the past, with other datasets). The main contribution of the paper is, therefore, the choice of applying this existing method to this specific dataset with this specific neural network model. The quantitative results appear to go beyond previous attempts at video input reconstruction (although measured with distinct datasets). The conclusions have potential practical interest for the field of brain decoding, and theoretical interest for possible future uses in functional brain exploration. 

Strengths: 

The authors use a validated optimization method on a recent large-scale dataset, with a state-of-the-art brain encoding model. The use of an ensemble of 7 distinct model instances (trained on distinct subsets of the dataset, with distinct random initializations) significantly improves the reconstructions. The exploration of the relation between reconstruction quality and the number of recorded neurons will be useful to those planning future experiments. 

Weaknesses: 

The main contribution is methodological, and the methodology combines pre-existing components without any new original components. 

We thank the reviewer for taking the time to review our paper and for their overall positive assessment. We would like to emphasise that combining pre-existing machine learning techniques to achieve top results in a new modality does require iteration and innovation. While gradient-based input optimization by backpropagating the brain-encoding error through a neural encoding model has been used in 2D static image optimization to generate maximally exciting images and reconstruct static images, we are the first to have applied it to movies which required accounting for the time domain. Previous methods used time averaged responses and were limited to the reconstruction of static images presented with fixed image intervals.

The movie reconstructions include a learned "transparency mask" to concentrate on the most informative area of the frame; it is not clear how this choice impacts the comparison with prior experiments. Did they all employ this same strategy? If not, shouldn't the quantitative results also be reported without masking, for a fair comparison? 

Yes, absolutely. All reconstruction approaches limit the field of view in some way, whether this is due to the size of the screen, the size of the image on the screen, or cropping of the presented/reconstructed images during analysis due to the retinotopic coverage of the recorded neurons. Note that we reconstruct a larger field of view than Yoshida et al. In Yoshida et al., the reconstructed field of view was 43 by 43 retinal degrees. we show the size of an example evaluation mask in comparison. 

To address the reviewer’s concern more specifically, we performed additional simulations and now also show the performance using a variety of different training and evaluation masks, including different alpha thresholds for training and evaluation masks as well as the effective retinotopic coverage at different alpha thresholds. Despite these comparisons, we would also like to highlight that the comparison to the benchmark is problematic itself. This is because image and movie reconstruction are not directly comparable. It does not make sense to train and apply a dynamic model on a static image dataset where neural activity is time averaged. Conversely, it does not make sense to train or apply a static model that expects time-averaged neural responses on continuous neural activity unless it is substantially augmented to incorporate temporal dynamics, which in turn would make it a new method. This puts us in the awkward position of being expected to compare our video reconstruction performance to previous image reconstruction methods without a fair way of doing so. We have therefore de-emphasised the phrasing comparing our method to previous publications in the abstract, results, and discussion. 

Abstract: “We achieve a ~2-fold increase in pixel-by-pixel correlation compared to previous state-of-the-art reconstructions of static images from mouse V1, while also capturing temporal dynamics.” with “We achieve a pixel-level correction of 0.57 between the ground truth movie and the reconstructions from single-trial neural responses.”

Results: “This represents a ~2x higher pixel-level correlation over previous single-trial static image reconstructions from V1 in awake mice (image correlation 0.238 +/- 0.054 s.e.m for awake mice) [Yoshida et al., 2020] over a similar retinotopic area (~43° x 43°) while also capturing temporal dynamics. However, we would like to stress that directly comparing static image reconstruction methods with movie reconstruction approaches is fundamentally problematic, as they rely on different data types both during training and evaluation (temporally averaged vs continuous neural activity, images flashed at fixed intervals vs continuous movies).”

Discussion: “In conclusion, we reconstruct videos presented to mice based on the activity of neurons in the mouse visual cortex, with a ~2-fold improvement in pixel-by-pixel correlation compared to previous static image reconstruction methods.” with “In conclusion, we reconstruct videos presented to mice based on single-trial activity of neurons in the mouse visual cortex.”

We have also removed the performance table and have instead added supplementary figure 3 with in-depth comparison across different versions of our reconstruction method (variations of masking, ensembling, contrast & luminance matching, and Gaussian blurring). 

We believe that we have given enough information in our paper now so that readers can make an informed decision whether our movie reconstruction method is appropriate for the questions they are interested in.

Recommendations for the authors:

Reviewer #2 (Recommendations for the authors): 

(1) "Reconstructions have been luminance (mean pixel value across video) and contrast (standard deviation of pixel values across video) matched to ground truth." This was not clear: was it done by the investigating team? I imagine that one of the most easily captured visual features is luminance and contrast, why wouldn't the optimization titrate these well? 

The contrast and luminance matching of the reconstructions to the ground truth videos was done by us, but this was only done to help readers assess the quality of the reconstructions by eye. Our performance metrics (frame and video correlation) are contrast and luminance insensitive. To clarify this, we have also added examples of non-adjusted frames in Supplementary Figure 3A, and added a sentence in the results, line 103: 

“When presenting videos in this paper we normalize the mean and standard deviation of the reconstructions to the average and standard deviation of the corresponding ground truth movie before applying the evaluation masks, but this is not done for quantification except in Supplementary Figure 3D.”

We were also initially surprised that contrast and luminance are not captured well by our reconstruction method, but this makes sense as V1 is largely luminance invariant (O’Shea et al., 2025 https://doi.org/10.1016/j.celrep.2024.115217 ) and contrast only has a gain effect on V1 activity (Tring et al., 2024 https://journals.physiology.org/doi/full/10.1152/jn.00336.2024). Decoding absolute contrast is likely unreliable because it is probably not the only factor modulating the overall gain of the neural population.

To address the reviewer’s comment more fully, we ran additional experiments. More specifically, to test why contrast and luminance are not recovered in the reconstructions, we checked how the predicted activity between the reconstruction and the contrast/luminance corrected reconstructions differs. Contrast and luminance adjustment had little impact on predicted response similarity on average. This makes the reconstruction optimization loss function insensitive to overall contrast and luminance so it cannot be decoded. There is a small effect on activity correlation, however, so we cannot completely rule out that contrast and luminance could be reconstructed with a different loss function. 

(2) The authors attempted to investigate the variability in reconstruction quality across different movies and 10-second snippets of a movie by correlating various visual features, such as video motion energy, contrast, luminance, and behavioral factors like running speed, pupil diameter, and eye movement, with reconstruction success. However, it would also be beneficial if the authors correlated the response loss (Poisson loss between neural responses) with reconstruction quality (video correlation) for individual videos, as these metrics are expected to be correlated if the reconstruction captures neural variance. 

We thank the reviewer for this suggestion. We have now included this analysis and find that if the neural activity was better predicted by the DNEM then the reconstruction of the video was also more similar to the ground truth video. We further found that this effect is shift-dependent (in time), meaning the prediction of activity based on proximal video frames is more influential on reconstruction performance. 

Reviewer #3 (Recommendations for the authors): 

(1) I was confused about the choice of applying a transparency mask thresholded with alpha>0.5 during training and alpha>1 during evaluation. Why treat the two situations differently? Also, shouldn't we expect alpha to be in the [0,1] range, in which case, what is the meaning of alpha>1? (And finally, as already described in "Weaknesses", how does this choice impact the comparison with prior experiments? Did they also employ a similar masking strategy?) 

We found that applying a mask during training increased performance regardless of the size of the evaluation mask. Using a less stringent mask during training than during evaluation increases performance slightly, but also allows inspection of the reconstruction in areas where the model will be less confident without sacrificing performance, if this is desired. The thresholds of 0.5 and 1 were chosen through trial and error, but the exact values do not hold intrinsic meaning. The alpha mask values can go above 1 during their optimization. We could have clipped alpha during the training procedure (algorithm 1), but we decided this was not worth redoing at this stage, as the alphas used for testing were not above 1. All reconstruction approaches in previous publications limit the field of view in some form, whether this is due to the size of the screen, the size of the image on the screen, or the cropping of the presented/reconstructed images during analysis. 

To address the reviewer’s comment in detail, we have added extensive additional analysis to evaluate the coverage of the reconstruction achieved in this paper and how different masking strategies affect performance, as well as how the mask relates to more traditional receptive field mapping.  

(2) I would not use the word "imagery" in the first sentence of the abstract, because this might be interpreted by some readers as reconstruction of mental imagery, a very distinct question. 

We changed imagery to images in the abstract.

(3) Line 145-146: "<1 frame, or <30Hz" should be "<1 frame, or >30Hz". 

We have corrected the error.

(4) Algorithm 1, Line 5, a subscript variable 'g' should be changed to 'h'

We have corrected the error.

Additional Changes

(1) Minor grammatical errors

(2) Addition of citations: We were previously not aware of a bioRxiv preprint from 2022 (Cobos et al., 2022), which used gradient descent-based input optimization to reconstruct static images but without the addition of a diffusion model. Instead, we had cited for this method Pierzchlewicz et al., 2023 bioRxiv/NeurIPS. In Cobos et al., 2022, they compare static image reconstruction similarity to ground truth images and the similarity of the in vivo evoked activity across multiple reconstruction methods. Performance values are only given for reconstructions from trial-averaged responses across ~40 trials (in the absence of original data or code we are also not able to retrospectively calculate single-trial performance). The authors find that optimizing for evoked activity rather than image similarity produces image reconstructions that evoke more similar in vivo responses compared to reconstructions optimized for image similarity itself. We have now added and discussed the citation in the main text. 

(3) Workaround for error in the open-source code from https://github.com/lRomul/sensorium for video hashing function in the SOTA DNEM: By checking the most correlated first frame for each reconstructed movie, we discovered there was a bug in the open-source code and 9/50 movies we originally used for reconstruction were not properly excluded from the training data between DNEM instances. The reason for this error was that some of the movies are different by only a few pixels, and the video hashing function used to split training and test set folds in the original DNEM code classified these movies as different and split them across folds. We have replaced these 9 movies and provide a figure below showing the next closest first frame for every movie clip we reconstruct. This does not affect our claims. Excluding these 9 movie clips, did not affect the reconstruction performance (video correlation went from 0.563 to 0.568), so there was no overestimation of performance due to test set contamination. However, they should still be removed so some of the values in the paper have changed slightly. The only statistical test that was affected was the correlation between video correlation and mean motion energy (Supplementary Figure 4A), which went from p = 0.043 to 0.071. 

Author response image 2.

exclusion of movie clips with duplicates in the DNEM training data. A) example frame of a reconstructed movie (ground truth) and the most correlated first frame from the training data. b) all movie clips and their corresponding most correlated clip from the training data. Red boxes indicate excluded duplicates. 

Ja en nee, grootste reden dan .loadtxt sneller is is omdat het minder checks doet en ervan uitgaat dat de data al schoon is

Dit is te absoluut. “Open het bestand kort in een teksteditor en check of elke regel hetzelfde aantal waardes heeft en of er geen rare lege velden staan.” of iets dergelijks zou een betere student instructie zijn.

saturation perceptive

L’avatar comme support identitaire → difficulté de séparation.

Thèse centrale : ce n’est pas le numérique qui rend addict, mais la problématique interne

consequences usages excessifs

conséquences usage excessifs

Dispositifs incitatifs ≠ cause suffisante de l’addiction.

refus du changement interne

fuite face à la puberté

L’adolescence amplifie les usages : période de vulnérabilité narcissique.

Ambivalence du virtuel : peut nourrir la créativité ou fragiliser le rapport à la réalité.

Déconstruction du fantasme de nouveauté : le virtuel prolonge une histoire ancienne des images.

Idée clé : le numérique peut être créatif OU destructif → pas de déterminisme technologique.

On retrouve pour la première fois ici le terme « dits-documents ». Il serait préférable d'opter pour un autre terme (simplement « documents » ?).

Plutôt : « Galerie d'Orléans » ?

La répétition est-elle volontaire ? On s'attendrait à lire seulement « Car l'utilité sociale du voyage se mesure à... »

Il faudrait indiquer le numéro de page de cette citation, si l'information est disponible.

La citation est ici donnée sans numéro de page.

La conclusion souligne les implications organisationnelles de l’apprentissage informel numérique, en invitant les entreprises à s’adapter aux pratiques réelles des professionnels.

Cette typologie permet de structurer et de mieux comprendre la diversité des pratiques d’apprentissage informel numérique.

Ce passage inscrit l’apprentissage informel numérique dans une logique plus large d’apprentissage continu tout au long de la vie.

Ce résultat remet en question la place centrale accordée aux formations formelles, en soulignant le poids majoritaire de l’apprentissage informel.

La référence à Bandura ancre l’apprentissage informel dans des théories reconnues de l’apprentissage social.

Ce passage permet de relativiser la nouveauté du phénomène en montrant que le numérique amplifie des pratiques d’apprentissage déjà existantes.

La conclusion réaffirme la thèse défendue dans l’article : seule une formation explicite et structurée permet de développer durablement les compétences d’évaluation de l’information

Ce passage montre que les difficultés des élèves ne relèvent pas uniquement de leurs capacités, mais aussi du cadre pédagogique proposé.

Ce résultat de recherche souligne un problème majeur : les élèves évaluent le contenu sans questionner l’origine ni les intérêts des sources.

Cet exemple montre que les élèves s’appuient souvent sur des indices superficiels pour évaluer l’information, ce qui peut conduire à des erreurs d’interprétation.

Ce passage met en évidence la controverse centrale : face à la complexité de l’évaluation de l’information, l’école est-elle en mesure de former efficacement les élèves ?

L’auteur souligne ici que la désinformation ne concerne pas seulement certains publics, mais l’ensemble des citoyens exposés aux contenus numériques.

Muy buena la conclusión. Solamente dos comentarios: 1.- me hace un poco de ruido definir como neutralización de clase el hecho de que estas condiciones objetivas no logren predecir pension mjp. Siguiendo la idea, me referiría más bien a un patrón generalizado de creencias entre clases 2.- En el párrafo de trabajos futuros mencionaría integrar nuevas mediciones de meritocracia (Castillo et al., 2023) en los estudios, pues se podrían develar otras configuraciones de creencias entre mjp y merit que la escala actual no permite observar

@andreas.lafferte dejar solo bajo la diagonal

Incluiría la referencia de Castillo et al. 2025 (Frontiers publicado hace poco), porque es el mayor antecedente de que la meritocracia se asocia con justicia de mercado en Chile, con los mismos datos

Le numérique augmente le risque de figer la relation dans le virtuel

Rvo correspond à un refus de l'ajustement au réel

Rov est une composante normale de toute relation

avatare médiatise et déforme la relation à autrui

immersion renforce la plausibilité perceptive des représentations

numérique transforme pronfondement le rapport aux représentations

usages numériques compensent un retrait déjà présent

numérique n'est pas la cause du Hikikomori

retrait implique une neutralisation affective défensive

désarticulation touche la capacité d'auto-représentation

retrait évite la responsabilité de l'action et de la transformation

dynamique actualisation / visualisation bloquée

Retrait durable socialement et temporellement défini

retrait = défense sontre tension psychique

virtuel psychique appartient à la réalité du sujet

relation dominée par les attentes internes au détriment de l'altérité réelle

travail psychique implique acceptation du décalage entre attente et réalité

démarche scientifique = dynamique fonctionnement cognitif

adaptation psychique / circulation continue entre virtuel et perceptif

absence d'actualisation => présence relationnelle appauvrie ou inadéquate.

sans virtuel psychique , sujet submergé par immédiateté sensorielle.

notion actualisation désigne passage de la représentation à l'action concrète. Virtualisation permet la révision des représentations à partir de l'expériences vécue.

Lévy conceptualise la dynamique psychique = va et vient entre virtuel et action

relation humaine oscille entre ettentes internes et ajustement perceptif au réel

imagination articule représentation et action en vue d'une modification concrète du réel

mécanisme évitement psychique face à la souffrance

processus dynamique

virtuel psychique composé anticipations qui orientent l'action anvant la rencontre avec le réel

tripartition des représentations

virtuel s'oppose à l'actuel non au réel

virtuel psychique n'existe pas en l'absence de l'objet

conception deleuzienne du virtuel

Spécificité interactive du numérique => fort pouvoir d'engagement cognitif et émotionnel. Eclaire les mécanismes décrits par young.

Complexité des rapports virtuels numérique => idée usages numériques ne peuvent être évalués indépendamment du contexte psychique et social.

Concept de virtuel ne renvoie pas au support mais à une construction psychique. Clarification conceptuelle essentielle pour éviter une pathologisation abusive des usages numériques. Virtuel désigne un espace de représentation et de symbolisation, non un artefact technologique.

désarticulation sociale

perceptif

imaginaire

virtuel psychique

ethymologie du mot virtuel

Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.

Learn more at Review Commons

Reply to the reviewers

1. General Statements

We thank the reviewers for their overall support, thorough review, and thoughtful comments. The points raised were all warranted and we feel that addressing them has improved the quality of our manuscript. Below we respond to each of the points raised.

2. Point-by-point description of the revisions

Reviewer #1

Minor comments:

Are the lgl-1; pac-1 M-Z- double mutants dead? Only the phenotype of pac-1(M-Z-); lgl-1 (M+Z-) is shown. In figures and text throughout, it should be clear whether mutants are referring to zygotic loss or both maternal and zygotic loss, as this distinction could have major implications on the interpretation of experiments.

Almost all experiments we performed used a combination of RNAi of lgl-1 in a homozygous pac-1 null mutant background, or the other way around. RNAi should eliminate maternal product, but we hesitate to use the terminology M/Z since it has previously been used for protein degradation strategies.

We have updated the text and figure 1 to address the potential of maternal product masking earlier phenotypes, and performed additional RNAi experiments to demonstrate that the phenotypes obtained by RNAi for either pac-1 or lgl-1 in a homozygous mutant background for the other are the same as for the genetic double mutant. The results are shown as additional images and quantifications in figure 1B,C. We also updated the legend to figure 1 to make it clear that double genetic mutants are obtained from heterozygous lgl-1/+ parents.

Regarding the phenotype of lgl-1; pac-1 M-Z- double mutants: assuming the reviewer refers to M-Z- double genetic mutants, we cannot make such embryos as the pac-1(M-Z-); lgl-1(M+Z-) animals are already lethal.

In Figure 1C, it would be more appropriate to show a fully elongated WT embryo to contrast with arrested elongation in mutant embryos.

We agree with the reviewer and have replaced the 2-fold WT embryo with a 3-fold embryo.

Is the lateral spread of DLG-1 in double mutant embryos a result of failure to polarize DLG-1, or failure to maintain polarity? This should be straightforward to address in higher time resolution movies.

We have analyzed additional embryos at early stages of development. In lgl-1; pac-1 embryos we never see the appearance of complete junctions: defects are apparent already at dorsal intercalation. We interpret these results as a failure to properly polarize DLG-1. We have added additional images to Figure S2 and added this sentence to the text: Imaging of embryos from early stages of development on showed that normal continuous junctional DLG-1 bands are never established in pac-1(RNAi); lgl-1(mib201) embryos (Fig. S2B).

The lack of enhancement of hmp-1(fe4) by lgl-1(RNAi) is quite interesting, given that pac-1 does enhance hmp-1(fe4). To rule out the possibility that this result stems from incomplete lgl-1 RNAi, this experiment should be repeated using the lgl-1 null mutant.

We have done this experiment by recreating the fe4 S823F mutation in the lgl-1(null) mutant background as well as in the wild-type CGC1 background using CRISPR/Cas9. The phenotype of both was similar, but differs from that of the original PE97 strain. In the original strain, there is ~50% embryonic lethality but worms that complete embryogenesis grow up to be fertile adults. In our new "fe4" strains, nearly all animals are severely malformed with little to no elongation taking place. We are able to maintain both strains (with and without lgl-1) homozygous but with difficulty as only ~5% of animals grow up and give progeny. Apparently, there are genetic differences between PE97 and our CGC1 background that cause phenotypic differences despite having the same amino acid change in HMP-1.

Nevertheless, using our original embryonic viability criterium of 'hatching', loss of lgl-1 does not enhance the S823F mutation. We have included the following text in the manuscript:

To rule out that the lack of enhancement by lgl-1(RNAi) is due to incomplete inactivation of lgl-1, we also re-created the hmp-1(fe4) mutation (S823F) by CRISPR in lgl-1(mib201) mutant animals and wild-type controls. The phenotype of the S823F mutant we created is more severe than that of the original PE97 hmp-1(fe4) strain, with only ~5% of animals becoming fertile adults (Fig. S2F). This likely represents the presence of compensatory changes that have accumulated over time in PE97. Nevertheless, consistent with our RNAi results, the presence of lgl-1(mib201) did not further exacerbate the phenotype of HMP-1(S823F) (Fig. S2E, F). Taken together, the lack of enhancement of hmp-1(S823F) mutants by inactivation of loss of lgl-1 This observation argues against a primary role for lgl-1 in regulating cell junctions.

• Related to point 4, do pac-1 or lgl-1 null mutants enhance partial knockdown of junction protein DLG-1, or is this effect (of pac-1) specific to HMP-1/AJs?*

We have attempted to address this point using feeding RNAi against dlg-1. However, we were not able to obtain partial depletion of DLG-1. On RNAi feeding plates, control, pac-1, and lgl-1 animals did not show significant embryonic lethality. We checked RNAi effectiveness with a DLG-1::mCherry strain and found RNAi by feeding to be very ineffective. Since we could not deplete DLG-1 to a level that results in partial embryonic lethality, we were not able to address this question properly.

Does lgl-1 loss affect PAC-1 protein localization and vice versa?

It does not. We have added the following text and a figure panel: Loss-of-function mutants that strongly enhance a phenotype are often interpreted as acting in parallel pathways. We therefore examined whether loss of lgl-1 or pac-1 alters the localization of endogenously GFP-tagged LGL-1 or PAC-1. In neither null background did we detect changes in the subcellular localization of the other protein, consistent with LGL-1 and PAC-1 functioning in parallel pathways (Fig. S1D).

Reviewer #2

Very little of the imaging data are analyzed quantitatively, and in many cases it is not clear how many embryos were analyzed. While the images that are presented show clear defects, readers cannot determine how reproducible, strong or significant the phenotypes are.

We completely agree with the reviewer that interpretation of our data requires this information and apologize for the omission in the first manuscript version. The phenotypes are highly penetrant and consistent (timing of arrest, % lethality, junctional defects), and we have now added quantifications throughout the manuscript.

In particular, the data below should be quantified and, where possible, analyzed statistically:

• The frequency of the various junctional phenotypes shown in 2C

We have now quantified the junctional phenotypes. The junctional defects are highly penetrant: >90% of lgl-1; pac-1 embryos have junctional defects (new Fig. 2B). We used airy-scan confocal imaging to analyze the distribution of the different phenotypes (unaffected, spread laterally, and ring-like pattern). The results are shown in Fig. 2G.

• The expansion of DLG-1::mCherry in pac-1 lgl-1 embryos should be quantified (related to Figure 2B). For example, the percentage of membrane (marked by PH::GFP) occupied by DLG-1 could be quantified.

We have performed this quantification, shown in Fig. 2D.

- Similarly, the expansion of the aPKC domain should be quantified (Figure 3A).

An objective quantification of aPKC signal is difficult due to the relatively weak expression of aPKC::GFP and the lack of a clear demarcating boundary. This is part of the reason we measured tortuosity as a more quantifyable indicator of apical domain expansion. We have now added a qualitative observation table as Figure 3B. In addition, we have expanded the quantification of cell geometry by measuring lateral and basal surfaces. Lateral surfaces were decreased. We added the following text:

To better understand the reason for the change in geometry, we also measured the lengths of the lateral and basal surfaces (Fig. 3F). We found that the absolute lengths of the apical surfaces were not significantly different between pac-1(RNAi); lgl-1(mib201) and control animals. Instead, the lengths of the lateral domain were reduced (Fig. 3F). Hence, the more dome-shaped appearance of epidermal cells in pac-1; lgl-1 double mutant animals is due to a decrease in lateral domain size, which is consistent with the observed lateral spreading of aPKC.

• How many embryos were analyzed for each marker shown in Figure 2A, and what proportion showed the described phenotypes? This could be given in the text or in a panel.

We have added these numbers to panel 2B, and indicated the percentage in the text.

• The frequency of the various junctional phenotypes shown in 4F.

To address this, we have changed figure 4F to show three types of phenotype (strong, mild, no phenotype) and added how frequently we observed each to the panels. In rescue experiments, 18/24 embryos showed no junctional defects, while 6/24 showed a mild defect (compared to 100% severe in non-rescued embryos). To make room for this and other quantifications in Figure 4, we moved the demonstration that PAC-1 is depleted by RNAi to supplemental figure S4.

Because the genetic perturbations used are global (either deletions or RNAi), it is not established whether PAC-1/LGL-1 act in epidermal epithelial cells per se (versus an earlier requirement that manifests in epidermal epithelial cells). While I agree that this is the most likely scenario, other mechanisms are possible.

Our experiments indeed use global depletion/deletion of lgl-1 and pac-1. We cannot exclude therefore that other tissues do not contribute to the epithelial phenotypes. We assume that other tissues would be affected as well, and in fact have observed abnormal looking pharynx tissue (see our response to reviewer 3 below for examples). As the epidermis is one of the first tissue to develop it is likely the first in which phenotypes become apparent.

In particular, the overall GFP::aPKC levels appear notably higher in pac-1 lgl-1 embryos in Figure 3A. aPKC levels should be quantified to determine if this is true of pac-1 lgl-1 embryos. If so, couldn't that explain (or at least contribute to) the observed phenotypes?

Overall higher levels could indeed contribute to the phenotype. However, we have now quantified total aPKC levels in control and pac-1; lgl-1 embryos found no difference between them. We have added the following text to the manuscript: To determine if increased expression of aPKC might explain the broadened apical localization, we measured total intensity levels of aPKC::GFP. However, we detected no differences in fluorescence levels between control and pac-1(RNAi); lgl-1(mib201) animals (Fig. S3B, C).

Minor

Figure 4: For completeness, please include the embryonic viability of pac-1 lgl-1 +/- embryos treated with EV and cdc-42(RNAi), as was done for pac-1 lgl-1 pkc-3(ts) in Figure 4E. Presumably the increased proportion of viable embryos with the lgl-1 deletion allele is reflected in an overall increase in embryonic viability.

The embryonic viability indeed increases, but not as much as one might think because 15% of embryos die from the cdc-42 RNAi itself. The most important rescue argument is that we can obtain adult pac-1; lgl-1 animals with cdc-42 RNAi.

We have now included the overall rescue and the following text: Overall, cdc-42 RNAi caused a mild increase in embryonic viability (Fig. 4A). However, total embryonic viability may underestimate rescue of pac-1; lgl-1 embryonic lethality, because it also includes the ~15% lethality caused by cdc-42 inactivation itself, even among animals wild type for lgl-1.

The orientation of the inset images in Figures 2C, 3A and 3D is confusing. An illustration showing how these images are oriented relative to each other would be helpful.

We have added a figure showing how the junctions are oriented in the figures (Fig. 2E). We have also added supplemental videos S3 and S4 that should illustrate the phenotype more clearly as well.

For completeness, it would be good to test whether lgl-1(delta) is also synthetically lethal with picc-1(RNAi) (Zilberman 2017).

We like this idea and had already looked into this. Lgl-1 and picc-1 are not synthetic lethal (see graph in word file submitted). However, PICC-1 is not the only junctional localization signal for PAC-1, as demonstrated by the Nance lab. We find the data interesting but feel that it deserves a more thorough structure/function investigation of PAC-1 than we can provide here. Therefore we would prefer not to include this data.

Reviewer #3

We thank the reviewer for their support of our manuscript.

A few small areas to improve this manuscript:

p. 6 like 139: "remain" should be "remaining"

We have fixed this typo.

Could the authors mention what is the phenotype of the 10% of pac-1 animals that die?

Yes. They die with pleotropic phenotypes not resembling those of our pac-1; lgl-1 double mutant embryos. We have added examples of these to Figure S1.

Based on the Supplemental figures, it made me curious to ask: Did the authors notice changes in dorsal epidermal fusions? Cadherin normally disappears in the dorsal hyp7 cells at this time. Did the timing of the fusions change at all?

We haven't analyzed this in detail but our time-lapse videos show that dorsal fusions still take place and do not seem to be particularly delayed (overall development is slightly delayed but the delay in fusion is consistent with overall delay).

Again, curiosity driven by the Supplemental figures: did the authors notice defects in apical regions of internal organs, like the pharynx or intestine? The CDC-42 biosensor is asymmetrical in the developing intestine. See: DOI: 10.1242/bio.056911

We did not pay much attention to the intestine as PAC-1 is barely detectable in this tissue. The pharynx is formed, which we can easily detect in arrested embryos as we use GFP or BFP expressed under the myo-2 promoter to mark the deletion of pac-1. While we did not look closely, we do observe defects in pharynx development.

Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

Learn more at Review Commons

Referee #2

Evidence, reproducibility and clarity

Summary: This study focuses on the polarization of epidermal epithelial cells in C. elegans. Whereas the basolateral polarity protein is LGL-1 is required for epithelial polarity in flies, LGL-1 is dispensable for polarization and viability in C. elegans. Through a whole-genome RNAi screen, Jarosinska et al discover that the depletion of the RhoGAP PAC-1 is synthetically lethal with an lgl-1 deletion mutant. pac-1 lgl-1 double mutants have significant polarity defects in the epidermal epithelial, including mislocalization of junctional markers and expansion of the apical aPKC domain. As a result pac-1 lgl-1 double mutants fail to maintain surface epithelial and arrest development. Genetic interaction data suggest that increased CDC42 and aPKC activity in pac-1 lgl-1 contributes, as least in part, to the polarity defects and resulting embryonic lethality.

Major comments:

Very little of the imaging data are analyzed quantitatively, and in many cases it is not clear how many embryos were analyzed. While the images that are presented show clear defects, readers cannot determine how reproducible, strong or significant the phenotypes are. In particular, the data below should be quantified and, where possible, analyzed statistically:

• The frequency of the various junctional phenotypes shown in 2C
• The expansion of DLG-1::mCherry in pac-1 lgl-1 embryos should be quantified(related to Figure 2B). For example, the percentage of membrane (marked by PH::GFP) occupied by DLG-1 could be quantified.
• Similarly, the expansion of the aPKC domain should be quantified (Figure 3A).
• How many embryos were analyzed for each marker shown in Figure 2A, and what proportion showed the described phenotypes? This could be given in the text or in a panel.
• The frequency of the various junctional phenotypes shown in 4F.

Because the genetic perturbations used are global (either deletions or RNAi), it is not established whether PAC-1/LGL-1 act in epidermal epithelial cells per se (versus an earlier requirement that manifests in epidermal epithelial cells). While I agree that this is the most likely scenario, other mechanisms are possible. In particular, the overall GFP::aPKC levels appear notably higher in pac-1 lgl-1 embryos in Figure 3A. aPKC levels should be quantified to determine if this is true of pac-1 lgl-1 embryos. If so, couldn't that explain (or at least contribute to) the observed phenotypes?

Minor

Figure 4: For completeness, please include the embryonic viability of pac-1 lgl-1 +/- embryos treated with EV and cdc-42(RNAi), as was done for pac-1 lgl-1 pkc-3(ts) in Figure 4E. Presumably the increased proportion of viable embryos with the lgl-1 deletion allele is reflected in an overall increase in embryonic viability.

The orientation of the inset images in Figures 2C, 3A and 3D is confusing. An illustration showing how these images are oriented relative to each other would be helpful.

For completeness, it would be good to test whether lgl-1(delta) is also synthetically lethal with picc-1(RNAi) (Zilberman 2017).

Significance

LGL-1 is a conserved polarity protein that is essential for viability in Drosophila. In contrast, lgl-1 mutants are viable and have weak polarity phenotypes in C. elegans. A previous study showed that LGL-1 acts redundantly with the posterior polarity proteins PAR-2 during establishment of anterior/posterior polarity in the one-cell worm embryo. Here, Jarosinska et al show that LGL-1 acts redundantly with another protein, the RhoGAP protein PAC-1, in the polarization of the embryonic epidermal epithelial. The strength of this study is the identification of redundant roles for PAC-1 and LGL-1, the apparent strength of the polarity defects in the double mutant and the broader implication that LGL-1 may act in a range of redundant, cell/tissue specific pathways to regulate polarity. The primary weakness of this study is the lack of quantification. Additionally, the aPKC and CDC42 genetic interaction data hint at potential pathways, but fall short of establishing LGL-1's or PAC-1's mechanism of action.

Advance: This works identifies a redundant genetic interaction between LGL-1 and PAC-1. While the data require additional quantification, the phenotypes presented appear clear and strong. Although the molecular mechanism by which LGL-1 and PAC-1 act is not well established in the current work, the core observation is significant and should provide a foundation for future studies dissecting the molecular mechanisms.

Audience: This work will be of interest to a broad audience. LGL-1 is conserved and its role in cell polarization and epithelial polarity is very actively studied, including in mammalian systems.

Field of expertise. C elegans embryonic development; cell polarity.

Esta clasificación es útil como guía para posteriormente desarrollar un análisis de la corporalidad, considerando los tres aspectos.

Añadir que de esta clasificación, también consideramos lasa portaciones de Butler, y otras contemporáneas . Añadir un párrafo que explique la tesis busca comprender a cuerpo conectad en esas distintas dimensiones,

Añadir Con qué perspectivas de aquí atraviesan a la tesis. Merleau Ponty, pero también las reflexiones de artistas que trabajan sobre este mismo tema.

Author response:

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

Public Reviews:

Reviewer #1 (Public review): 

Summary: 

The authors analyzed the expression of ATAD2 protein in post-meiotic stages and characterized the localization of various testis-specific proteins in the testis of the Atad2 knockout (KO). By cytological analysis as well as the ATAC sequencing, the study showed that increased levels of HIRA histone chaperone, accumulation of histone H3.3 on post-meiotic nuclei, defective chromatin accessibility and also delayed deposition of protamines. Sperm from the Atad2 KO mice reduces the success of in vitro fertilization. The work was performed well, and most of the results are convincing. However, this manuscript does not suggest a molecular mechanism for how ATAD2 promotes the formation of testis-specific chromatin. 

We would like to take this opportunity to highlight that the present study builds on our previously published work, which examined the function of ATAD2 in both yeast S. pombe and mouse embryonic stem (ES) cells (Wang et al., 2021). In yeast, using genetic analysis we showed that inactivation of HIRA rescues defective cell growth caused by the absence of ATAD2. This rescue could also be achieved by reducing histone dosage, indicating that the toxicity depends on histone over-dosage, and that HIRA toxicity, in the absence of ATAD2, is linked to this imbalance.

Furthermore, HIRA ChIP-seq performed in mouse ES cells revealed increased nucleosome-bound HIRA, particularly around transcription start sites (TSS) of active genes, along with the appearance of HIRA-bound nucleosomes within normally nucleosome-free regions (NFRs). These findings pointed to ATAD2 as a major factor responsible for unloading HIRA from nucleosomes. This unloading function may also apply to other histone chaperones, such as FACT (see Wang et al., 2021, Fig. 4C).

In the present study, our investigations converge on the same ATAD2 function in the context of a physiologically integrated mammalian system—spermatogenesis. Indeed, in the absence of ATAD2, we observed H3.3 accumulation and enhanced H3.3-mediated gene expression. Consistent with this functional model of ATAD2— unloading chaperones from histone- and non-histone-bound chromatin—we also observed defects in histone-toprotamine replacement.

Together, the results presented here and in Wang et al. (2021) reveal an underappreciated regulatory layer of histone chaperone activity. Previously, histone chaperones were primarily understood as factors that load histones. Our findings demonstrate that we must also consider a previously unrecognized regulatory mechanism that controls assembled histone-bound chaperones. This key point was clearly captured and emphasized by Reviewer #2 (see below).

Strengths:

The paper describes the role of ATAD2 AAA+ ATPase in the proper localization of sperm-specific chromatin proteins such as protamine, suggesting the importance of the DNA replication-independent histone exchanges with the HIRA-histone H3.3 axis. 

Weaknesses: 

(1) Some results lack quantification. 

We will consider all the data and add appropriate quantifications where necessary.

(2) The work was performed well, and most of the results are convincing. However, this manuscript does not suggest a molecular mechanism for how ATAD2 promotes the formation of testis-specific chromatin. 

Please see our comments above.

Reviewer #2 (Public review): 

Summary:

This manuscript by Liakopoulou et al. presents a comprehensive investigation into the role of ATAD2 in regulating chromatin dynamics during spermatogenesis. The authors elegantly demonstrate that ATAD2, via its control of histone chaperone HIRA turnover, ensures proper H3.3 localization, chromatin accessibility, and histone-toprotamine transition in post-meiotic male germ cells. Using a new well-characterized Atad2 KO mouse model, they show that ATAD2 deficiency disrupts HIRA dynamics, leading to aberrant H3.3 deposition, impaired transcriptional regulation, delayed protamine assembly, and defective sperm genome compaction. The study bridges ATAD2's conserved functions in embryonic stem cells and cancer to spermatogenesis, revealing a novel layer of epigenetic regulation critical for male fertility. 

Strengths:

The MS first demonstration of ATAD2's essential role in spermatogenesis, linking its expression in haploid spermatids to histone chaperone regulation by connecting ATAD2-dependent chromatin dynamics to gene accessibility (ATAC-seq), H3.3-mediated transcription, and histone eviction. Interestingly and surprisingly, sperm chromatin defects in Atad2 KO mice impair only in vitro fertilization but not natural fertility, suggesting unknown compensatory mechanisms in vivo. 

Weaknesses:

The MS is robust and there are not big weaknesses 

Reviewer #3 (Public review): 

Summary: 

The authors generated knockout mice for Atad2, a conserved bromodomain-containing factor expressed during spermatogenesis. In Atad2 KO mice, HIRA, a chaperone for histone variant H3.3, was upregulated in round spermatids, accompanied by an apparent increase in H3.3 levels. Furthermore, the sequential incorporation and removal of TH2B and PRM1 during spermiogenesis were partially disrupted in the absence of ATAD2, possibly due to delayed histone removal. Despite these abnormalities, Atad2 KO male mice were able to produce offspring normally. 

Strengths:

The manuscript addresses the biological role of ATAD2 in spermatogenesis using a knockout mouse model, providing a valuable in vivo framework to study chromatin regulation during male germ cell development. The observed redistribution of H3.3 in round spermatids is clearly presented and suggests a previously unappreciated role of ATAD2 in histone variant dynamics. The authors also document defects in the sequential incorporation and removal of TH2B and PRM1 during spermiogenesis, providing phenotypic insight into chromatin transitions in late spermatogenic stages. Overall, the study presents a solid foundation for further mechanistic investigation into ATAD2 function. 

Weaknesses:

While the manuscript reports the gross phenotype of Atad2 KO mice, the findings remain largely superficial and do not convincingly demonstrate how ATAD2 deficiency affects chromatin dynamics. Moreover, the phenotype appears too mild to elucidate the functional significance of ATAD2 during spermatogenesis. 

We respectfully disagree with the statement that our findings are largely superficial. Based on our investigations of this factor over the years, it has become evident that ATAD2 functions as an auxiliary factor that facilitates mechanisms controlling chromatin dynamics (see, for example, Morozumi et al., 2015). These mechanisms can still occur in the absence of ATAD2, but with reduced efficiency, which explains the mild phenotype we observed.

This function, while not essential, is nonetheless an integral part of the cell’s molecular biology and should be studied and brought to the attention of the broader biological community, just as we study essential factors. Unfortunately, the field has tended to focus primarily on core functional actors, often overlooking auxiliary factors. As a result, our decade-long investigations into the subtle yet important roles of ATAD2 have repeatedly been met with skepticism regarding its functional significance, which has in turn influenced editorial decisions.

We chose eLife as the venue for this work specifically to avoid such editorial barriers and to emphasize that facilitators of essential functions do exist. They deserve to be investigated, and the underlying molecular regulatory mechanisms must be understood.

(1) Figures 4-5: The analyses of differential gene expression and chromatin organization should be more comprehensive. First, Venn diagrams comparing the sets of significantly differentially expressed genes between this study and previous work should be shown for each developmental stage. Second, given the established role of H3.3 in MSCI, the effect of Atad2 knockout on sex chromosome gene expression should be analyzed. Third, integrated analysis of RNA-seq and ATAC-seq data is needed to evaluate how ATAD2 loss affects gene expression. Finally, H3.3 ChIP-seq should be performed to directly assess changes in H3.3 distribution following Atad2 knockout.  

(1) In the revised version, we will include Venn diagrams to illustrate the overlap in significantly differentially expressed genes between this study and previous work. However, we believe that the GSEAs presented here provide stronger evidence, as they indicate the statistical significance of this overlap (p-values). In our case, we observed p-value < 0.01 (**) and p < 0.001 (***).

(2) Sex chromosome gene expression was analyzed and is presented in Fig. 5C.

(3) The effect of ATAD2 loss on gene expression is shown in Fig. 4A, B, and C as histograms, with statistical significance indicated in the middle panels.

(4) Although mapping H3.3 incorporation across the genome in wild-type and Atad2 KO cells would have been informative, the available anti-H3.3 antibody did not work for ChIP-seq, at least in our hands. The authors of Fontaine et al., 2022, who studied H3.3 during spermatogenesis in mice, must have encountered the same problem, since they tagged the endogenous H3.3 gene to perform their ChIP experiments.

(2) Figure 3: The altered distribution of H3.3 is compelling. This raises the possibility that histone marks associated with H3.3 may also be affected, although this has not been investigated. It would therefore be important to examine the distribution of histone modifications typically associated with H3.3. If any alterations are observed, ChIP-seq analyses should be performed to explore them further.

Based on our understanding of ATAD2’s function—specifically its role in releasing chromatin-bound HIRA—in the absence of ATAD2 the residence time of both HIRA and H3.3 on chromatin increases. This results in the detection of H3.3 not only on sex chromosomes but across the genome. Our data provide clear evidence of this phenomenon. The reviewer is correct in suggesting that the accumulated H3.3 would carry H3.3-associated histone PTMs; however, we are unsure what additional insights could be gained by further demonstrating this point.

(3) Figure 7: While the authors suggest that pre-PRM2 processing is impaired in Atad2 KO, no direct evidence is provided. It is essential to conduct acid-urea polyacrylamide gel electrophoresis (AU-PAGE) followed by western blotting, or a comparable experiment, to substantiate this claim. 

Figure 7 does not suggest that pre-PRM2 processing is affected in Atad2 KO; rather, this figure—particularly Fig. 7B—specifically demonstrates that pre-PRM2 processing is impaired, as shown using an antibody that recognizes the processed portion of pre-PRM2. ELISA was used to provide a more quantitative assessment; however, in the revised manuscript we will also include a western blot image.

(4) HIRA and ATAD2: Does the upregulation of HIRA fully account for the phenotypes observed in Atad2 KO? If so, would overexpression of HIRA alone be sufficient to phenocopy the Atad2 KO phenotype? Alternatively, would partial reduction of HIRA (e.g., through heterozygous deletion) in the Atad2 KO background be sufficient to rescue the phenotype? 

These are interesting experiments that require the creation of appropriate mouse models, which are not currently available.

(5) The mechanism by which ATAD2 regulates HIRA turnover on chromatin and the deposition of H3.3 remains unclear from the manuscript and warrants further investigation. 

The Reviewer is absolutely correct. In addition to the points addressed in response to Reviewer #1’s general comments (see above), it would indeed have been very interesting to test the segregase activity of ATAD2 (likely driven by its AAA ATPase activity) through in vitro experiments using the Xenopus egg extract system described by Tagami et al., 2004. This system can be applied both in the presence and absence (via immunodepletion) of ATAD2 and would also allow the use of ATAD2 mutants, particularly those with inactive AAA ATPase or bromodomains. However, such experiments go well beyond the scope of this study, which focuses on the role of ATAD2 in chromatin dynamics during spermatogenesis.

References:

(1) Wang T, Perazza D, Boussouar F, Cattaneo M, Bougdour A, Chuffart F, Barral S, Vargas A, Liakopoulou A, Puthier D, Bargier L, Morozumi Y, Jamshidikia M, Garcia-Saez I, Petosa C, Rousseaux S, Verdel A, Khochbin S. ATAD2 controls chromatin-bound HIRA turnover. Life Sci Alliance. 2021 Sep 27;4(12):e202101151. doi: 10.26508/lsa.202101151. PMID: 34580178; PMCID: PMC8500222.

(2) Morozumi Y, Boussouar F, Tan M, Chaikuad A, Jamshidikia M, Colak G, He H, Nie L, Petosa C, de Dieuleveult M, Curtet S, Vitte AL, Rabatel C, Debernardi A, Cosset FL, Verhoeyen E, Emadali A, Schweifer N, Gianni D, Gut M, Guardiola P, Rousseaux S, Gérard M, Knapp S, Zhao Y, Khochbin S. Atad2 is a generalist facilitator of chromatin dynamics in embryonic stem cells. J Mol Cell Biol. 2016 Aug;8(4):349-62. doi: 10.1093/jmcb/mjv060. Epub 2015 Oct 12. PMID: 26459632; PMCID: PMC4991664.

(3) Fontaine E, Papin C, Martinez G, Le Gras S, Nahed RA, Héry P, Buchou T, Ouararhni K, Favier B, Gautier T, Sabir JSM, Gerard M, Bednar J, Arnoult C, Dimitrov S, Hamiche A. Dual role of histone variant H3.3B in spermatogenesis: positive regulation of piRNA transcription and implication in X-chromosome inactivation. Nucleic Acids Res. 2022 Jul 22;50(13):7350-7366. doi: 10.1093/nar/gkac541. PMID: 35766398; PMCID: PMC9303386.

(4) Tagami H, Ray-Gallet D, Almouzni G, Nakatani Y. Histone H3.1 and H3.3 complexes mediate nucleosome assembly pathways dependent or independent of DNA synthesis. Cell. 2004 Jan 9;116(1):51-61. doi: 10.1016/s0092-8674(03)01064-x. PMID: 14718166.

Recommendations for the authors:

Reviewing Editor Comments:

I note that the reviewers had mixed opinions about the strength of the evidence in the manuscript. A revision that addresses these points would be welcome.

Reviewer #1 (Recommendations for the authors):  

Major points: 

(1) No line numbers: It is hard to point out the issues.

The revised version harbors line numbers.

(2) Given the results shown in Figure 3 and Figure 4, it is nice to show the chromosomal localization of histone H3.3 in spermatocytes or post-meiotic cells by Chromatin-immunoprecipitation sequencing (ChIP-seq).

Although mapping H3.3 incorporation across the genome in wild-type and Atad2 KO cells would have been informative, the available anti-H3.3 antibody did not work for ChIP-seq in our hands. In fact, this antibody is not well regarded for ChIP-seq. For example, Fontaine et al. (2022), who investigated H3.3 during spermatogenesis in mice, circumvented this issue by tagging the endogenous H3.3 genes for their ChIP experiments.

(3) Figure 7B and 8: Why the authors used ELISA for the protein quantification. At least, western blotting should be shown.

ELISA is a more quantitative method than traditional immunoblotting. Nevertheless, as requested by the reviewer, we have now included a corresponding western blot in Fig. S3.

(4) For readers, please add a schematic pathway of histone-protamine replacement in sperm formation in Fig.1 and it would be nice to have a model figure, which contains the authors' idea in the last figure.

As requested by this reviewer, we have now included a schematic model in Figure 9 to summarize the main conclusions of our work.

Minor points: 

(1) Page 2, the second paragraph, "pre-PRM2: Please explain more about pre-PRM2 and/or PRM2 as well as PRM1 (Figure 6).

More detailed descriptions of PRM2 processing are now given in this paragraph. 

(2) Page 3, bottom paragraph, line 1: "KO" should be "knockout (KO)".

Done.

(3) Page 4, second paragraph bottom: Please explain more about the protein structure of germ-line-specific ATAD2S: how it is different from ATAD2L. Germ-line specific means it is also expressed in ovary?

As Atad2 is predominantly expressed in embryonic stem cells and in spermatogenic cells, we replaced all through the text germ-line specific by more appropriate terms.

(4) Figure 1C, western blotting: Wild-type testis extracts, both ATAD2L and -S are present. Does this mean that ATADS2L is expressed in both germ line as well as supporting cells. Please clarify this and, if possible, show the western blotting of spermatids well as spermatocytes.

Figure 1D shows sections of seminiferous tubules from Atad2 KO mice, in which lacZ expression is driven by the endogenous Atad2 promoter. The results indicate that Atad2 is expressed mainly in post-meiotic cells. Most labeled cells are located near the lumen, whereas the supporting Sertoli cells remain unlabeled. Sertoli cells, which are anchored to the basal lamina, span the entire thickness of the germinal epithelium from the basal lamina to the lumen. Their nuclei, however, are usually positioned closer to the basal membrane. Thus, the observed lacZ expression pattern argues against substantial Atad2 expression in Sertoli cells. 

(5) Figure 1C: Please explain a bit more about the reduction of ATAD2 proteins in heterozygous mice.

Done

(6) Figure 1C: Genotypes of the mice should be shown in the legend.

Done 

(7) Figure 1D: Please add a more magnified image of the sections to see the staining pattern in the seminiferous tubules.

The magnification does not bring more information since we lose the structure of cells within tubules due the nature of treatment of the sections for X-gal staining. Please see comments to question 1C to reviewer 2

(8) Page 5, first paragraph, line 2, histone dosage: What do the authors meant by the histone dosage? Please explain more or use more appropriate word.

"Histone dosage" refers to the amount or relative abundance of histone proteins in a cell.

(9) Figure 2A: Figure 2A: Given the result in Figure 1C, it is interesting to check the amount of HIRA in Atad2 heterozygous mice.

In Atad2 heterozygous mice, we would expect an increase in HIRA, but only to about half the level seen in the Atad2 homozygous knockout shown in Figure 2A, which is relatively modest. Therefore, we doubt that detecting such a small change—approximately half of that in Figure 2A—would yield clear or definitive results. 

(10) Figure 2A, legend (n=5): What does this "n" mean? The extract of testes from "5" male mice like Figure 2B. Or 5 independent experiments. If the latter is true, it is important to share the other results in the Supplements.

“n” refers to five WT and five Atad2 KO males. The legend has been clarified as suggested by the reviewer.

(11) Figure 2A, legend, line 2, Atad2: This should be italicized.

Done

(12) Figure 2B: Please show the quantification of amounts of HIRA protein like Fig. 2A.

As indicated in the legend, what is shown is a pool of testes from 3 individuals per genotype.

(13) Figure 2B shows an increased level of HIRA in Atad2 KO testis. This suggests the role of ATAD2 in the protein degradation of HIRA. This possibility should be mentioned or tested since ATAD2 is an AAA+ ATPase. 

The extensive literature on ATAD2 provides no indication that it is involved in protein degradation. In our early work on ATAD2 in the 2000s, we hypothesized that, as a member of the AAA ATPase family, ATAD2 might associate with the 19S proteasome subunit (through multimerization with the other AAA ATPase member of this regulatory subunit). However, both our published pilot studies (Caron et al., PMID: 20581866) and subsequent unpublished work ruled out this possibility. Instead, since the amount of nucleosome-bound HIRA increases in the absence of ATAD2, we propose that chromatin-bound HIRA is more stable than soluble HIRA once it has been released from chromatin by ATAD2.

(14) Page 6, second paragraph, line 5, ko: KO should be capitalized.

Done

(15) Page 6, second paragraph, line 2 from the bottom, chromatin dynamics: Throughout the text, the authors used "chromatin dynamics". However, all the authors analyzed in the current study is the localization of chromatin protein.  So, it is much easier to explain the results by using "chromatin status," etc. In this context, "accessibility" is better. 

We changed the term “chromatin dynamics” into a more precise term according to the context used all through the text.

(16) Figure 3: Please provide the quantification of signals of histone H3.3 in a nucleus or nuclear cytoplasm.

This request is not clear to us since we do not observe any H3.3 signal in the cytoplasm.

(17) Figure 3: As the control of specificity in post-meiotic cells, please show the image and quantification of the H3.3 signals in spermatocyte, for example.

This request is not clear to us. What specificity is meant? 

(18) Figure 3, bottom panels: Please show what the white lines indicate? 

The white lines indicate the limit of cell nucleus and estimated by Hoechst staining. This is now indicated in the legend of the figure. 

(19) Figure 4A: Please explain more about what kind of data is here. Is this wild-type and/or Atad2 KO? The label of the Y-axis should be "mean expression level". What is the standard deviation (SD) here on the X-axis. Moreover, there is only one red open circle, but the number of this class is 5611. All 5611 genes in this group show NO expression. Please explain more.

The plot displays the mean expression levels (y-axis, labeled as "mean expression level") versus the corresponding standard deviations (x-axis), both calculated from three independent biological replicates of isolated round spermatids (Atad2 wild-type and Atad2 KO). The standard deviation reflects the variability of gene expression across biological replicates. Genes were grouped into four categories (grp1: blue, grp2: cyan, grp3: green, grp4: orange) according to the quartile of their mean expression. For grp4, all genes have no detectable expression, resulting in a mean expression of zero and a standard deviation of zero; consequently, the 5611 genes in this group are represented by a single overlapping point (red open circle) at the origin. 

(20) Figure 4C: If possible, it would be better to have a statistical comparison between wild-type and the KO.  

The mean profiles are displayed together with their variability (± 2 s.e.m.) across the four replicates for both ATAD2 WT (blue) and ATAD2 KO (red). For groups 1, 2, and 3, the envelopes of the curves remain clearly separated around the peak, indicating a consistent difference in signal between the two conditions. In contrast, group 4 does not present a strong signal and, accordingly, no marked difference is observed between WT and KO in this group.

(21) Figure 5, GSEA panels: Please explain more about what the GSEA is in the legend.  The legend has been updated as follows:

(A) Expression profiles of post-meiotic H3.3-activated genes. The heatmap (left panel) displays the normalized expression levels of genes identified by Fontaine and colleagues as upregulated in the absence of histone H3.3 (Fontaine et al. 2022) for Atad2 WT (WT) and Atad2 KO (KO) samples at days 20, 22, 24, and 26 PP (D20 to D26). The colour scale represents the z-score of log-transformed DESeq2-normalized counts. The middle panel box plots display, pooled, normalized expression levels, aggregated across replicates and genes, for each condition (WT and KO) and each time point (D20 to D26). Statistical significance between WT and KO conditions was determined using a two-sided t-test, with p-values indicated as follows: * for p-value<0.05, ** for p-value<0.01 and *** for p-value<0.001. The right panel shows the results of gene set enrichment analysis (GSEA), which assesses whether predefined groups of genes show statistically significant differences between conditions. Here, the post-meiotic H3.3-activated genes set, identified by Fontaine et al. (2022), is significantly enriched in Atad2 KO compared with WT samples at day 26 (p < 0.05, FDR < 0.25). Coloured vertical bars indicate the “leading edge” genes (i.e., those contributing most to the enrichment signal), located before the point of maximum enrichment score.  (B) As shown in (A) but for the "post-meiotic H3.3-repressed genes" gene set. (C) As shown in (A) but for the " sex chromosome-linked genes " gene set.

(22) Figure 6. In the KO, the number of green cells is more than red and yellow cells, suggesting the delayed maturation of green (TH2B-positive) cells. It is essential to count the number of each cell and show the quantification.

The green cells correspond to those expressing TH2B but lacking transition proteins (TP) and protamine 1 (Prm1), indicating that they are at earlier stages than elongating–condensing spermatids. Counting these green cells simply reflects the ratio of elongating/condensing spermatids to earlier-stage cells, which varies depending on the field examined. The key point in this experiment is that in wild-type mice, only red cells (elongating/condensing spermatids) and green cells (earlier stages) are observed. By contrast, in Atad2 KO testes, a significant proportion of yellow cells appears, which are never seen in wild-type tissue. The crucial metric here is the percentage of yellow cells relative to the total number of elongating/condensing spermatids (red cells). In wild-type testes, this value is consistently 0%, whereas in Atad2 KO testes it always ranges between 50% and 100% across all fields containing substantial numbers of elongating/condensing spermatids.

(23) Figure 8A: Please show the images of sperm (heads) in the KO mice with or without decompaction.

The requested image is now displayed in Figure S5.

(24) Figure 8C: In the legend, it says n=5. However, there are more than 5 plots on the graph. Please explain the experiment more in detail.

The experiment is now better explained in the legend of this Figure.

Reviewer #2 (Recommendations for the authors): 

While the study is rigorous and well performed, the following minor points could be addressed to strengthen the manuscript: 

Figure 1C should indicate each of the different types of cells present in the sections. It would be of interest to show specifically the different post-meiotic germ cells.

With this type of sample preparation, it is difficult to precisely distinguish the different cell types within the sections. Nevertheless, the staining pattern strongly indicates that most of the intensely stained cells are post-meiotic, situated near the tubule lumens and extending roughly halfway toward the basal membrane.

In the absence of functional ATAD2, the accumulation of HIRA primarily occurs in round spermatids (Fig. 2B). If technically possible, it would be of great interest to show this by IHC of testis section. 

Unfortunately, our antibody did not satisfactorily work in IHC.

The increased of H3.3 signal in Atad2 KO spermatids (Fig. 3) is interpreted because of a reduced turnover. However, alternative explanations (e.g., H3.3 misincorporation or altered chaperone affinity) should not be ruled out. 

The referee is correct that alternative explanations are possible. However, based on our previous work (Wang et al., 2021; PMID: 34580178), we demonstrated that in the absence of ATAD2, there is reduced turnover of HIRAbound nucleosomes, as well as reduced nucleosome turnover, evidenced by the appearance of nucleosomes in regions that are normally nucleosome-free at active gene TSSs. We have no evidence supporting any other alternative hypothesis.

In the MS the reduced accessibility at active genes (Fig. 4) is attributed to H3.3 overloading. However, global changes in histone acetylation (e.g., H4K5ac) or other remodelers in KO cells could be also consider.

In fact, we meant that histone overloading could be responsible for the altered accessibility. This has been clearly demonstrated in case of S. cerevisiae in the absence of Yta7 (S.  cerevisiae’ ATAD2) (PMID: 25406467).

In relation with the sperm compaction assay (Fig. 8A), the DTT/heparin/Triton protocol may not fully reflect physiological decompaction. This could be validated with alternative methods (e.g., MNase sensitivity). 

The referee is right, but since this is a subtle effect as it can be judged by normal fertility, we doubt that milder approaches could reveal significant differences between wildtype and Atad2 KO sperms.

It is surprising that despite the observed alterations in the genome organization of the sperm, the natural fertility of the KO mice is not affected (Fig. 8C). This warrants deeper discussion: Is functional compensation occurring (e.g., by p97/VCP)? Analysis of epididymal sperm maturation or uterine environment could provide insights.

As detailed in the Discussion section, this work, together with our previous study (Wang et al., 2021; PMID: 34580178), highlights an overlooked level of regulation in histone chaperone activity: the release of chromatinbound factors following their interaction with chromatin. This is an energy-dependent process, driven by ATP and the associated ATPase activity of these factors. Such activity could be mediated by various proteins, such as p97/VCP or DNAJC9–HSP70, as discussed in the manuscript, or by yet unidentified factors. However, most of these mechanisms are likely to occur during the extensive histone-to-histone variant exchanges of meiosis and post-meiotic stages. To the best of our knowledge, epididymal sperm maturation and the uterine environment do not involve substantial histone-to-histone or histone-to-protamine exchanges.

The authors showed that MSCI genes present an enhancement of repression in the absence of ATAD2 by enhancing H3.3 function. It would be also of interest to analyze the behavior of the Sex body during its silencing (zygotene to pachytene) by looking at different markers (i.e., gamma-H2AX phosphorylation, Ubiquitylation etc). 

The referee is correct that this is an interesting question. Accordingly, in our future work, we plan to examine the sex body in more detail during its silencing, using a variety of relevant markers, including those suggested by the reviewer. However, we believe that such investigations fall outside the scope of the present study, which focuses on the molecular relationship between ATAD2 and H3.3, rather than on the role of H3.3 in regulating sex body transcription. For a comprehensive analysis of this aspect, studies should primarily focus on the H3.3 mouse models reported by Fontaine and colleagues (PMID: 35766398).

Fig. 6: Co-staining of TH2B/TP1/PRM1 is convincing but would benefit from quantification (% cells with overlapping signals).

The green cells correspond to those expressing TH2B but lacking transition proteins (TP) and protamine 1 (Prm1), indicating that they are at earlier stages than elongating–condensing spermatids. Counting these green cells simply reflects the ratio of elongating/condensing spermatids to earlier-stage cells, which varies depending on the field examined. The key point is that in wild-type mice, only red cells (elongating/condensing spermatids) and green cells (earlier stages) are observed. By contrast, in Atad2 KO testes, a significant proportion of yellow cells appears, which are never seen in wild-type tissue. The crucial metric is the percentage of yellow cells relative to the total number of elongating/condensing spermatids (red cells). In wild-type testes, this value is consistently 0%, whereas in Atad2 KO testes it always ranges between 50% and 100% across all fields containing substantial numbers of elongating/condensing spermatids.

légitime scientifiquement la notion d'addiction non chimique

Notion de retentissement durable constitue un critère clinique : altération des fonctions exécutives (inhibition, planification) et priorisation excessive des stimulations numériques des stimulations numériques dans le système attentionnel. Permet de distinguer un usage intensif d'un usage pathologique

Definition de l'addiction : rapproche les usages numériques des addictions comportementales. Rejoint la conceptualisation de Young : addiction à Internet repose moins sur l’objet que sur l’altération des processus de régulation cognitive et motivationnelle. Prudence terminologique adoptée par Mouren anticipe la sur-pathologisation des usages numériques chez les jeunes.

retentissement fonctionnel

DOI: 10.1111/imcb.70061

Resource: Nonhuman Primate Reagent Resource (RRID:SCR_012986)

Curator: @giovanni.decastro

SciCrunch record: RRID:SCR_012986

What is this?

DOI: 10.1111/jnc.70320

Resource: (Vector Laboratories Cat# BA-1000, RRID:AB_2313606)

Curator: @scibot

SciCrunch record: RRID:AB_2313606

What is this?

DOI: 10.1016/j.cell.2025.11.022

Resource: (BioLegend Cat# 135025, RRID:AB_2562114)

Curator: @scibot

SciCrunch record: RRID:AB_2562114

What is this?

Cette conclusion ouvre la controverse vers des enjeux plus larges, en soulignant l’impact des fake news sur l’organisation de l’espace public et médiatique.

L’auteure explique ici que la faible responsabilité perçue dans certains espaces favorise la diffusion d’informations fausses sans réelle intention de tromper

Cette phrase met en évidence que la diffusion des informations varie selon les contextes sociaux, ce qui nuance l’idée d’une réception homogène des fake news.

Ce passage présente la démarche méthodologique adoptée afin d’analyser les contextes de circulation des informations, ce qui renforce la crédibilité scientifique de l’argumentation.

le partage d’une fake news ne signifie pas nécessairement l’adhésion à son contenu.

Ce passage remet en cause une approche purement quantitative de la désinformation et invite à analyser la réception et l’interprétation des fake news par les individus.

L’auteure montre ici que, dans les contextes de communication informelle, la vérification de l’information passe au second plan au profit de logiques sociales ou relationnelles.

Están al revés, pero PISA logra la invarianza. Lo dejamos así?

2ª Aula - A mais nova hábilidade "aprender a aprender" [criatividade tambem]

• Ele fala sobre sua visão do futuro do brasil e do ensino.

Caderno da Mastria

https://youtu.be/oyoVN3z7ifQ?list=PLpNPFZNi1cmU_3aJZ4PiEw_9yClxVHjPm&t=999

nessa parte do vídeo ele fala sobre o CADERNO DA MESTRIA

1ª Aula - Nessa aula ele apresenta o curso e fala sobre ele e seus projetos.

Các thuốc chống đông đường tiêm có thể được sử dụng để ức chế quá trình đông máu huyết tương cả dự phòng lẫn điều trị, ví dụ nhằm phòng ngừa và điều trị huyết khối.

Heparin là nhóm thuốc được sử dụng phổ biến nhất; các thuốc khác thường chỉ được dùng khi có không dung nạp. Bên cạnh tăng nguy cơ biến chứng chảy máu, tác dụng phụ đáng sợ nhất là giảm tiểu cầu do heparin (HIT) — một biến chứng hiếm gặp, trong đó tự kháng thể gây kết tập tiểu cầu, dẫn đến thuyên tắc huyết khối nguy hiểm. Do biến chứng này có thể được nhận biết qua sự giảm mạnh số lượng tiểu cầu, nên với việc sử dụng heparin rất thường xuyên trong thực hành lâm sàng, cần theo dõi công thức máu.

A menos que cada uno tuviera su propio servidor y los servidores centralizados fueran usados sólo para coordinar comunicaciones, como ocurre con Fossil y ocurrirá con Cardumem. Así, la centralización ofrece conveniencia pero no usa asimetría fundamental de capacidad o poder, como ocurre actualmente y servicios/protocolos de descubrimiento de servidores podrían ser implementados sobre la infraestructura cliente servidor actual, en caso de que algún servidor sea dado de baja.

Acá hay transiciones en los siguientes cuadrantes:

Cardumem toma una ruta alterna y más sencilla para explorar transiciones similares.

1. Inicia por el wiki, como software documental asíncrono.
2. Se conectará con HedgeDoc como software documetal síncrono.
3. Se conectará con Hypothesis como software dialógico asíncrono.
4. Implementará progresivamente funcionalidades síncronas vía sistemas hipermedia en tiempo real.

La idea de local primero ocurrirá debido a que el servidor puede correr de manera local o remota.

Es interesante esta generalidad desde los bloques (objetos, tipos y relaciones, que se juntan en un grafo). Los Dumems en Cardumem son otra forma de generalización desde el hipertexto programable (gracias al scripting en YueScript) y los metadatos personalizables que permiten las tablas de Lua.

Sin embargo, para disminuir la complejidad y aumentar la practicidad, en Cardumem no apuntamos a tecnologías de la llamada web 3.0, sino que usamos las buenas y confliables web 2.0 con algo de retrofuturismo en los sistemas hipermedia.

lo definiste hasta aca bajo y esta sigla la tienes mas arriba, checar tus siglas y que estén definidas

estas partículas ya pueden ingresar a los alveolos

mencionas varias veces estudio podrias cambiar por investigación o se ha documentado etc

aqui podría ir esa imagen que menciono

te recomiendo que quites la palabra de título y solo dejes tu tema. OJO los títulos no llevan punto al final

alguien que me ayude con eso

Reviewer #2 (Public review):

Summary:

Siddiqui et al. show that C. elegans prefers certain bacterial strains that have been supplemented with the essential amino acid (EEA) leucine. They convincingly show that some leucine enriched bacteria stimulate the production of isoamyl alcohol (IAA). IAA is an attractive odorant that is sensed by the AWC. The authors an identify a receptor, SRD-12, that is expressed in the AWC chemosensory neurons and is required for chemotaxis to IAA. The authors propose that IAA is a predominant olfactory cue that determines diet preference in C. elegans. Since leucine is an EAA, the authors propose that worm IAA sensing allows the animal provides a proxy mechanism to identify EAA rich diets.

Strengths:

The authors propose IAA as a predominant olfactory cue that determines diet preference in C. elegans providing molecular mechanism underlying diet selection. They show that wild isolates of C. elegans have strong chemotactic response to IAA indicating that IAA is an ecologically relevant odor for the worm. The paper is well written, and the presented data are convincing and well organized. This is an interesting paper that connects chemotactic response with bacterially produced odors and thus provides an understanding how animals adapt their foraging behavior through the perception of molecules that may indicate the nutritional value.

Weaknesses:

Major: While I do like the way the authors frame C. elegans IAA sensing as mechanisms to identify leucine (EAA) rich diets, it is not fully clear whether bacterial IAA production is a proxy for bacterial leucine levels.

(1) Can the authors measure leucine (or other EAA) content of the different CeMbio strains? This would substantiate the premise in the way they frame this in the introduction. While the authors convincingly show that leucine supplementation induces IAA production in some strains, it is not clear if there are lower leucine levels in the different in the non-preferred strains.

(2) It is not clear whether the non-preferred bacteria in Figure 1A and 1B have the ability to produce IAA. To substantiate the claim that C. elegans prefers CEent1, JUb66, and BIGb0170 due to their ability to generate IAA from leucine, it would be measure IAA levels in non-preferred bacteria (+ and - leucine supplementation). If the authors have these data it would be good to include this.

(3) The authors would strengthen their claim if they could show that deletion or silencing ilvE enzyme reduces IAA levels and eliminates the increased preference upon leucine supplementation.

(4) While the three preferred bacteria possess the ilvE gene, it is not clear whether this enzyme is present in the other non-preferred bacterial strains. As far as I know, the CeMbio strains have been sequenced, so it should be easy to determine if the non-preferred bacteria possess the capacity to make IAA. Does expression of ilvE in e.g. E. coli increase its preference index or are the other genes in the biosynthesis pathway missing?

(5) It is strongly implied that leucine rich diets are beneficial to the worm. Do the authors have data to show the effect on leucine supplementation on C. elegans healthspan, life-span or broodsize?

Comments on revisions:

(1) The authors have addressed most of the earlier questions. The main unresolved issue is the link between iaa production is a reflection of bacterial leucine levels. It is not clear if there are lower leucine levels in the different in non-preferred strains.

The main conclusions that: 1. some bacterial strains can convert exogenous leucine into IAA which is an attractant to C. elegans. 2. The identification of a GPCR required for IAA responses are solid. These are important results that carry the paper. My outstanding concern remains with the overinterpretation of the framing that C. elegans IAA sensing is used as a mechanism to identify leucine (EAA) rich diets. It is fine to leave this a favorite hypothesis in the discussion but statements throughout the paper need to be nuanced without leucine measurement of the different bacterial strains. (Also since for the bacterial chemotaxis assays there were only done with a single concentration of leucine makes it difficult to infer bacterial leucine concentrations). I recommend softening claims related to leucine-rich diet detection unless quantitative measurements are provided.

Part of the issue in the text lies in the difference between "supplemented" and "chemotaxis" (lab based constructs) and enriched and foraging (natural environment based). This is also the way it is set up in the introduction "Do animals use specific sensing mechanisms to find an EAA-enriched diet?". If enriched is used strictly the same as supplemented then it would be fine but in the text this distinction gets blurred and enriched drifts to the more ethological explanation.

Then it is more than just semantics since leucine-supplemented diets are not something that occurs in the natural environment. IAA production by bacteria could be a signal for a leucine rich environment and it is fine to speculate about this in the discussion.

Examples where the wording needs to be more precise to reflect the experimental results rather than the possible impact in its natural environment:

The title:' The olfactory receptor SNIF-1 mediates foraging for leucine-rich diets in C. elegans"

The intro:"Taken together, SNIF-1 regulates the dietary preference of worms to IAA-producing bacteria and thereby mediates the foraging behavior of C. elegans to leucine-enriched diets. Thus, IAA produced by bacteria is a dietary quality code for leucine-enriched bacteria."

Results "Figure 1. C. elegans relies on odors to select leucine-enriched bacteria"

Supplementation is used more in the text and the figure legends whereas headings and abstract use enriched. The experiments in the paper only describe leucine-supplemented experiments. So use I would supplemented instead of enriched when describing experiments for clarity.

For instance:

Page 4:"Microbial odors drive the preference of C. elegans for leucine-enriched diet"

Page 5: "Altogether, these findings suggested that worms rely on odors to distinguish various bacteria and find leucine-enriched bacteria"

Page 7: "Isoamyl alcohol odor is a signature for a leucine-enriched diet"

Page 9: AWC odor sensory neurons facilitate the diet preference of C. elegans for leucine-enriched diets"

page 20 "Leucine-enriched diets produce significantly higher levels of IAA odor, making up to 90% of their headspace"

(2) As suggested in the first round of review the authors now add data IAA levels in non-preferred bacteria (+ and - leucine supplementation) in table S2. While it is good to have this data, the table is not very clear. Not clear what ND stands for in the table S2. Not determined or not detected? I assume not determined since some strains Jub44, BiGb0393 Jub134 produce IAA even in the absence of LEU. The authors mention that "the abundance of IAA in these strains is significantly less". However, the table just reflects yes or no. Can the authors give an indication of the concentration to understand what significantly less means? Fig. 2c at least gives a heat map.

(3) On wormbase the gene is still called srd-12. The authors should seek permission to rename srd-12 to snif-1.

Author response:

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

eLife Assessment:

This is an important study, supported by solid to convincing data, that suggests a model for diet selection in C. elegans. The significance is that while C. elegans has long been known to be attracted to bacterial volatiles, what specific bacterial volatiles may signify to C. elegans is largely unknown. This study also provides evidence for a possible odorant/GPCR pairing.

Public Reviews:

Reviewer #1 (Public review):

Summary:

Siddiqui et al., investigate the question of how bacterial metabolism contributes to the attraction of C. elegans to specific bacteria. They show that C. elegans prefers three bacterial species when cultured in a leucine-enriched environment. These bacterial species release more isoamyl alcohol, a known C. elegans attractant, when cultured with leucine supplement than without leucine supplement. The study shows correlative evidence that isoamyl alcohol is produced from leucine by the Ehrlich pathway. In addition, they show that SRD-12 (SNIF-1) is likely a receptor for isoamyl alcohol because a null mutant of this receptor exhibits lower chemotaxis to isoamyl alcohol and lower preference for leucine-enriched bacteria.

Strengths:

(1) This study takes a creative approach to examine the question of what specific volatile chemicals released by bacteria may signify to C. elegans by examining both bacterial metabolism and C. elegans preference behavior. Although C. elegans has long been known to be attracted to bacterial metabolites, this study may be one of the first to examine the role of a specific bacterial metabolic pathway in mediating attraction.

(2)  A strength of the paper is the identification of SRD-12 (SNIF-1) as a likely receptor for isoamyl alcohol. The ligands for very few olfactory receptors have been identified in C. elegans and so this is a significant addition to the field. The srd-12 (snif-1) null mutant strain will likely be a useful reagent for many labs examining olfactory and foraging behaviors.

Weaknesses:

(1) The authors write that the leucine metabolism via the Ehrlich pathway is required for the production of isoamyl alcohol by three bacteria (CEent1, JUb66, BIGb0170), but their evidence for this is correlation and not causation. They write that the gene ilvE is a bacterial homolog of the first gene in the yeast Ehrlich pathway (it would be good to include a citation for this) and that the gene is present in these three bacterial strains. In addition, they show that this gene, ilvE, is upregulated in CEent1 bacteria upon exposure to leucine. To show causation, they need to knockout ilvE from one of these strains, show that the bacteria does not have increased isoamyl alcohol production when cultured on leucine, and that the bacteria is no longer attractive to C. elegans.

Thank you for the comment. We have added the appropriate citation [1,2]. We agree that worms’ diet preference for the preferred strains upon ilvE knockout will further strengthen the claim for IAA being used as a proxy for leucine-enriched diet. Currently, protocols and tools for genetic manipulations for CeMbio strains are not available, making this experiment not feasible at this time.  

(2) The authors examine three bacterial strains that C. elegans showed increased preference when grown with leucine supplementation vs. without leucine supplementation. However, there also appears to be a strong preference for another strain, JUb0393, when grown on plus leucine (Figure 1B). It would be good to include statistics and criteria for selecting the three strains.

Thanks for your comment. We agree that for Pantoea nemavictus, JUb393, worms seem to prefer the leucine supplemented (+ LEU) bacteria over unsupplemented (-LEU). However, when given a choice between the individual CeMbio bacteria and E. coli OP50, worms showed preference for only CEent1, JUb66, and BIGb0170 (Figure 1F). Consequently, CEent1, JUb66, and BIGb0170 were selected for further analyses. We have included statistics for Figure 1B-C and Figure S1A-G with details mentioned in the figure legend. 

(3) Although the behavioral evidence that srd-12 (snif-1) gene encodes a receptor for isoamyl alcohol is compelling, it does not meet the standard for showing that it is an olfactory receptor in C. elegans. To show it is indeed a likely receptor one or more of the following should be done:

(a) Calcium imaging of AWC neurons in response to isoamyl alcohol in the receptor mutant with the expectation that the response would be reduced or abolished in the mutant compared to wildtype.

(b)"A receptor swap" experiment where the SRD-12 (SNIF-1) receptor is expressed in AWB repulsive neuron in SRD-12 (SNIF-1) receptor mutant background with the expectation that with receptor swap C. elegans will now be repulsed from isoamyl alcohol in chemotaxis assays (experiment from Sengupta et al., 1996 odr-10 paper).

Thanks for all your comments and suggestions. While the lab currently does not have the necessary expertise to conduct calcium imaging of neurons, we have performed additional experiments to confirm the requirements of AWC neurons for SNIF-1 function. We generated transgenic worms with extrachromosomal array expressing snif-1 under (a) AWC-specific promoter, odr-1, and (b) AWB-specific promoter, str-1. As shown in new panel 6H in the revised manuscript and Author response image 1, we found that overexpression of snif-1 in AWC neurons completely rescues the chemotaxis defect of snif-1 mutant (referred at VSL2401), whereas upon the “receptor swap" in AWB neurons IAA is sensed as a repellent.  

Author response image 1.

(A) Chemotaxis index (CI) of WT, VSL2401, VSL2401 [AWCp::snif-1] and VSL2401 [AWBp::snif-1] worms to IAA at 1:1000 dilution. Significant differences are indicated as **** P ≤ 0.0001 determined by one-way ANOVA followed by post hoc Dunnett’s multiple comparison test. Error bars indicate SEM (n≥15).

(4) The authors conclude that C. elegans cannot detect leucine in chemotaxis assays. It is important to add the method for how leucine chemotaxis assay was done in order to interpret these results. Because leucine is not volatile if leucine is put on the plates immediately before the worms are added (as in a traditional odor chemotaxis assay), there is no leucine gradient for the worm to detect. It would be good to put leucine on the plate several hours before worms are introduced so worms have the possibility to be able to detect the gradient of leucine (for example, see Wakabayashi et al., 2009).

Previously, the chemotaxis assays with leucine were performed like traditional odor chemotaxis assays. We also performed chemotaxis assay as detailed in Shingai et al 2005[3]. Leucine was spotted on the assay plates 5 hours prior to the introduction of worms on the plates. As shown in new panel S1H in the revised manuscript, wild-type worms do not show response to leucine in the modified chemotaxis assay.

We have included the experimental details for leucine chemotaxis assays in the revised manuscript.  

(5) The bacterial preference assay entitled "odor-only assay" is a misleading name. In the assay, C. elegans is exposed to both volatile chemicals (odors) and non-volatile chemicals because the bacteria are grown on the assay plate for 12 hours before the worms are introduced to the assay plate. In that time, the bacteria is likely releasing non-volatile metabolites into the plate which may affect the worm's preference. A true odor-only assay would have the bacteria on the lid and the worms on the plate.

The ‘odor-only’ diet preference assay does not allow for non-volatile chemicals to reach worms. We achieved this by using tripartite dishes where the compartments containing worms and bacterial odors are separated by polystyrene barriers. At the time of the assay, worms were spotted in a separate compartment from that of bacteria (as shown in schematic 1A). The soluble metabolites released by the bacteria during their growth will accumulate in the agar within the bacterial compartment alone such that worms only encounter the volatile metabolites produced by bacteria wafting past the polystyrene barrier.

(6) The findings of the study should be discussed more in the context of prior literature. For example, AWC neurons have been previously shown to be involved in bacterial preference (Harris et al., 2014; Worthy et al., 2018). In addition, CeMbio bacterial strains (the strains examined in this study) have been previously shown to release isoamyl alcohol (Chai et al. 2024).

Thanks for the suggestion. We have modified the Discussion section to discuss the study in the light of relevant prior literature.  

Reviewer #2 (Public review):

Summary:

Siddiqui et al. show that C. elegans prefers certain bacterial strains that have been supplemented with the essential amino acid (EEA) leucine. They convincingly show that some leucine enriched bacteria stimulate the production of isoamyl alcohol (IAA). IAA is an attractive odorant that is sensed by the AWC. The authors an identify a receptor, SRD-12 (SNIF-1), that is expressed in the AWC chemosensory neurons and is required for chemotaxis to IAA. The authors propose that IAA is a predominant olfactory cue that determines diet preference in C. elegans. Since leucine is an EAA, the authors propose that worm IAA sensing allows the animal provides a proxy mechanism to identify EAA rich diets.

Strengths:

The authors propose IAA as a predominant olfactory cue that determines diet preference in C. elegans providing molecular mechanism underlying diet selection. They show that wild isolates of C. elegans have a strong chemotactic response to IAA indicating that IAA is an ecologically relevant odor for the worm. The paper is well written, and the presented data are convincing and well organized. This is an interesting paper that connects chemotactic response with bacterially produced odors and thus provides an understanding of how animals adapt their foraging behavior through the perception of molecules that may indicate the nutritional value.

Weaknesses:

Major:

While I do like the way the authors frame C. elegans IAA sensing as mechanisms to identify leucine (EAA) rich diets it is not fully clear whether bacterial IAA production is a proxy for bacterial leucine levels.

(1) Can the authors measure leucine (or other EAA) content of the different CeMbio strains? This would substantiate the premise in the way they frame this in the introduction. While the authors convincingly show that leucine supplementation induces IAA production in some strains, it is not clear if there are lower leucine levels in the different in non-preferred strains.

Thanks for your suggestion. Estimating leucine levels in various bacteria will provide useful information, and we hope to do so in future studies.

(2) It is not clear whether the non-preferred bacteria in Figure 1A and 1B have the ability to produce IAA. To substantiate the claim that C. elegans prefers CEent1, JUb66, and BIGb0170 due to their ability to generate IAA from leucine, it would measure IAA levels in non-preferred bacteria (+ and - leucine supplementation). If the authors have these data it would be good to include this.

Thanks for the suggestion. We have included the table indicating the presence or absence of IAA production by all the bacteria under + LEU and – LEU conditions (Table S2). Some of the nonpreferred bacteria indeed produce isoamyl alcohol. However, the abundance of IAA in these strains is significantly less than in the preferred bacteria.  

Using the available genomic sequence data, we found that all CeMbio strains encode IlvE-like transaminase enzymes[4]. This suggests that presumably all the bacteria have the metabolic capacity to make alpha-ketoisocaproate (an intermediate in IAA biosynthetic pathway) from leucine. However, the regulation of metabolic flux is likely to be quite complex in various bacteria.  

(3) The authors would strengthen their claim if they could show that deletion or silencing ilvE enzyme reduces IAA levels and eliminates the increased preference upon leucine supplementation.

We agree that testing worms’ diet preference for the preferred strains upon ilvE knockout will further strengthen the claim for IAA being crucial for finding leucine-enriched diet. Currently the lab does not have the necessary expertise and standardize protocols to do genetic manipulations for the CeMbio strains.

(4) While the three preferred bacteria possess the ilvE gene, it is not clear whether this enzyme is present in the other non-preferred bacterial strains. As far as I know, the CeMbio strains have been sequenced so it should be easy to determine if the non-preferred bacteria possess the capacity to make IAA. Does the expression of ilvE in e.g. E. coli increase its preference index or are the other genes in the biosynthesis pathway missing?

Thanks for the suggestion. Using the available genomic sequence data, we find that all the bacteria in the CeMbio collection possess IlvE-like transaminase necessary for synthesis of alphaketoisocaproate, key metabolite in leucine turn over as well as precursor for IAA [4]. E. coli has an IlvE encoding gene in its genome [2]. However, we do not find IAA in the headspace of E. coli either with or without leucine supplementation. This indicates either (i) E. coli lacks enzymes for subsequent steps in IAA biosynthesis or (ii) leucine provided under the experimental regime is not sufficient to shift the metabolic flux to IAA production.  

Previous studies have suggested that in yeast, the final two steps leading to IAA production are catalyzed by decarboxylase and dehydrogenase enzymes1. The genomic and metabolic flux data available for CeMbio do not describe specific enzymes leading up to IAA synthesis [4].  

(5) It is strongly implied that leucine-rich diets are beneficial to the worm. Do the authors have data to show the effect on leucine supplementation on C. elegans healthspan, life-span or broodsize?

Edwards et al. 2015 reported a 15% increase in the lifespan of worms upon 1 mM leucine supplementation [5]. Wang et al 2018 also showed lifespan extension upon 1 mM and 10 mM leucine supplementation. They also reported that while leucine supplementation did not have any effect on brood size, it did make worms more resistant to heat, paraquat, and UV-stress [6]. These studies have been included in the discussion section.

Other comments:

Page 6. Figure 2c. While the authors' conclusions are correct based on AWC expts. it would be good at this stage to include the possibility that odors that enriched in the absence of leucine may be aversive.

Thanks for the comment. We have tested the chemotaxis response of the worms for most of the odors produced by CeMbio strains without leucine supplementation. We did not find any odor that is aversive to worms. However, we cannot completely rule out the possibility that a low abundance of aversive odor in the headspace of the bacteria was missed.

Interestingly, we did identify 2-nonanone, a known repellent, in the headspace of the preferred bacteria upon leucine supplementation. However, the abundance of 2-nonanone in headspace of bacteria is relatively low (less than 1% for CEent1, and JUb66, and ~10% for BIGb0170). This suggests that the relative abundance of odors in an odor bouquet may be a relevant factor in determining worms’ reference.  

Page 6. IAA increases 1.2-4 folds upon leucine supplementation. If the authors perform a chemotaxis assay with just IAA with 1-2-4 fold differences do you get the shift in preference index as seen with the bacteria? i.e. is the difference in IAA concentration sufficient to explain the shift in bacterial PI upon leucine supplementation? Other attractants such as Acetoin and isobutanol go up in -Leu conditions.

Thanks for the suggestion. As shown in Figure S2H and S2I, when given a choice between a concentration of IAA (1:1000 dilution) attractive to worms and a 4-fold higher amount of IAA, worms chose the latter. This result suggests that worms can distinguish between relatively small difference in concentrations of IAA.

We agree that the relative abundance of Acetoin and Isobutanol is high in -LEU conditions. The presence of other attractants in - LEU conditions should skew the preference of worms for – LEU bacteria. However, we found that worms prefer + LEU bacteria (Figure 1B), suggesting that the abundance of IAA mainly influences the diet preference of the worms.  

Page 14-15. The authors identify a putative IAA receptor based on expression studies. I compliment the authors for isolating two CRISPR deletion alleles. They show that the srd-12 (snif-1) mutants have obvious defects in IAA chemotaxis. Very few ligand-odorant receptors combinations have been identified so this is an important discovery. CenGen data indicate that srd-12 (snif-1) is expressed in a limited set of neurons. Did the authors generate a reporter to show the expression of srd-12 (snif-1)? This is a simple experiment that would add to the characterization of the SRD-12 (SNIF-1) receptor. Rescue experiments would be nice even though the authors have independent alleles. To truly claim that SRD-12 (SNIF-1) is the ligand for IAA and activates the AWC neurons would require GCamp experiments in the AWC neuron or heterologous expression system. I understand that GCamp imaging might not be part of the regular arsenal of the lab but it would be a great addition (even in collaboration with one of the many labs that do this regularly). Comparing AWC activity using GCaMP in response IAA-producing bacteria with high leucine levels in both wild-type and SRD-12 (SNIF-1) deficient backgrounds, would further support their narrative. I leave that to the authors.

Thanks for your comments and suggestions. To address this comment, we rescued snif-1 mutant (referred as VSL2401) with extrachromosomal array expressing snif-1 under AWC-specific promoter as well as its native promoter. As shown in Figure 6H and Author response image 2, we find that both transgenic lines show a complete rescue of chemotaxis response to isoamyl alcohol. To find where snif-1 is expressed, we generated a transgenic line of worms expressing GFP under snif-1 promoter, and mCherry under odr-1 promoter (to mark AWC neurons). As shown in Figure 6I, we found that snif-1 is expressed faintly in many neurons, with strong expression in one of the two AWC neurons marked by odr-1::mCherry. This result suggests that SNIF-1 is expressed in AWC neuron.

We hope to perform GCaMP assay and further characterization of SNIF-1 in the future.

Author response image 2.

Chemotaxis index (CI) of WT, VSL2401, VSL2401 [AWCp:: snif-1] and VSL2401 [snif-1p::snif-1] worms to IAA at 1:1000 dilution. Significant differences are indicated as **** P ≤ 0.0001 determined by one-way ANOVA followed by post hoc Dunnett’s multiple comparison test. Error bars indicate SEM (n≥15).

Minor:

Page 4 "These results suggested that worms can forage for diets enriched in specific EAA, leucine...." More precise at this stage would be to state " These results indicated that worms can forage for diets supplemented with specific EAA...".

We have changed the statement in the revised manuscript.

Page 5."these findings suggested that worms not only rely on odors to choose between two bacteria but also to find leucine enriched bacteria" This statement is not clear to me and doesn't follow the data in Fig. S2. Preferred diets in odorant assays are the IAA producing strains.

Thanks for your comment. We have revised the manuscript to make it clear. “Altogether, these findings suggested that worms rely on odors to distinguish different bacteria and find leucineenriched bacteria”. This statement concludes all the data shown in Figure 1 and Figure S1.  

Page 5. Figure S2A provides nice and useful data that can be part of the main Figure 1.

Thanks for the comment. We have incorporated the data from Figure S2A to main Figure 1.

Reviewer #3 (Public review):

Summary:

The authors first tested whether EAA supplementation increases olfactory preference for bacterial food for a variety of bacterial strains. Of the EAAs, they found only leucine supplementation increased olfactory preference (within a bacterial strain), and only for 3 of the bacterial strains tested. Leucine itself was not found to be intrinsically attractive.

They determined that leucine supplementation increases isoamyl alcohol (IAA) production in the 3 preferred bacterial strains. They identify the biochemical pathway that catabolizes leucine to IAA, showing that a required enzyme for this pathway is upregulated upon supplementation.

Consistent with earlier studies, they find that AWC olfactory neuron is primarily responsible for increased preference for IAA-producing bacteria.

Testing volatile compounds produced by bacteria and identified by GC/MS, and identified several as attractive, most of them require AWC for the full effect. Adaptation assays were used to show that odorant levels produced by bacterial lawns were sufficient to induce olfactory adaptation, and adaptation to IAA reduced chemotaxis to leucine-supplemented lawns. They then showed that IAA attractiveness is conserved across wild strains, while other compounds are more variable, suggesting IAA is a principal foraging cue.

Finally, using the CeNGEN database, they developed a list of candidate IAA receptors. Using behavioral tests, they show that mutation of srd-12 (snif-1) greatly impairs IAA chemotaxis without affecting locomotion or attraction to another AWC-sensed odor, PEA.

Comments

This study will be of great interest in the field of C. elegans behavior, chemical senses and chemical ecology, and understanding of the sensory biology of foraging.

Strengths:

The identification of a receptor for IAA is an excellent finding. The combination of microbial metabolic chemistry and the use of natural bacteria and nematode strains makes an extremely compelling case for the ecological and adaptive relevance of the findings.

Weaknesses:

AWC receives synaptic input from other chemosensory neurons, and thus could potentially mediate navigation behaviors to compounds detected in whole or in part by those neurons. Language concluding detection by AWC should be moderated (e.g. p9 "worms sense an extensive repertoire...predominantly using AWC") unless it has been demonstrated.

Thanks for your comment. We have modified the manuscript to incorporate the suggestion.

srd-12 (snif-1) is not exclusively expressed in AWC. Normally, cell-specific rescue or knockdown would be used to demonstrate function in a specific cell. The authors should provide such a demonstration or explain why they are confident srd-12 (snif-1) acts in AWC.

Thanks for the comment. We have performed AWC-specific rescue of snif-1 in mutant worms. As shown in Figure 6H, we found that AWC neurons specific rescue completely recovered the chemotaxis defect of the snif-1 mutant (referred as VSL2401) for IAA. In addition, snif-1 is expressed in one of the AWC neurons.

A comparison of AWC's physiological responses between WT and srd-12 (snif-1), preferably in an unc13 background, would be nice. Even further, the expression of srd-12 (snif-1) in a different neuron type and showing that it confers responsiveness to IAA (in this case, inhibition) would be very convincing.

Thanks for the suggestion. We have performed a receptor swap experiment, where snif-1 is misexpressed in AWB neurons. We find that these worms show slight but significant repulsion to IAA compared to WT and snif-1 mutant worms (Author response image 1).

Recommendations for the authors:

Reviewing Editor:

Please consider all of the reviewer comments. In particular, as noted in the individual reviews, the strength of the evidence would be bolstered by additional experiments to demonstrate that the iLvE enzyme affects IAA levels in the preferred bacteria. The reviewers note that the authors haven't shown that IAA production is a reflection of leucine content. Are the non-preferred bacteria low on leucine or lack iLvE or IAA synthesis pathways? Further, more direct evidence that SRD-12 (SNIF-1) is in fact the primary IAA receptor would further strengthen the study. The authors should also be aware that geographic distance for wild isolate C. elegans may not directly correlate with phylogenetic distance. This should be assessed/discussed for the strains used.

Thanks for the suggestions. Some of these have been addressed in response to reviewers. Thanks for your comments about possible disconnect between geographical and phylogenetic distances amongst natural isolates used here.

By analyzing the phylogenetic tree generated using neighbor-joining algorithm available at CaeNDR database, we found that QX1211 and JU3226 are phylogenetically close, but the remaining isolates fall under different clades separated by long phylogenetic distances [7,8].  

Reviewer #1 (Recommendations for the authors):

(1) In the first sentence of the third paragraph of the introduction, C. elegans are described as "soildwelling." Although C. elegans has been described as soil-dwelling in the past, current research indicates they are most often found on rotten fruit, compost heaps and other bacterial-rich environments, not soil. "All Caenorhabditis species are colonizers of nutrient- and bacteria-rich substrates and none of them is a true soil nematode." from Kiontke, K. and Sudhaus, W. Ecology of Caenorhabditis species (WormBook).

Your specific comment about C. elegans’ habitat is well received. However, in that sentence we are referring to the chemosensory system of soil-dwelling animals in general, and not particularly C. elegans.

(2) Figure 3K, the model would be clearer if leucine-rich diet -> volatile chemicals ->AWC (instead of leucine-rich diet -> AWC <- volatile chemicals). The leucine-rich diet results in the production of volatile chemicals which are detected by AWC.

We have modified the figure to make it clearer.

(3) Figure 4 - it would help to include a table summarizing the volatile chemicals that each bacteria releases. Then the reader could more easily evaluate whether the adaptation to each specific odor is consistent with the change in preference for the specific bacteria based on what it releases in its headspace. In addition, Figure 4 would help to clarify whether bacteria in these experiments were cultured with or without leucine supplementation.

Table S2 summarizes the odors released by all the bacteria under + LEU and – LEU conditions.

In Figure 4, adaptation was performed by odors of bacteria when cultured under leucineunsupplemented conditions.

Reviewer #2 (Recommendations for the authors):

Page 9. Previous studies e.g. Bargmann Hartwieg and Horvitz have shown IAA is sensed by the AWC. Would be good to cite appropriately.

Thanks for the comment. The reference has been cited at p9 and p16.

References:

(1) Yuan, J., Mishra, P., and Ching, C.B. (2017). Engineering the leucine biosynthetic pathway for isoamyl alcohol overproduction in Saccharomyces cerevisiae. Journal of Industrial Microbiology and Biotechnology 44, 107-117. 10.1007/s10295-016-1855-2 %J Journal of Industrial Microbiology and Biotechnology.

(2) Kanehisa, M., Furumichi, M., Sato, Y., Matsuura, Y., and Ishiguro-Watanabe, M. (2025). KEGG: biological systems database as a model of the real world. Nucleic Acids Res 53, D672-d677. 10.1093/nar/gkae909.

(3) Shingai, R., Wakabayashi, T., Sakata, K., and Matsuura, T. (2005). Chemotaxis of Caenorhabditis elegans during simultaneous presentation of two water-soluble attractants, llysine and chloride ions. Comparative biochemistry and physiology. Part A, Molecular & integrative physiology 142, 308-317. 10.1016/j.cbpa.2005.07.010.

(4) Dirksen, P., Assié, A., Zimmermann, J., Zhang, F., Tietje, A.M., Marsh, S.A., Félix, M.A., Shapira, M., Kaleta, C., Schulenburg, H., and Samuel, B.S. (2020). CeMbio - The Caenorhabditis elegans Microbiome Resource. G3 (Bethesda, Md.) 10, 3025-3039. 10.1534/g3.120.401309.

(5) Edwards, C., Canfield, J., Copes, N., Brito, A., Rehan, M., Lipps, D., Brunquell, J., Westerheide, S.D., and Bradshaw, P.C. (2015). Mechanisms of amino acid-mediated lifespan extension in Caenorhabditis elegans. BMC genetics 16, 8. 10.1186/s12863-015-0167-2.

(6) Wang, H., Wang, J., Zhang, Z.J.J.o.F., and Research, N. (2018). Leucine Exerts Lifespan Extension and Improvement in Three Types of Stress Resistance (Thermotolerance, AntiOxidation and Anti-UV Irradiation) in C. elegans. 6, 665-673.

(7) Crombie, T.A., McKeown, R., Moya, N.D., Evans, Kathryn S., Widmayer, Samuel J., LaGrassa, V., Roman, N., Tursunova, O., Zhang, G., Gibson, Sophia B., et al. (2023). CaeNDR, the Caenorhabditis Natural Diversity Resource. Nucleic Acids Research 52, D850-D858. 10.1093/nar/gkad887 %J Nucleic Acids Research.

(8) Cook, D.E., Zdraljevic, S., Roberts, J.P., and Andersen, E.C. (2017). CeNDR, the Caenorhabditis elegans natural diversity resource. Nucleic Acids Res 45, D650-d657. 10.1093/nar/gkw893.

公式ドキュメントによると、*, はないようです。

https://docs.python.org/ja/3/library/concurrent.futures.html#concurrent.futures.ProcessPoolExecutor

Python 3.13からサポートされていますが、ここでは3.14のことしか書かない?

https://docs.python.org/ja/3.14/howto/free-threading-python.html

haganme una cita de esto

Reviewer #1 (Public review):

Summary:

This work by Al-Jezani et al. focused on characterizing clonally derived MSC populations from the synovium of normal and osteoarthritis (OA) patients. This included characterizing the cell surface marker expression in situ (at time of isolation), as well as after in vitro expansion. The group also tried to correlate marker expression with trilineage differential potential. They also tested the ability of the different sub-populations for their efficacy in repairing cartilage in a rat model of OA. The main finding of the study is that CD47hi MSCs may have a greater capacity to repair cartilage than CD47lo MSCs, suggesting that CD47 may be a novel marker of human MSCs that have enhanced chondrogenic potential.

Strengths:

Studies on cell characterization of the different clonal populations isolated indicate that the MSC are heterogenous and traditional cell surface markers for MSCs do not accurately predict the differentiation potential of MSCs. While this has been previously established in the field of MSC therapy, the authors did attempt to characterize clones derived from single cells, as well as evaluate the marker profile at the time of isolation. While the outcome of heterogeneity is not surprising, the methods used to isolate and characterze the cells were well developed. The interesting finding of the study is the identification of CD47 as a potential MSC marker that could be related to chondrogenic potential. The authors suggest that MSCs with high CD47 repaired cartilage more effectively than MSC with low CD47 in a rat OA model.

Comments on revisions:

Thank you for addressing the comments from the first review. No additional revisions.

Author response:

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

Public Reviews: 

Reviewer #1 (Public review): 

Summary: 

This work by Al-Jezani et al. focused on characterizing clonally derived MSC populations from the synovium of normal and osteoarthritis (OA) patients. This included characterizing the cell surface marker expression in situ (at time of isolation), as well as after in vitro expansion. The group also tried to correlate marker expression with trilineage differential potential. They also tested the ability of the different subpopulations for their efficacy in repairing cartilage in a rat model of OA. The main finding of the study is that CD47hi MSCs may have a greater capacity to repair cartilage than CD47lo MSCs, suggesting that CD47 may be a novel marker of human MSCs that have enhanced chondrogenic potential. 

Strengths: 

Studies on cell characterization of the different clonal populations isolated indicate that the MSC are heterogenous and traditional cell surface markers for MSCs do not accurately predict the differentiation potential of MSCs. While this has been previously established in the field of MSC therapy, the authors did attempt to characterize clones derived from single cells, as well as evaluate the marker profile at the time of isolation. While the outcome of heterogeneity is not surprising, the methods used to isolate and characterize the cells were well developed. The interesting finding of the study is the identification of CD47 as a potential MSC marker that could be related to chondrogenic potential. The authors suggest that MSCs with high CD47 repaired cartilage more effectively than MSC with low CD47 in a rat OA model. 

Weaknesses: 

While the identification of CD47 as a novel MSC marker could be important to the field of cell therapy and cartilage regeneration, there was a lack of robust data to support the correlation of CD47 expression to chondrogenesis. The authors indicated that the proteomics suggested that the MSC subtype expressed significantly more CD47 than the non-MSC subtype. However, it was difficult to appreciate where this was shown. It would be helpful to clearly identify where in the figure this is shown, especially since it is the key result of the study. The authors were able to isolate CD47hi and CD47 low cells. While this is exciting, it was unclear how many cells could be isolated and whether they needed to be expanded before being used in vivo. Additional details for the CD47 studies would have strengthened the paper. Furthermore, the CD47hi cells were not thoroughly characterized in vitro, particularly for in vitro chondrogenesis. More importantly, the in vivo study where the CD47hi and CD47lo MSCs were injected into a rat model of OA lacked experimental details regarding how many cells were injected and how they were labeled. No representative histology was presented and there did not seem to be a statistically significant difference between the OARSI score of the saline injected and MSC injected groups. The repair tissue was stained for Sox9 expression, which is an important marker of chondrogenesis but does not show production of cartilage. Expression of Collagen Type II would be needed to more robustly claim that CD47 is a marker of MSCs with enhanced repair potential. 

Reviewer #2 (Public review): 

Summary: 

This is a compelling study that systematically characterized and identified clonal MSC populations derived from normal and osteoarthritis human synovium. There is immense growth in the focus on synovial-derived progenitors in the context of both disease mechanisms and potential treatment approaches, and the authors sought to understand the regenerative potential of synovial-derived MSCs. 

Strengths: 

This study has multiple strengths. MSC cultures were established from an impressive number of human subjects, and rigorous cell surface protein analyses were conducted, at both pre-culture and post-culture timepoints. In vivo experiments using a rat DMM model showed beneficial therapeutic effects of MSCs vs non-MSCs, with compelling data demonstrating that only "real" MSC clones incorporate into cartilage repair tissue and express Prg4. Proteomics analysis was performed to characterize non-MSC vs MSC cultures, and high CD47 expression was identified as a marker for MSC. Injection of CD47-Hi vs CD47-Low cells in the same rat DMM model also demonstrated beneficial effects, albeit only based on histology. A major strength of these studies is the direct translational opportunity for novel MSC-based therapeutic interventions, with high potential for a "personalized medicine" approach. 

Weaknesses: 

Weaknesses of this study include the rather cursory assessment of the OA phenotype in the rat model, confined entirely to histology (i.e. no microCT, no pain/behavioral assessments, no molecular readouts). It is somewhat unclear how the authors converged on CD47 vs the other factors identified in the proteomics screen, and additional information is needed to understand whether true MSCs only engraft in articular cartilage or also in ectopic cartilage (in the context of osteophyte/chondrophyte formation). Some additional discussion and potential follow-up analyses focused on other cell surface markers recently described to identify synovial progenitors is also warranted. A conceptual weakness is the lack of discussion or consideration of the multiple recent studies demonstrating that DPP4+ PI16+ CD34+ stromal cells (i.e. the "universal fibroblasts") act as progenitors in all mesenchymal tissues, and their involvement in the joint is actively being investigated. Thus, it seems important to understand how the MSCs of the present study are related to these DPP4+ progenitors. Despite these areas for improvement, this is a strong paper with a high degree of rigor, and the results are compelling, timely, and important. 

Overall, the authors achieved their aims, and the results support not just the therapeutic value of clonally-isolated synovial MSCs but also the immense heterogeneity in stromal cell populations (containing true MSCs and non-MSCs) that must be investigated further. Of note, the authors employed the ISCT criteria to characterize MSCs, with mixed results in pre-culture and post-culture assessments. This work is likely to have a longterm impact on methodologies used to culture and study MSCs, in addition to advancing the field's knowledge about how synovial-derived progenitors contribute to cartilage repair in vivo.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors): 

In all figures, it would be beneficial to report the sample number used for the data reported. It is difficult to appreciate the statistical analysis without that information.

Understood, the sample number and replicates have been added to each figure legend.

Please check that Table S7 is part of the manuscript. It could not be found.

It was added as an additional excel file since it was too large to fit in the word document.

Lines 377-379 (Figure 2E): the authors write that rats receiving MSCs had a significantly lower OARSI and Krenn score vs. rats injected with non-MSCs. However, none of the bars indicating statistical significance run between these two groups. Please verify the text and figure.

This has been corrected

The details surrounding the labeling of the cells with tdTomato were not presented in the methods. 

This has been added to the methods

The fluorescent antibodies used should be listed and more details provided in the methods rather than a general statement that fluorescent antibodies were used.

Our apologies, the clones and companies have been added.

Additional information on the CD47 experiments (# cells, # animals) would have strengthened the study.

This has been added to the methods and figure legend.

Reviewer #2 (Recommendations for the authors): 

My comments span minor corrections, requests for additional analyses, some suggestions for additional experiments, and requests for additional discussion of recent important studies. 

Introduction: 

The introduction is thorough and well-written. I recommend a brief discussion about the emerging evidence demonstrating that DPP4+ PI16+ CD34+ synovial cells, i.e. the "universal fibroblasts", act as stromal progenitors in development, homeostasis, and disease. Relevant osteoarthritis-related papers encompass human and mouse studies (PMIDs: 39375009, 38266107, 38477740, 36175067, 36414376).

This has been added.

Relatedly, as DPP4 is CD26 and therefore useful as a cell-surface antigen for flow cytometry, sorting, etc, it would be interesting to understand the relationship and similarities between the CD47-High cells identified in this study and the DPP4/PI16+ cells previously described. Do they overlap in phenotype/identity?

We have added a new flow cytometry figure for address this question. 

Results: 

Note type-o on Line 311: "preformed" instead of "performed". Line 313 "prolife" instead of "profile"

Thank you for catching these.

The identified convergence of the cell surface marker profile of all normal and OA clones in culture is a highly intriguing result. Do the authors have stored aliquots of these cells to demonstrate whether this would also occur in soft substrate, i.e. low stiffness culture conditions? This could be done with standard dishes coated with bulk collagen or with commercially available low-stiffness dishes (1 kPa). This is relevant to multiple studies demonstrating the induction of a myofibroblast-like phenotype by stromal cells cultured on high-stiffness plastic or glass. This is also the experiment where assessment of DPP4/CD26 could be added, if possible.

While we agree it would be interesting to determine the mechanism by which the cells phenotypes converge, we would argue that it is outside of the scope of the current manuscript. We have instead added a sentence to the discussion. 

Line 353 regarding the use of CD68 as a negative gate: can the authors pleasecomment on why they employed CD68 here and not CD45? While monocytes/macs/myeloid cells are the most abundant immune cells in synovium, CD45 would more comprehensively exclude all immune cells. 

That is a fair point, and we really don’t have any reason to have picked CD68 over CD45. In our opinion either would be a fair negative marker to use based on the literature. 

Fig 2, minor suggestion: consider adding "MSC vs non-MSC" on the experimental schematic to more comprehensively summarize the experiment. 

This has been modified 

Fig 2E should show all individual datapoints, not just bar graphs. 

This has been modified

Fig 2: Given the significant reduction in Krenn score in DMM-MSC injected knees compared to DMM-saline knees, Fig 2 should also show representative images of the synovial phenotype to demonstrate which aspects of synovial pathology were mitigated. Was the effect related to lining hyperplasia, subsynovial infiltrate, fibrosis, etc? Similarly, can the authors narrate which aspects of the OARSI score drove the treatment effect (proteoglycans vs structure vs osteophytes, etc). 

We have added a new sup figure breaking down the Krenn score as well as higher magnification images of representative synovium.

Fig 2: In the absence of microCT imaging, can the authors quantify subchondral bone morphometrics using multiple histological sections? The tibial subchondral bone in Fig 2D appears protected from sclerosis/thickening.

Unfortunately, this is beyond what are able to add to the manuscript. 

The Fig 3 results are highly compelling and interesting. Congratulations.

Thank you very much.

Fig 4A: the cell highlighted in the high-mag zoom box in Fig 4A appears to be localized within the joint capsule or patellar tendon (it is unclear which anatomic region this image represents). The highly aligned nature of the tissue and cells along a fibrillar geometry indicates that this is not synovium. The interface between synovium and the tissue in question can be clearly observed in this image. Please choose an image more representative of synovium.

We completely agree with the reviewers assessment. However, it is the synovium that overlays this tissue (Fig 4A arrow). We are simply showing that there were very few MSCs that took up residence in the synovium or the adjacent tissues. 

Fig 4C and F: please show individual data points.

This has been added

Fig 5D: I see DPP4 and ITGA5 were also hits in the proteomics analysis, which is intriguing. Besides my comments/suggestions regarding DPP4 above, please note this recent paper identifying a ITGA5+ synovial fibroblast subset that orchestrates pathological crosstalk with lymphocytes in RA, PMID: 39486872

Thank you for the information. We have added the reference in the results section. 

Fig 5B-D: How did the authors converge on CD47 as the target for follow-up study? It does not appear to be a differentially-expressed protein based on the Volcano plot in Fig 5B, and it's unclear why it is a more important factor than any of the other proteins shown in the network diagram in Fig 5D, e.g. CTSL, ITGA5, DPP4. Can the authors add a quantitative plot supporting their statement "the MSC sub-type expressed significantly more CD47 than the non-MSCs" on Line 458? 

We have re-written this line. It was incorrect to discuss amount of CD47. That was shown later with the flow analysis.  

Fig 6D: Please show individual data points and also representative histology images to demonstrate the nature of the phenotypic effect.

This has been added. 

Fig 6E-F: In what anatomic region are these images? Please add anatomic markers to clarify the location and allow the reader to interpret whether this is articular cartilage or ectopic cartilage

We have redone the figure to show the area as requested.

Relevant to this, do the authors observe this type of cellular engraftment in ectopic cartilage/osteophytes or only in articular cartilage? Understanding the contribution of these cells to the formation/remodeling of various cartilage types in the context of OA is a critical aspect of this line of investigation.

We didn’t see any contribution of these cells to ectopic cartilage formation and are actively working on a follow up study discussing this point specifically. 

Discussion: 

Besides my comments regarding DPP4 and ITGA5 above, the authors may also consider discussing PMID: 37681409 (JCI Insight 2023), which demonstrates that adult Prg4+ progenitors derived from synovium contribute to articular cartilage repair in vivo. 

We agree that there are numerous markers we could look at in future studies and that other people in the field are actively studying.

Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.

Learn more at Review Commons

Reply to the reviewers

The PDF version of point-by-point response includes figures (I, II, III,... IX) that are not included in the manuscript nor in this post but serve to illustrate and clarify our replies to the reviewers' comments.

Dear Editor,

Many thanks for forwarding the comments from reviewers #1-#4 regarding our manuscript (Preprint #RC-2025-03087144), entitled "HIV-1 Envelope glycoprotein modulates CXCR4 clustering and dynamics on the T cell membrane", by Quijada-Freire A. et al.

We have carefully reviewed all reviewer comments and prepared our specific, detailed responses. Alongside this, we have created a revised version of the manuscript to post them on BioRxiv, and we are pleased to announce that we will transfer this new version to an affiliate journal for consideration.

Reviewer #1

Thank you very much for considering that our manuscript evaluates an important question and that the reagents used are well prepared and characterized. We also much appreciate that you consider the information generated as potentially useful for those studying HIV infection processes and strategies to prevent infection.

• While a single particle tracking routine was applied to the data, it's not clear how the signal from a single GFP was defined and if movement during the 100 ms acquisition time impacts this. My concern would be that the routine is tracking fluctuations, and these are related to single particle dynamics, it appears from the movies that the density or the GFP tagged receptors in the cells is too high to allow clear tracking of single molecules. SPT with GFP is very difficult due to bleaching and relatively low quantum yield. Current efforts in this direction that are more successful include using SNAP tags with very photostable organic fluorophores. The data likely does mean something is happening with the receptor, but they need to be more conservative about the interpretation. *

Some of the paradoxical effects might be better understood through deeper analysis of the SPT data, particularly investigation of active transport and more detailed analysis of "immobile" objects. Comments on early figures illustrate how this could be approached. This would require selecting acquisitions where the GFP density is low enough for SPT and performing a more detailed analysis, but this may be difficult to do with GFP.

When the authors discuss clusters of 3, how do they calibrate the value of GFP and the impact of diffusion on the measurement. One way to approach this might be single molecules measurements of dilute samples on glass vs in a supported lipid bilayer to map the streams of true immobility to diffusion at >1 µm2/sec.

We fully understand the reviewer's apprehensions regarding the application of these high-end biophysical techniques, in particular the associated complexity of the data analysis. We provide below extensive explanations on our methodology, which we hope will satisfactorily address all of the reviewer's concerns.

We would first like to emphasize that the experimental conditions and the quantitative analysis used in our current experiments are similar to the established protocols and methodologies applied by our group previously (Martinez-Muñoz et al. Mol. Cell, 2018; García-Cuesta et al. PNAS, 2022; Gardeta et al. Frontiers in Immunol., 2022; García-Cuesta et al.eLife, 2024; Gardeta et al. Cell. Commun. Signal., 2025) and by others (Calebiro et al. PNAS, 2013; Jaqaman et al. Cell,2011; Mattila et al. Immunity, 2013; Torreno-Pina et al. PNAS, 2014; Torreno-Pina et al. PNAS, 2016).

As SPT (single-particle tracking) experiments require low-expressing conditions in order to follow individual trajectories (Manzo & García-Parajo Rep. Prog. Phys., 2015), we transiently transfected Jurkat CD4+ cells with CXCR4-AcGFP or CXCR4R334X-AcGFP. At 24 h post-transfection, cells expressing low CXCR4-AcGFP levels were selected by a MoFlo Astrios Cell Sorter (Beckman-Coulter) to ensure optimal conditions for SPT. Using Dako Qifikit (DakoCytomation), we quantified the number of CXCR4 receptors and found ∼8,500 - 22,000 CXCR4-AcGFP receptors/cell, which correspond to a particle density ∼2 - 4.5 particles/mm2 (Figure I, only for review purposes) and are similar to the expression levels found in primary human lymphocytes.

These cells were resuspended in RPMI supplemented with 2% FBS, NaPyr and L-glutamine and plated on 96-well plates for at least 2 h. Cells were centrifuged and resuspended in a buffer with HBSS, 25 mM HEPES, 2% FBS (pH 7.3) and plated on glass-bottomed microwell dishes (MatTek Corp.) coated with fibronectin (FN) (Sigma-Aldrich, 20 mg/ml, 1 h, 37{degree sign}C). To observe the effect of the ligand, we coated dishes with FN + CXCL12; FN + X4-gp120 or FN + VLPs, as described in material and methods; cells were incubated (20 min, 37{degree sign}C, 5% CO2) before image acquisition.

For SPT measurements, we use a total internal reflection fluorescence (TIRF) microscope (Leica AM TIRF inverted) equipped with an EM-CCD camera (Andor DU 885-CS0-#10-VP), a 100x oil-immersion objective (HCX PL APO 100x/1.46 NA) and a 488-nm diode laser. The microscope was equipped with incubator and temperature control units; experiments were performed at 37{degree sign}C with 5% CO2. To minimize photobleaching effects before image acquisition, cells were located and focused using the bright field, and a fine focus adjustment in TIRF mode was made at 5% laser power, an intensity insufficient for single-particle detection that ensures negligible photobleaching. Image sequences of individual particles (500 frames) were acquired at 49% laser power with a frame rate of 10 Hz (100 ms/frame). The penetration depth of the evanescent field used was 90 nm.

We performed automatic tracking of individual particles using a very well established and common algorithm first described by Jaqaman (Jaqaman et al. Nat. Methods, 2008). Nevertheless, we would stress that we implemented this algorithm in a supervised fashion, i.e., we visually inspect each individual trajectory reconstruction in a separate window. Indeed, this algorithm is not able to quantify merging or splitting events.

We follow each individual fluorescence spot frame-by-frame using a three-by-three matrix around the centroid position of the spot, as it diffuses on the cell membrane. To minimize the effect of photon fluctuations, we averaged the intensity over 20 frames. Nevertheless, to assure the reviewer that most of the single molecule traces last for at least 50 frames (i.e., 5 seconds), we provide the following data and arguments. We currently measure the photobleaching times from individual CD86-AcGFP spots exclusively having one single photobleaching step to guarantee that we are looking at individual CD86-AcGFP molecules. The distribution of the photobleaching times is shown below (Figure II, only for review purposes). Fitting of the distribution to a single exponential decay renders a t0 value of ~5 s. Thus, with 20 frames averaging, we are essentially measuring the whole population of monomers in our experiments. As the survival time of a molecule before photobleaching will strongly depend on the excitation conditions, we used low excitation conditions (2 mW laser power, which corresponds to an excitation power density of ~0.015 kW/cm2 considering the illumination region) and longer integration times (100 ms/frame) to increase the signal-to-background for single GFP detection while minimizing photobleaching.

To infer the stoichiometry of receptor complexes, we also perform single-step photobleaching analysis of the TIRF trajectories to establish the existence of different populations of monomers, dimers, trimers and nanoclusters and extract their percentage. Some representative trajectories of CXCR4-AcGFP with the number of steps detected are shown in new Supplementary Figure 1.

The emitted fluorescence (arbitrary units, a.u.) of each spot in the cells is quantified and normalized to the intensity emitted by monomeric CD86-AcGFP spots that strictly showed a single photobleaching step (Dorsch et al. Nat. Methods,2009). We have preferred to use CD86-AcGFP in cells rather than AcGFP on glass to exclude any potential effect on the different photodynamics exhibited by AcGFP when bound directly to glass. We have also previously shown pharmacological controls to exclude CXCL12-mediated receptor clustering due to internalization processes (Martinez-Muñoz et al. Mol. Cell, 2018) that, together with the evaluation of single photobleaching steps and intensity histograms, allow us to exclude the presence of vesicles in our data. Thus, the dimers, trimers and nanoclusters found in our data do correspond to CXCR4 molecules on the cell surface. Finally, distribution of monomeric particle intensities, obtained from the photobleaching analysis, was analyzed by Gaussian fitting, rendering a mean value of 980 {plus minus} 86 a.u. This value was then used as the monomer reference to estimate the number of receptors per particle in both cases, CXCR4-AcGFP and CXCR4R334X-AcGFP (new Supplementary Figure 1).

• I understand that the CXCL12 or gp120 are attached to the substrate with fibronectin for adhesion. I'm less clear how how that VLPs are integrated. Were these added to cells already attached to FN?*

For TIRF-M experiments, cells were adhered to glass-bottomed microwell dishes coated with fibronectin, fibronectin + CXCL12, fibronectin + X4-gp120, or fibronectin + VLPs. As for CXCL12 and X4-gp120, the VLPs were attached to fibronectin taking advantage of electrostatic interactions. To clarify the integration of the VLPs in these assays, we have stained the microwell dishes coated with fibronectin and those coated with fibronectin + VLPs with wheat germ agglutinin (WGA) coupled to Alexa647 (Figure III, only for review purposes) and evaluated the staining by confocal microscopy. These results indicate the presence of carbohydrates on the VLPs and are, therefore, indicative of the presence of VLPs on the fibronectin layer.

Moreover, it is important to remark that the effect of the VLPs on CXCR4 behavior at the cell surface observed by TIRF-M confirmed that the VLPs remained attached to the substrate during the experiment.

• Fig 1A- The classification of particle tracks into mobile and immobile is overly simplistic description that goes back to bulk FRAP measurements and it not really applicable to single molecule tracking data, where it's rare to see anything that is immobile and alive. An alternative classification strategy uses sub-diffusion, normal diffusion and active diffusion (or active transport) to descriptions and particles can transition between these classes over the tracking period. Fig 1B- this data might be better displayed as histograms showing distributions within the different movement classes.*

In agreement with the reviewer's commentary, the majority of the particles detected in our TIRF-M experiments were indeed mobile. However, we also detected a variable, and biologically appreciable, percentage of immobile particles depending on the experimental condition analyzed (Figure 1A in the main manuscript). To establish a stringent threshold for identifying these immobile particles under our specific experimental conditions, we used purified monomeric AcGFP proteins immobilized on glass coverslips. Our analysis demonstrated that 95% of these immobilized proteins showed a diffusion coefficient £0.0015 mm2/s; consequently, this value was established as the cutoff to distinguish immobile from mobile trajectories. While the observation of truly immobile entities in a dynamic, living system is rare, the presence of these particles under our conditions is biologically significant. For instance, the detection of large, immobile receptor nanoclusters at the plasma membrane is entirely consistent with facilitating key cellular processes, such as enabling the robust signaling cascade triggered by ligand binding or promoting the crucial events required for efficient viral entry into the cells.

Regarding the mobile receptors (defined as those with D1-4 values exceeding 0.0015 mm2/s), we observed distinct diffusion profiles derived from mean square displacement (MSD) plots (Figure V) (Manzo & García-Parajo Rep. Prog. Phys., 2015), which were further classified based on motion, using the moment scaling spectrum (MSS) (Ewers et al. PNAS, 2005). Under all experimental conditions, the majority of mobile particles, ∼85%, showed confined diffusion: for example under basal conditions, without ligand addition, ∼90% of mobile particles showed confined diffusion, ∼8.5% showed Brownian-free diffusion and ∼1.5% exhibited directed motion (new Supplementary Figure 5A in the main manuscript). These data have been also included in the revised manuscript to show, in detail, the dynamic parameters of CXCR4.

Due to the space constraints, it is very difficult to include all the figures generated. However, to ensure comprehensive assessment and transparency (for the purpose of this review), we have included below representative plots of the MSD values as a function of time from individual trajectories, showing different types of motion obtained in our experiments (Figure IV, only for review purposes).

• Fig 1C,D- It would be helpful to see a plot of D vs MSI at a single particle level. In comparing C and D I'm surprised there is not a larger difference between CXCL12 and X4-gp120. It would also be very important to see the behaviour of X4-gp120 on the CXCR4 deficient Jurkat that would provide a picture of CD4 diffusion. The CXCR4 nanoclustering related to the X4-gp120 could be dominated by CD4 behaviour.*

As previously described, all analyses were performed under SPT conditions (see previous response to point 1 in this reply). Figure 1C details the percentage of oligomers (>3 receptors/particle) calibrated using Jurkat CD4+ cells electroporated with monomeric CD86-AcGFP (Dorsch et al. Nat. Methods, 2009). The monomer value was determined by analyzing photobleaching steps as described in our previous response to point 1.

In our experiments, we observed a trend towards a higher number of oligomers upon activation with CXCL12 compared with X4-gp120. This trend was further supported by measurements of Mean Spot Intensity. However, the values are also influenced by the number of larger spots, which represents a minor fraction of the total spots detected.

The differences between the effect triggered by CXCL12 or X4-gp120 might also be attributed to a combination of factors related to differences in ligand concentration, their structure, and even to the technical requirements of TIRF-M. Both ligands are in contact with the substrate (fibronectin) and the specific nature of this interaction may differ between both ligands and influence their accessibility to CXCR4. Moreover, the requirement of the prior binding of gp120 to CD4 before CXCR4 engagement, in contrast to the direct binding of CXCL12 to CXCR4, might also contribute to the differences observed.

We previously reported that CXCL12-mediated CXCR4 dynamics are modulated by CD4 co-expression (Martinez-Muñoz et al. Mol. Cell, 2018). We have now detected the formation of CD4 heterodimers with both CXCR4 and CXCR4R334X, and found that these conformations are influenced by gp120-VLPs. In the present manuscript, we did not focus on CD4 clustering as it has been extensively characterized previously (Barrero-Villar et al. J. Cell Sci., 2009; Jiménez-Baranda et al. Nat. Cell. Biol., 2007; Yuan et al. Viruses, 2021). Regarding the investigation of the effects of X4-gp120 on CXCR4-deficient Jurkat cells, which would provide a picture of CD4 diffusion, we would note that a previous report has already addressed this issue using single-molecule super-resolution imaging, and revealed that CD4 molecules on the cell membrane are predominantly found as individual molecules or small clusters of up to 4 molecules, and that the size and number of these clusters increases upon virus binding or gp120 activation (Yuan et al. Viruses, 2021).

• Fig S1D- This data is really interesting. However, if both the CD4 and the gp120 have his tags they need to be careful as poly-His tags can bind weakly to cells and increasing valency could generate some background. So, they should make the control is fair here. Ideally, using non-his tagged person of sCD4 and gp120 would be needed ideal or they need a His-tagged Fab binding to gp120 that doesn't induce CXCR4 binding.*

New Supplementary Figure 2D shows that X4-gp120 does not bind Daudi cells (these cells do not express CD4) in the absence of soluble CD4. While the reviewer is correct to state that both proteins contain a Histidine Tag, cell binding is only detected if X4-gp120 binds sCD4. Nonetheless, we have included in the revised Supplementary Figure 2D a control showing the negative binding of sCD4 to Daudi cells in the absence of X4-gp120. Altogether, these results confirm that only sCD4/X4-gp120 complexes bind these cells.

• Fig S4- Panel D needs a scale bar. I can't figure out what I'm being shown without this.*

Apologies. A scale bar has been included in this panel (new Supplementary Figure 6D).

Reviewer #2

• This study is well described in both the main text and figures. Introduction provides adequate background and cites the literature appropriately. Materials and Methods are detailed. Authors are careful in their interpretations, statistical comparisons, and include necessary controls in each experiment. The Discussion presents a reasonable interpretation of the results. Overall, there are no major weaknesses with this manuscript.*

We very much appreciate the positive comments of the reviewer regarding the broad interest and strength of our work.

• NL4-3deltaIN and immature HIV virions are found to have less associated gp120 relative to wild-type particles. It is not obvious why this is the case for the deltaIN particles or genetically immature particles. Can the authors provide possible explanations? (A prior paper was cited, Chojnacki et al Science, 2012 but can the current authors provide their own interpretation.)*

Our conclusion from the data is actually exactly the opposite. As shown in Figure 2D, the gp120 staining intensity was higher for NL4-3DIN particles (1,786 a.u.) than for gp120-VLPs (1,223 a.u.), indicating lower expression of Env proteins in the latter. Furthermore, analysis of gp120 intensity per particle (Figure 2E) confirmed that gp120-VLPs contained fewer gp120 molecules per particle than NL4-3DIN virions. These levels were comparable with, or even lower than, those observed in primary HIV-1 viruses (Zhu et al. Nature, 2006). This reduction was a direct consequence of the method used to generate the VLPs, as our goal was to produce viral particles with minimal gp120 content to prevent artifacts in receptor clustering that might occur using high levels of Env proteins in the VLPs to activate the receptors.

This misunderstanding may arise from the fact that we also compared Gag condensation and Env distribution on the surface of gp120-VLPs with those observed in genetically immature particles and integrase-defective NL4-3ΔIN virions, which served as controls. STED microscopy data revealed differences in Env distribution between gp120-VLPs and NL4-3ΔIN virions, supporting the classification of gp120-VLPs as mature particles (Figure 2 A,B).

Reviewer #3

We thank the reviewer for considering that our work offers new insights into the spatial organization of receptors during HIV-1 entry and infection and that the manuscript is well written, and the findings significant.

• For mechanistic basis of gp120-CXCR4 versus CXCL12-CXCR4 differences. Provide additional structural or biochemical evidence to support the claim that gp120 stabilises a distinct CXCR4 conformation compared to CXCL12. If feasible, include molecular modelling, mutagenesis, or cross-linking experiments to corroborate the proposed conformational differences.*

We appreciate the opportunity to clarify this point. The specific claim that gp120 stabilizes a conformation of CXCR4 that is distinct from the CXCL12-bound state was not explicitly stated in our manuscript, although we agree that our data strongly support this possibility. It is important to consider that CXCL12 binds directly to CXCR4, whereas gp120 requires prior sequential binding to CD4, and its subsequent interaction is with a CXCR4 molecule that is already forming part of the CD4/CXCR4 complex, as demonstrated by our FRET experiments and supported by previous studies (Zaitseva et al. J. Leuk. Biol., 2005; Busillo & Benovic Biochim. Biophys. Acta, 2007; Martínez-Muñoz et al. PNAS, 2014). This difference makes it inherently complex to compare the conformational changes induced by gp120 and CXCL12 on CXCR4.

However, our findings show that both stimuli induce oligomerization of CXCR4, a phenomenon not observed when mutant CXCR4R334X was exposed to the chemokine CXCL12 (García-Cuesta et al. PNAS, 2022).

1. CXCL12 induced oligomerization of CXCR4 but did not affect the dynamics of CXCR4R334X (Martinez-Muñoz et al. Mol. Cell, 2018; García-Cuesta et al. PNAS, 2022). By contrast, X4-gp120 and the corresponding VLPs-which require initial binding to CD4 to engage the chemokine receptor-stabilized oligomers of both CXCR4 and CXCR4R334X.

FRET analysis revealed distinct FRET50 values for CD4/CXCR4 (2.713) and CD4/CXCR4R334X (0.399) complexes, suggesting different conformations for each complex. Consistent with previous reports (Balabanian et al. Blood, 2005; Zmajkovicova et al. Front. Immunol., 2024; García-Cuesta et al. PNAS, 2022), the molecular mechanisms activated by CXCL12 are distinct when comparing CXCR4 with CXCR4R334X. For instance, CXCL12 induces internalization of CXCR4, but not of mutant CXCR4R334X. Conversely, X4-gp120 triggers approximately 25% internalization of both receptors. Similarly, CXCL12 does not promote CD4 internalization in cells co-expressing CXCR4 or CXCR4R334X, whereas X4-gp120 does, although CD4 internalization was significantly higher in cells co-expressing CXCR4.

These findings suggest that CD4 influences the conformation and the oligomerization state of both co-receptors. To further support this hypothesis, we have conducted new in silico molecular modeling of CD4 in complex with either CXCR4 or its mutant CXCR4R334X using AlphaFold 3.0 (Abramson et al. Nature, 2024). The server was provided with both sequences, and the interaction between the two molecules for each protein was requested. It produced a number of solutions, which were then analyzed using the software ChimeraX 1.10 (Meng et al. Protein Sci., 2023). CXCR4 and its mutant, CXCR4R334X bound to CD4, were superposed using one of the CD4 molecules from each complex, with the aim of comparing the spatial positioning of CD4 molecules when interacting with CXCR4.

As illustrated in Figure V (only for review purposes), the superposition of the CD4/CXCR4 complexes was complete. However, when CD4/CXCR4 complexes were superimposed with CD4/CXCR4R334X complexes using the same CD4 molecule as a reference, indicated by an arrow in the figure, a clear structural deviation became evident. The main structural difference detected was the positioning of the CD4 transmembrane domains when interacting with either the wild-type or mutant CXCR4. While in complexes with CXCR4, the angle formed by the lines connecting residues E416 at the C-terminus end of CD4 with N196 in CXCR4 was 12{degree sign}, for the CXCR4R334X complex, this angle increased to 24{degree sign}, resulting in a distinct orientation of the CD4 extracellular domain (Figure VI, only for review purposes).

To further analyze the models obtained, we employed PDBsum software (Laskowski & Thornton Protein Sci., 2021) to predict the CD4/CXCR4 interface residues. Data indicated that at least 50% of the interaction residues differed when the CD4/CXCR4 interaction surface was compared with that of the CD4/CXCR4R334X complex (Figure VII, only for review purposes). It is important to note that while some hydrogen bonds were present in both complex models, others were exclusive to one of them. For instance, whereas Cys394(CD4)-Tyr139 and Lys299(CD4)-Glu272 were present in both CD4/CXCR4 and CD4/CXCR4R334X complexes, the pairs Asn337(CD4)-Ser27(CXCR4R334X) and Lys325(CD4)-Asp26(CXCR4R334X) were only found in CD4/CXCR4R334X complexes.

These findings, which are consistent with our FRET results, suggest distinct interaction surfaces between CD4 and the two chemokine receptors. Overall, these results are compatible with differences in the spatial conformation adopted by these complexes.

• For Empty VLP effects on CXCR4 dynamics: Explore potential causes for the observed effects of Env-deficient VLPs. It's valuable to include additional controls such as particles from non-producer cells, lipid composition analysis, or blocking experiments to assess nonspecific interactions. *

As VLPs are complex entities, we thought that the relevant results should be obtained comparing the effects of Env(-) VLPs with gp120-VLPs. Therefore, we would first remark that regardless of the effect of Env(-) VLPs on CXCR4 dynamics, the most evident finding in this study is the strong effect of gp120-VLPs compared with control Env(-) VLPs. Nevertheless, regarding the effect of the Env(-) VLPs compared with medium, we propose several hypotheses. As several virions can be tethered to the cell surface via glycosaminoglycans (GAGs), we hypothesized that VLPs-GAGs interactions might indirectly influence the dynamics of CXCR4 and CXCR4R334X at the plasma membrane. Additionally, membrane fluidity is essential for receptor dynamics, therefore VLPs interactions with proteins, lipids or any other component of the cell membrane could also alter receptor behavior. It is well known that lipid rafts participate in the interaction of different viruses with target cells (Nayak & Hu Subcell. Biochem., 2004; Manes et al. Nat. Rev. Immunol., 2003; Rioethmullwer et al. Biochim. Biophys. Acta, 2006) and both the lipid composition and the presence of co-expressed proteins modulate ligand-mediated receptor oligomerization (Gardeta et al. Frontiers in Immunol., 2022; Gardeta et al. Cell. Commun. Signal., 2025). We have thus performed Raster Image Correlation Spectroscopy (RICS) analysis to assess membrane fluidity through membrane diffusion measurements on cells treated with Env(-) VLPs.

Jurkat cells were labeled with Di-4-ANEPPDHG and seeded on FN and on FN + VLPs prior to analysis by RICS on confocal microscopy. The results indicated no significant differences in membrane diffusion under the treatment tested, thereby discarding an effect of VLPs on overall membrane fluidity (Figure VIII, only for review purposes).

Nonetheless, these results do not rule out other non-specific interactions of Env(-) VLPs with membrane proteins that could affect receptor dynamics. For instance, it has been reported that C-type lectin DC-SIGN acts as an efficient docking site for HIV-1 (Cambi et al. J. Cell. Biol., 2004; Wu & KewalRamani Nat. Rev. Immunol., 2006). However, a detailed investigation of these possible mechanisms is beyond the scope of this manuscript.

• For Direct link between clustering and infection efficiency - Test whether disruption of CXCR4 clustering (e.g., using actin cytoskeleton inhibitors, membrane lipid perturbants, or clustering-deficient mutants) alters HIV-1 fusion or infection efficiency*.

Designing experiments using tools that disrupt receptor clustering by interacting with the receptors themselves is difficult and challenging, as these tools bind the receptor and can therefore alter parameters such as its conformation and/or its distribution at the cell membrane, as well as affect some cellular processes such as HIV-1 attachment and cell entry. Moreover, effects on actin polymerization or lipids dynamics can affect not only receptor clustering but also impact on other molecular mechanisms essential for efficient infection.

Many previous reports have, nonetheless, indirectly correlated receptor clustering with cell infection efficiency. Cholesterol plays a key role in the entry of several viruses. Its depletion in primary cells and cell lines has been shown to confer strong resistance to HIV-1-mediated syncytium formation and infection by both CXCR4- and CCR5-tropic viruses (Liao et al. AIDS Res. Hum. Retrovisruses, 2021). Moderate cholesterol depletion also reduces CXCL12-induced CXCR4 oligomerization and alters receptor dynamics (Gardeta et al. Cell. Commun. Signal., 2025). By restricting the lateral diffusion of CD4, sphingomyelinase treatment inhibits HIV-1 fusion (Finnegan et al. J. Virol., 2007). Depletion of sphingomyelins also disrupts CXCL12-mediated CXCR4 oligomerization and its lateral diffusion (Gardeta et al. Front Immunol., 2022). Additional reports highlight the role of actin polymerization at the viral entry site, which facilitates clustering of HIV-1 receptors, a crucial step for membrane fusion (Serrano et al. Biol. Cell., 2023). Blockade of actin dynamics by Latrunculin A treatment, a drug that sequesters actin monomers and prevents its polymerization, blocks CXCL12-induced CXCR4 dynamics and oligomerization (Martínez-Muñoz et al. Mol. Cell, 2018).

Altogether, these findings strongly support our hypothesis of a direct link between CXCR4 clustering and the efficiency of HIV-1 infection.

• CD4/CXCR4 co-endocytosis hypothesis - Support the proposed model with direct evidence from live-cell imaging or co-localization experiments during viral entry. Clarification is needed on whether internalization is simultaneous or sequential for CD4 and CXCR4.*

When referring to endocytosis of CD4 and CXCR4, we only hypothesized that HIV-1 might promote the internalization of both receptors either sequentially or simultaneously. The hypothesis was based in several findings:

1) Previous studies have suggested that HIV-1 glycoproteins can reduce CD4 and CXCR4 levels during HIV-1 entry (Choi et al. Virol. J., 2008; Geleziunas et al. FASEB J, 1994; Hubert et al. Eur. J. Immunol., 1995).

2) Receptor endocytosis has been proposed as a mechanism for HIV-1 entry (Daecke et al. J. Virol., 2005; Aggarwal et al.Traffick, 2017; Miyauchi et al. Cell, 2009; Carter et al. Virology, 2011).

3) Our data from cells activated with X4-gp120 demonstrated internalization of CD4 and chemokine receptors, which correlated with HIV-1 infection in PBMCs from WHIM patients and healthy donors.

4) CD4 and CXCR4 have been shown to co-localize in lipid rafts during HIV-1 infection (Manes et al. EMBO Rep., 2000; Popik et al. J. Virol., 2002)

5) Our FRET data demonstrated that CD4 and CXCR4 form heterocomplexes and that FRET efficiency increased after gp120-VLPs treatment.

We agree with the reviewer that further experiments are required to test this hypothesis, however, we believe that this is beyond the scope of the current manuscript.

Minor Comments:

• The conclusions rely solely on the HXB2 X4-tropic Env. It would strengthen the study to assess whether other X4 or dual-tropic strains induce similar receptor clustering and dynamics.*

The primary goal of our current study was to investigate the dynamics of the co-receptor CXCR4 during HIV-1 infection, motivated by previous reports showing CD4 oligomerization upon HIV-1 binding and gp120 stimulation (Yuan et al.Viruses, 2021). We initially used a recombinant X4-gp120, a soluble protein that does not fully replicate the functional properties of the native HIV-1 Env. Previous studies have shown that Env consists of gp120 trimers, which redistribute and cluster on the surface of virions following proteolytic Gag cleavage during maturation (Chojnacki et al. Nat. Commun., 2017). An important consideration in receptor oligomerization studies is the concentration of recombinant gp120 used, as it does not accurately reflect the low number of Env trimers present on native HIV-1 particles (Hart et al. J. Histochem. Cytochem., 1993; Zhu et al. Nature, 2006). To address these limitations, we generated virus-like particles (VLPs) containing low levels of X4-gp120 and repeated the dynamic analysis of CXCR4. The use of primary HIV-1 isolates was limited, in this project, to confirm that PBMCs from both healthy donors and WHIM patients were equally susceptible to infection. This result using a primary HIV-1 virus supports the conclusion drawn from our in vitroapproaches. We thus believe that although the use of other X4- and dual-tropic strains may complement and reinforce the analysis, it is far beyond the scope of the current manuscript.

• Given the observed clustering effects, it would be valuable to explore whether gp120-induced rearrangements alter epitope exposure to broadly neutralizing antibodies like 17b or 3BNC117. This would help connect the mechanistic insights to therapeutic relevance.*

As 3BNC117, VRC01 and b12 are broadly neutralizing mAbs that recognize conformational epitopes on gp120 (Li et al. J. Virol., 2011; Mata-Fink et al. J. Mol. Biol., 2013), they will struggle to bind the gp120/CD4/CXCR4 complex and therefore may not be ideal for detecting changes within the CD4/CXCR4 complex. The experiment suggested by the reviewer is thus challenging but also very complex. It would require evaluating antibody binding in two experimental conditions, in the absence and in the presence of oligomers. However, our data indicate that receptor oligomerization is promoted by X4-gp120 binding, and the selected antibodies are neutralizing mAbs, so they should block or hinder the binding of gp120 and, consequently, receptor oligomerization. An alternative approach would be to study the neutralizing capacity of these mAbs on cells expressing CD4/CXCR4 or CD4/CXCR4R334X complexes. Variations in their neutralizing activity could be then extrapolated to distinct gp120 conformations, which in turn may reflect differences between CD4/CXCR4 and CD4/CXCR4R334X complexes.

We thus assessed the ability of the VRC01 and b12, anti-gp120 mAbs, which were available in our laboratory, to neutralize gp120 binding on cells expressing CD4/CXCR4 or CD4/CXCR4R334X. Specifically, increasing concentrations of each antibody were preincubated (60 min, 37ºC) with a fixed amount of X4-gp120 (0.05 mg/ml). The resulting complexes were then incubated with Jurkat cells expressing CD4/CXCR4 or CD4/CXCR4R334X (30 min, 37ºC) and, finally, their binding was analyzed by flow cytometry. Although we did not observe statistically significant differences in the neutralization capacity of b12 or VRC01 for the binding of X4-gp120 depending on the presence of CXCR4 or CXCR4334X, we observed a trend for greater concentrations of both mAbs to neutralize X4-gp120 binding in Jurkat CD4/CXCR4 cells than in Jurkat CD4/CXCR4R334X cells (Figure IX, only for review purposes).

These slight alterations in the neutralizing capacity of b12 and VRC01 mAbs may thus suggest minimal differences in the conformations of gp120 depending of the coreceptor used. We also detected that X4-gp120 and VLPs expressing gp120, which require initial binding to CD4 to engage the chemokine receptor, stabilized oligomers of both CXCR4 and CXCR4R334X, but FRET data indicated distinct FRET50 values between the partners, (2.713) for CD4/CXCR4 and (0.399) for CD4/CXCR4R334X (Figure 5A,B in the main manuscript). Moreover, we also detected significantly more CD4 internalization mediated by X4-gp120 in cells co-expressing CD4 and CXCR4 than in those co-expressing CD4 and CXCR4R334X (Figure 6 in the main manuscript). Overall these latter data and those included in Figures V, VI and VII of this reply, indicate distinct conformations within each receptor complexes.

• TIRF imaging limits analysis to the cell substrate interface. It would be useful to clarify whether CXCR4 receptor clustering occurs elsewhere, such as at immunological synapses or during cell-to-cell contact.*

In recent years, chemokine receptor oligomerization has gained significant research interest due to its role in modulating the ability of cells to sense chemoattractant gradients. This molecular organization is now recognized as a critical factor in governing directed cell migration (Martínez-Muñoz et al. Mol. Cell, 2018; García-Cuesta et al. PNAS, 2022, Hauser et al.Immunity, 2016). In addition, advanced imaging techniques such as single-molecule and super-resolution microscopy have been used to investigate the spatial distribution and dynamic behaviour of CXCR4 within the immunological synapse in T cells (Felce et al. Front. Cell Dev. Biol., 2020). Building on these findings, we are currently conducting a project focused on characterizing CXCR4 clustering specifically within this specialized cellular region.

• In LVP experiments, it would be useful to report transduction efficiency (% GFP+ cells) alongside MSI data to relate VLP infectivity with receptor clustering functionally.*

These experiments were designed to validate the functional integrity of the gp120 conformation on the LVPs, confirming their suitability for subsequent TIRF microscopy. Our objective was to establish a robust experimental tool rather than to perform a high-throughput quantification of transduction efficiency. It is for that reason that these experiments were included in new Supplementary Figure S6, which also contains the complete characterization of gp120-VLPs and LVPs. In such experimental conditions, quantifying the percentage of GFP-positive cells relative to the total number of cells plated in each well is very difficult. However, in line with the reviewer's commentary and as we used the same number of cells in each experimental condition, we have included, in the revised manuscript, a complementary graph illustrating the GFP intensity (arbitrary units) detected in all the wells analyzed (new Supplementary Fig. 6E).

• To ensure that differences in fusion events (Figure 7B) are attributable to target cell receptor properties, consider confirming that effector cells express similar levels of HIV-1 Env. Quantifying gp120 expression by flow cytometry or western blot would rule out the confounding effects of variable Env surface density.*

In these assays (Figure 7B), we used the same effector cells (cells expressing X4-gp120) in both experimental conditions, ensuring that any observed differences should be attributable solely to the target cells, either JKCD4X4 or JKCD4X4R334X. For this reason, in Figure 7A we included only the binding of X4-gp120 to the target cells which demonstrated similar levels of the receptors expressed by the cells.

• HIV-mediated receptor downregulation may occur more slowly than ligand-induced internalization. Including a 24-hour time point would help assess whether gp120 induces delayed CD4 or CXCR4 loss beyond the early effects shown and to better capture potential delayed downregulation induced by gp120.*

The reviewer suggests using a 24-hour time point to facilitate detection of receptor internalization. However, such an extended incubation time may introduce some confounding factors, including receptor degradation, recycling and even de novo synthesis, which could affect the interpretation of the results. Under our experimental conditions, we observed that CXCL12 did not trigger CD4 internalization whereas X4-gp120 did. Interestingly, CD4 internalization depended on the co-receptor expressed by the cells.

• Increase label font size in microscopy panels for improved readability.*

Of course; the font size of these panels has been increased in the revised version.

• Consider adding more references on ligand-induced co-endocytosis of CD4 and chemokine receptors during HIV-1 entry.*

We have added more references to support this hypothesis (Toyoda et al. J. Virol., 2015; Venzke et al. J. Virol., 2006; Gobeil et al J. Virol., 2013).

• For Statistical analysis. Biological replicates are adequate, and statistical tests are generally appropriate. For transparency, report n values, exact p-values, and the statistical test used in every figure legend and discussed in the results.*

Thank you for highlighting the importance of transparency in statistical reporting. We confirm that the n values for all experiments have been included in the figure legends. The statistical tests used for each analysis are also clearly indicated in the figure legends, and the interpretation of these results is discussed in detail in the Results section. Furthermore, the Methods section specifies the tests applied and the thresholds for significance, ensuring full transparency regarding our analytical approach.

In accordance with established conventions in the field, we have utilized categorical significance indicators (e.g., n.s., *, **, ***) within our figures to enhance readability and focus on biological trends. This approach is widely adopted in high-impact literature to prevent visual clutter. However, to ensure full transparency and reproducibility, we have ensured that the underlying statistical tests and thresholds are clearly defined in the respective figure legends and Methods section.

Reviewer #4

We thank the reviewer for considering that this work is presented in a clear fashion, and the main findings are properly highlighted, and for remarking that the paper is of interest to the retrovirology community and possibly to the broader virology community.

We also agree on the interest that X4-gp120 clusters CXCR4R334X suggests a different binding mechanism for X4-gp120 from that of the natural ligand CXCL12, an aspect that we are now evaluating. These data also indicate that WHIM patients can be infected by HIV-1 similarly to healthy people.

• The observation that "empty VLPs" reduce CXCR4 diffusivity is potentially interesting. However, it is not supported by the data owing to insufficient controls. The authors correctly discuss the limitations of that observation in the Discussion section (lines 702-704). However, they overinterpret the observation in the Results section (lines 509-512), suggesting non-specific interactions between empty VLPs, CD4 and CXCR4. I suggest either removing the sentence from the Results section or replacing it with a sentence similar to the one in the Discussion section.*

In accordance with the reviewer`s suggestion, the sentence in the result section has been replaced with one similar to that found in the discussion section. In addition, we have performed Raster Image Correlation Spectroscopy (RICS) analysis using the Di-4-ANEPPDHQ lipid probe to assess membrane fluidity by means of membrane diffusion, and compared the results with those of cells treated with Env(-) VLPs. The results indicated that VLPs did not modulate membrane fluidity (Figure VIII in this reply). Nonetheless, these results do not rule out other potential non-specific interactions of the Env(-) VLPs with other components of the cell membrane that might affect receptor dynamics (see our response to point 2 of reviewer #3 p. 14-15 of this reply).

• In the case of the WHIM mutant CXCR4-R334X, the addition of "empty VLPs" did not cause a significant change in the diffusivity of CXCR4-R334X (Figure 4B). This result is in contrast with the addition of empty VLPs to WT CXCR4. However, the authors neither mention nor comment on that result in the results section. Please mention the result in the paper and comment on it in relation to the addition of empty VLPs to WT CXCR4.*

We would remark that the main observation in these experiments should focus on the effect of gp120-VLPs, and the results indicates that gp120-VLPs promoted clustering of CXCR4 and of CXCR4R334X and reduced their diffusion at the cell membrane. The Env(- ) VLPs were included as a negative control in the experiments, to compare the data with those obtained using gp120-VLPs. However, once we observed some residual effect of the Env(-) VLPs, we decided to give a potential explanation, formulated as a hypothesis, that the Env(-) VLPs modulated membrane fluidity. We have now performed a RICS analysis using Di-4-ANEPPDHQ as a lipid probe (Figure IX only for review purposes). The results suggest that Env(-) VLPs do not modulate cell membrane fluidity, although we do not rule out other potential interactions with membrane proteins that might alter receptor dynamics. We appreciate the reviewer's observation and agree that this result can be noted. However, since the main purpose of Figure 4B is to show that gp120-VLPs modulate the dynamics of CXCR4R334X rather than to remark that the Env(-) VLPs also have some effects, we consider that a detailed discussion of this specific aspect would detract from the central finding and may dilute the primary narrative of the study.

Minor comments

• It would be helpful for the reader to combine thematically or experimentally linked figures, e.g., Figures 3 and 4.*

• Figures 3 and 4 are very similar. Please unify the colours in them and the order of the panels (e.g. Figure 3 panel A shows diffusivity of CXCR4, while Figure 4 panel A shows MSI of CXCR4-R334X).*

While we considered consolidating Figures 3 and 4, we believe that maintaining them as separate entities enhances conceptual clarity. Since Figure 3 establishes the baseline dynamics for wild-type CXCR4 and Figure 4 details the distinct behavior of the CXCR4R334X mutant, keeping them separate allows the reader to fully appreciate the specificities of each system before making a cross-comparison.

• Some parts of the Discussion section could be shortened, moved to the Introduction (e.g.,lines 648-651), or entirely removed (e.g.,lines 633-635 about GPCRs).*

In accordance, the Discussion section has been reorganized and shortened to improve clarity.

• I suggest renaming "empty VLPs" to "Env(−) VLPs" (or similar). The name empty VLPs can mislead the reader into thinking that these are empty vesicles.*

The term empty VLPs has been renamed to Env(−) VLPs throughout the manuscript to more accurately reflect their composition. Many thanks for this suggestion.

• Line 492 - please rephrase "...lower expression of Env..." to "...lower expression of Env or its incorporation into the VLPs...".*

The sentence has been rephrased

• Line 527 - The data on CXCL12 modulating CXCR4-R334X dynamics and clustering are not present in Figure 4 (or any other Figure). Please add them or rephrase the sentence with an appropriate reference. Make clear which results are yours.*

• Line 532 - Do the data in the paper really support a model in which CXCL12 binds to CXCR4-R334X? If not, please rephrase with an appropriate reference.*

Previous studies support the association of CXCL12 with CXCR4R334X (Balabanian et al. Blood, 2005; Hernandez et al. Nat Genet., 2003; Busillo & Benovic Biochim. Biophys. Acta, 2007). In fact, this receptor has been characterized as a gain-of-function variant for this ligand (McDermott et al. J. Cell. Mol. Med., 2011). The revised manuscript now includes these bibliographic references to support this commentary. In any case, our previous data indicate that CXCL12 binding does not affect CXCR4R334X dynamics (García-Cuesta et al. PNAS, 2022).

• Line 695 - "...lipid rafts during HIV-1 (missing word?) and their ability to..." During what?*

Many thanks for catching this mistake. The sentence now reads: "Although direct evidence for the internalization of CD4 and CXCR4 as complexes is lacking, their co-localization in lipid rafts during HIV-1 infection (97-99) and their ability to form heterocomplexes (22) strongly suggest they could be endocytosed together."

Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.

Learn more at Review Commons

Referee #1

Evidence, reproducibility and clarity

The authors investigate the impact of surface bound HIV gp120 and VLPs on CXCR4 dynamics in Jurkat T cells expressing WT or WHIM syndrome mutated CXCR4, which has a defective response to CXCL12. Jurkat cells were transfected with CXCR4-AcGFP. Images were acquired and a single particle tracking routine was applied to generate information about nanoclustering and diffusion, and FRET was used to investigate CD4-CXCR4 proximity. They compare effects of soluble gp120 to immature and mature VLPs, which include varying degrees of gp120 clustering. They find that solid phase gp120 or VLP can increase CXCR4 clustering size and decrease diffusion in Jurkat cells. Surprisingly, VLP lacking gp120 could increase CXCR4 clustering and speed, which is paradoxical as there were no known ligands on the VLPs, but they likely carry many cellular proteins with potential interactions. The impact of CXCL12 and gp120 binding to CXCR4 was different in terms of clustering and receptor down-regulation.

While a single particle tracking routine was applied to the data, it's not clear how the signal from a single GFP was defined and if movement during the 100 ms acquisition time impacts this. My concern would be that the routine is tracking fluctuations, and these are related to single particle dynamics, it appears from the movies that the density or the GFP tagged receptors in the cells is too high to allow clear tracking of single molecules. SPT with GFP is very difficult due to bleaching and relatively low quantum yield. Current efforts in this direction that are more successful include using SNAP tags with very photostable organic fluorophores. The data likely does mean something is happening with the receptor, but they need to be more conservative about the interpretation.

Some of the paradoxical effects might be better understood through deeper analysis of the SPT data, particularly investigation of active transport and more detailed analysis of "immobile" objects. Comments on early figures illustrate how this could be approached. This would require selecting acquisitions where the GFP density is low enough for SPT and performing a more detailed analysis, but this may be difficult to do with GFP.

When the authors discuss clusters of <2 or >3, how do they calibrate the value of GFP and the impact of diffusion on the measurement. One way to approach this might be single molecules measurements of dilute samples on glass vs in a supported lipid bilayer to map the streams of true immobility to diffusion at >1 µm2/sec.

I understand that the CXCL12 or gp120 are attached to the substrate with fibronectin for adhesion. I'm less clear how how that VLPs are integrated. Were these added to cells already attached to FN? Fig 1A- The classification of particle tracks into mobile and immobile is overly simplistic description that goes back to bulk FRAP measurements and it not really applicable to single molecule tracking data, where it's rare to see anything that is immobile and alive. An alternative classification strategy uses sub-diffusion, normal diffusion and active diffusion (or active transport) to descriptions and particles can transition between these classes over the tracking period. Fig 1B- this data might be better displayed as histograms showing distributions within the different movement classes. Fig 1C,D- It would be helpful to see a plot of D vs MSI at a single particle level. In comparing C and D I'm surprised there is not a larger difference between CXCL12 and X4-gp120. It would also be very important to see the behaviour of X4-gp120 on the CXCR4 deficient Jurkat that would provide a picture of CD4 diffusion. The CXCR4 nanoclustering related to the X4-gp120 could be dominated by CD4 behaviour.

Fig S1D- This data is really interesting. However, if both the CD4 and the gp120 have his tags they need to be careful as poly-His tags can bind weakly to cells and increasing valency could generate some background. So, they should make the control is fair here. Ideally, using non-his tagged person of sCD4 and gp120 would be needed ideal or they need a His-tagged Fab binding to gp120 that doesn't induce CXCR4 binding.

Fig S4- Panel D needs a scale bar. I can't figure out what I'm being shown without this.

Significance

The strengths are that its an important question and the reagents are well prepared and characterised. They are detecting quantitative effects that will likely be reproducible. The information generated is potentially useful for those studying HIV infection processes and strategies to prevent infection.

The major weakness is that the conditions for the SPT experiments are not ideal in that the density of particles is too high for SPT and the single molecule basis for assessing nanoclusters is not clear. This means that the data is getting at complex molecules phenomena and less likely be generating pure single molecules measurements.

Reviewer #1 (Public review):

Summary:

The manuscript by Lu and colleagues demonstrates convincingly that PRRT2 interacts with brain voltage-gated sodium channels to enhance slow inactivation in vitro and in vivo. The work is interesting and rigorously conducted. The relevance to normal physiology and disease pathophysiology (e.g., PRRT2-related genetic neurodevelopmental disorders) seems high. Some simple additional experiments could elevate the impact and make the study more complete.

Strengths:

Experiments are conducted rigorously, including experimenter blinding and appropriate controls. Data presentation is excellent and logical. The paper is well written for a general scientific audience.

Weaknesses:

There are a few missing experiments and one place where data are over-interpreted.

(1) An in vitro study of Nav1.6 is conspicuously absent. In addition to being a major brain Na channel, Nav1.6 is predominant in cerebellar Purkinje neurons, which the authors note lack PRRT2 expression. They speculate that the absence of PRRT2 in these neurons facilitates the high firing rate. This hypothesis would be strengthened if PRRT2 also enhanced slow inactivation of Nav1.6. If a stable Nav1.6 cell were not available, then simple transient co-transfection experiments would suffice.

(2) To further demonstrate the physiological impact of enhanced slow inactivation, the authors should consider a simple experiment in the stable cell line experiments (Figure 1) to test pulse frequency dependence of peak Na current. One would predict that PRRT2 expression will potentiate 'run down' of the channels, and this finding would be complementary to the biophysical data.

(3) The study of one K channel is limited, and the conclusion from these experiments represents an over-interpretation. I suggest removing these data unless many more K channels (ideally with measurable proxies for slow inactivation) were tested. These data do not contribute much to the story.

(4) In Figure 2, the authors should confirm that protein is indeed expressed in cells expressing each truncated PRRT2 construct. Absent expression should be ruled out as an explanation for absent enhancement of slow inactivation.

Author response:

Public Reviews:

Reviewer #1 (Public review):

Summary:

The manuscript by Lu and colleagues demonstrates convincingly that PRRT2 interacts with brain voltage-gated sodium channels to enhance slow inactivation in vitro and in vivo. The work is interesting and rigorously conducted. The relevance to normal physiology and disease pathophysiology (e.g., PRRT2-related genetic neurodevelopmental disorders) seems high. Some simple additional experiments could elevate the impact and make the study more complete.

Strengths:

Experiments are conducted rigorously, including experimenter blinding and appropriate controls. Data presentation is excellent and logical. The paper is well written for a general scientific audience.

Weaknesses:

There are a few missing experiments and one place where data are over-interpreted.

(1) An in vitro study of Nav1.6 is conspicuously absent. In addition to being a major brain Na channel, Nav1.6 is predominant in cerebellar Purkinje neurons, which the authors note lack PRRT2 expression. They speculate that the absence of PRRT2 in these neurons facilitates the high firing rate. This hypothesis would be strengthened if PRRT2 also enhanced slow inactivation of Nav1.6. If a stable Nav1.6 cell were not available, then simple transient co-transfection experiments would suffice.

We thank the reviewer for this suggestion. In the revised manuscript, we will examine whether PRRT2 modulates slow inactivation of Nav1.6 channels using heterologous co-expression experiments.

(2) To further demonstrate the physiological impact of enhanced slow inactivation, the authors should consider a simple experiment in the stable cell line experiments (Figure 1) to test pulse frequency dependence of peak Na current. One would predict that PRRT2 expression will potentiate 'run down' of the channels, and this finding would be complementary to the biophysical data.

We agree that examining pulse frequency-dependent changes in peak sodium current would provide a functional readout linking PRRT2-mediated enhancement of slow inactivation to use-dependent channel availability. In the revision, we will include a pulse-train protocol to quantify use-dependent attenuation (“run-down”) of peak sodium current across stimulation trains and will compare this adaptation between control and PRRT2-expressing conditions.

(3) The study of one K channel is limited, and the conclusion from these experiments represents an over-interpretation. I suggest removing these data unless many more K channels (ideally with measurable proxies for slow inactivation) were tested. These data do not contribute much to the story.

We agree with the reviewer’s assessment. To avoid over-interpretation and to maintain focus on PRRT2-dependent regulation of Nav channel slow inactivation, we will remove potassium channel dataset and the associated conclusions from the revised manuscript.

(4) In Figure 2, the authors should confirm that protein is indeed expressed in cells expressing each truncated PRRT2 construct. Absent expression should be ruled out as an explanation for the enhancement of slow inactivation.

We appreciate the reviewer’s concern regarding expression of the truncated PRRT2 constructs in the Nav1.2 stable cell line, particularly PRRT2(1-266), which shows little effect on slow inactivation of Nav1.2 channels. In the revision, we will include expression controls for each truncation construct in the Nav1.2-expressing cells to rule out lack of expression as an explanation for the observed functional differences.

Reviewer #2 (Public review):

Summary:

As a member of DspB subfamily, PRRT2 is primarily expressed in the nervous system and has been associated with various paroxysmal neurological disorders. Previous studies have shown that PRRT2 directly interacts with Nav1.2 and Nav1.6, modulating channel properties and neuronal excitability.

In this study, Lu et al. reported that PRRT2 is a physiological regulator of Nav channel slow inactivation, promoting the development of Nav slow inactivation and impeding the recovery from slow inactivation. This effect can be replicated by the C-terminal region (256-346) of PRRT2, and is highly conserved across species from zebrafish, mouse, to human PRRT2. TRARG1 and TMEM233, the other two DspB family members, showed similar effects on Nav1.2 slow inactivation. Co-IP data confirms the interaction between Nav channels and PRRT2. Prrt2-mutant mice, which lack PRRT2 expression, require lower stimulation thresholds for evoking after-discharges when compared to WT mice.

Strengths:

(1) This study is well designed, and data support the conclusion that PRRT2 is a potent regulator of slow inactivation of Nav channels.

(2) This study reveals similar effects on Nav1.2 slow inactivation by PRRT2, TMEM233, and TRARG1, indicating a common regulation of Nav channels by DspB family members (Supplemental Figure 2). A recent study has shown that TMEM233 is essential for ExTxA (a plant toxin)-mediated inhibition on fast inactivation of Nav channels; and PRRT2 and TRARG1 could replicate this effect (Jami S, et al. Nat Commun 2023). It is possible that all three DspB members regulate Nav channel properties through the same mechanism, and exploring molecules that target PRRT2/TRARG1/TMEM233 might be a novel strategy for developing new treatments of DspB-related neurological diseases.

Weaknesses:

(1) Previously, the authors have reported that PRRT2 reduces Nav1.2 current density and alters biophysical properties of both Nav1.2 and Nav1.6 channels, including enhanced steady-state inactivation, slower recovery, and stronger use-dependent inhibition (Lu B, et al. Cell Rep 2021, Fig 3 & S5). All those changes are expected to alter neuronal excitability and should be discussed.

We agree that PRRT2 has been reported to exert multiple effects on Nav channels which are all expected to influence neuronal excitability (Fruscione et al., 2018; Lu et al., 2021; Valente et al., 2023). In the revised manuscript, we will expand the Discussion to integrate these prior findings and to clarify how these PRRT2-dependent changes may interact with (and potentially converge on) modulation of slow inactivation to shape neuronal excitability.

(2) In this study, the fast inactivation kinetics was examined by a single stimulus at 0 mV, which may not be sufficient for the conclusion. Inactivation kinetics at more voltage potentials should be added.

We thank the reviewer for this suggestion. In the revision, we will extend our analysis of Nav1.2 fast-inactivation kinetics across a range of test potentials (e.g., -20, -10, 0, +10 and +20 mV) in the presence and absence of PRRT2.

(3) It is a little surprising that there is no difference in Nav1.2 current density in axon-blebs between WT and Prrt2-mutant mice (Figure 7B). PRRT2 significantly shifts steady-state slow inactivation curve to hyperpolarizing direction, at -70 mV, nearly 70% of Nav1.2 channels are inactivated by slow inactivation in cells expressing PRRT2 when compared to less than 10% in cells expressing GFP (Figure supplement 1B); with a holding potential of -70 mV, I would expect that most of Nav channels are inactivated in axon-blebs from WT mice but not in axon-blebs from Prrt2-mutant mice, and therefore sodium current density should be different in Figure 7B, which was not. Any explanation?

We appreciate the reviewer for raising this point. In our axonal bleb recordings, although the holding potential was -70 mV, sodium current density was measured after a hyperpolarizing pre-pulse (-110 mV) to relieve inactivation immediately prior to the test depolarization (as described in the Methods). Thus, the current density measurement in Figure 7B reflects the maximal available current following this recovery step, rather than the steady-state availability at -70 mV. In the revision, we will state this explicitly in the Results and/or figure legend to avoid confusion.

(4) Besides Nav channels, PRRT2 has been shown to act on Cav2.1 channels as well as molecules involved in neurotransmitter release, which may also contribute to abnormal neuronal activity in Prrt2-mutant mice. These should be mentioned when discussing PRRT2's role in neuronal resilience.

We agree with the reviewer. In the revised manuscript, we will broaden the Discussion to acknowledge PRRT2 functions beyond Nav channels, including reported roles in Cav2.1 regulation and neurotransmitter release. We will frame the in vivo phenotypes in Prrt2-mutant mice as likely arising from convergent mechanisms—altered intrinsic excitability together with changes in synaptic transmission.

Reviewer #3 (Public review):

This paper reveals that the neuronal protein PRRT2, previously known for its association with paroxysmal dyskinesia and infantile seizures, modulates the slow inactivation of voltage-gated sodium ion (Nav) channels, a gating process that limits excitability during prolonged activity. Using electrophysiology, molecular biology, and mouse models, the authors show that PRRT2 accelerates entry of Nav channels into the slow-inactivated state and slows their recovery, effectively dampening excessive excitability. The effect seems evolutionarily conserved, requires the C-terminal region of PRRT2, and is recapitulated in cortical neurons, where PRRT2 deficiency leads to hyper-responsiveness and reduced cortical resilience in vivo. These findings extend the functional repertoire of PRRT2, identifying it as a physiological brake on neuronal excitability. The work provides a mechanistic link between PRRT2 mutations and episodic neurological phenotypes.

Comments:

(1) The precise structural interface and the molecular basis of gating modulation remain inferred rather than demonstrated.

We thank the reviewer for this comment. In the revision, we will make it explicit that our structural modeling are based on prediction rather than evidential. We will also expand the Limitations section to highlight that direct structural and biochemical mapping of the PRRT2-Nav interface (e.g., through targeted mutagenesis, crosslinking, and/or structural determination) will be required to define the binding interface and establish the molecular basis of gating modulation.

(2) The in vivo phenotype reflects a complex circuit outcome and does not isolate slow-inactivation defects per se.

We agree with the reviewer. In the revision, we will refine the Discussion to avoid over-attributing the in vivo phenotype to slow-inactivation defects alone and to explicitly state that impaired slow inactivation in Prrt2-mutant mice represents one plausible contributing mechanism to reduced cortical resilience, alongside other PRRT2-dependent process.

(3) Expression of PRRT2 in muscle or heart is low, so the cross-isoform claims are likely of limited physiological significance.

We thank the review for your comment about physiological relevance. In the revised manuscript, we will clarify that our Nav isoform panel was designed to assess mechanistic generality at the channel level rather than to imply broad in vivo relevance across tissues. We will also expand the Discussion to emphasize that any therapeutic strategy involving PRRT2 delivery should consider its consistent effect on slow inactivation across multiple Nav isoforms.

(4) The mechanistic separation between the trafficking effect of PRRT2 and its gating effects is not clearly resolved.

We appreciate the reviewer for raising this important point. In the revision, we will expand the Discussion to clarify why we interpret the effect of PRRT2 on slow inactivation as a gating modulation rather than a secondary consequence of altered channel abundance or localization. First, our slow inactivation measurements are expressed as the fraction of available channels after depolarization conditioning relative to baseline availability within the same cell (post-/pre-conditioning), which minimizes confounding by differences in initial surface expression. Second, the slow inactivation of Nav channel occurs on a rapid, activity-dependent timescale (seconds), whereas remarkable changes in trafficking and surface abundance generally develop over longer intervals (minutes to hours).

(5) Additional studies with Nav1.6 should be carried out.

We thank the reviewer’s suggestion. We will include Nav1.6 slow inactivation experiments in the revised manuscript.

DOI: 10.1016/j.xcrm.2025.102510

Resource: (Proteintech Cat# 66006-2-Ig, RRID:AB_2881490)

Curator: @scibot

SciCrunch record: RRID:AB_2881490

What is this?

DOI: 10.1016/j.xcrm.2025.102478

Resource: (Cell Signaling Technology Cat# 58981, RRID:AB_2732809)

Curator: @scibot

SciCrunch record: RRID:AB_2732809

What is this?

Nejvyšší kvalita bez kompromisů Na naše výrobky se můžete plně spolehnout i za nepříznivého počasí.

Bezkonkurenční záruka Na stanové konstrukce poskytujeme záruku až 5 let. I po jejím uplynutí držíme náhradní díly skladem.

Proč zvolit MITKO? +CZ text in video Vyrábíme reklamní stany a vybavení, které přitahují zákazníky. Hlavní důraz je kladen na odolnost, spolehlivost a komfortní užívání. Naše výrobky nejsou na jednu sezónu, ale budou Vás provázet na akcích mnoho let. (next paragraph) Veškeré naše výrobky můžete potisknout libovolným designem! Využívám digitální metody potisku s UV odolností barev. (next paragraph) Jako jediní na CZ a SK trhu vyrábíme vybrané typy stanů s odolností vůči větru až do 100 km/h, což máme podloženo statickými výpočty.

Vyrábíme reklamní stany a eventové doplňky s důrazem na odolnost a spolehlivost. Skupina MITKO je přímým výrobcem všech nabízených produktů! Výrobní proces máme pod kontrolou od A do Z a nečekáme tak na dodávky subdodavatelů. (next paragraph)Na naše výrobky a celkovou spolupráci se spolehají velké nadnárodní společnosti, ale i střední firmy a mikrofirmy.

Pozáruční servis

I po skončení záruky provedeme jeho kompletní prohlídku a zkontrolujeme, které části se opotřebovaly a vyžadují výměnu. Díly pak po odsouhlasení ceny vyměníme a celý stan Vám pošleme zpět - připravený do akce!

dílů svépomocí

Creo que hace falta más hincapié en el aspecto regresivo y hacen falta más que comillas para desarticular la potencia de la palabra "modernización" que, no por nada, es la elegida para presentar esta reforma. Sugiero no usar la palabra modernización a menos que este acompañada por nuestra propia definición que podría ser, simplemente, "retroceso" laboral.

Si es la conclusión final de todo el análisis debería estar desatacado. De suerte lo vi :P

Al mostrar estos números, entiendo que lo que están suponiendo es que cada año se van a despedir casi 1 millón de trabajadores con antigüedad. De manera similar, en el siguiente punto, suponen que cada año se van a desvincular a 1 millón y medio de trabajadores en periodo de prueba. Por supuesto que en los casos en los que haya despidos significa una transferencia de ingresos y un ataque grave al ingreso, y aun si no los hubiera, la amenaza de que podría haberlos es un golpe a la capacidad organizativa de los trabajadores. Pero si están haciendo lo que creo que están haciendo, no sería así como debería calcularse.

Para que se den una idea, un informe retomado en esta nota calcula 200.000 despidos en los primeros 5 meses de la era Milei https://www.ambito.com/economia/advierten-que-hubo-casi-200000-despidos-los-primeros-5-meses-la-era-javier-milei-n6036930

Ya es malo de por si, pero los números que ustedes presentan exceden esto por muchísimo. Quizás me equivoco, pero me parece que están tomando los totales de trabajadores que pertenecen a ciertas categorías, cuando en estos casos tendrían que tomar números de despidos reales y proyectarlos para 2025.

Quizás si incluyesen un apartado metodológico explicando de dónde sacan esos números se podría entender mejor

Il manque le " ² " sur le oméga_0 dans la fonction

Alternativa: https://web.archive.org/web/20220708064806if_/https://otca.org/wp-content/uploads/2021/02/Pueblos-Indigenas-de-la-Amazonia-Boliviana.pdf

Rótulo da Taxa Pico deve aparecer abaixo. Bola azul deveria ser um triângulo. Rótulo da bola vermelha não deveria tocar na bola.

Retirar casa decimal da variação

Reviewer #2 (Public review):

Summary:

This manuscript investigates how olfactory representations are transformed along the cortico-hippocampal pathway in mice during a non-associative learning paradigm involving novel and familiar odors. By recording single-unit activity in several key brain regions (AON, aPCx, LEC, CA1, and SUB), the authors aim to elucidate how stimulus identity and experience are encoded and how these representations change across the pathway.

The study addresses an important question in sensory neuroscience regarding the interplay between sensory processing and signaling novelty/familiarity. It provides insights into how the brain processes and retains sensory experiences, suggesting that the earlier stations in the olfactory pathway, the AON aPCx, play a central role in detecting novelty and encoding odor, while areas deeper into the pathway (LEC, CA1 & Sub) are more sparse and encodes odor identity but not novelty/familiarity. However, there are several concerns related to methodology, data interpretation, and the strength of the conclusions drawn.

Strengths:

The authors combine the use of modern tools to obtain high-density recordings from large populations of neurons at different stages of the olfactory system (although mostly one region at a time) with elegant data analyses to study an important and interesting question.

Weaknesses:

The first and biggest problem I have with this paper is that it is very confusing, and the results seem to be all over the place. In some parts, it seems like the AON and aPCx are more sensitive to novelty; in others, it seems the other way around. I find their metrics confusing and unconvincing. For example, the example cells in Figure 1C shows an AON neuron with a very low spontaneous firing rate and a CA1 with a much higher firing rate, but the opposite is true in Fig. 2A. So, what are we to make of Fig. 2C that shows the difference in firing rates between novel vs. familiar odors measured as a difference in spikes/sec. The meaning of this is unclear. The authors could have used a difference in Z-scored responses to normalize different baseline activity levels. (This is just one example of a problem with the methodology.)

There are a lot of high-level data analyses (e.g., decoding, analyzing decoding errors, calculating mutual information, calculating distances in state space, etc.) but very little neural data (except for Fig. 2C, and see my comment above about how this is flawed). So, if responses to novel vs. familiar odors are different in the AON and aPCx, how are they different? Why is decoding accuracy better for novel odors in CA1 but better for familiar odors in SUB (Fig. 3A)? The authors identify a small subset of neurons that have unusually high weights in the SVM analyses that contribute to decoding novelty, but they don't tell us which neurons these are and how they are responding differently to novel vs. familiar odors.

The authors call AON and aPCx "primary sensory cortices" and LEC, CA1, and Sub "multisensory areas". This is a straw man argument. For example, we now know that PCx encodes multimodal signals (Poo et al. 2021, Federman et al., 2024; Kehl et al., 2024), and LEC receives direct OB inputs, which has traditionally been the criterion for being considered a "primary olfactory cortical area". So, this terminology is outdated and wrong, and although it suits the authors' needs here in drawing distinctions, it is simplistic and not helpful moving forward.

Why not simply report z-scored firing rates for all neurons as a function of trial number? (e.g., Jacobson & Friedrich, 2018). Fig. 2C is not sufficient. For example, in the Discussion, they say, "novel stimuli caused larger increases in firing rates than familiar stimuli" (L. 270), but what does this mean? Odors typically increase the firing in some neurons and suppress firing in others. Where does the delta come from? Is this because novel odors more strongly activate neurons that increase their firing or because familiar odors more strongly suppress neurons?

Ls. 122-124 - If cells in AON and aPCx responded the same way to novel and familiar odors, then we would say that they only encode for odor and not at all for experience. So, I don't understand why the authors say these areas code for a "mixed representation of chemical identity and experience." "On the other hand," if LEC, CA1, and SUB are odor selective and only encode novel odors, then these areas, not AON and aPCx, are the jointly encoding chemical identity and experience. Also, I do not understand why, here, they say that AON and PCx respond to both while LEC, CA1, and SUB were selective for novel stimuli, but the authors then go on to argue that novelty is encoded in the AON and PCx, but not in the LEC, CA1, and SUB.

Ls. 132-140 - As presented in the text and the figure, this section is unclear and confusing. Their use of the word "shuffled" is a major source of this confusion, because this typically is the control that produces outcomes at chance level. More importantly, it seems as though they did the wrong analysis here. A better way to do this analysis is to train on some of the odors and test on an untrained odor (i.e., what Bernardi et al., 2021 called "cross-condition generalization performance"; CCGP).

Comments on revisions:

I think the authors have done an adequate job addressing the reviewers' concerns. Most importantly, I found the first version of the manuscript quite confusing, and the consequent clarifications have addressed this issue.

In several cases, I see their point, while I still disagree with whether they made the best decisions. However, the issues here do not fundamentally change the big-picture outcome, and if they want to dig in with their approaches (e.g., only using auROC or just reporting delta firing rates without any normalization), it's their choice.

Author response:

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

Public reviews:

Reviewer #1 (Public review):

In this important study, the authors characterized the transformation of neural representations of olfactory stimuli from the primary sensory cortex to multisensory regions in the medial temporal lobe and investigated how they were affected by non-associative learning. The authors used high-density silicon probe recordings from five different cortical regions while familiar vs. novel odors were presented to a head-restrained mouse. This is a timely study because unlike other sensory systems (e.g., vision), the progressive transformation of olfactory information is still poorly understood. The authors report that both odor identity and experience are encoded by all of these five cortical areas but nonetheless some themes emerge. Single neuron tuning of odor identity is broad in the sensory cortices but becomes narrowly tuned in hippocampal regions. Furthermore, while experience affects neuronal response magnitudes in early sensory cortices, it changes the proportion of active neurons in hippocampal regions. Thus, this study is an important step forward in the ongoing quest to understand how olfactory information is progressively transformed along the olfactory pathway.

The study is well-executed. The direct comparison of neuronal representations from five different brain regions is impressive. Conclusions are based on single neuronal level as well as population level decoding analyses. Among all the reported results, one stands out for being remarkably robust. The authors show that the anterior olfactory nucleus (AON), which receives direct input from the olfactory bulb output neurons, was far superior at decoding odor identity as well as novelty compared to all the other brain regions. This is perhaps surprising because the other primary sensory region - the piriform cortex - has been thought to be the canonical site for representing odor identity. A vast majority of studies have focused on aPCx, but direct comparisons between odor coding in the AON and aPCx are rare. The experimental design of this current study allowed the authors to do so and the AON was found to convincingly outperform aPCx. Although this result goes against the canonical model, it is consistent with a few recent studies including one that predicted this outcome based on anatomical and functional comparisons between the AON-projecting tufted cells vs. the aPCx-projecting mitral cells in the olfactory bulb (Chae, Banerjee et. al. 2022). Future experiments are needed to probe the circuit mechanisms that generate this important difference between the two primary olfactory cortices as well as their potential causal roles in odor identification.

The authors were also interested in how familiarity vs. novelty affects neuronal representation across all these brain regions. One weakness of this study is that neuronal responses were not measured during the process of habituation. Neuronal responses were measured after four days of daily exposure to a few odors (familiar) and then some other novel odors were introduced. This creates a confound because the novel vs. familiar stimuli are different odorants and that itself can lead to drastic differences in evoked neural responses. Although the authors try to rule out this confound by doing a clever decoding and Euclidian distance analysis, an alternate more straightforward strategy would have been to measure neuronal activity for each odorant during the process of habituation.

Reviewer #2 (Public review):

This manuscript investigates how olfactory representations are transformed along the cortico-hippocampal pathway in mice during a non-associative learning paradigm involving novel and familiar odors. By recording single-unit activity in several key brain regions (AON, aPCx, LEC, CA1, and SUB), the authors aim to elucidate how stimulus identity and experience are encoded and how these representations change across the pathway.

The study addresses an important question in sensory neuroscience regarding the interplay between sensory processing and signaling novelty/familiarity. It provides insights into how the brain processes and retains sensory experiences, suggesting that the earlier stations in the olfactory pathway, the AON aPCx, play a central role in detecting novelty and encoding odor, while areas deeper into the pathway (LEC, CA1 & Sub) are more sparse and encodes odor identity but not novelty/familiarity. However, there are several concerns related to methodology, data interpretation, and the strength of the conclusions drawn.

Strengths:

The authors combine the use of modern tools to obtain high-density recordings from large populations of neurons at different stages of the olfactory system (although mostly one region at a time) with elegant data analyses to study an important and interesting question.

Weaknesses:

(1) The first and biggest problem I have with this paper is that it is very confusing, and the results seem to be all over the place. In some parts, it seems like the AON and aPCx are more sensitive to novelty; in others, it seems the other way around. I find their metrics confusing and unconvincing. For example, the example cells in Figure 1C show an AON neuron with a very low spontaneous firing rate and a CA1 with a much higher firing rate, but the opposite is true in Figure 2A. So, what are we to make of Figure 2C that shows the difference in firing rates between novel vs. familiar odors measured as a difference in spikes/sec. This seems nearly meaningless. The authors could have used a difference in Z-scored responses to normalize different baseline activity levels. (This is just one example of a problem with the methodology.)

We appreciate the reviewer’s concerns regarding clarity and methodology. It is less clear why all neurons in a given brain area should have similar firing rates. Anatomically defined brain areas typically comprise of multiple cell types, which can have diverse baseline firing rates. Since we computed absolute firing rate differences per neuron (i.e., novel vs. familiar odor responses within the same neuron), baseline differences across neurons do not have a major impact.

The suggestion to use Z-scores instead of absolute firing rate differences is well taken. However, Z-scoring assumes that the underlying data are normally distributed, which is not the case in our dataset. Specifically, when analyzing odor-evoked firing rates on a per-neuron basis, only 4% of neurons exhibit a normal distribution. In cases of skewed distributions, Z-scoring can distort the data by exaggerating small variations, leading to misleading conclusions. We acknowledge that different analysis methods exist, we believe that our chosen approach best reflects the properties of the dataset and avoids potential misinterpretations introduced by inappropriate normalization techniques.

(2) There are a lot of high-level data analyses (e.g., decoding, analyzing decoding errors, calculating mutual information, calculating distances in state space, etc.) but very little neural data (except for Figure 2C, and see my comment above about how this is flawed). So, if responses to novel vs. familiar odors are different in the AON and aPCx, how are they different? Why is decoding accuracy better for novel odors in CA1 but better for familiar odors in SUB (Figure 3A)? The authors identify a small subset of neurons that have unusually high weights in the SVM analyses that contribute to decoding novelty, but they don't tell us which neurons these are and how they are responding differently to novel vs. familiar odors.

We performed additional analyses to address the reviewer’s feedback (Figures 2C-E and lines 118-132) and added more single-neuron data (Figures 1, S3 and S4).

(3) The authors call AON and aPCx "primary sensory cortices" and LEC, CA1, and Sub "multisensory areas". This is a straw man argument. For example, we now know that PCx encodes multimodal signals (Poo et al. 2021, Federman et al., 2024; Kehl et al., 2024), and LEC receives direct OB inputs, which has traditionally been the criterion for being considered a "primary olfactory cortical area". So, this terminology is outdated and wrong, and although it suits the authors' needs here in drawing distinctions, it is simplistic and not helpful moving forward.

We appreciate the reviewer’s concern regarding the classification of brain regions as “primary sensory” versus “multisensory.” Of note, the cited studies (Poo et al., 2021; Federman et al., 2024; Kehl et al., 2024) focus on posterior PCx (pPCx), while our recordings were conducted in very anterior section of anterior PCx. The aPCx and pPCx have distinct patterns of connectivity, both anatomically and functionally. To the best of our knowledge, there is no evidence for multimodal responses in aPCx, whereas there is for LEC, CA1 and SUB. Furthermore, our distinction is not based on a connectivity argument, as the reviewer suggests, but on differences in the α-Poisson ratio (Figure 1E and F).

To avoid confusion due to definitions of what constitutes a “primary sensory” region, we adopted a more neutral description throughout the manuscript.

(4) Why not simply report z-scored firing rates for all neurons as a function of trial number? (e.g., Jacobson & Friedrich, 2018). Figure 2C is not sufficient.

Regarding z-scores, please see response to 1). We further added a figure showing responses of all neurons to novel stimuli (using ROC instead of z-scoring, as described previously (e.g. Cohen et al. Nature 2012). We added the following figure to the supplementary for the completeness of the analysis (S2E).

For example, in the Discussion, they say, "novel stimuli caused larger increases in firing rates than familiar stimuli" (L. 270), but what does this mean?

This means that on average, the population of neurons exhibit higher firing rates in response to novel odors compared to familiar ones.

Odors typically increase the firing in some neurons and suppress firing in others. Where does the delta come from? Is this because novel odors more strongly activate neurons that increase their firing or because familiar odors more strongly suppress neurons?

We thank the reviewer for this valuable feedback and extended the characterization of firing rate properties, including a separate analysis of neurons i) significantly excited by odorants, ii) significantly inhibited by odorants and iii) not responsive to odorants. We added the analysis and corresponding discussion to the main manuscript (Figures 2C-E and lines 118-132)

(5) Lines 122-124 - If cells in AON and aPCx responded the same way to novel and familiar odors, then we would say that they only encode for odor and not at all for experience. So, I don't understand why the authors say these areas code for a "mixed representation of chemical identity and experience." "On the other hand," if LEC, CA1, and SUB are odor selective and only encode novel odors, then these areas, not AON and aPCx, are the jointly encoding chemical identity and experience. Also, I do not understand why, here, they say that AON and PCx respond to both while LEC, CA1, and SUB were selective for novel stimuli, but the authors then go on to argue that novelty is encoded in the AON and PCx, but not in the LEC, CA1, and SUB.

We appreciate the reviewer’s request for clarification. Throughout the brain areas we studied, odorant identity and experience can be decoded. However, the way information is represented is different between regions. We acknowledge that that “mixed” representation is a misleading term and removed it from the manuscript.

In AON and aPCx, neurons significantly respond to both novel and familiar odors. However, the magnitude of their responses to novel and familiar odors is sufficiently distinct to allow for decoding of odor experience (i.e., whether an odor is novel or familiar). Moreover, novelty engages more neurons in encoding the stimulus (Figure 2D). In neural space, the position of an odor’s representation in AON and aPCx shifts depending on whether it is novel or familiar, meaning that experience modifies the neural representation of odor identity. This suggests that in these regions the two representations are intertwined.

In contrast, some neurons in LEC, CA1, and SUB exhibit responses to novel odors, but few neurons respond to familiar odors at all. This suggests a more selective encoding of novelty.

(6) Lines 132-140 - As presented in the text and the figure, this section is poorly written and confusing. Their use of the word "shuffled" is a major source of this confusion, because this typically is the control that produces outcomes at the chance level. More importantly, they did the wrong analysis here. The better and, I think, the only way to do this analysis correctly is to train on some of the odors and test on an untrained odor (i.e., what Bernardi et al., 2021 called "cross-condition generalization performance"; CCGP).

We appreciate the feedback and thank the reviewer for the recommendation to implement cross-condition generalization performance (CCGP) as used in Bernardi et al., 2020. We acknowledge that the term "shuffled" may have caused confusion, as it typically refers to control analyses producing chance-level outcomes. In our case, by "shuffling" we shuffled the identity of novel and familiar odors to assess how much the decoder relies on odor identity when distinguishing novelty. This test provided insight into how novelty-based structure exists within neural activity beyond random grouping but does not directly assess generalization.

As suggested, we used CCGP to measure how well novelty-related representations generalize across different odors. Our findings show that in AON and aPCx, novelty-related information is indeed highly generalizable, supporting the idea that these regions encode novelty in a less odor-selective manner (Figure 2K).

Reviewer #3 (Public review):

In this manuscript, the authors investigate how odor-evoked neural activity is modulated by experience within the olfactory-hippocampal network. The authors perform extracellular recordings in the anterior olfactory nucleus (AON), the anterior piriform (aPCx) and lateral entorhinal cortex (LEC), the hippocampus (CA1), and the subiculum (SUB), in naïve mice and in mice repeatedly exposed to the same odorants. They determine the response properties of individual neurons and use population decoding analyses to assess the effect of experience on odor information coding across these regions.

The authors' findings show that odor identity is represented in all recorded areas, but that the response magnitude and selectivity of neurons are differentially modulated by experience across the olfactory-hippocampal pathway.

Overall, this work represents a valuable multi-region data set of odor-evoked neural activity. However, limitations in the interpretability of odor experience of the behavioral paradigm, and limitations in experimental design and analysis, restrict the conclusions that can be drawn from this study.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Some suggestions, in no particular order, to further improve the manuscript:

(1) The example neuronal responses for CA1 and SUB in Figure 1 are not very inspiring. To my eyes, the odor period response is not that different from the baseline period. In general, a thorough characterization of firing rate properties during the odor period between the different brain regions would be informative.

We thank the reviewer for this valuable feedback. We have replaced the example neurons from CA1 and SUB in Figure 1C. We further extended the characterization of firing rate properties, including a separate analysis of neurons i) significantly excited by odorants, ii) significantly inhibited by odorants and iii) not responsive to odorants. We added the analysis and corresponding discussion to the main manuscript (Figures 2C-E and lines 118-132)

(2) For the summary in Figure 1, why not show neuronal responses as z-scored firing rates as opposed to auROC?

We chose to use auROC instead of z-scored firing rates due to the non-normality of the dataset, which can distort results when using z-scores. Specifically, z-scoring can exaggerate small deviations in neurons with low responsiveness, potentially leading to misleading conclusions. auROC provides a more robust measure of response change that is less sensitive to these distortions because it does not assume any specific distribution. This approach has been used previously (e.g. Cohen et al. 2012, Nature).

(3) To study novelty, the authors presented odorants that were not used during four days of habituation. But this design makes it hard to dissociate odor identity from novelty. Why not track the response of the same odorants during the habituation process itself?

We respectfully disagree with the argument that using different stimuli as novel and familiar constitutes a confound in our analysis. In our study, we used multiple different, structurally dissimilar single molecule chemicals which were randomly assigned to novel and familiar categories in each animal. If individual stimuli did cause “drastic differences in evoked neural responses”, these would be evenly distributed between novel and familiar stimuli. It is therefore extremely unlikely that the clear differences we observed between novel and familiar conditions and between brain areas can be attributed to the contribution of individual stimuli, in particular given our analyses was performed at the population level. In fact, we observed that responses between novel and familiar conditions were qualitatively very similar in the short time window after odor onset (Figure 1G and H).

Importantly, the goal of this study was to investigate the impact of long-term habituation over more than 4 days, rather than short term habituation during one behavioral session. However, tracking the activity of large numbers of neurons across multiple days presents a significant technical challenge, due to the difficulty of identifying stable single-unit recordings over extended periods of time with sufficient certainty. Tools that facilitate tracking have recently been developed (e.g. Yuan AX et al., Elife. 2024) and it will be interesting to apply them to our dataset in the future.

(4) Since novel odors lead to greater sniffing and sniffing strongly influences firing rates in the olfactory system, the authors decided to focus on a 400 ms window with similar sniffing rates for both novel vs. familiar odors. Although I understand the rationale for this choice, I worry that this is too restrictive, and it may not capture the full extent of the phenomenology.

Could the authors model the effect of sniffing on firing rates of individual neurons from the data, and then check whether the odor response for novel context can be fully explained just by increased sniffing or not?

It is an interesting suggestion to extend the window of analysis and observe how responses evolve with sniffing (and other behavioral reactions). To address this, we added an additional figure to the supplementary material, showing the mean responses of all neurons to novel stimuli during the entire odor presentation window (Fig. S1B).

As suggested, we further created a Generalized Linear Model (GLM) for the entire 2s odor stimulation period, incorporating sniffing and novelty as independent variables. As expected, sniffing had a dominant impact on firing rate in all brain areas. A smaller proportion of neurons was modulated by novelty or by the interaction between novelty x breathing, suggesting the entrainment of neural activity by sniffing during the response to novel odors. These results support our decision to focus the analysis on the early 400ms window in order to dissociate the effects of novelty and behavioral responses. Taken together, our results suggest that odorant responses are modulated by novelty early during odorant processing, whereas at later stages sniffing becomes the predominant factor driving firing (Figure S2C-D).

(5) The authors conclude that aPCx has a subset of neurons dedicated to familiar odors based on the distribution of SVM weights in Figure 3D. To me, this is the weakest conclusion of the paper because although significant, the effect size is paltry; the central tendencies are hardly different for the two conditions in aPCx. Could the authors show the PSTHs of some of these neurons to make this point more convincing?

We appreciate the reviewer’s concern regarding the effect size. To strengthen our conclusion, we now include PSTHs of representative neurons in the least 10% and best 10% of neuronal population based on the SVM analysis (Figures S3 and S4). We hope this provides more clarity and support for the interpretation that there is a subset of neurons in aPCx that show greater sensitivity to familiar odors, despite the relatively modest central tendency differences.

In the revised manuscript, we discuss the effect size more explicitly in the text to provide context for its significance (lines 193 - 195).

Reviewer #2 (Recommendations for the authors):

(1) The authors only talk about "responsive" neurons. Does this include neurons whose activity increases significantly (activated) and neurons whose activity decreases (suppressed)?

Yes, the term "responsive" refers to neurons whose activity either increases significantly (excited) or decreases (inhibited) in response to the odor stimuli. We performed additional analyses to characterize responses separately for the different groups (Figure 2C-E and lines 118-132).

(2) Line 54 - The Schoonover paper doesn't show that cells lose their responses to odors, but rather that the population of cells that respond to odors changes with time. That is, population responses don't become more sparse

The fact that “the population of cells that respond to odors changes with time”, implies that some neurons lose their responsiveness (e.g. unit 2 in Figure 1 of Schoonover et al., 2021), while others become responsive (e.g. unit 1 in Figure 1 of Schoonover et al., 2021). Frequent responses reduce drift rate (Figure 4 of Schoonover et al., 2021), thus fewer neurons loose or gain responsiveness. We have revised the manuscript to clarify this.

(3) Line 104 - "Recurrent" is incorrectly used here. I think the authors mean "repeated" or something more like that.

Thank you for pointing this out. We replaced "recurrent" with "repeated".

(4) Figure 3D - What is the scale bar here?

We apologize for the accidental omission. The scale bar was be added to Figure 3D in the revised version of the manuscript.

(5) Line 377 - They say they lowered their electrodes to "200 um/s per second." This must be incorrect. Is this just a typo, or is it really 200 um/s, because that's really fast?

Thank you for pointing this out. It was 20 to 60 um/s, the change has been made in the manuscript.

(6) Line 431: The authors say they used auROC to calculate changes in firing rates (which I think is only shown in Figure 1D). Note that auROC measures the discriminability of two distributions, not the strength or change in the strength of response.

Indeed we used auROC to measure the discriminability of firing between baseline and during stimulus response. We have corrected the wording in the methods.

(7) Figure 1B: The anatomical locations of the five areas they recorded from are straightforward, and this figure is not hugely helpful. However, the reader would benefit tremendously by including an experimental schematic. As is, we needed to scour the text and methods sections to understand exactly what they did when.

We thank the reviewer for this suggestion. We included an experimental schematic in the supplementary material.

(8) Figure 1F(left): This plot is much less useful without showing a pre-odor window, even if only times after the odor onset were used for calculation alpha

We appreciate this concern, however the goal of Figure 1F is to illustrate the meaning of the alpha value itself. We chose not to include a pre-odor window comparison to avoid confusing the reader.

(9) Figure 2A: What are the bar plots above the raster plots? Are these firing rates? Are the bars overlaid or stacked? Where is the y-axis scale bar?

The bar plots above the raster plots represent a histogram of the spike count/trials over time, with a bin width of 50 ms. These bars are overlaid on the raster plot. We will include a y-axis scale bar in the revised figure to clarify the presentation.

(10) Figure 4G: This makes no sense. First, the Y axis is supposed to measure standard deviation, but the axis label is spikes/s. Second, if responses in the AON are much less reliable than responses in "deeper" areas, why is odor decoding in AON so much better than in the other areas?

We acknowledge the error in the axis label, and we will correct it to indicate the correct units. AON has a larger response variability but also larger responses magnitudes, which can explain the higher decoding accuracy.

(11) From the model and text, one predicts that the lifetime sparseness increases along the pathway. The authors should use this metric as well/instead of "odor selectivity" because of problems with arbitrary thresholding.

We acknowledge that lifetime sparseness, often computed using lifetime kurtosis, can be an informative measure of selectivity. However, we believe it has limitations that make it less suitable for our analysis. One key issue is that lifetime sparseness does not account for the stability of responses across multiple presentations of the same stimulus. In contrast, our odor selectivity measure incorporates trial-to-trial variability by considering responses over 10 trials and assessing significance using a Wilcoxon test compared to baseline. While the choice of a p-value threshold (e.g., 0.05) is somewhat arbitrary, it is a widely accepted statistical convention. Additionally, lifetime sparseness does not account for excitatory and inhibitory responses. For example, if a neuron X is strongly inhibited by odor A, strongly excited by odor B, and unresponsive to odors C and D, lifetime sparseness would classify it as highly selective for odor B, without capturing its inhibitory selectivity for odor A. The lifetime sparseness will be higher than if X was simply unresponsive for A.

Our odor selectivity measure addresses this by considering both excitation and inhibition as potential responses. Thus, while lifetime sparseness could provide a useful complementary perspective in another type of dataset, it does not fully capture the dynamics of odor selectivity here.

Author response 1.

Lifetime Kurtosis distribution per region.

Reviewer #3 (Recommendations for the authors):

Main points:

(1) The authors use a non-associative learning paradigm - repeated odor exposure - to test how experience modulates odor responses along the olfactory-hippocampal pathway. While repeated odor exposure clearly modulates odor-evoked neural activity, the relevance of this modulation and its differential effect across different brain areas are difficult to assess in the absence of any behavioral read-outs.

Our experimental paradigm involves a robust, reliable behavioral readout of non-associative learning. Novel olfactory stimuli evoke a well-characterized orienting reaction, which includes a multitude of physiological reactions, including exploratory sniffing, facial movements and pupil dilation (Modirshanechi et al., Trends Neuroscience 2023). In our study, we focused on exploration sniffing.

Compared to associative learning, non-associative learning might have received less attention. However, it is critically important because it forms the foundation for how organisms adapt to their environment through experience without forming associations. This is highlighted by the fact that non-instrumental stimuli can be remembered in large number (Standing, 1973) and with remarkable detail (Brady et al., 2008). While non-associative learning can thus create vast, implicit memory of stimuli in the environment, it is unclear how stimulus representations reflect this memory. Our study contributes to answering this question. We describe the impact of experience on olfactory sensory representations and reveal a transformation of representations from olfactory cortical to hippocampal structures. Our findings also indicate that sensory responses to familiar stimuli persist within sensory cortical and hippocampal regions, even after spontaneous orienting behaviors habituated. Further studies involving experimental manipulation techniques are needed to elucidate the causal mechanisms underlying the formation of stimulus memory during non-associative learning.

(2) The authors discuss the olfactory-hippocampal pathway as a transition from primary sensory (AON, aPCx) to associative areas (LEC, CA1, SUB). While this is reasonable, given the known circuit connectivity, other interpretations are possible. For example, AON, aPCx, and LEC receive direct inputs from the olfactory bulb ('primary cortex'), while CA1 and SUB do not; AON receives direct top-down inputs from CA1 ('associative cortex'), while aPCx does not. In fact, the data presented in this manuscript does not appear to support a consistent, smooth transformation from sensory to associative, as implied by the authors (e.g. Figure 4A, F, and G).

Thank you for this insightful comment. Indeed, there are complexities in the circuitry, and the relationships between different areas are not linear. We believe that AON and aPCx are distinctly different from LEC, CA1 and SUB, as the latter areas have been shown to integrate multimodal sensory information. To avoid confusion due to definitions of what constitutes a “primary sensory” region, we adopted a more neutral description throughout the manuscript. We also removed the term “gradual” to describe the transition of neural representations from olfactory cortical to hippocampal areas.

(3) The analysis of odor-evoked responses is focused on a 400 ms window to exclude differences in sniffing behavior. This window spans 200 ms before and after the first inhalation after odor onset. Inhalation onset initiates neural odor responses - why do the authors include neural data before inhalation onset?

The reason to include a brief time window prior to odor onset is to account for what is often called “partical” sniffs. In our experimental setup, odor delivery is not triggered by the animal’s inhalation. Therefore, it can happen that an animal has just begun to inhale when the stimulus is delivered. In this case, the animal is exposed to odorant molecules prior to the first complete inhalation after odor onset. We acknowledge that this limits the temporal resolution of our measurements, but it does not affect the comparison of sensory representations between different brain areas.

It would also be interesting to explore the effect of sniffing behavior (see point 2) on odor-evoked neural activity.

Thank you for your comment, we performed additional analysis including a GLM to address this question (Figure S2C-D).

Minor points:

(4) Figure 2A represents raster plots for 2 neurons per area - it is unclear how to distinguish between the 2 neurons in the plots.

Figure 2A shows one example neuron per brain area. Each neurons has two raster plot which indicate responses to either a novel (orange) or a familiar stimulus (blue). We have revised the figure caption for clarity.

(5) Overall, axes should be kept consistent and labeled in more detail. For example, Figure 2H and I are difficult to compare, given that the y-axis changes and that decoding accuracies are difficult to estimate without additional marks on the y-axis.

Axes are indeed different, because chance level decoding accuracy is different between those two figures. The decoding between novel and familiar odors has a chance level of 0.5, while chance level decoding odors is 0.1 (there are 10 odors to decode the identity from).

(6) Some parts of the discussion seem only loosely related to the data presented in this manuscript. For example, the statement that 'AON rather than aPCx should be considered as the primary sensory cortex in olfaction' seems out of context. Similarly, it would be helpful to provide data on the stability of subpopulations of neurons tuned to familiar odors, rather than simply speculate that they could be stable. The authors could summarize more speculative statements in an 'Ideas and Speculation' subsection.

Thank you for your comment. We appreciate your perspective on our hypotheses. We have revised the discussion accordingly. Specifically, we removed the discussion of stable subpopulations, since we have not performed longitudinal tracking in this study.

(7) The authors should try to reference relevant published work more comprehensively.

Thank you for your comment. We attempted to include relevant published work without exceeding the limit for references but might have overseen important contributions. We apologize to our colleagues, whose relevant work might not have been cited.

Reviewer #1 (Public review):

Summary:

This manuscript describes critical intermediate reaction steps of a HA synthase at the molecular level; specifically, it examines the 2nd step, polymerization, adding GlcA to GlcNAc to form the initial disaccharide of the repeating HA structure. Unlike the vast majority of known glycosyltransferases, the viral HAS (a convenient proxy extrapolated to resemble the vertebrate forms) uses a single pocket to catalyze both monosaccharide transfer steps. The authors' work illustrates the interactions needed to bind & proof-read the UDP-GlcA using direct and '2nd layer' amino acid residues. This step also allows the HAS to distinguish the two UDP-sugars; this is very important as the enzymes are not known or observed to make homopolymers of only GlcA or GlcNAc, but only make the HA disaccharide repeats GlcNAc-GlcA.

Strengths:

Overall, the strengths of this paper lie in its techniques & analysis.

The authors make significant leaps forward towards understanding this process using a variety of tools and comparisons of wild-type & mutant enzymes. The work is well presented overall with respect to the text and illustrations (especially the 3D representations), and the robustness of the analyses & statistics is also noteworthy.

Furthermore, the authors make some strides towards creating novel sugar polymers using alternative primers & work with detergent binding to the HAS. The authors tested a wide variety of monosaccharides and several disaccharides for primer activity and observed that GlcA could be added to cellobiose and chitobiose, which are moderately close structural analogs to HA disaccharides. Did the authors also test the readily available HA tetramer (HA4, [GlcA-GlcNAc]2) as a primer in their system? This is a highly recommended experiment; if it works, then this molecule may also be useful for cryo-EM studies of CvHAS as well.

Weaknesses:

In the past, another report describing the failed attempt of elongating short primers (HA4 & chitin oligosaccharides larger than the cello- or chitobiose that have activity in this report) with a vertebrate HAS, XlHAS1, an enzyme that seems to behave like the CvHAS ( https://pubmed.ncbi.nlm.nih.gov/10473619/); this work should probably be cited and briefly discussed. It may be that the longer primers in the 1999 paper and/or the different construct or isolation specifics (detergent extract vs crude) were not conducive to the extension reaction, as the authors extracted recombinant enzyme.

There are a few areas that should be addressed for clarity and correctness, especially defining the class of HAS studied here (Class I-NR) as the results may (Class I-R) or may not (Class II) align (see comment (a) below), but overall, a very nicely done body of work that will significantly enhance understanding in the field.

Reviewer #2 (Public review):

Summary:

The paper by Stephens and co-workers provides important mechanistic insight into how hyaluronan synthase (HAS) coordinates alternating GlcNAc and GlcA incorporation using a single Type-I catalytic centre. Through cryo-EM structures capturing both "proofreading" and fully "inserted" binding poses of UDP-GlcA, combined with detailed biochemical analysis, the authors show how the enzyme selectively recognizes the GlcA carboxylate, stabilizes substrates through conformational gating, and requires a priming GlcNAc for productive turnover.

These findings clarify how one active site can manage two chemically distinct donor sugars while simultaneously coupling catalysis to polymer translocation.

The work also reports a DDM-bound, detergent-inhibited conformation that possibly illuminates features of the acceptor pocket, although this appears to be a purification artefact (it is indeed inhibitory) rather than a relevant biological state.

Overall, the study convincingly establishes a unified catalytic mechanism for Type-I HAS enzymes and represents a significant advance in understanding HA biosynthesis at the molecular level.

Strengths:

There are many strengths.

This is a multi-disciplinary study with very high-quality cryo-EM and enzyme kinetics (backed up with orthogonal methods of product analysis) to justify the conclusions discussed above.

Weaknesses:

There are few weaknesses.

The abstract and introduction assume a lot of detailed prior knowledge about hyaluronan synthases, and in doing so, risk lessening the readership pool.

A lot of discussion focuses on detergents (whose presence is totally inhibitory) and transfer to non-biological acceptors (at high concentrations). This risks weakening the manuscript.

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

The main contributions of this paper are: (1) a replication of the surprising prior finding that information about peripherally-presented stimuli can be decoded from foveal V1 (Williams et al 2008), (2) a new demonstration of cross-decoding between stimuli presented in the periphery and stimuli presented at the fovea, (3) a demonstration that the information present in the fovea is based on shape not semantic category, and (4) a demonstration that the strength of foveal information about peripheral targets is correlated with the univariate response in the same block in IPS.

Strengths:

The design and methods appear sound, and finding (2) above is new, and importantly constrains our understanding of this surprising phenomenon. The basic effect investigated here is so surprising that even though it has been replicated several times since it was first reported in 2008, it is useful to replicate it again.

We thank the reviewer for their summary. While we agree with many points, we would like to respectfully push back on the notion that this work is a replication of Williams et al. (2008). What our findings share with those of Williams is a report of surprising decoding at the fovea without foveal stimulation. Beyond this similarity, we treat these as related but clearly separate findings, for the following reasons:

(1) Foveal feedback, as shown by Williams et al. (2008) and others during fixation, was only observed during a shape discrimination task, specific to the presented stimulus. Control experiments without such a task (or a color-related task) did not show effects of foveal feedback. In contrast, in the present study, the participants’ task was merely to perform saccades towards stimuli, independently of target features. We thus show that foveal feedback can occur independently of a task related to stimulus features. This dissociation demonstrates that our study must be tapping into something different than reported by Williams.

(2) In a related study, Kroell and Rolfs (2022, 2025) demonstrated a connection between foveal feedback and saccade preparation, including the temporal details of the onset of this effect before saccade execution, highlighting the close link of this effect to saccade preparation. Here we used a very similar behavioral task to capture this saccade-related effect in neural recordings and investigate how early it occurs and what its nature is. Thus, there is a clear motivation for this study in the context of eye movement preparation that is separate from the previous work by Williams.

(3) Lastly, decoding in the experimental task was positively associated with activity in FEF and IPS, areas that have been reliably linked to saccade preparation. We have now also performed an additional analysis (see our response to Specific point 2 of Reviewer 2) showing that decoding in the control condition did not show the same association, further supporting the link of foveal feedback to saccade preparation. 

Despite our emphasis on these critical differences in studies, covert peripheral attention, as required by the task in Williams et al., and saccade preparation in natural vision, as in our study, are tightly coupled processes. Indeed, the task in Williams et al. would, during natural vision, likely involve an eye movement to the peripheral target. While speculative, a parsimonious and ecologically valid explanation is that both ours and earlier studies involve eye movement preparation, for which execution is suppressed, however, in studies enforcing fixation (e.g., Williams et al., 2008). We now discuss this idea of a shared underlying mechanism more extensively in the revised manuscript (pg 8 ln 228-240). 

Weaknesses:

(1) The paper, including in the title ("Feedback of peripheral saccade targets to early foveal cortex") seems to assume that the feedback to foveal cortex occurs in conjunction with saccade preparation. However, participants in the original Williams et al (2008) paper never made saccades to the peripheral stimuli. So, saccade preparation is not necessary for this effect to occur. Some acknowledgement and discussion of this prior evidence against the interpretation of the effect as due to saccade preparation would be useful. (e.g., one might argue that saccade preparation is automatic when attending to peripheral stimuli.)

We agree that the effects Williams et al. showed were not sufficiently discussed in the first version of this manuscript. To more clearly engage with these findings we now introduce saccade related foveal feedback (foveal prediction) and foveal feedback during fixation separately in the introduction (pg 2 ln 46-59).

We further added another section in the discussion called “Foveal feedback during saccade preparation” in which we discuss how our findings are related to Williams et al. and how they differ (pg 8 ln 211-240). 

As described in our previous response, we believe that our findings go beyond those described by Williams et al. (2008) and others in significant ways. However, during natural vision, the paradigm used by Williams et al. (2008) would likely be solved using an eye movement. Thus, while participants in Williams et al. (2008) did not execute saccades, it appears plausible that they have prepared saccades. Given the fact that covert peripheral attention and saccade preparation are tightly coupled processes (Kowler et al., 1995, Vis Res; Deubel & Schneider, 1996, Vis Res; Montagnini & Castet, 2007, J Vis; Rolfs & Carrasco, 2012, J Neurosci; Rolfs et al., 2011, Nat Neurosci), their results are parsimoniously explained by saccade preparation (but not execution) to a behaviorally relevant target.

(2) The most important new finding from this paper is the cross-decodability between stimuli presented in the fovea and stimuli presented in the periphery. This finding should be related to the prior behavioral finding (Yu & Shim, 2016) that when a foveal foil stimulus identical to a peripheral target is presented 150 ms after the onset of the peripheral target, visual discrimination of the peripheral target is improved, and this congruency effect occurred even though participants did not consciously perceive the foveal stimulus (Yu, Q., & Shim, W. M., 2016). Modulating foveal representation can influence visual discrimination in the periphery (Journal of Vision, 16(3), 15-15).

We thank the reviewer for highlighting this highly relevant reference. In the revised version of the manuscript, we now put more emphasis on the finding of cross-decodability (pg 2 ln 60-61). We now also discuss Yu et al.’s finding, which support our conclusion that foveal feedback and direct stimulus presentation share representational formats in early visual areas (pg 9 ln 277-279).

(3) The prior literature should be laid out more clearly. For example, most readers will not realize that the basic effect of decodability of peripherally-presented stimuli in the fovea was first reported in 2008, and that that original paper already showed that the effect cannot arise from spillover effects from peripheral retinotopic cortex because it was not present in a retinotopic location between the cortical locus corresponding to the peripheral target and the fovea. (For example, this claim on lines 56-57 is not correct: "it remains unknown 1) whether information is fed back all the way to early visual areas".) What is needed is a clear presentation of the prior findings in one place in the introduction to the paper, followed by an articulation and motivation of the new questions addressed in this paper. If I were writing the paper, I would focus on the cross-decodability between foveal and peripheral stimuli, as I think that is the most revealing finding.

We agree that the structure of the introduction did not sufficiently place our work in the context of prior literature. We have now expanded upon our Introduction section to discuss past studies of saccade- and fixation-related foveal feedback (pg 2 ln 49-59), laying out how this effect has been studied previously. We also removed the claim that "it remains unknown 1) whether information is fed back all the way to early visual areas", where our intention was to specifically focus on foveal prediction. We realize that this was not clear and hence removed this section. Instead, we now place a stronger focus on the cross-decodability finding (pg 2 ln 60-61).

Reviewer #2 (Public review):

Summary:

This study investigated whether the identity of a peripheral saccade target object is predictively fed back to the foveal retinotopic cortex during saccade preparation, a critical prediction of the foveal prediction hypothesis proposed by Kroell & Rolfs (2022). To achieve this, the authors leveraged a gaze-contingent fMRI paradigm, where the peripheral saccade target was removed before the eyes landed near it, and used multivariate decoding analysis to quantify identity information in the foveal cortex. The results showed that the identity of the saccade target object can be decoded based on foveal cortex activity, despite the fovea never directly viewing the object, and that the foveal feedback representation was similar to passive viewing and not explained by spillover effects. Additionally, exploratory analysis suggested IPS as a candidate region mediating such foveal decodability. Overall, these findings provide neural evidence for the foveal cortex processing the features of the saccade target object, potentially supporting the maintenance of perceptual stability across saccadic eye movements.

Strengths:

This study is well-motivated by previous theoretical findings (Kroell & Rolfs, 2022), aiming to provide neural evidence for a potential neural mechanism of trans-saccadic perceptual stability. The question is important, and the gaze-contingent fMRI paradigm is a solid methodological choice for the research goal. The use of stimuli allowing orthogonal decoding of stimulus category vs stimulus shape is a nice strength, and the resulting distinctions in decoded information by brain region are clean. The results will be of interest to readers in the field, and they fill in some untested questions regarding pre-saccadic remapping and foveal feedback.

We thank the reviewer for the positive assessment of our study.

Weaknesses:

The conclusions feel a bit over-reaching; some strong theoretical claims are not fully supported, and the framing of prior literature is currently too narrow. A critical weakness lies in the inability to test a distinction between these findings (claiming to demonstrate that "feedback during saccade preparation must underlie this effect") and foveal feedback previously found during passive fixation (Williams et al., 2008). Discussions (and perhaps control analysis/experiments) about how these findings are specific to the saccade target and the temporal constraints on these effects are lacking. The relationship between the concepts of foveal prediction, foveal feedback, and predictive remapping needs more thorough treatment. The choice to use only 4 stimuli is justified in the manuscript, but remains an important limitation. The IPS results are intriguing but could be strengthened by additional control analysis. Finally, the manuscript claims the study was pre-registered ("detailing the hypotheses, methodology, and planned analyses prior to data collection"), but on the OSF link provided, there is just a brief summary paragraph, and the website says "there have been no completed registrations of this project".

We thank the reviewer for these helpful considerations. We agree that some of the claims were not sufficiently supported by the evidence, and in the revised manuscript, we added nuance to those claims (pg 8 ln 211-240). Furthermore, we now address more directly the distinction between foveal feedback during fixation and foveal feedback (foveal prediction) during saccade preparation. In particular, we now describe the literature about these two effects separately in the introduction (pg 2 ln 46-59), and we have added a new section in the discussion (“Foveal feedback during saccade preparation”) that more thoroughly explains why a passive fixation condition would have been unlikely to produce the same results we find (pg 8 ln 211-227). We also adapted the section about “Saccadic remapping or foveal prediction”, clearly delineating foveal prediction from feature remapping and predictive updating of attention pointers. As recommended by the reviewer, we conducted the parametric modulation analyses on the control condition, strengthening the claim that our findings are saccade-related. These results were added as Supplementary Figure 2 and are discussed in (pg 7 ln 190-191) and (pg 8 ln 224-227). 

Lastly, we would like to apologize about a mistake we made with the pre-registration. We realized that the pre-registration had indeed not been submitted. We have now done so without changing the pre-registration itself, which can be seen from the recent activity of the preregistration (screenshot attached in the end). After consulting an open science expert at the University of Leipzig, we added a note of this mistake to the methods section of the revised manuscript (pg 10 ln 326-332). We could remove reference to this preregistration altogether, but would keep it at the discretion of the editor. 

Specifics:

(1) In the eccentricity-dependent decoding results (Figure 2B), are there any statistical tests to support the results being a U-shaped curve? The dip isn't especially pronounced. Is 4 degrees lower than the further ones? Are there alternative methods of quantifying this (e.g., fitting it to a linear and quadratic function)?

We statistically tested the U-shaped relationship using a weighted quadratic regression, which showed significant positive curvature for decoding between fovea and periphery in all early visual areas (V1: t(27) = 3.98, p = 0.008, V2: t(27) = 3.03, p = 0.02, V3: t(27)= 2.776, p = 0.025, one-sided). We now report these results in the revised manuscript (pg 5 ln 137-138).

(2) In the parametric modulation analysis, the evidence for IPS being the only region showing stronger fovea vs peripheral beta values was weak, especially given the exploratory nature of this analysis. The raw beta value can reflect other things, such as global brain fluctuations or signal-to-noise ratio. I would also want to see the results of the same analysis performed on the control condition decoding results.

We appreciate the reviewer’s suggestion and repeated the same parametric modulation analysis on the control condition to assess the influence of potential confounds on the overall beta values (Supplementary Figure 2). The results show a negative association between foveal decoding and FEF and IPS (likely because eye movements in the control condition lead to less foveal presentation of the stimulus) and a positive association with LO. Peripheral decoding was not associated with significant changes in any of the ROIs, indicating that global brain fluctuations alone are not responsible for the effects reported in the experimental condition. The results of this analysis thus show a specific positive association of IPS activity with the experimental condition, not the control condition, which is in line with the idea that the foveal feedback effect reported in this study may be related to saccade preparation.

(3) Many of the claims feel overstated. There is an emphasis throughout the manuscript (including claims in the abstract) that these findings demonstrate foveal prediction, specifically that "image-specific feedback during saccade preparation must underlie this effect." To my understanding, one of the key aspects of the foveal prediction phenomenon that ties it closely to trans-saccadic stability is its specificity to the saccade target but not to other objects in the environment. However, it is not clear to what degree the observed findings are specific to saccade preparation and the peripheral saccade target. Should the observers be asked to make a saccade to another fixation location, or simply maintain passive fixation, will foveal retinotopic cortex similarly contain the object's identity information? Without these control conditions, the results are consistent with foveal prediction, but do not definitively demonstrate that as the cause, so claims need to be toned down.

We fully agree with the reviewer and toned down claims about foveal prediction. We engage with the questions raised by the reviewer more thoroughly in the new discussion section “Foveal feedback during saccade preparation”.

In addition, we agree that another condition in which subjects make a saccade towards a different location would have been a great addition that we also considered, but due to concerns with statistical power did not add. While including such a condition exceeds the scope of the current study, we included this limitation in the Discussion section (pg 10 ln 316) and hope that future studies will address this question.

(4) Another critical aspect is the temporal locus of the feedback signal. In the paradigm, the authors ensured that the saccade target object was never foveated via the gaze-contingent procedure and a conservative data exclusion criterion, thus enabling the test of feedback signals to foveal retinotopic cortex. However, due to the temporal sluggishness of fMRI BOLD signals, it is unclear when the feedback signal arrives at the foveal retinotopic cortex. In other words, it is possible that the feedback signal arrives after the eyes land at the saccade target location. This possibility is also bolstered by Chambers et al. (2013)'s TMS study, where they found that TMS to the foveal cortex at 350-400 ms SOA interrupts the peripheral discrimination task. The authors should qualify their claims of the results occurring "during saccade preparation" (e.g., pg 1 ln 22) throughout the manuscript, and discuss the importance of temporal dynamics of the effect in supporting stability across saccades.

We fully agree that the sluggishness of the fMRI signal presents an important challenge in investigating foveal feedback. We have now included this limitation in the discussion (pg 10 ln 306-318). We also clarify that our argument connects to previous studies investigating the temporal dynamics of foveal feedback using similar tasks (pg 10 ln 313-316). Specifically, in their psychophysical work, Kroell and Rolfs (2022) and (2025) showed that foveal feedback occurs before saccade execution with a peak around 80 ms before the eye movement. 

(5) Relatedly, the claims that result in this paradigm reflect "activity exclusively related to predictive feedback" and "must originate from predictive rather than direct visual processes" (e.g., lines 60-65 and throughout) need to be toned down. The experimental design nicely rules out direct visual foveal stimulation, but predictive feedback is not the only alternative to that. The activation could also reflect mental imagery, visual working memory, attention, etc. Importantly, the experiment uses a block design, where the same exact image is presented multiple times over the block, and the activation is taken for the block as a whole. Thus, while at no point was the image presented at the fovea, there could still be more going on than temporally-specific and saccade-specific predictive feedback.

We agree that those claims could have misled the reader. Our intention was to state that the activation originates from feedback rather than direct foveal stimulation because of the nature of the design. We have now clarified these statements (pg 2 ln 65) and also included a discussion of other effects including imagery and working memory in the limitations section (pg 10 ln 306-313).

(6) The authors should avoid using the terms foveal feedback and foveal prediction interchangeably. To me, foveal feedback refers to the findings of Williams et al. (2008), where participants maintained passive fixation and discriminated objects in the periphery (see also Fan et al., 2016), whereas foveal prediction refers to the neural mechanism hypothesized by Kroell & Rolfs (2022), occurring before a saccade to the target object and contains task irrelevant feature information.

We agree, and we have now adopted a clearer distinction between these terms, referring to foveal prediction only when discussing the distinct predictive nature of the effect discovered by Kroell and Rolfs (2022). Otherwise we referred to this effect as foveal feedback.

(7) More broadly, the treatment of how foveal prediction relates to saccadic remapping is overly simplistic. The authors seem to be taking the perspective that remapping is an attentional phenomenon marked by remapping of only attentional/spatial pointers, but this is not the classic or widely accepted definition of remapping. Within the field of saccadic remapping, it is an ongoing debate whether (/how/where/when) information about stimulus content is remapped alongside spatial location (and also whether the attentional pointer concept is even neurophysiologically viable). This relationship between saccadic remapping and foveal prediction needs clarification and deeper treatment, in both the introduction and discussion.

We thank the reviewer for their remarks. We reformulated the discussion section on “Saccadic remapping or foveal prediction” to include the nuances about spatial and feature remapping laid out in the reviewer’s comment (pg 8-9 ln 241-269). We also put a stronger focus on the special role the fovea seems to be playing regarding the feedback of visual features (pg 8-9 ln 265-269).

(8) As part of this enhanced discussion, the findings should be better integrated with prior studies. E.g., there is some evidence for predictive remapping inducing integration of non-spatial features (some by the authors themselves; Harrison et al., 2013; Szinte et al., 2015). How do these findings relate to the observed results? Can the results simply be a special case of non-spatial feature integration between the currently attended and remapped location (fovea)? How are the results different from neurophysiological evidence for facilitation of the saccade target object's feature across the visual field (Burrow et al., 2014)? How might the results be reconciled with a prior fMRI study that failed to find decoding of stimulus content in remapped responses (Lescroart et al, 2016)? Might this reflect a difference between peripheral-to-peripheral vs peripheral-to-foveal remapping? A recent study by Chiu & Golomb (2025) provided supporting evidence for peripheral-to-fovea remapping (but not peripheral-to-peripheral remapping) of object-location binding (though in the post-saccadic time window), and suggested foveal prediction as the underlying mechanism.

We thank the reviewer for raising these intriguing questions. We now address them in the revised discussion. We argue that the findings by Harrison et al., 2013 and Szinte et al., 2015 of presaccadic integration of features across two peripheral locations can be explained by presaccadic updating of spatial attention pointers rather than remapping of feature information (pg 8 ln 248-253). The lack of evidence for periphery-to-periphery remapping (Lescroart et al, 2016) and the recent study by Chiu & Golomb (2025) showing object-location binding from periphery to fovea nicely align with our characterization of foveal processing as unique in predicting feature information of upcoming stimuli (pg 8-9 ln 265-269). Finally, we argue that the global (i.e., space-invariant) selection task-irrelevant saccadic target features (Burrows et al., 2014) is well-established at the neural level, but does not suffice to explain the spatially specific nature of foveal prediction (pg 8 ln 220-224). We now include these studies in the revised discussion section.

Reviewer #3 (Public review):

Summary:

In this paper, the authors used fMRI to determine whether peripherally viewed objects could be decoded from the foveal cortex, even when the objects themselves were never viewed foveally. Specifically, they investigated whether pre-saccadic target attributes (shape, semantic category) could be decoded from the foveal cortex. They found that object shape, but not semantic category, could be decoded, providing evidence that foveal feedback relies on low-mid-level information. The authors claim that this provides evidence for a mechanism underlying visual stability and object recognition across saccades.

Strengths:

I think this is another nice demonstration that peripheral information can be decoded from / is processed in the foveal cortex - the methods seem appropriate, and the experiments and analyses are carefully conducted, and the main results seem convincing. The paper itself was very clear and well-written.

We thank the reviewer for this positive evaluation of our work. As discussed in our response to Reviewer 1, we now elaborate on the differences between previous work showing decoding of peripheral information from foveal cortex from the effect shown here. While there are important similarities between these findings, foveal prediction in our study occurs in a saccade condition and in the absence of a task that is specific to stimulus features. 

Weaknesses:

There are a couple of reasons why I think the main theoretical conclusions drawn from the study might not be supported, and why a more thorough investigation might be needed to draw these conclusions.

(1) The authors used a blocked design, with each object being shown repeatedly in the same block. This meant that the stimulus was entirely predictable on each block, which weakens the authors' claims about this being a predictive mechanism that facilitates object recognition - if the stimulus is 100% predictable, there is no aspect of recognition or discrimination actually being tested. I think to strengthen these claims, an experiment would need to have unpredictable stimuli, and potentially combine behavioural reports with decoding to see whether this mechanism can be linked to facilitating object recognition across saccades.

We appreciate the reviewer’s point and would like to highlight that it was not our intention to claim a behavioral effect on object recognition. We believe that an ambiguous formulation in the original abstract may have been interpreted this way, and we thus removed this reference. We also speculated in our Discussion that a potential reason for foveal prediction could be a headstart in peripheral object recognition and in the revised manuscript more clearly highlight that this is a  potential future direction only.

(2)  Given that foveal feedback has been found in previous studies that don't incorporate saccades, how is this a mechanism that might specifically contribute to stability across saccades, rather than just being a general mechanism that aids the processing/discrimination of peripherally-viewed stimuli? I don't think this paper addresses this point, which would seem to be crucial to differentiate the results from those of previous studies.

We fully agree that this point had not been sufficiently addressed in the previous version of the manuscript. As described in our responses to similar comments from reviewers 1 and 2, we included an additional section in the Discussion (“Foveal feedback during saccade preparation”) to more clearly delineate the present study from previous findings of foveal feedback. Previous studies (Williams et al., 2008) only found foveal feedback during narrow discrimination tasks related to spatial features of the target stimulus, not during color-discrimination or fixation-only tasks, concluding that the observed effect must be related to the discrimination behavior. In contrast, we found foveal feedback (as evidenced by decoding of target features) during a saccade condition that was independent of the target features, suggesting a different role of foveal feedback than hypothesized by Williams et al. (2008).

Recommendations for the authors:  

Reviewer #2 (Recommendations for the authors):

(A) Minor comments:

(1)  The task should be clarified earlier in the manuscript.

We now characterise the task in the abstract and clarified its description in the third paragraph, right after introducing the main literature.

(2) Is there actually only 0.5 seconds between saccades? This feels very short/rushed.

The inter-trial-interval was 0.5 seconds, though effectively it varied because the target only appeared once participants fixated on the fixation dot. Note that this pacing is slower than the rate of saccades in natural vision (about 3 to 4 saccades per second).Participants did not report this paradigm as rushed.

(3) Typo on pg2 ln64 (whooe).

Fixed.

(4)  Can the authors also show individual data points for Figures 3 and 4?

We added individual data points for Figures 4 and S2

(5) The MNI coordinates on Figure 4A seem to be incorrect.

We took out those coordinates.

(6) Pg4 ln126 and pg6 ln194, why cite Williams et al. (2008)?

We included this reference here to acknowledge that Williams et al. raised the same issues. We added a “cf.” before this reference to clarify this.

(7) Pg7 ln207 Fabius et al. (2020) showed slow post-saccadic feature remapping, rather than predictive remapping of spatial attention.

We have corrected this mistake.

(8) The OSF link is valid, but I couldn't find a pre-registration.

The issue with the OSF link has been resolved. The pre-registration had been set up but not published. We now published it without changing the original pre-registration (see the screenshot attached).

(9) I couldn't access the OpenNeuro repository.

The issue with the OpenNeuro link has been resolved.

(B) Additional references you may wish to include:

(1) Burrows, B. E., Zirnsak, M., Akhlaghpour, H., Wang, M., & Moore, T.  (2014). Global selection of saccadic target features by neurons in area v4. Journal of Neuroscience.

(2) Chambers, C. D., Allen, C. P., Maizey, L., & Williams, M. A. (2013). Is delayed foveal feedback critical for extra-foveal perception?. Cortex.

(3) Chiu, T. Y., & Golomb, J. D. (2025). The influence of saccade target status on the reference frame of object-location binding. Journal of Experimental Psychology. General.

(4) Harrison, W. J., Retell, J. D., Remington, R. W., & Mattingley, J. B. (2013). Visual crowding at a distance during predictive remapping. Current Biology.

(5) Lescroart, M. D., Kanwisher, N., & Golomb, J. D. (2016). No evidence for automatic remapping of stimulus features or location found with fMRI. Frontiers in Systems Neuroscience.

(6) Moran, C., Johnson, P. A., Hogendoorn, H., & Landau, A. N. (2025). The representation of stimulus features during stable fixation and active vision. Journal of Neuroscience.

(7) Szinte, M., Jonikaitis, D., Rolfs, M., Cavanagh, P., & Deubel, H. (2016). Presaccadic motion integration between current and future retinotopic locations of attended objects. Journal of Neurophysiology.

We thank the reviewer for pointing out these references. We have included them in the revised version of the manuscript.

Reviewer #3 (Recommendations for the authors):

I just have a few minor points where I think some clarifications could be made.

(1) Line 64 - "whooe" should be "whoose" I think.

Fixed.

(2) Around line 53 - you might consider citing this review on foveal feedback - https://doi.org/10.1167/jov.20.12.2

We included the reference (pg 2 ln 55).

(3) Line 129 - you mention a u-shaped relationship for decoding - I wasn't quite sure of the significance/relevance of this relationship - it would be helpful to expand on this / clarify what this means.

We have expanded this section and added statistical tests of the u-shaped relationship in decoding using a weighted quadratic regression. We found significant positive curvature in all early visual areas between fovea and periphery (V1: t(27) = 3.98, p = 0.008, V2: t(27) = 3.03, p = 0.02, V3: t(27)= 2.776, p = 0.025). These findings support a u-shaped relationship. We now report these results in the revised manuscript (pg 5 ln 137-138).

(4) Figure 1 - it would be helpful to indicate how long the target was viewed in the "stim on" panels - I assume it was for the saccade latency, but it would be good to include those values in the main text.

We included that detail in the text (pg 3 ln 96-97).

Author response:

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

Public Reviews:

Reviewer #1 (Public review):

Summary:

This study examines whether changes in pupil size index prediction-error-related updating during associative learning, formalised as information gain via Kullback-Leibler (KL) divergence. Across two independent tasks, pupil responses scaled with KL divergence shortly after feedback, with the timing and direction of the response varying by task. Overall, the work supports the view that pupil size reflects information-theoretic processes in a context-dependent manner.

Strengths:

This study provides a novel and convincing contribution by linking pupil dilation to informationtheoretic measures, such as KL divergence, supporting Zénon's hypothesis that pupil responses reflect information gain during learning. The robust methodology, including two independent datasets with distinct task structures, enhances the reliability and generalisability of the findings. By carefully analysing early and late time windows, the authors capture the timing and direction of prediction-error-related responses, oPering new insights into the temporal dynamics of model updating. The use of an ideal-learner framework to quantify prediction errors, surprise, and uncertainty provides a principled account of the computational processes underlying pupil responses. The work also highlights the critical role of task context in shaping the direction and magnitude of these ePects, revealing the adaptability of predictive processing mechanisms. Importantly, the conclusions are supported by rigorous control analyses and preprocessing sanity checks, as well as convergent results from frequentist and Bayesian linear mixed-ePects modelling approaches.

Weaknesses:

Some aspects of directionality remain context-dependent, and on current evidence cannot be attributed specifically to whether average uncertainty increases or decreases across trials. DiPerences between the two tasks (e.g., sensory modality and learning regime) limit direct comparisons of ePect direction and make mechanistic attribution cautious. In addition, subjective factors such as confidence were not measured and could influence both predictionerror signals and pupil responses. Importantly, the authors explicitly acknowledge these limitations, and the manuscript clearly frames them as areas for future work rather than settled conclusions.

Reviewer #2 (Public review):

Summary:

The authors investigate whether pupil dilation reflects information gain during associative learning, formalised as Kullback-Leibler divergence within an ideal observer framework. They examine pupil responses in a late time window after feedback and compare these to informationtheoretic estimates (information gain, surprise, and entropy) derived from two diPerent tasks with contrasting uncertainty dynamics.

Strength:

The exploration of task evoked pupil dynamics beyond the immediate response/feedback period and then associating them with model estimates was interesting and inspiring. This oPered a new perspective on the relationship between pupil dilation and information processing.

Weakness:

However, the interpretability of the findings remains constrained by the fundamental diPerences between the two tasks (stimulus modality, feedback type, and learning structure), which confound the claimed context-dependent ePects. The later time-window pupil ePects, although intriguing, are small in magnitude and may reflect residual noise or task-specific arousal fluctuations rather than distinct information-processing signals. Thus, while the study oPers valuable methodological insight and contributes to ongoing debates about the role of the pupil in cognitive inference, its conclusions about the functional significance of late pupil responses should be treated with caution.

Reviewer #3 (Public review):

Summary:

Thank you for inviting me to review this manuscript entitled "Pupil dilation oPers a time-window on prediction error" by Colizoli and colleagues. The study examines prediction errors, information gain (Kullback-Leibler [KL] divergence), and uncertainty (entropy) from an information-theory perspective using two experimental tasks and pupillometry. The authors aim to test a theoretical proposal by Zénon (2019) that the pupil response reflects information gain (KL divergence). The conclusion of this work is that (post-feedback) pupil dilation in response to information gain is context dependent.

Strengths:

Use of an established Bayesian model to compute KL divergence and entropy.

Pupillometry data preprocessing and multiple robustness checks.

Weaknesses:

Operationalization of prediction errors based on frequency, accuracy, and their interaction:

The authors rely on a more model-agnostic definition of the prediction error in terms of stimulus frequency ("unsigned prediction error"), accuracy, and their interaction ("signed prediction error"). While I see the point, I would argue that this approach provides a simple approximation of the prediction error, but that a model-based approach would be more appropriate.

Model validation:

My impression is that the ideal learner model should work well in this case. However, the authors don't directly compare model behavior to participant behavior ("posterior predictive checks") to validate the model. Therefore, it is currently unclear if the model-derived terms like KL divergence and entropy provide reasonable estimates for the participant data.

Lack of a clear conclusion:

The authors conclude that this study shows for the first time that (post-feedback) pupil dilation in response to information gain is context dependent. However, the study does not oPer a unifying explanation for such context dependence. The discussion is quite detailed with respect to taskspecific ePects, but fails to provide an overarching perspective on the context-dependent nature of pupil signatures of information gain. This seems to be partly due to the strong diPerences between the experimental tasks.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

I highly appreciate the care and detail in the authors' response and thank them for the ePort invested in revising the manuscript. They addressed the core concerns to a high standard, and the manuscript has substantially improved in methodological rigour (through additional controls/sanity checks and complementary mixed-ePects analyses) and in clarity of interpretation (by explicitly acknowledging context-dependence and tempering stronger claims). The present version reads clearly and is much strengthened overall. I only have a few minor points below:

Minor suggestions:

Abstract:

In the abstract KL is introduced as abbreviation, but at first occurence it should be written out as "Kullback-Leibler (KL)" for readers not familiar with it.

We thank the reviewer for catching this error. It has been correct in the version of record.

Methods:

I appreciate the additional bayesian LME analysis. I only had a few things that I thought were missing from knowing the parameters: 1) what was the target acceptance rate (default of .95?), 2) which family was used to model the response distribution: (default) "gaussian" or robust "student-t"? Depending on the data a student-t would be preferred, but since the author's checked the fit & the results corroborate the correlation analysis, using the default would also be fine! Just add the information for completeness.

Thank you for bringing this to our attention. We have now noted that default parameters were used in all cases unless otherwise mentioned. 

Thank you once again for your time and consideration.

Reviewer #2 (Recommendations for the authors):

Thanks to the authors' ePort on revision. I am happy with this new version of manuscript.

Thank you once again for your time and consideration.

Reviewer #3 (Recommendations for the authors):

(1) Regarding comments #3 and #6 (first round) on model validation and posterior predictive checks, the authors replied that since their model is not a "generative" one, they can't perform posterior predictive checks. Crucially, in eq. 2, the authors present the p{tilde}^j_k variable denoting the learned probability of event k on trial j. I don't see why this can't be exploited for simulations. In my opinion, one could (and should) generate predictions based on this variable. The simplest implementation would translate the probability into a categorical choice (w/o fitting any free parameter). Based on this, they could assess whether the model and data are comparable.

We thank the reviewer for this clarification. The reviewer suggests using the probability distributions at each trial to predict which event should be chosen on each trial. More specifically, the event(s) with the highest probability on trial j could be used to generate a prediction for the choice of the participant on trial j. We agree that this would indeed be an interesting analysis. However, the response options of each task are limited to two-alternatives. In the cue-target task, four events are modeled (representing all possible cue-target conditions) while the participants’ response options are only “left” and “right”. Similarly, in the letter-color task, 36 events are modeled while the participants’ response options are “match” and “no-match”. In other words, we do not know which event (either four or 36, for the two tasks) the participant would have indicated on each trial. As an approximation to this fine-grained analysis, we investigated the relationship between the information-theoretic variables separately for error and correct trials. Our rationale was that we would have more insight into how the model fits depended on the participants’ actual behavior as compared with the ideal learner model.

(2) I recommend providing a plot of the linear mixed model analysis of the pupil data. Currently, results are only presented in the text and tables, but a figure would be much more useful.

We thank the reviewer for the suggestion to add a plot of the linear mixed model results. We appreciate the value of visualizing model estimates; however, we feel that the current presentation in the text and tables clearly conveys the relevant findings. For this reason, and to avoid further lengthening the manuscript, we prefer to retain the current format.

(3) I would consider only presenting the linear mixed ePects for the pupil data in the main results, and the correlation results in the supplement. It is currently quite long.

We thank the reviewer for this recommendation. We agree that the results section is detailed; however, we consider the correlation analyses to be integral to the interpretation of the pupil data and therefore prefer to keep them in the main text rather than move them to the supplement.

The following is the authors’ response to the original reviews

eLife Assessment

This important study seeks to examine the relationship between pupil size and information gain, showing opposite effects dependent upon whether the average uncertainty increases or decreases across trials. Given the broad implications for learning and perception, the findings will be of broad interest to researchers in cognitive neuroscience, decision-making, and computational modelling. Nevertheless, the evidence in support of the particular conclusion is at present incomplete - the conclusions would be strengthened if the authors could both clarify the differences between model-updating and prediction error in their account and clarify the patterns in the data.

Public Reviews:

Reviewer #1 (Public review):

Summary:

This study investigates whether pupil dilation reflects prediction error signals during associative learning, defined formally by Kullback-Leibler (KL) divergence, an information-theoretic measure of information gain. Two independent tasks with different entropy dynamics (decreasing and increasing uncertainty) were analyzed: the cue-target 2AFC task and the lettercolor 2AFC task. Results revealed that pupil responses scaled with KL divergence shortly after feedback onset, but the direction of this relationship depended on whether uncertainty (entropy) increased or decreased across trials. Furthermore, signed prediction errors (interaction between frequency and accuracy) emerged at different time windows across tasks, suggesting taskspecific temporal components of model updating. Overall, the findings highlight that pupil dilation reflects information-theoretic processes in a complex, context-dependent manner.

Strengths:

This study provides a novel and convincing contribution by linking pupil dilation to informationtheoretic measures, such as KL divergence, supporting Zénon's hypothesis that pupil responses reflect information gained during learning. The robust methodology, including two independent datasets with distinct entropy dynamics, enhances the reliability and generalisability of the findings. By carefully analysing early and late time windows, the authors capture the temporal dynamics of prediction error signals, offering new insights into the timing of model updates. The use of an ideal learner model to quantify prediction errors, surprise, and entropy provides a principled framework for understanding the computational processes underlying pupil responses. Furthermore, the study highlights the critical role of task context - specifically increasing versus decreasing entropy - in shaping the directionality and magnitude of these effects, revealing the adaptability of predictive processing mechanisms.

Weaknesses:

While this study offers important insights, several limitations remain. The two tasks differ significantly in design (e.g., sensory modality and learning type), complicating direct comparisons and limiting the interpretation of differences in pupil dynamics. Importantly, the apparent context-dependent reversal between pupil constriction and dilation in response to feedback raises concerns about how these opposing effects might confound the observed correlations with KL divergence. 

We agree with the reviewer’s concerns and acknowledge that the speculation concerning the directional effect of entropy across trials can not be fully substantiated by the current study. As the reviewer points out, the directional relationship between pupil dilation and information gain must be due to other factors, for instance, the sensory modality, learning type, or the reversal between pupil constriction and dilation across the two tasks. Also, we would like to note that ongoing experiments in our lab already contradict our original speculation. In line with the reviewer’s point, we noted these differences in the section on “Limitations and future research” in the Discussion. To better align the manuscript with the above mentioned points, we have made several changes in the Abstract, Introduction and Discussion summarized below: 

We have removed the following text from the Abstract and Introduction: “…, specifically related to increasing or decreasing average uncertainty (entropy) across trials.”

We have edited the following text in the Introduction (changes in italics) (p. 5):

“We analyzed two independent datasets featuring distinct associative learning paradigms, one characterized by increasing entropy and the other by decreasing entropy as the tasks progressed. By examining these different tasks, we aimed to identify commonalities (if any) in the results across varying contexts. Additionally, the contrasting directions of entropy in the two tasks enabled us to disentangle the correlation between stimulus-pair frequency and information gain in the postfeedback pupil response.

We have removed the following text from the Discussion:

“…and information gain in fact seems to be driven by increased uncertainty.”

“We speculate that this difference in the direction of scaling between information gain and the pupil response may depend on whether entropy was increasing or decreasing across trials.” 

“…which could explain the opposite direction of the relationship between pupil dilation and information gain”

“… and seems to relate to the direction of the entropy as learning progresses (i.e., either increasing or decreasing average uncertainty).” 

We have edited the following texts in the Discussion (changes in italics):

“For the first time, we show that the direction of the relationship between postfeedback pupil dilation and information gain (defined as KL divergence) was context dependent.” (p. 29):

Finally, we have added the following correction to the Discussion (p. 30):

“Although it is tempting to speculate that the direction of the relationship between pupil dilation and information gain may be due to either increasing or decreasing entropy as the task progressed, we must refrain from this conclusion. We note that the two tasks differ substantially in terms of design with other confounding variables and therefore cannot be directly compared to one another. We expand on these limitations in the section below (see Limitations and future research).”

Finally, subjective factors such as participants' confidence and internal belief states were not measured, despite their potential influence on prediction errors and pupil responses.

Thank you for the thoughtful comment. We agree with the reviewer that subjective factors, such as participants' confidence, can be important in understanding prediction errors and pupil responses. As per the reviewer’s point, we have included the following limitation in the Discussion (p. 33): 

“Finally, while we acknowledge the potential relevance of subjective factors, such as the participants’ overt confidence reports, in understanding prediction errors and pupil responses, the current study focused on the more objective, model-driven measure of information-theoretic variables. This approach aligns with our use of the ideal learner model, which estimates information-theoretic variables while being agnostic about the observer's subjective experience itself. Future research is needed to explore the relationship between information-gain signals in pupil dilation and the observer’s reported experience of or awareness about confidence in their decisions.” 

Reviewer #2 (Public review):

Summary:

The authors proposed that variability in post-feedback pupillary responses during the associative learning tasks can be explained by information gain, which is measured as KL divergence. They analysed pupil responses in a later time window (2.5s-3s after feedback onset) and correlated them with information-theory-based estimates from an ideal learner model (i.e., information gain-KL divergence, surprise-subjective probability, and entropy-average uncertainty) in two different associative decision-making tasks.

Strength:

The exploration of task-evoked pupil dynamics beyond the immediate response/feedback period and then associating them with model estimates was interesting and inspiring. This offered a new perspective on the relationship between pupil dilation and information processing.

Weakness:

However, disentangling these later effects from noise needs caution. Noise in pupillometry can arise from variations in stimuli and task engagement, as well as artefacts from earlier pupil dynamics. The increasing variance in the time series of pupillary responses (e.g., as shown in Figure 2D) highlights this concern.

It's also unclear what this complicated association between information gain and pupil dynamics actually means. The complexity of the two different tasks reported made the interpretation more difficult in the present manuscript.

We share the reviewer’s concerns. To make this point come across more clearly, we have added the following text to the Introduction (p. 5):

“The current study was motivated by Zenon’s hypothesis concerning the relationship between pupil dilation and information gain, particularly in light of the varying sources of signal and noise introduced by task context and pupil dynamics. By demonstrating how task context can influence which signals are reflected in pupil dilation, and highlighting the importance of considering their temporal dynamics, we aim to promote a more nuanced and model-driven approach to cognitive research using pupillometry.”

Reviewer #3 (Public review):

Summary:

This study examines prediction errors, information gain (Kullback-Leibler [KL] divergence), and uncertainty (entropy) from an information-theory perspective using two experimental tasks and pupillometry. The authors aim to test a theoretical proposal by Zénon (2019) that the pupil response reflects information gain (KL divergence). In particular, the study defines the prediction error in terms of KL divergence and speculates that changes in pupil size associated with KL divergence depend on entropy. Moreover, the authors examine the temporal characteristics of pupil correlates of prediction errors, which differed considerably across previous studies that employed different experimental paradigms. In my opinion, the study does not achieve these aims due to several methodological and theoretical issues.

Strengths:

(1)  Use of an established Bayesian model to compute KL divergence and entropy.

(2)  Pupillometry data preprocessing, including deconvolution.

Weaknesses:

(1) Definition of the prediction error in terms of KL divergence:

I'm concerned about the authors' theoretical assumption that the prediction error is defined in terms of KL divergence. The authors primarily refer to a review article by Zénon (2019): "Eye pupil signals information gain". It is my understanding that Zénon argues that KL divergence quantifies the update of a belief, not the prediction error: "In short, updates of the brain's internal model, quantified formally as the Kullback-Leibler (KL) divergence between prior and posterior beliefs, would be the common denominator to all these instances of pupillary dilation to cognition." (Zénon, 2019).

From my perspective, the update differs from the prediction error. Prediction error refers to the difference between outcome and expectation, while update refers to the difference between the prior and the posterior. The prediction error can drive the update, but the update is typically smaller, for example, because the prediction error is weighted by the learning rate to compute the update. My interpretation of Zénon (2019) is that they explicitly argue that KL divergence defines the update in terms of the described difference between prior and posterior, not the prediction error.

The authors also cite a few other papers, including Friston (2010), where I also could not find a definition of the prediction error in terms of KL divergence. For example [KL divergence:] "A non-commutative measure of the non-negative difference between two probability distributions." Similarly, Friston (2010) states: Bayesian Surprise - "A measure of salience based on the Kullback-Leibler divergence between the recognition density (which encodes posterior beliefs) and the prior density. It measures the information that can be recognized in the data." Finally, also in O'Reilly (2013), KL divergence is used to define the update of the internal model, not the prediction error.

The authors seem to mix up this common definition of the model update in terms of KL divergence and their definition of prediction error along the same lines. For example, on page 4: "KL divergence is a measure of the difference between two probability distributions. In the context of predictive processing, KL divergence can be used to quantify the mismatch between the probability distributions corresponding to the brain's expectations about incoming sensory input and the actual sensory input received, in other words, the prediction error (Friston, 2010; Spratling, 2017)."

Similarly (page 23): "In the current study, we investigated whether the pupil's response to decision outcome (i.e., feedback) in the context of associative learning reflects a prediction error as defined by KL divergence."

This is problematic because the results might actually have limited implications for the authors' main perspective (i.e., that the pupil encodes prediction errors) and could be better interpreted in terms of model updating. In my opinion, there are two potential ways to deal with this issue:

(a) Cite work that unambiguously supports the perspective that it is reasonable to define the prediction error in terms of KL divergence and that this has a link to pupillometry. In this case, it would be necessary to clearly explain the definition of the prediction error in terms of KL divergence and dissociate it from the definition in terms of model updating.

(b) If there is no prior work supporting the authors' current perspective on the prediction error, it might be necessary to revise the entire paper substantially and focus on the definition in terms of model updating.

We thank the reviewer for pointy out these inconsistencies in the manuscript and appreciate their suggestions for improvement. We take approach (a) recommended by the reviewer, and provide our reasoning as to why prediction error signals in pupil dilation are expected to correlate with information gain (defined as the KL divergence between posterior and prior belief distributions). This can be found in a new section in the introduction, copied here for convenience (p. 3-4):

“We reasoned that the link between prediction error signals and information gain in pupil dilation is through precision-weighting. Precision refers to the amount of uncertainty (inverse variance) of both the prior belief and sensory input in the prediction error signals [6,64–67]. More precise prediction errors receive more weighting, and therefore, have greater influence on model updating processes. The precisionweighting of prediction error signals may provide a mechanism for distinguishing between known and unknown sources of uncertainty, related to the inherent stochastic nature of a signal versus insufficient information of the part of the observer, respectively [65,67,68]. In Bayesian frameworks, information gain is fundamentally linked to prediction error, modulated by precision [65,66,69–75]. In non-hierarchical Bayesian models, information gain can be derived as a function of prediction errors and the precision of the prior and likelihood distributions, a relationship that can be approximately linear [70]. In hierarchical Bayesian inference, the update in beliefs (posterior mean changes) at each level is proportional to the precision-weighted prediction error; this update encodes the information gained from new observations [65,66,69,71,72]. Neuromodulatory arousal systems are well-situated to act as precision-weighting mechanisms in line with predictive processing frameworks [76,77]. Empirical evidence suggests that neuromodulatory systems broadcast precisionweighted prediction errors to cortical regions [11,59,66,78]. Therefore, the hypothesis that feedback-locked pupil dilation reflects a prediction error signal is similarly in line with Zenon’s main claim that pupil dilation generally reflects information gain, through precision-weighting of the prediction error. We expected a prediction error signal in pupil dilation to be proportional to the information gain.”

We have referenced previous work that has linked prediction error and information gain directly (p. 4): “The KL divergence between posterior and prior belief distributions has been previously considered to be a proxy of (precision-weighted) prediction errors [68,72].”

We have taken the following steps to remedy this error of equating “prediction error” directly with the information gain.

First, we have replaced “KL divergence” with “information gain” whenever possible throughout the manuscript for greater clarity. 

Second, we have edited the section in the introduction defining information gain substantially (p. 4): 

“Information gain can be operationalized within information theory as the KullbackLeibler (KL) divergence between the posterior and prior belief distributions of a Bayesian observer, representing a formalized quantity that is used to update internal models [29,79,80]. Itti and Baldi (2005)81 termed the KL divergence between posterior and prior belief distributions as “Bayesian surprise” and showed a link to the allocation of attention. The KL divergence between posterior and prior belief distributions has been previously considered to be a proxy of (precision-weighted) prediction errors[68,72]. According to Zénon’s hypothesis, if pupil dilation reflects information gain during the observation of an outcome event, such as feedback on decision accuracy, then pupil size will be expected to increase in proportion to how much novel sensory evidence is used to update current beliefs [29,63]. ” 

Finally, we have made several minor textual edits to the Abstract and main text wherever possible to further clarify the proposed relationship between prediction errors and information gain.

(2) Operationalization of prediction errors based on frequency, accuracy, and their interaction:

The authors also rely on a more model-agnostic definition of the prediction error in terms of stimulus frequency ("unsigned prediction error"), accuracy, and their interaction ("signed prediction error"). While I see the point here, I would argue that this approach offers a simple approximation to the prediction error, but it is possible that factors like difficulty and effort can influence the pupil signal at the same time, which the current approach does not take into account. I recommend computing prediction errors (defined in terms of the difference between outcome and expectation) based on a simple reinforcement-learning model and analyzing the data using a pupillometry regression model in which nuisance regressors are controlled, and results are corrected for multiple comparisons.

We agree with the reviewer’s suggestion that alternatively modeling the data in a reinforcement learning paradigm would be fruitful. We adopted the ideal learner model as we were primarily focused on Information Theory, stemming from our aim to test Zenon’s hypothesis that information gain drives pupil dilation. However, we agree with the reviewer that it is worthwhile to pursue different modeling approaches in future work. We have now included a complementary linear mixed model analysis in which we controlled for the effects of the information-theoretic variables on one another, while also including the nuisance regressors of pre-feedback baseline pupil dilation and reaction times (explained in more detail below in our response to your point #4). Results including correction for multiple comparisons was reported for all pupil time course data as detailed in Methods section 2.5. 

(3) The link between model-based (KL divergence) and model-agnostic (frequency- and accuracy-based) prediction errors:

I was expecting a validation analysis showing that KL divergence and model-agnostic prediction errors are correlated (in the behavioral data). This would be useful to validate the theoretical assumptions empirically.

The model limitations and the operalization of prediction error in terms of post-feedback processing do not seem to allow for a comparison of information gain and model-agnostic prediction errors in the behavioral data for the following reasons. First, the simple ideal learner model used here is not a generative model, and therefore, cannot replicate or simulate the participants responses (see also our response to your point #6 “model validation” below). Second, the behavioral dependent variables obtained are accuracy and reaction times, which both occur before feedback presentation. While accuracy and reaction times can serve as a marker of the participant’s (statistical) confidence/uncertainty following the decision interval, these behavioral measures cannot provide access to post-feedback information processing. The pupil dilation is of interest to us because the peripheral arousal system is able to provide a marker of post-feedback processing. Through the analysis presented in Figure 3, we indeed aimed to make the comparison of the model-based information gain to the model-agnostic prediction errors via the proxy variable of post-feedback pupil dilation instead of behavioral variables. To bridge the gap between the “behaviorally agnostic” model parameters and the actual performance of the participants, we examined the relationship between the model-based information gain and the post-feedback pupil dilation separately for error and correct trials as shown in Figure 3D-F & Figure 3J-L. We hope this addresses the reviewers concern and apologize in case we did not understand the reviewers suggestion here.

(4) Model-based analyses of pupil data:

I'm concerned about the authors' model-based analyses of the pupil data. The current approach is to simply compute a correlation for each model term separately (i.e., KL divergence, surprise, entropy). While the authors do show low correlations between these terms, single correlational analyses do not allow them to control for additional variables like outcome valence, prediction error (defined in terms of the difference between outcome and expectation), and additional nuisance variables like reaction time, as well as x and y coordinates of gaze.

Moreover, including entropy and KL divergence in the same regression model could, at least within each task, provide some insights into whether the pupil response to KL divergence depends on entropy. This could be achieved by including an interaction term between KL divergence and entropy in the model.

In line with the reviewer’s suggestions, we have included a complementary linear mixed model analysis in which we controlled for the effects of the information-theoretic variables on one another, while also including the nuisance regressors of pre-feedback baseline pupil dilation and reaction times. We compared the performance of two models on the post-feedback pupil dilation in each time window of interest: Modle 1 had no interaction between information gain and entropy and Model 2 included an interaction term as suggested. We did not include the x- and y- coordinates of gaze in the mixed linear model analysis, as there are multiple values of these coordinates per trial. Furthermore, regressing out the x and y- coordinates of gaze can potentially remove signal of interest in the pupil dilation data in addition to the gaze-related confounds and we did not measure absolute pupil size (Mathôt, Melmi & Castet, 2015; Hayes & Petrov, 2015). We present more sanity checks on the pre-processing pipeline as recommended by Reviewer 1.  

This new analysis resulted in several additions to the Methods (see Section 2.5) and Results. In sum, we found that including an interaction term for information gain and entropy did not lead to better model fits, but sometimes lead to significantly worse fits. Overall, the results of the linear mixed model corroborated the “simple” correlation analysis across the pupil time course while accounting for the relationship to the pre-feedback baseline pupil and preceeding reaction time differences. There was only one difference to note between the correlation and linear mixed modeling analyses: for the error trials in the cue-target 2AFC task, including entropy in the model accounted for the variance previously explained by surprise.

(5) Major differences between experimental tasks:

More generally, I'm not convinced that the authors' conclusion that the pupil response to KL divergence depends on entropy is sufficiently supported by the current design. The two tasks differ on different levels (stimuli, contingencies, when learning takes place), not just in terms of entropy. In my opinion, it would be necessary to rely on a common task with two conditions that differ primarily in terms of entropy while controlling for other potentially confounding factors. I'm afraid that seemingly minor task details can dramatically change pupil responses. The positive/negative difference in the correlation with KL divergence that the authors interpret to be driven by entropy may depend on another potentially confounding factor currently not controlled.

We agree with the reviewer’s concerns and acknowledge that the speculation concerning the directional effect of entropy across trials can not be fully substantiated by the currect study. We note that Review #1 had a similar concern. Our response to Reviewer #1 addresses this concern of Reviewer #3 as well. To better align the manuscript with the above mentioned points, we have made several changes that are detailed in our response to Reviewer #1’s public review (above). 

(6) Model validation:

My impression is that the ideal learner model should work well in this case. However, the authors don't directly compare model behavior to participant behavior ("posterior predictive checks") to validate the model. Therefore, it is currently unclear if the model-derived terms like KL divergence and entropy provide reasonable estimates for the participant data.

Based on our understanding, posterior predictive checks are used to assess the goodness of fit between generated (or simulated) data and observed data. Given that the “simple” ideal learner model employed in the current study is not a generative model, a posterior predictive check would not apply here (Gelman, Carlin, Stern, Dunson, Vehtari, & Rubin (2013). The ideal learner model is unable to simulate or replicate the participants’ responses and behaviors such as accuracy and reaction times; it simply computes the probability of seeing each stimulus type at each trial based on the prior distribution and the exact trial order of the stimuli presented to each participant. The model’s probabilities are computed directly from a Dirichlet distribution of values that represent the number of occurences of each stimulus-pair type for each task. The information-theoretic variables are then directly computed from these probabilities using standard formulas. The exact formulas used in the ideal learner model can be found in section 2.4.

We have now included a complementary linear mixed model analysis which also provides insight into the amount of explained variance of these information-theoretic predictors on the post-feedback pupil response, while also including the pre-feedback baseline pupil and reaction time differences (see section 3.3, Tables 3 & 4). The R² values ranged from 0.16 – 0.50 across all conditions tested.

(7) Discussion:

The authors interpret the directional effect of the pupil response w.r.t. KL divergence in terms of differences in entropy. However, I did not find a normative/computational explanation supporting this interpretation. Why should the pupil (or the central arousal system) respond differently to KL divergence depending on differences in entropy?

The current suggestion (page 24) that might go in this direction is that pupil responses are driven by uncertainty (entropy) rather than learning (quoting O'Reilly et al. (2013)). However, this might be inconsistent with the authors' overarching perspective based on Zénon (2019) stating that pupil responses reflect updating, which seems to imply learning, in my opinion. To go beyond the suggestion that the relationship between KL divergence and pupil size "needs more context" than previously assumed, I would recommend a deeper discussion of the computational underpinnings of the result.

Since we have removed the original speculative conclusion from the manuscript, we will refrain from discussing the computational underpinnings of a potential mechanism. To note as mentioned above, we have preliminary data from our own lab that contradicts our original hypothesis about the relationship between entropy and information gain on the post-feedback pupil response. 

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Apart from the points raised in the public review above, I'd like to use the opportunity here to provide a more detailed review of potential issues, questions, and queries I have:

(1) Constriction vs. Dilation Effects:

The study observes a context-dependent relationship between KL divergence and pupil responses, where pupil dilation and constriction appear to exhibit opposing effects. However, this phenomenon raises a critical concern: Could the initial pupil constriction to visual stimuli (e.g., in the cue-target task) confound correlations with KL divergence? This potential confound warrants further clarification or control analyses to ensure that the observed effects genuinely reflect prediction error signals and are not merely a result of low-level stimulus-driven responses.

We agree with the reviewers concern and have added the following information to the limitations section in the Discussion (changes in italics below; p. 32-33).

“First, the two associative learning paradigms differed in many ways and were not directly comparable. For instance, the shape of the mean pupil response function differed across the two tasks in accordance with a visual or auditory feedback stimulus (compare Supplementary Figure 3A with Supplementary Figure 3D), and it is unclear whether these overall response differences contributed to any differences obtained between task conditions within each task. We are unable to rule out whether so-called “low level” effects such as the initial constriction to visual stimuli in the cue-target 2AFC task as compared with the dilation in response auditory stimuli in letter-color 2AFC task could confound correlations with information gain. Future work should strive to disentangle how the specific aspects of the associative learning paradigms relate to prediction errors in pupil dilation by systematically manipulating design elements within each task.”

Here, I also was curious about Supplementary Figure 1, showing 'no difference' between the two tones (indicating 'error' or 'correct'). Was this the case for FDR-corrected or uncorrected cluster statistics? Especially since the main results also showed sig. differences only for uncorrected cluster statistics (Figure 2), but were n.s. for FDR corrected. I.e. can we be sure to rule out a confound of the tones here after all?

As per the reviewer’s suggestion, we verified that there were also no significant clusters after feedback onset before applying the correction for multiple comparisons. We have added this information to Supplemenatary section 1.2 as follows: 

“Results showed that the auditory tone dilated pupils on average (Supplementary Figure 1C). Crucially, however, the two tones did not differ from one another in either of the time windows of interest (Supplementary Figure 1D; no significant time points after feedback onset were obtained either before or after correcting for multiple comparisons using cluster-based permutation methods; see Section 2.5.” 

Supplementary Figure 1 is showing effects cluster-corrected for multiple comparisons using cluster-based permutation tests from the MNE software package in Python (see Methods section 2.5). We have clarified that the cluster-correction was based on permutation testing in the figure legend. 

(2) Participant-Specific Priors:

The ideal learner models do not account for individualised priors, assuming homogeneous learning behaviour across participants. Could incorporating participant-specific priors better reflect variability in how individuals update their beliefs during associative learning?

We have clarified in the Methods (see section 2.4) that the ideal learner models did account for participant-specific stimuli including participant-specific priors in the letter-color 2AFC task. We have added the following texts: 

“We also note that while the ideal learner model for the cue-target 2AFC task used a uniform (flat) prior distribution for all participants, the model parameters were based on the participant-specific cue-target counterbalancing conditions and randomized trial order.” (p. 13)

“The prior distributions used for the letter-color 2AFC task were estimated from the randomized letter-color pairs and randomized trial order presentation in the preceding odd-ball task; this resulted in participant-specific prior distributions for the ideal learner model of the letter-color 2AFC task. The model parameters were likewise estimated from the (participant-specific) randomized trial order presented in the letter-color 2AFC task.” (p. 13)

(3) Trial-by-Trial Variability:

The analysis does not account for random effects or inter-trial variability using mixed-effects models. Including such models could provide a more robust statistical framework and ensure the observed relationships are not influenced by unaccounted participant- or trial-specific factors.

We have included a complementary linear mixed model analysis in which “subject” was modeled as a random effect on the post-feedback pupil response in each time window of interest and for each task. Across all trials, the results of the linear mixed model corroborated the “simple” correlation analysis across the pupil time course while accounting for the relationship to the prefeedback baseline pupil and preceeding reaction time differences (see section 3.3, Tables 3 & 4).

(4) Preprocessing/Analysis choices:

Before anything else, I'd like to highlight the authors' effort in providing public code (and data) in a very readable and detailed format!

We appreciate the compliment - thank you for taking the time to look at the data and code provided.

I found the idea of regressing the effect of Blinks/Saccades on the pupil trace intriguing. However, I miss a complete picture here to understand how well this actually worked, especially since it seems to be performed on already interpolated data. My main points here are:

(4.1) Why is the deconvolution performed on already interpolated data and not on 'raw' data where there are actually peaks of information to fit?

To our understanding, at least one critical reason for interpolating the data before proceeding with the deconvolution analysis is that the raw data contain many missing values (i.e., NaNs) due to the presence of blinks. Interpolating over the missing data first ensures that there are valid numerical elements in the linear algebra equations. We refer the reviewer to the methods detailed in Knapen et al. (2016) for more details on this pre-processing method. 

(4.2) What is the model fit (e.g. R-squared)? If this was a poor fit for the regressors in the first place, can we trust the residuals (i.e. clean pupil trace)? Is it possible to plot the same Pupil trace of Figure 1D with a) the 'raw' pupil time-series, b) after interpolation only (both of course also mean-centered for comparison), on top of the residuals after deconvolution (already presented), so we can be sure that this is not driving the effects in a 'bad' way? I'd just like to make sure that this approach did not lead to artefacts in the residuals rather than removing them.

We thank the reviewer for this suggestion. In the Supplementary Materials, we have included a new figure (Supplementary Figure 2, copied below for convience), which illustrates the same conditions as in Figure 1D and Figure 2D, with 1) the raw data, and 2) the interpolated data before the nuisance regression. Both the raw data and interpolated data have been band-pass filtered as was done in the original pre-processing pipeline and converted to percent signal change. These figures can be compared directly to Figure 1D and Figure 2D, for the two tasks, respectively. 

Of note is that the raw data seem to be dominated by responses to blinks (and/or saccades). Crucially, the pattern of results remains overall unchaged between the interpolated-only and fully pre-processed version of the data for both tasks. 

In the Supplementary Materials (see Supplementary section 2), we have added the descriptives of the model fits from the deconvolution method. Model fits (R²) for the nuisance regression were generally low: cue-target 2AFC task, M = 0.03, SD = 0.02, range = [0.00, 0.07]; letter-color visual 2AFC, M = 0.08, SD = 0.04, range = [0.02, 0.16].

Furthermore, a Pearson correlation analysis between the interpolated and fully pre-processed data within the time windows of interest for both task indicated high correspondence: 

Cue-target 2AFC task

Early time window: M = 0.99, SD = 0.01, range = [0.955, 1.000]

Late time window: M = 0.99, SD = 0.01, range = [0.971, 1.000]

Letter-color visual 2AFC

Early time window: M = 0.95, SD = 0.04, range = [0.803, 0.998]

Late time window: M = 0.97, SD = 0.02, range = [0.908, 0.999]

In hindsight, including the deconvolution (nuisance regression) method may not have changed the pattern of results much. However, the decision to include this deconvolution method was not data-driven; instead, it was based on the literature establishing the importance of removing variance (up to 5 s) of these blinks and saccades from cognitive effects of interest in pupil dilation (Knapen et al., 2016). 

(4.3) Since this should also lead to predicted time series for the nuisance-regressors, can we see a similar effect (of what is reported for the pupil dilation) based on the blink/saccade traces of a) their predicted time series based on the deconvolution, which could indicate a problem with the interpretation of the pupil dilation effects, and b) the 'raw' blink/saccade events from the eye-tracker? I understand that this is a very exhaustive analysis so I would actually just be interested here in an averaged time-course / blink&saccade frequency of the same time-window in Figure 1D to complement the PD analysis as a sanity check.

Also included in the Supplementary Figure 2 is the data averaged as in Figure 1D and Figure 2D for the raw data and nuisance-predictor time courses (please refer to the bottom row of the sub-plots). No pattern was observed in either the raw data or the nuisance predictors as was shown in the residual time courses. 

(4.4) How many samples were removed from the time series due to blinks/saccades in the first place? 150ms for both events in both directions is quite a long bit of time so I wonder how much 'original' information of the pupil was actually left in the time windows of interest that were used for subsequent interpretations.

We thank the reviewer for bringing this issue to our attention. The size of the interpolation window was based on previous literature, indicating a range of 100-200 ms as acceptable (Urai et al., 2017; Knapen et al., 2016; Winn et al., 2018). The ratio of interpolated-to-original data (across the entire trial) varied greatly between participants and between trials: cue-target 2AFC task, M = 0.262, SD = 0.242, range = [0,1]; letter-color 2AFC task, M = 0.194, SD = 0.199, range = [0,1]. 

We have now included a conservative analysis in which only trials with more than half (threshold = 60%) of original data are included in the analyses. Crucially, we still observe the same pattern of effects as when all data are considered across both tasks (compare the second to last row in the Supplementary Figure 2 to Figure 1D and Figure 2D).

(4.5) Was the baseline correction performed on the percentage change unit?

Yes, the baseline correction was performed on the pupil timeseries after converting to percentsignal change. We have added that information to the Methods (section 2.3).

(4.6) What metric was used to define events in the derivative as 'peaks'? I assume some sort of threshold? How was this chosen?

The threshold was chosen in a data-driven manner and was kept consistent across both tasks. The following details have been added to the Methods:

“The size of the interpolation window preceding nuisance events was based on previous literature [13,39,99]. After interpolation based on data-markers and/or missing values, remaining blinks and saccades were estimated by testing the first derivative of the pupil dilation time series against a threshold rate of change. The threshold for identifying peaks in the temporal derivative is data-driven, partially based on past work[10,14,33]. The output of each participant’s pre-processing pipeline was checked visually. Once an appropriate threshold was established at the group level, it remained the same for all participants (minimum peak height of 10 units).” (p. 8 & 11).

(5) Multicollinearity Between Variables:

Lastly, the authors state on page 13: "Furthermore, it is expected that these explanatory variables will be correlated with one another. For this reason, we did not adopt a multiple regression approach to test the relationship between the information-theoretic variables and pupil response in a single model". However, the very purpose of multiple regression is to account for and disentangle the contributions of correlated predictors, no? I might have missed something here.

We apologize for the ambiguity of our explanation in the Methods section. We originally sought to assess the overall relationship between the post-feedback response and information gain (primarily), but also surprise and entropy. Our reasoning was that these variables are often investigated in isolation across different experiments (i.e., only investigating Shannon surprise), and we would like to know what the pattern of results would look like when comparing a single information-theoretic variable to the pupil response (one-by-one). We assumed that including additional explanatory variables (that we expected to show some degree of collinearity with each other) in a regression model would affect variance attributed to them as compared with the one-on-one relationships observed with the pupil response (Morrissey & Ruxton 2018). We also acknowledge the value of a multiple regression approach on our data. Based on the suggestions by the reviewers we have included a complementary linear mixed model analysis in which we controlled for the effects of the information-theoretic variables on one another, while also including the nuisance regressors of pre-feedback baseline pupil dilation and reaction times.  

This new analysis resulted in several additions to the Methods (see Section 2.5) and Results (see Tables 3 and 4). Overall, the results of the linear mixed model corroborated the “simple” correlation analysis across the pupil time course while accounting for the relationship to the prefeedback baseline pupil and preceeding reaction time differences. There was only one difference to note between the correlation and linear mixed modeling analyses: for the error trials in the cue-target 2AFC task, including entropy in the model accounted for the variance previously explained by surprise. 

Reviewer #2 (Recommendations for the authors):

(1) Given the inherent temporal dependencies in pupil dynamics, characterising later pupil responses as independent of earlier ones in a three-way repeated measures ANOVA may not be appropriate. A more suitable approach might involve incorporating the earlier pupil response as a covariate in the model.

We thank the reviewer for bringing this issue to our attention. From our understanding, a repeated-measures ANOVA with factor “time window” would be appropriate in the current context for the following reasons. First, autocorrelation (closely tied to sphericity) is generally not considered a problem when only two timepoints are compared from time series data (Field, 2013; Tabachnick & Fidell, 2019). Second, the repeated-measures component of the ANOVA takes the correlated variance between time points into account in the statistical inference. Finally, as a complementary analysis, we present the results testing the interaction between the frequency and accuracy conditions across the full time courses (see Figures 1D and 2D); in these pupil time courses, any difference between the early and late time windows can be judged by the reader visually and qualitatively. 

(2) Please clarify the correlations between KL divergence, surprise, entropy, and pupil response time series. Specifically, state whether these correlations account for the interrelationships between these information-theoretic measures. Given their strong correlations, partialing out these effects is crucial for accurate interpretation.

As mentioned above, based on the suggestions by the reviewers we have included a complementary linear mixed model analysis in which we controlled for the effects of the information-theoretic variables on one another, while also including the nuisance regressors of pre-feedback baseline pupil dilation and reaction times.  

This new analysis resulted in several additions to the Methods (see Section 2.5) and Results (see Tables 3 and 4). Overall, the results of the linear mixed model corroborated the “simple” correlation analysis across the pupil time course while accounting for the relationship to the prefeedback baseline pupil and preceeding reaction time differences. There was only one difference to note between the correlation and linear mixed modeling analyses: for the error trials in the cue-target 2AFC task, including entropy in the model accounted for the variance previously explained by surprise. 

(3) The effects observed in the late time windows appear weak (e.g., Figure 2E vs. 2F, and the generally low correlation coefficients in Figure 3). Please elaborate on the reliability and potential implications of these findings.

We have now included a complementary linear mixed model analysis which also provides insight into the amount of explained variance of these information-theoretic predictors on the post-feedback pupil response, while also including the pre-feedback baseline pupil and reaction time differences (see section 3.3, Tables 3 & 4). The R² values ranged from 0.16 – 0.50 across all conditions tested. Including the pre-feedback baseline pupil dilation as a predictor in the linear mixed model analysis consistently led to more explained variance in the post-feedback pupil response, as expected.  

(4) In Figure 3 (C-J), please clarify how the trial-by-trial correlations were computed (averaged across trials or subjects). Also, specify how the standard error of the mean (SEM) was calculated (using the number of participants or trials).

The trial-by-trial correlations between the pupil signal and model parameters were computed for each participant, then the coefficients were averaged across participants for statistical inference. We have added several clarifications in the text (see section 2.5 and legends of Figure 3 and Supplementary Figure 4).

We have added “the standard error of the mean across participants” to all figure labels.

(5) For all time axes (e.g., Figure 2D), please label the ticks at 0, 0.5, 1, 1.5, 2, 2.5, and 3 seconds. Clearly indicate the duration of the feedback on the time axes. This is particularly important for interpreting the pupil dilation responses evoked by auditory feedback.

We have labeled the x-ticks every 0.5 seconds in all figures and indicated the duration of the auditory feedback in the letter-color decision task and as well as the stimuli presented in the control tasks in the Supplementary Materials. 

Reviewer #3 (Recommendations for the authors):

(1) Introduction page 3: "In information theory, information gain quantifies the reduction of uncertainty about a random variable given the knowledge of another variable. In other words, information gain measures how much knowing about one variable improves the prediction or understanding of another variable."

(2) In my opinion, the description of information gain can be clarified. Currently, it is not very concrete and quite abstract. I would recommend explaining it in the context of belief updating.

We have removed these unclear statements in the Introduction. We now clearly state the following:

“Information gain can be operationalized within information theory as the KullbackLeibler (KL) divergence between the posterior and prior belief distributions of a Bayesian observer, representing a formalized quantity that is used to update internal models [29,79,80].” (p. 4)

(3) Page 4: The inconsistencies across studies are described in extreme detail. I recommend shortening this part and summarizing the inconsistencies instead of listing all of the findings separately.

As per the reviewer’s recommendation, we have shortened this part of the introduction to summarize the inconsistencies in a more concise manner as follows: 

“Previous studies have shown different temporal response dynamics of prediction error signals in pupil dilation following feedback on decision outcome: While some studies suggest that the prediction error signals arise around the peak (~1 s) of the canonical impulse response function of the pupil [11,30,41,61,62,90], other studies have shown evidence that prediction error signals (also) arise considerably later with respect to feedback on choice outcome [10,25,32,41,62]. A relatively slower prediction error signal following feedback presentation may suggest deeper cognitive processing, increased cognitive load from sustained attention or ongoing uncertainty, or that the brain is integrating multiple sources of information before updating its internal model. Taken together, the literature on prediction error signals in pupil dilation following feedback on decision outcome does not converge to produce a consistent temporal signature.” (p. 5)

We would like to note some additional minor corrections to the preprint:

We have clarified the direction of the effect in Supplementary Figure 3 with the following: 

“Participants who showed a larger mean difference between the 80% as compared with the 20% frequency conditions in accuracy also showed smaller differences (a larger mean difference in magnitude in the negative direction) in pupil responses between frequency conditions (see Supplementary Figure 4).”

The y-axis labels in Supplementary Figure 3 were incorrect and have been corrected as the following: “Pupil responses (80-20%)”.

We corrected typos, formatting and grammatical mistakes when discovered during the revision process. Some minor changes were made to improve clarity. Of course, we include a version of the manuscript with Tracked Changes as instructed for consideration.

there has to be CZ texts, not "Flagi"

Odolnost

Reviewer #2 (Public review):

Summary

This study investigates the role of GMCL1 in regulating the mitotic surveillance pathway (MSP), a protective mechanism that activates p53 following prolonged mitosis. The authors identify a physical interaction between 53BP1 and GMCL1, but not with GMCL2. They propose that the ubiquitin ligase complex CRL3-GMCL1 targets 53BP1 for degradation during mitosis, thereby preventing the formation of the "mitotic stopwatch" complex (53BP1-USP28-p53) and subsequent p53 activation. The authors show that high GMCL1 expression correlates with resistance to paclitaxel in cancer cell lines that express wild-type p53. Importantly, loss of GMCL1 restores paclitaxel sensitivity in these cells, but not in p53-deficient lines. They propose that GMCL1 overexpression enables cancer cells to bypass MSP-mediated p53 activation, promoting survival despite mitotic stress. Targeting GMCL1 may thus represent a therapeutic strategy to re-sensitize resistant tumors to taxane-based chemotherapy.

Strengths

This manuscript presents potentially interesting observations. The major strength of this article is the identification of GMCL1 as 53BP1 interaction partner. The authors identified relevant domains and show that GMCL1 controls 53BP1 stability. The authors further show a potentially interesting link between GMCL1 status and sensitivity to Taxol.

Weaknesses

A major limitation of the original manuscript was that the functional relevance of GMCL1 in regulating 53BP1 within an appropriate model system was not clearly demonstrated. In the revised version, the authors attempt to address this point. However, the new experiment is insufficiently controlled, making it difficult to interpret the results. State-of-the-art approaches would typically rely on single-cell tracking to monitor cell fate following release from a moderately prolonged mitosis.

In contrast, the authors use a population-based assay, but the reported rescue from arrest is minimal. If the assay were functioning robustly, one would expect that nearly all cells depleted of USP28 or 53BP1 should have entered S-phase at a defined time after release. Thus, the very small rescue effect of siTP53BP1 suggests that the current assay is not suitable. It is also likely that release from a 16-hour mitotic arrest induces defects independent of the 53BP1-dependent p53 response.

Furthermore, the cell-cycle duration of RPE1 cells is less than 20 hours. It is therefore unclear why cells are released for 30 hours before analysis. At this time point, many cells are likely to have progressed into the next cell cycle, making it impossible to draw conclusions regarding the immediate consequences of prolonged mitosis. As a result, the experiment cannot be evaluated due to inadequate controls.

To strengthen this part of the study, I recommend that the authors first establish an assay that reliably rescues the mitotic-arrest-induced G1 block upon depletion of p53, 53BP1, or USP28. Once this baseline is validated, GMCL1 knockout can then be introduced to quantify its contribution to the response.

A broader conceptual issue is that the evidence presented does not form a continuous line of reasoning. For example, it is not demonstrated that GMCL1 interacts with or regulates 53BP1 in RPE1 cells-the system in which the limited functional experiments are conducted.

There are also a number of inconsistencies and issues with data presentation that need to be addressed:

(1) Figure 2C: p21 levels appear identical between GMCL1 KO and WT rescue. If GMCL1 regulates p53 through 53BP1, p21 should be upregulated in the KO.

(2) Figure 2A vs. 2C: GMCL1 KO affects chromatin-bound 53BP1 in Figure 2A, yet in Figure 2C it affects 53BP1 levels specifically in G1-phase cells. This discrepancy requires clarification.

(3) Figure 2C quantification: The three biological repeats show an unusual pattern, with one repeat's data points lying exactly between the other two. It is unclear what the line represents; please clarify.

(4) Figure nomenclature: Some abbreviations (e.g., FLAG-KI in Fig. 1F, WKE in Fig. 1C-D, ΔMFF in Fig. 1E) are not defined in the figure legends. All abbreviations must be explained.

(5) Figure 2D: Please indicate how many times the experiment was reproduced. Quantification with statistical testing would strengthen the result. Pull-downs of 53BP1 with calculation of the ubiquitinated/total ratio could also support the conclusion.

(6) Figures 3A and 3C: The G1 bars share the same color as the error bars, making the graphs difficult to interpret. Please adjust the color scheme.

Author response:

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

Public Reviews:

Reviewer #1(Public review):

In this manuscript, Pagano and colleagues test the idea that the protein GMCL1 functions as a substrate receptor for a Cullin RING 3 E3 ubiquitin ligase (CUL3) complex. Using a pulldown approach, they identify GMCL1 binding proteins, including the DNA damage scaffolding protein 53BP1. They then focus on the idea that GMCL1 recruits 53BP1 for CUL3-dependent ubiquitination, triggering subsequent proteasomal degradation of ubiquitinated 53BP1.

In addition to its DNA damage signalling function, in mitosis, 53BP1 is reported to form a stopwatch complex with the deubiquitinating enzyme USP28 and the transcription factor p53 (PMID: 38547292). These 53BP1-stopwatch complexes generated in mitosis are inherited by G1 daughter cells and help promote p53-dependent cell cycle arrest independent from DNA damage (PMID: 38547292). Several studies show that knockout of 53BP1 overcomes G1 cell cycle arrest after mitotic delays caused by anti-mitotic drugs or centrosome ablation (PMID: 27432897, 27432896). In this model, it is crucial that 53BP1 remains stable in mitosis and more stopwatch complex is formed after delayed mitosis.

Major concerns:

Pagano and coworkers suggest that 53BP1 levels can sometimes be suppressed in mitosis if the cells overexpress GMCL1. They carry out a bioinformatic analysis of available public data for p53 wild-type cancer cell lines resistant to the anti-mitotic drug paclitaxel and related compounds. Stratifying GMCL1 into low and high expression groups reveals a weak (p = 0.05 or ns) correlation with sensitivity to taxanes. It is unclear on what basis the authors claim paclitaxel-resistant and p53 wild-type cancer cell lines bypass the mitotic surveillance/timer pathway. They have not tested this. Figure 3 is a correlation assembled from public databases but has no experimental tests. Figure 4 looks at proliferation but not cell cycle progression or the length of mitosis. The main conclusions relating to cell cycle progression and specifically the link to mitotic delays are therefore not supported by experimental data. There is no imaging of the cell cycle or cell fate after mitotic delays, or analysis of where the cells arrest in the cell cycle. Most of the cell lines used have been reported to lack a functional mitotic surveillance pathway in the recent work by Meitinger. To support these conclusions, the stability of endogenous 53BP1 under different conditions in cells known to have a functional mitotic surveillance pathway needs to be examined. A key suggestion in the work is that the level of GMCL1 expression correlates with resistance to taxanes. For the mitotic surveillance pathway, the type of drug (nocodazole, taxol, etc) used to induce a delay isn't thought to be relevant, only the length of the delay. Do GMCL1-overexpressing cells show resistance to anti-mitotics in general?

We thank the reviewer for this insightful comment. We propose that GMCL1 promotes CUL3-dependent ubiquitination of 53BP1 during prolonged mitotic arrest, thereby facilitating its proteasome-dependent degradation. To evaluate the potential clinical relevance of this mechanism, we stratified cancer cell lines based on GMCL1 mRNA expression using publicly available datasets from DepMap (PMID: 39468210). We observed correlations between GMCL1 expression levels and taxane sensitivity that appear to reflect specific cancer type-drug combinations. To experimentally evaluate this correlation and obtain mechanistic insights, we performed knockdown experiments in hTERT-RPE1 cells, which are known to possess an intact mitotic surveillance pathway. Silencing of GMCL1 alone inhibited cell proliferation and induced apoptosis, while co-depletion of either TP53BP1 or USP28 significantly rescued these effects. These results suggest that GMCL1 modulates the stability of 53BP1 and therefore the availability of the 53BP1-USP28-p53 ternary complex in cells with a functional mitotic surveillance pathway (MSP) (new Figure 5I,J) directly linking GMCL1 to the regulation of the MSP complex. Moreover, to further support our mechanism, we assessed the effect of GMCL1 levels on cell cycle progression. Briefly, following nocodazole synchronization and release, we treated cells with EdU and performed FACS analyses at different times. Knockdown of GMCL1 alone led to a delayed cell cycle progression, but co-depletion of either TP53BP1 or USP28 restored this phenotype (new Figure 3A and new Supplementary Figure 3A-C). These results are consistent with our proliferation data and suggest that the observed effects of GMCL1 are specific to mitotic exit. Finally, overexpression of GMCL1 accelerates cell cycle progression (as assessed by FACS analyses) upon release from prolonged mitotic arrest (new Figure 3B and new Supplementary Figure 3D-E). 

Importantly, if GMCL1 specifically degrades 53BP1 during prolonged mitotic arrests, the authors should show what happens during normal cell divisions without any delays or drug treatments. How much 53BP1 is destroyed in mitosis under those conditions? Does 53BP1 destruction depend on the length of mitosis, drug treatment, or does 53BP1 get degraded every mitosis regardless of length? Testing the contribution of key mitotic E3 ligase activities on mitotic 53BP1 stability, such as the anaphase-promoting complex/cyclosome (APC/C) is important in this regard. One previous study reported an analysis of putative APC/C KEN-box degron motifs in 53BP1 and concluded these play a role in 53BP1 stability in anaphase (PMID: 28228263).

Physiological mitosis under unperturbed conditions is typically brief (approximately 30 minutes), making protein quantification during this window challenging. Despite this, we tried by synchronizing cells using RO-3306 and releasing them into drug-free medium to assess GMCL1 dynamics during normal mitosis. Under these conditions, GMCL1 expression was similar to that in asynchronous cells and higher than the levels upon extended mitosis. However, when we attempted to measure the half-life of proteins using cycloheximide, most cells died, likely due to the toxic effect of cycloheximide in cells subjected to co-treatment with RO-3306 or nocodazole. This is the same reasons why in Figure 2C, we assessed 53BP1 in daughter cells rather than mitotic cells. 

There is no direct test of the proposed mechanism, and it is therefore unclear if 53BP1 is ubiquitinated by a GMCL1-CUL3 ligase in cells, and how efficient this process would be at different cell cycle stages. A key issue is the lack of experimental data explaining why the proposed mechanism would be restricted to mitosis. Indirect effects, such as loss of 53BP1 from the chromatin fraction during M phase upon GMCL1 overexpression, do not necessarily mean that 53BP1 is degraded. PLK1-dependent chromatin-cytoplasmic shuttling of 53BP1 during mitotic delays has been described previously (PMID: 38547292, 37888778). These papers are cited in the text, but the main conclusions of those papers on 53BP1 incorporation into a stopwatch complex during mitotic delays have been ignored. Are the authors sure that 53BP1 is destroyed in mitosis and not simply re-localised between chromatin and non-chromatin fractions? At the very least, these reported findings should be discussed in the text.

To examine whether GMCL1 promotes 53BP1 ubiquitination in cells, we expressed in cells Trypsin-Resistant Tandem Ubiquitin-Binding Entity (TR-TUBE), a protein that binds polyubiquitin chains. Abundant, endogenous ubiquitinated 53BP1 co-precipitated with TR-TUBE constructs only when wild-type GMCL1 but not the E142K GMCL1 mutant, was expressed (new Figure 2D).  The PLK1-dependent incorporation of 53BP1 into the stopwatch complex and the chromatin-cytoplasmic shuttling of 53BP1 during mitotic delays is now discussed in the text. That said, compared to parental cells, 53BP1 levels in the chromatin fraction are high in two different GMCL1 KO clones in M phase arrested cells (Figure 2A-B).  This increase does not correspond to a decrease in the 53BP1 soluble fraction (Figure 2A and new Supplementary Figure 2D), suggesting decreased 53BP1 is not due to re-localization. The increased half-life of 53BP1 in daughter cells (Figure 2C), also supports this hypothesis. 

The authors use a variety of cancer cell line models throughout their study, most of which have been reported to lack a functional mitotic surveillance pathway. U2OS and HCT116 cells do not respond normally to mitotic delays, despite being annotated as p53 WT. Other studies have used p53 wild-type hTERT RPE-1 cells to study the mitotic surveillance pathway. If the model is correct, then over-expressing GMCL1 in hTERT-RPE1 cells should suppress cell cycle arrest after mitotic delays, and GMCL1 KO should make the cells more sensitive to delays. These experiments are needed to provide an adequate test of the proposed model.

We greatly appreciate the reviewer’s suggestion regarding overexpression of GMCL1 in hTERT-RPE1 cells. To address this, we generated stable RPE1 cells expressing V5-tagged GMCL1 and conducted EdU incorporation assays following nocodazole synchronization and release. Overexpression of GMCL1 enhanced cell cycle progression compared to control cells (new Figure 3B and new Supplementary Figure 3D-E) after mitotic arrest, consistent with our model. We, therefore, propose that GMCL1 controls 53BP1 stability to suppress p53-dependent cell cycle arrest.

We also want to point out that while some papers suggest that HCT116 and U2OS cells do not have an intact mitotic surveillance pathway, others have shown that the MSP is indeed functioning in HCT116 cells and can be triggered with variable efficiency in U2OS cells (PMID: 38547292). This is likely due to high heterogeneity and extensive clonal diversity of cancer cell lines grown in different labs. Please see examples in PMIDs: 3620713, 30089904, and 30778230. In particular, PMID: 30089904 shows that this heterogeneity correlates with considerably different drug responses. 

To conclude, while the authors propose a potentially interesting model on how GMCL1 overexpression could regulate 53BP1 stability to limit p53-dependent cell cycle arrest, it is unclear what triggers this pathway or when it is relevant. 53BP1 is known to function in DNA damage signalling, and GMCL1 might be relevant in that context. The manuscript contains the initial description of GMCL1-53BP1 interaction but lacks a proper analysis of the function of this interaction and is therefore a preliminary report.

We hope that the new experiments, along with the clarifications provided in this response letter and revised manuscript, offer the reviewer increased confidence in the robustness and validity of our proposed model.

Reviewer #2 (Public review):

This study investigates the role of GMCL1 in regulating the mitotic surveillance pathway (MSP), a protective mechanism that activates p53 following prolonged mitosis. The authors identify a physical interaction between 53BP1 and GMCL1, but not with GMCL2. They propose that the ubiquitin ligase complex CRL3-GMCL1 targets 53BP1 for degradation during mitosis, thereby preventing the formation of the "mitotic stopwatch" complex (53BP1-USP28-p53) and subsequent p53 activation. The authors show that high GMCL1 expression correlates with resistance to paclitaxel in cancer cell lines that express wild-type p53. Importantly, loss of GMCL1 restores paclitaxel sensitivity in these cells, but not in p53-deficient lines. They propose that GMCL1 overexpression enables cancer cells to bypass MSP-mediated p53 activation, promoting survival despite mitotic stress. Targeting GMCL1 may thus represent a therapeutic strategy to re-sensitize resistant tumors to taxane-based chemotherapy.

Strengths:

This manuscript presents potentially interesting observations. The major strength of this article is the identification of GMCL1 as a 53BP1 interaction partner. The authors identified relevant domains and showed that GMCL1 controls 53BP1 stability. The authors further show a potentially interesting link between GMCL1 status and sensitivity to Taxol.

Weaknesses:

However, the manuscript is significantly weakened by unsubstantiated mechanistic claims, overreliance on a non-functional model system (U2OS), and overinterpretation of correlative data. To support the conclusions of the manuscript, the authors must show that the GMCL1-dependent sensitivity to Taxol depends on the mitotic surveillance pathway.

To demonstrate that GMCL1-dependent taxane sensitivity is mediated through the mitotic surveillance pathway (MSP), we now performed experiments using hTERT-RPE1 (RPE1) cells, a widely used, non-transformed cell line known to possess a functional MSP.  We compared RPE1 cells with knockdown of GMCL1 alone to those with simultaneous knockdown of GMCL1 and either TP53BP1 or USP28. Upon paclitaxel (Taxol) treatment, cells with GMCL1 knockdown exhibited suppressed proliferation and increased apoptosis. Notably, these phenotypes were rescued by co-depletion of TP53BP1 or USP28 (new Figure 5I,J). These results support the notion that GMCL1 contributes to MSP activity, at least in part, through its regulation of 53BP1.       

To further strengthen our mechanistic experiments, we assessed the effect of GMCL1 levels on cell cycle progression. Following nocodazole synchronization and release, we treated cells with EdU and performed FACS analyses at different times. Knockdown of GMCL1 alone led to a delay in cell cycle progression, but co-depletion of either TP53BP1 or USP28 alleviate this phenotype (new Figure 3A and new Supplementary Figure 3A, B). These results are consistent with our proliferation data.

Reviewer #3 (Public review):

Summary:

In this study, Kito et al follow up on previous work that identified Drosophila GCL as a mitotic substrate recognition subunit of a CUL3-RING ubiquitin ligase (CRL3) complex.

Here they characterize mutants of the human ortholog of GCL, GMCL1, that disrupt the interaction with CUL3 (GMCL1E142K) and that lack the substrate interaction domain (GMCL1 BBO). Immunoprecipitation followed by mass spectrometry identified 9 proteins that interacted with wild-type FLAG-GMCL1 and GMCL1 EK but not GMCL1 BBO. These proteins included 53BP1, which plays a well-characterized role in double-strand break repair but also functions in a USP28-p53-53BP1 "mitotic stopwatch" complex that arrests the cell cycle after a substantially prolonged mitosis. Consistent with the IP-MS results, FLAG-GMCL1 immunoprecipitated 53BP1. Depletion of GMCL1 during mitotic arrest increased protein levels of 53BP1, and this could be rescued by wild-type GMCL1 but not the E142K mutant or a R433A mutant that failed to immunoprecipitate 53BP1.

Using a publicly available dataset, the authors identified a relatively small subset of cell lines with high levels of GMCL1 mRNA that were resistant to the taxanes paclitaxel, cabazitaxel, and docetaxel. This type of analysis is confounded by the fact that paclitaxel and other microtubule poisons accumulate to substantially different levels in various cell lines (DOI: 10.1073/pnas.90.20.9552 , DOI: 10.1091/mbc.10.4.947 ), so careful follow-up experiments are required to validate results. The correlation between increased GMCL1 mRNA and taxane resistance was not observed in lung cancer cell lines. The authors propose this was because nearly half of lung cancers harbor p53 mutations, and lung cancer cell lines with wild-type but not mutant p53 showed the correlation between increased GMCL1 mRNA and taxane resistance. However, the other cancer cell types in which they report increased GMCL1 expression correlates with taxane sensitivity also have high rates of p53 mutation. Furthermore, p53 status does not predict taxane response in patients (DOI: 10.1002/1097-0142(20000815)89:4<769::aid-cncr8>3.0.co;2-6 , DOI: 10.1002/(SICI)1097-0142(19960915)78:6<1203::AID-CNCR6>3.0.CO;2-A , PMID: 10955790).

The authors then depleted GMCL1 and reported that it increased apoptosis in two cell lines with wild-type p53 (MCF7 and U2OS) due to activation of the mitotic stopwatch. This is surprising because the mitotic stopwatch paper they cite (DOI: 10.1126/science.add9528 ) reported that U2OS cells have an inactive stopwatch and that activation of the stopwatch results in cell cycle arrest rather than apoptosis in most cell types, including MCF7. Beyond this, it has recently been shown that the level of taxanes and other microtubule poisons achieved in patient tumors is too low to induce mitotic arrest (DOI: 10.1126/scitranslmed.3007965 , DOI: 10.1126/scitranslmed.abd4811 , DOI: 10.1371/journal.pbio.3002339 ), raising concerns about the relevance of prolonged mitosis to paclitaxel response in cancer. The findings here demonstrating that GMCL1 mediates degradation of 53BP1 during mitotic arrest are solid and of interest to cell biologists, but it is unclear that these findings are relevant to paclitaxel response in patients.

Strengths:

This study identified 53BP1 as a target of CRL3GMCL1-mediated degradation during mitotic arrest. AlphaFold3 predictions of the binding interface, followed by mutational analysis, identified mutants of each protein (GMCL1 R433A and 53BP1 IEDI1422-1425AAAA) that disrupted their interaction. Knock-in of a FLAG tag into the C-terminus of GMCL1 in HCT116 cells, followed by FLAG immunoprecipitation, confirmed that endogenous GMCL1 interacts with endogenous CUL3 and 53BP1 during mitotic arrest.

Weaknesses:

The clinical relevance of the study is overinterpreted. The authors have not taken relevant data about the clinical mechanism of taxanes into account. Supraphysiologic doses of microtubule poisons cause mitotic arrest and can activate the mitotic stopwatch. However, in physiologic concentrations of clinically useful microtubule poisons, cells proceed through mitosis and divide their chromosomes on mitotic spindles that are at least transiently multipolar. Though these low concentrations may result in a brief mitotic delay, it is substantially shorter than the arrest caused by high concentrations of microtubule poisons, and the one mimicked here by 16 hours of 0.4 mg/mL nocodazole, which is not used clinically and does not induce multipolar spindles. Resistance to mitotic arrest occurs through different mechanisms than resistance to multipolar spindles. No evidence is presented in the current version of the manuscript that GMCL1 affects cellular response to clinically relevant doses of paclitaxel.

We agree that it would be an overstatement to claim that GMCL1 and p53 regulates paclitaxel sensitivity in cancer patients in a clinical context. The correlations we observed were based on publicly available cancer cell lines from datasets catalogued in CCLE and DepMap, which do not fully account for clinical heterogeneity and patient-specific factors. In response to this important point, we have revised the text accordingly. 

In the experiments shown in former Figure 4A-H (now Figure 5A-H) and in those shown in the new Figure 5I-J, we used 100 nM paclitaxel to test the hypothesis that low GMCL1 levels sensitizes cancer cells in a p53-dependent manner. Here, paclitaxel was chosen to mimic the conditions reported in the PRISM dataset (PMID: 32613204), which compiles the proliferation inhibitory activity of 4,518 compounds tested across 578 cancer cell lines. Consistent with our cell cycle findings, the paclitaxel sensitivity caused by GMCL1 depletion was reverted by silencing 53BP1 or USP28 (new Figure 5I-J), again supporting the involvement of the stopwatch complex. We are unsure about how to model the “physiologic concentrations of clinically useful microtubule poisons” in cell-based studies. A recent review notes that “The time above a threshold paclitaxel plasma concentration (0.05 mmol/L) is important for the efficacy and toxicity of the drug” (PMID: 28612269).  Two other reviews mention that the clinically relevant concentration of paclitaxel is considered to be plasma levels between 0.05–0.1 μmol/L (approximately 50–100 nM) and that in clinical dosing, typical patient plasma concentrations after paclitaxel infusion range from 80–280 nM, with corresponding intratumoral concentrations between 1.1–9.0 μM, due to drug accumulation in tumor tissue (PMIDs: 24670687 and  29703818).  We have now emphasized in the revised text the rationale for using 100 nM paclitaxel in our experiments.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

General comments on the Figures:

(1) Western blots lack molecular weight markers on most panels and are often over-exposed and over-contrasted, rendering them hard to interpret.

We have now included molecular weight markers in all Western blot panels. We have also reprocessed the images to avoid overexposure and excessive contrast, ensuring that the bands are clearly visible and interpretable.

(2) Input and IP samples do not show percentage loading, so it is hard to interpret relative enrichments.

In the revised figures, we have indicated what % of the input was loaded.

(3) The authors change between cell line models for their experiments, and this is not clear in the figures. These are important details for interpreting the data, as many of the cell lines used are not functional for the mitotic surveillance pathway.

In the revised manuscript, we have clearly indicated the specific cell lines used in each experiment in the figure legends. Additionally, to address concerns regarding the mitotic surveillance pathway, we have included new experiments using hTERT-RPE1 cells, which have been reported to possess a functional mitotic surveillance pathway (MSP) (Figure 4I-J).

(4) No n-numbers are provided in the figure legends. Are the Western blots provided done once, or are they reproducible? Many of the blots would benefit from quantification and presentation via graphs to test for reproducible changes to 53BP1 levels under the different conditions.

As now indicated in the methods section, we have conducted each Western blot no less than three times, yielding results that exhibit a high degree of reproducibility. A representative Western blot has been selected for each figure. We did not include densiometric quantification of immunoblots, given that the semi-quantitative nature of this technique would lead to an overinterpretation of our data; unfortunately, this is a limitation of the technique. In fact, eLife and other similar scientific journals do not adhere to the practice of quantifying Western blots. One exception to this norm is for protein half-life studies, which is done to measure the kinetics of decay rates and their internal comparisons. Accordingly, the experiments in Figure 2C were quantified.

(5) Graphs displayed in the supplementary figures are blacked out, and individual data points cannot be visualised. All graphs should have individual data points clearly visible.

We revised the quantified graphs and replaced them with scatter plots to clearly display individual data points, showing sample distribution.

Additional experiments with specific comments on Figures:

(1) Figure 1C-D: the relative amount of 53BP1 co-precipitating with FLAG-tagged GMCL1 WT appears very different between the two experiments. If the idea is that MLN4924 (Cullin neddylation inhibitor) makes the interaction easier to capture, then this should be explained in the text, and ideally shown on the same gel/blot -/+ MLN4924.

We now present the samples treated with and without MLN4924 on the same gel/blot to allow direct comparison (new Figure 1D) and clarified this point in the text.

(2) Figure 1E: The figure legend states that GMCL1 was immunoprecipitated, but the Figure looks as though FLAG-tagged 53BP1 was the bait protein being immunoprecipitated? Can the authors clarify?

We thank the reviewer for pointing out the discrepancy between the figure and the figure legend in Figure 1E. The immunoprecipitation was indeed performed using FLAG-tagged 53BP1, and we have now rectified the figure legend accordingly. 

(3) Figure 1F: Rather than parental cell lysate, the better control would be to IP FLAG from another FLAG-tagged expressing cell line, to rule out non-specific binding with the FLAG tag at the non-overexpressed level. 

Figure 1F shows interaction at the endogenous level. The specificity of binding with overexpressed proteins is shown in Figures 1C and 1D.

The USP28 blot is over-exposed and makes it hard to see any changes in electrophoretic mobility - it looks as though there is a change between the parental and the KI cell line? It is surprising that USP28 would co-IP with GMCL1 (presumably because USP28 is bound to 53BP1) if the function of GMCL1-53BP1 interaction is to promote 53BP1 degradation. Can the authors reconcile this? Crucially, if the authors claim that the 53BP1-GMCL1 interaction is specific to prolonged mitosis, then this experiment should be repeated and performed with asynchronous, normal-length mitosis, and prolonged mitosis conditions. This is vital for supporting the claim that this interaction only occurs during prolonged mitoses and does not occur in every mitosis regardless of length.

This is a good point. Unfortunately, many of the protein-protein interactions occur post lysis. Therefore, we could not observe differences in asynchronous vs. mitotic cells.

(4) Figure S1F: Label on blot should be CUL3 not CUI3.

We thank the reviewer for pointing this out and we have corrected the typo.

(5) Figure 2A: The authors suggest an increase in chromatin-bound 53BP1 in GMCL1 KO U2OS cells, specifically in M phase. Again, is this time in mitosis dependent, or would this be evident in every mitosis, regardless of length? Such an experiment would benefit from repetition and quantification to test whether the observed effect is reproducibly consistent. If the authors' model is correct, simply treating U2OS WT mitotic cells with MG132 during the mitotic arrest and performing the same fractionation should bring 53BP1 levels up to that seen in GMCL1 KO cells under the same conditions.

The reviewer’s suggestion to assess 53BP1 accumulation in wild-type U2OS cells treated with MG132 during mitotic arrest is indeed highly relevant. However, treatment with MG132 during prolonged mitosis consistently led to significant cell death, making it technically challenging to evaluate 53BP1 levels under these conditions.

(6) Figure 2B: The authors restore GMCL1 expression in the KO U2OS cells using WT and 2 distinct mutant cDNAs. However, the expression of these constructs is not equivalent, and thus their effects cannot be directly compared. It is also surprising that GMCL1 is much higher in M phase samples in this experiment (shouldn't it be destroyed?), when no such behaviour has been observed in the other figures.

There is no evidence in our study or others that GMCL1 should be destroyed in M phase.  We show that the R433A mutant is expressed at a level very similar to the WT protein, yet it doesn’t promote the degradation of 53BP1. It is true that the E142K is expressed less in mitotic cells whereas is the most expressed in asynchronous cells. For some reason, this mutant has an inverse behavior compared to the WT, limiting the interpretation of this result. We now mention this in the text. 

(7) Figure 2C: The CHX experiment would benefit from inclusion of a control protein known to have a short half-life (e.g. c-myc, p53). Is GMCL1 known to have a relatively short half-life? It looks as though GMCL1 disappears after 1 h CHX treatment (although hard to definitively tell in the absence of molecular weight markers). 53BP1 appears to continue declining in the absence of GMCL1, which is surprising if p53BP1 degradation requires GMCL1. How can the authors reconcile this?

As a control for the CHX chase experiments, we included p21, whose protein levels decreased in a CHX-dependent. GMCL1 itself also appeared to undergo degradation upon CHX treatment, but it doesn’t disappear completely.

(8) Supplemental Figure 2:

Transcription is largely inhibited in M phase, so the p53 target gene transcripts present in M phase are inherited from the preceding G2 phase. The qPCR's thus need a reference sample to compare against. I.e., was p21/PUMA/NOXA mRNA already low in G2 in the GMCL1 KO + WT cells before they entered mitosis? Or is the mRNA stability affected during M phase specifically? Is this effect on the mRNA dependent on the time in mitosis?

It is well established that transcription is not entirely shut down during mitosis, particularly for a subset of genes involved in cell cycle regulation. For example, p21, PUMA, NOXA, and p53 mRNAs have been shown to remain actively transcribed during mitosis (see Table S5 in PMID: 28912132). However, we currently lack direct evidence that p53 activation during mitosis, specifically through the mitotic surveillance pathway, drives the transcription of p21, PUMA, or NOXA mRNAs during M phase. In the absence of such mechanistic data, we opted to exclude these analyses from the final figures.

Panel B: blots are too over-exposed to see differences in p53 stability under the different conditions. Mitotic samples should be included to show how these differ from the G1 samples.

The background of all blot images has been adjusted to ensure clarity and consistency.

Panel D: The authors show no significant difference in the cell cycle profiles of the GMCL1 KO and reconstituted cells compared to parental U2OS cells. This should also be performed in the G1 daughter cells following a prolonged mitosis, to test the effect of the different GMCL1 constructs on G1 cell cycle arrest. U2OS cells have been reported not to have a functional mitotic surveillance pathway (Meitinger et al, Science, 2024), so U2OS cells are perhaps not a good model for testing this.

We performed cell cycle profiling using EdU incorporation in hTERT-RPE1 cells, which possess a functional MSP, to evaluate cell cycle progression in daughter cells following prolonged mitosis. We observed that GMCL1 knockdown alone leads to G1-phase arrest. In contrast, co-depletion of GMCL1 with either 53BP1 or USP28 bypasses this arrest, indicating that GMCL1 regulates cell cycle progression in an MSP-dependent manner. Please see also the answer to the public review above. 

(9) Figure 3:

The authors show expression data for GMCL1 in the different cancer cell lines. This should be validated for a subset of cancer cell lines at the GMCL1 protein level, and cross-correlated to their MSP/mitotic timer status. Does GMCL1 depletion or knockout in p53 wild-type cancer cell lines overexpressing GMCL1 protein restore mitotic surveillance function?

We were unable to assess GMCL1 protein levels using publicly available proteomics datasets, as GMCL1 expression was not detected. In p53 wild-type hTERT-RPE1 cells, GMCL1 knockdown impaired the mitotic surveillance pathway, as evidenced by G1-phase arrest following prolonged mitosis (new Figure 3A and new Supplementary Figure 3A, B). This arrest was rescued by co-depletion of either TP53BP1 or USP28, indicating that GMCL1 acts upstream of the MSP.

(10) Figure 4:

The authors show siRNA experiments depleting GMCL1 and testing the effects of GMCL1 loss on cell viability and apoptosis induction. This is performed in different cell line backgrounds. However, there is no demonstration that any of the observed effects are due to a lack of GMCL1 activity on 53BP1. These experiments need to be repeated in 53BP1 co-depleted cells to test for rescue. Without this, the interpretation is purely correlative.

We assessed the effects of GMCL1 knockdown, alone or in combination with TP53BP1 or USP28 knockdown, on cell viability and apoptosis in hTERT-RPE1 cells using siRNA. Knockdown of GMCL1 alone led to a significant reduction in cell viability and an increase in apoptosis. However, co-depletion of GMCL1 with either TP53BP1 or USP28 restored both cell viability and apoptosis levels to those observed in control cells (new Figure 5I,J).

(11) Text comments:

Line 257: HeLa cells supress p53 through the E6 viral protein and are not "mutant" for p53.

The authors should cite early work by Uetake and Sluder describing the effects of spindle poisons on the mitotic surveillance pathway.

We appreciate the reviewer’s comments – We have now made the necessary corrections.

Reviewer #2 (Recommendations for the authors):

Major Points:

(1) Unsubstantiated Mechanistic Claims:

In Figures 3 and 4, the authors show correlations between GMCL1 expression and sensitivity to Taxol. However, they fail to demonstrate that the mitotic stopwatch is mechanistically involved. To support this conclusion, the authors must test whether deletion of 53BP1, USP28, or disruption of their interaction rescues Taxol sensitivity in GMCL1-depleted cells. Since 53BP1 also plays a role in DNA damage response, such rescue experiments are necessary to distinguish between mitotic surveillance-specific and broader stress-response effects. Deletion of USP28 would be particularly informative.

We sought to experimentally determine whether GMCL1 is involved in regulating the mitotic stopwatch. Knockdown of GMCL1 alone resulted in reduced cell proliferation and increased apoptosis. In contrast, co-depletion of GMCL1 with either TP53BP1 or USP28 restored both proliferation and apoptosis levels to those observed in control cells (new Figure 5I, J). To further strengthen our mechanistic experiments, we assessed the effect of GMCL1 levels on cell cycle progression. We conducted EdU incorporation assays following nocodazole synchronization and release. Knockdown of GMCL1 alone led to a delay in G1 progression, whereas co-depletion of either TP53BP1 or USP28 rescued normal cell cycle progression (new Figure 3A and new Supplementary Figure 3A, B). These results are consistent with our proliferation data and suggest that GMCL1 functions upstream of the ternary complex, likely by regulating 53BP1 protein levels.

(2) Model System Limitations (U2OS Cells):

The use of U2OS cells is highly problematic for investigating the mitotic surveillance pathway. U2OS cells lack a functional mitotic stopwatch and do not arrest following prolonged mitosis in a 53BP1/USP28-dependent manner (PMID: 38547292). Therefore, conclusions drawn from this model system about the function of the mitotic surveillance pathway are not substantiated. Key experiments should be repeated in a cell line with an intact pathway, such as RPE1.

We now performed all key experiments also hTERT-RPE1 cells (see above). We also would like to point out that while some papers suggest that HCT116 and U2OS cells do not have an intact mitotic surveillance pathway, others have showed that the MSP is indeed functioning in HCT116 cells and can be triggered with variable efficiency in U2OS cells (PMID: 38547292).  This is likely due to high heterogeneity and extensive clonal diversity of cancer cell lines grown in different labs. Please see examples in PMIDs: 3620713, 30089904, and 30778230. In particular, PMID: 30089904 shows that this heterogeneity correlates with considerably different drug responses. 

(3) Misinterpretation of p53 Activity Timing:

The manuscript states that "GMCL1 KO cells led to decreased mRNA levels of p21 and NOXA during mitosis" (line 194). However, it is well established that the mitotic surveillance pathway activates p53 in the G1 phase following prolonged mitosis-not during mitosis itself (PMID: 38547292). Therefore, the observed changes in mRNA levels during mitosis are unlikely to be relevant to this pathway.

We currently lack direct evidence that p53 activated during mitosis through the mitotic surveillance pathway directly influences the transcription of p21, PUMA, or NOXA mRNAs during M phase. Therefore, we have chosen to exclude these data from the final figures.

(4) Incorrect Interpretation of 53BP1 Chromatin Binding:

The authors claim that 53BP1 remains associated with chromatin during mitosis, which contradicts established literature. It is known that 53BP1 is released from chromatin during mitosis via mitosis-specific phosphorylation (PMID: 24703952), and this is supported by more recent findings (PMID: 38547292). A likely explanation for the discrepancy may be contamination of mitotic fractions with interphase cells. The chromatin fraction data in Figure 2C must be interpreted with caution.

Our method to synchronize in M phase is rather stringent (see Supplementary Figure 3D as an example). The literature indicates that the bulk of 53BP1 is released from chromatin during mitosis. Yet, even in the two publications mentioned by the reviewer, there is a difference in the observable amount of 53BP1 bound to chromatin (compare Figure 2B in PMID: 38547292 and Figure 5A in PMID: 24703952). The difference is likely due to the different biochemical approaches used to purify chromatin bound proteins (salt and detergent concentrations, sonication, etc.). Using our fractionation approach, we can reliably separate the soluble fraction (containing also the nucleoplasmic fraction) and chromatin associated proteins as indicated by the controls such as a-Tubulin and Histon H3.  We have now mentioned these limitations when comparing different fractionation methods in our discussion section.

(5) Inadequate Citation of Foundational Literature:

The literature on the mitotic surveillance pathway is relatively limited, and it is essential that the authors provide a comprehensive and accurate account of its development. The foundational work by the Sluder lab (PMID: 20832310), demonstrating a p53-dependent arrest following prolonged mitosis, must be cited. Furthermore, the three key 2016 papers (PMID: 27432896, 27432897, 27432896) that identified the involvement of USP28 and 53BP1 in this pathway are critical and should be cited as the basis of the mitotic surveillance pathway.

In contrast, the manuscript currently emphasizes publications that either contribute minimally or have been contradicted by prior and subsequent work. For example: PMID: 31699974, which proposes Ser15 phosphorylation of p53 as critical, has been contradicted by multiple groups (e.g., Holland, Oegema, and Tsou labs).

PMID: 37888778, which suggests that 53BP1 must be released from kinetochores, is inconsistent with findings that indicate kinetochore localization is not relevant.

The authors should thoroughly revise the Introduction to reflect what this reviewer would describe as a more accurate and scholarly approach to the literature.

We have substantially revised both the Introduction and Discussion sections to incorporate important references kindly suggested by the reviewer.

Minor Points:

(1) Overexposed Western Blots:

The Western blots throughout the manuscript are heavily overexposed and saturated, obscuring differences in protein levels and hindering data interpretation. The authors should provide properly exposed blots with quantification where appropriate.

We have provided Western blot images with appropriate exposure levels and included quantification where appropriate (i.e., to measure the kinetics of decay rates as in Figure 2C). For all the other immunoblots, we did not include densiometric quantification, given that the semi-quantitative nature of this technique would lead to overinterpretation of our data. This is, unfortunately, a limitation of the technique. In fact, eLife and other similar scientific journals do not adhere to the practice of quantifying Western blot analyses. 

(2) Missing information in the graphs in Figure 2C and 4; S2? How many repeats? What are the asterisks?

Panels referenced above have been repeated several times, and further details are now provided in the figure legends.

Reviewer #3 (Recommendations for the authors):

(1)   The claim that GMCL1 modulates paclitaxel sensitivity in cancer should be toned down

.

We agree that it would be an overstatement to claim that GMCL1 regulates paclitaxel sensitivity in cancer patients in a clinical context. The correlations we observed were based on publicly available, cell line–based datasets, which do not fully account for clinical heterogeneity and patient-specific factors. In response to this important point, we have revised our statements and corresponding text accordingly. We now placed greater emphasis on our molecular and cell biology studies.

(2) Additional experiments in low, physiologically relevant concentrations of paclitaxel would be interesting. It is possible that these concentrations activate the mitotic stopwatch in a portion of cells, in addition to inducing cell death due to chromosome loss, activation of an immune response, and chromothripsis. Results should be interpreted in the context of this complexity.

Please see the response to the public review. 

(3) It would be helpful to show that CUL3 interacts with 53BP1 only in the presence of GMCL1.

We show that the binding of 53BP1 to GMCL1 is independent of the ability of GMCL1 to bind CUL3 (Figure 1C, D). The binding between 53BP1 and CUL3 is difficult to detect (Figure 1F) likely because it’s not direct but mediated by GMCL1.

(4) The GMCL1 "KO" lines appear to still express a low level of GMCL1 (Figure 2A), which should be acknowledged

We have included the GMCL1 mRNA expression data, as measured by RT-PCR, in Supplementary Figure 1G, demonstrating that GMCL1 expression was undetectable under the tested conditions.

(5) Additional description of the methods is warranted. This is particularly true for the database analysis that forms the basis for the claim that GMCL1 overexpression causes resistance to paclitaxel and other taxanes presented in Figure 3, the methodology used to obtain M-phase cells, and the concentration and duration of taxol treatment.

We have now extensively revised the Methods section.  

(6) "Taxol" and "paclitaxel" are used interchangeably throughout the manuscript. Consistency would be preferable.

We have revised the manuscript to maintain consistency in the use of the terms “Taxol” and “paclitaxel” and now refer to “paclitaxel” when discussing that individual compound; “taxanes” when referring collectively to cabazitaxel, docetaxel and paclitaxel; and “Taxol” has been removed entirely to avoid redundancy or confusion.    

(7) It is unclear why it is claimed that GMCL1 interacts "specifically" with 53BP1 (line 176) since multiple interactors were identified in the IP-MS study

We meant that the GMCL1 R433A mutant loses its ability to bind 53BP1, suggesting that the GMCL1-53BP1 interaction is not an artifact. We have now clarified the text. 

(8) The bottom row in Figure S3 is misleading. Paclitaxel is not uniformly effective in every tumor of any given type, and so resistance occurs in every cancer type.

We fully agree that cancer is highly heterogeneous and that paclitaxel efficacy varies across tumors, even within the same histological subtype. Our intension was not to suggest uniform sensitivity/resistance, but rather to provide a high-level overview using aggregated data. We acknowledge that this coarse-grained representation may unintentionally imply overly generalized conclusions. To avoid potential misinterpretation, we have removed the corresponding panel in the revised paper.

delete

delete

Reviewer #1 (Public review):

Summary:

This paper introduces a dual-pathway model for reconstructing naturalistic speech from intracranial ECoG data. It integrates an acoustic pathway (LSTM + HiFi-GAN for spectral detail) and a linguistic pathway (Transformer + Parler-TTS for linguistic content). Output from the two components is later merged via CosyVoice2.0 voice cloning. Using only 20 minutes of ECoG data per participant, the model achieves high acoustic fidelity and linguistic intelligibility.

Strengths:

(1) The proposed dual-pathway framework effectively integrates the strengths of neural-to-acoustic and neural-to-text decoding and aligns well with established neurobiological models of dual-stream processing in speech and language.

(2) The integrated approach achieves robust speech reconstruction using only 20 minutes of ECoG data per subject, demonstrating the efficiency of the proposed method.

(3) The use of multiple evaluation metrics (MOS, mel-spectrogram R², WER, PER) spanning acoustic, linguistic (phoneme and word), and perceptual dimensions, together with comparisons against noise-degraded baselines, adds strong quantitative rigor to the study.

Weaknesses:

(1) It is unclear how much the acoustic pathway contributes to the final reconstruction results, based on Figures 3B-E and 4E. Including results from Baseline 2 + CosyVoice and Baseline 3 + CosyVoice could help clarify this contribution.

(2) As noted in the limitations, the reconstruction results heavily rely on pre-trained generative models. However, no comparison is provided with state-of-the-art multimodal LLMs such as Qwen3-Omni, which can process auditory and textual information simultaneously. The rationale for using separate models (Wav2Vec for speech and TTS for text) instead of a single unified generative framework should be clearly justified. In addition, the adaptor employs an LSTM architecture for speech but a Transformer for text, which may introduce confounds in the performance comparison. Is there any theoretical or empirical motivation for adopting recurrent networks for auditory processing and Transformer-based models for textual processing?

(3) The model is trained on approximately 20 minutes of data per participant, which raises concerns about potential overfitting. It would be helpful if the authors could analyze whether test sentences with higher or lower reconstruction performance include words that were also present in the training set.

(4) The phoneme confusion matrix in Figure 4A does not appear to align with human phoneme confusion patterns. For instance, /s/ and /z/ differ only in voicing, yet the model does not seem to confuse these phonemes. Does this imply that the model and the human brain operate differently at the mechanistic level?

(5) In general, is the motivation for adopting the dual-pathway model to better align with the organization of the human brain, or to achieve improved engineering performance? If the goal is primarily engineering-oriented, the authors should compare their approach with a pretrained multimodal LLM rather than relying on the dual-pathway architecture. Conversely, if the design aims to mirror human brain function, additional analysis, such as detailed comparisons of phoneme confusion matrices, should be included to demonstrate that the model exhibits brain-like performance patterns.

Le Sneakernet : Repenser le Partage de Données à l'Ère de la Big Tech

Résumé Exécutif

Face à la mainmise croissante des géants de la technologie (Big Tech) sur les données personnelles et les infrastructures numériques, un mouvement alternatif émerge : le sneakernet.

Ce concept, qui désigne le partage physique de fichiers hors ligne, s'oppose directement au modèle centralisé et commercial de l'internet actuel.

Des collectifs d'artistes et d'activistes développent des initiatives concrètes pour réhabiliter des pratiques d'échange de données autonomes, locales et matérielles.

Les principales conclusions de l'analyse des sources sont les suivantes :

• Le Problème Central : Les données hébergées sur des plateformes comme Google Drive, iCloud ou Instagram n'appartiennent pas aux utilisateurs mais aux entreprises qui les stockent.

Celles-ci contrôlent l'accès, peuvent en modifier les conditions et exploitent les informations de navigation, créant une forte dépendance et un risque de surveillance.

• La Solution Sneakernet : Ce "réseau basket" repose sur l'échange physique de données (via clés USB, par exemple) à la "vitesse des jambes".

Il représente une démarche de reprise de contrôle sur la circulation de l'information, en marge des infrastructures traditionnelles.

• Les Initiatives Clés :

◦ Les "Data-Foires" de l'Outdoor Computer Club : Des événements où les participants échangent des fichiers sur un ordinateur collectif, souvent alimenté par des sources d'énergie autonomes, promouvant une gestion consciente des ressources et une vision de la technologie comme un "commun".   

◦ Le projet "Dead Drops" d'Aram Bartholl : Un réseau mondial et participatif de clés USB scellées dans des murs, fonctionnant comme des "boîtes aux lettres mortes" anonymes pour l'échange de fichiers.   

◦ Les serveurs DIY du collectif Actinomy : Des ateliers pour construire ses propres mini-serveurs portables, permettant un hébergement local et privé, créant ainsi une "chambre à soi" numérique indépendante des grandes plateformes.

• La Philosophie Sous-jacente : Le mouvement critique "l'obésité de la donnée" et la course à la vitesse.

Il propose de retrouver un "affect par rapport aux données" en privilégiant des échanges plus lents, plus intentionnels et en réutilisant des technologies plus anciennes pour répondre à des enjeux politiques actuels comme la surveillance.

En conclusion, le sneakernet n'est pas une simple nostalgie technologique, mais une réponse politique et pratique à la structure de pouvoir de l'internet moderne.

Il démontre que des alternatives artisanales et autonomes existent déjà pour échapper au contrôle des plateformes et repenser notre rapport à la technologie et aux données.

1. La Problématique : Dépendance et Perte de Contrôle à l'Ère Numérique

Le modèle dominant de l'internet actuel, contrôlé par un nombre restreint de grandes entreprises technologiques, pose un problème fondamental de souveraineté sur les données personnelles.

• Propriété des Données : Une fois stockés en ligne sur des services comme Google Drive, iCloud ou postés sur des réseaux sociaux, les fichiers, photos et documents "deviennent la propriété des plateformes numériques qui les hébergent."

La perception de possession par l'utilisateur est une illusion, car ce dernier perd le contrôle direct sur ses propres créations.

• Contrôle de l'Accès : Les entreprises qui gèrent l'accès et le stockage (telles qu'Amazon, Microsoft, Oracle, Google et Meta) ont le pouvoir unilatéral de "décider de nous faire payer plus cher ou carrément de nous couper cet accès."

• Exploitation des Informations : L'acceptation des cookies autorise les entreprises à utiliser les informations de navigation des utilisateurs, transformant leurs goûts et intérêts en données monétisables.

• Dépendance Structurelle : La facilité d'utilisation de l'internet à haut débit et du stockage "infini" sur le cloud a créé une telle dépendance que l'on "n'imagine même plus comment faire sans".

Cette situation est décrite comme une "mainmise des big tech sur nos vies."

2. Le Sneakernet : Une Alternative au Réseau Global

En réponse à cette centralisation, le sneakernet propose un paradigme radicalement différent, fondé sur l'échange physique et déconnecté.

• Définition : Le terme "sneakernet" signifie littéralement "réseau basket".

Il désigne un réseau d'échange physique fonctionnant "à la vitesse des jambes".

C'est "l'antithèse de l'internet actuel", où les données transitent par des infrastructures de câbles et d'ondes.

• Contrôle et Matérialité : Le principal avantage est le contrôle total sur le chemin de l'information.

Comme le souligne un participant, "on a le contrôle sur par où l'information elle passe, dans ta poche, dans ta main et dans sa poche et donc en ça c'est hors du réseau."

• Une "Innovation" Rétro-Technologique : Le mouvement propose d'utiliser des technologies plus anciennes pour répondre aux problématiques contemporaines.

Un organisateur explique : "Nous, on imagine qu'en prenant peut-être des technologies plus anciennes, on propose une autre vision de l'innovation."

Cette approche est justifiée par le fait qu'elle a "du sens par rapport à ce qui se passe politiquement aujourd'hui autour de l'internet."

3. Initiatives et Projets Phares

Plusieurs collectifs d'artistes et d'activistes ont mis en place des projets concrets pour matérialiser les principes du sneakernet.

Les "Data-Foires" de l'Outdoor Computer Club

Ce collectif, dont les organisateurs utilisent les pseudonymes "Jeff Bisou" et "Xavier Nul", organise des événements d'échange de données hors ligne appelés "data-foires".

• Concept : Un ordinateur est installé dans un lieu (par exemple, une forêt), où les participants peuvent déposer et récupérer des données via des clés USB.

• Autonomie Énergétique : L'installation est souvent alimentée par des "batteries au lithium, de récupération" connectées à un convertisseur qui fournit un courant standard de 230 volts.

Cette démarche soulève des questions sur la gestion collective de l'électricité, perçue non "comme une ressource infinie" mais en fonction des besoins réels.

• Contenus Partagés : Les échanges sont hétéroclites, incluant :

◦ Musique, logiciels, brochures. 

◦ Scans 3D.   

◦ Un documentaire de Mathieu Rigouste, "Nous sommes des champs de bataille".  

◦ Une thèse scientifique sur un hydrogel supramoléculaire utilisé pour cultiver des cellules cancéreuses.  

◦ Le site d'une maison d'édition, présenté en avant-première.

• Limites et Modération : Le système n'est pas parfait, avec la présence de "pas mal de fichiers corrompus" et de transferts incomplets.

Cependant, la modération se fait "de manière fluide vu que tout le monde est là en présentiel", permettant de retrouver plus facilement une personne malveillante.

• Philosophie : L'initiative promeut l'idée de "penser l'ordinateur comme un commun".

Le partage d'une machine unique s'oppose à l'usage individuel habituel et transforme la relation à la technologie en une pratique collective.

Dead Drops par Aram Bartholl : Un Réseau d'Échange Anonyme

Initié en 2010 par l'artiste et professeur d'art numérique Aram Bartholl, ce projet est l'une des incarnations les plus connues du sneakernet.

• Concept : Des centaines de clés USB sont scellées dans des murs et autres lieux publics à travers le monde, formant un "réseau d'échanges ouvert".

Le nom "Dead Drops" est une référence aux "boîtes aux lettres mortes" utilisées en espionnage pour déposer des documents de façon anonyme.

• Caractère Participatif : Chacun peut installer une "dead drop" dans sa ville, contribuant ainsi à l'expansion du réseau.

• Évolution Politique : Initialement artistique, le projet a acquis une nouvelle signification militante avec la montée en puissance de Big Tech.

Il invite désormais à s'interroger sur les moyens d'échapper "à la dépendance vis-à-vis des plateformes numériques, mais aussi à leur surveillance".

Serveurs DIY du Collectif Actinomy : Reprendre le Contrôle de l'Hébergement

Pour atteindre une autonomie complète, le collectif Actinomy, basé à Brême, propose d'apprendre à construire soi-même des serveurs locaux et privés.

• Concept : Lors d'ateliers "do-it-yourself", les participants fabriquent de "petits serveurs informatiques portables en forme de porte-clés".

• Fonctionnement : Ces mini-serveurs ne peuvent héberger qu'une page internet légère, accessible uniquement via un réseau Wi-Fi local très restreint.

Le site n'est pas visible sur l'internet mondial mais sur un "petit réseau parallèle".

• Objectif : Cette démarche vise une "reprise de contrôle sur ses informations".

Elle est comparée à la création de "sa propre chambre à soi dans la grande maison de l'Internet mondiale", en opposition aux immenses fermes de serveurs centralisées qui tournent 24/7.

4. Principes et Philosophie du Mouvement Sneakernet

Au-delà des aspects techniques, le sneakernet est porteur d'une vision critique et d'une philosophie alternative de la technologie.

• Critique de "l'Obésité de la Donnée" : Le mouvement remet en question la logique du "toujours plus" (plus de vitesse, plus de stockage).

Il s'interroge : "Est-ce qu'on veut juste envoyer des fichiers hyper lourds le plus vite possible et tout, ou est-ce que on veut retrouver un certain affect par rapport aux données ?"

• Valorisation de la Donnée Précieuse : Dans un contexte de transfert physique, les participants ont tendance à apporter une "petite quantité de données", généralement celles qui "leur semblent précieuses", une sélectivité qui se perd avec le haut débit.

• L'Ordinateur comme un "Commun" : Le partage d'une seule machine lors des data-foires transforme l'ordinateur d'un objet personnel en une ressource collective, modifiant la relation individuelle à la technologie.

• Conscience Énergétique et Matérielle : L'utilisation de systèmes d'alimentation autonomes et de récupération force à une réflexion sur la gestion collective de l'énergie et sur l'empreinte matérielle des infrastructures numériques.

• Sécurité par la Proximité : La présence physique des participants lors des échanges crée une forme d'autorégulation et de responsabilité qui n'existe pas dans les interactions en ligne anonymes.

5. Conclusion : Une Vision Alternative pour l'Avenir Numérique

Le mouvement sneakernet, bien qu'il puisse paraître "utopique ou rétrograde", constitue une critique pertinente et une alternative tangible à l'écosystème numérique actuel.

Il démontre que l'autonomie face aux grandes plateformes est possible.

Le futur, selon cette perspective, pourrait impliquer de "moins stocker, moins partager et héberger en local pour vraiment échapper au contrôle des grandes plateformes".

Ces alternatives artisanales et autonomes ne sont pas de simples expérimentations ; elles représentent une proposition politique concrète face aux défis de la surveillance et de la centralisation du pouvoir numérique.

Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.

Learn more at Review Commons

Reply to the reviewers

Manuscript number: RC-2025-03206

Corresponding author(s____): Teresa M. Przytycka

General Statements

We thank all the reviewers for their time and their constructive criticism, based on which we have revised our manuscript. All review comments in are italics. Our responses are indicated in normal font except the excerpts from manuscript which are shown within double quote and in italics. The line numbers indicated here refer to those in the revised manuscript.

Point-by-point description of the revisions

Reviewer #1 (Evidence, reproducibility and clarity (Required)):

This paper addresses the interesting question of how cell size may scale with organ size in different tissues. The approach is to mine data from the fly single cell atlas (FCA) which despite its name is a database of gene expression levels in single isolated nuclei. Using this data, they infer cell size based on ribosomal protein gene expression, and based on this approach infer that there are tissue and sex specific differences in scaling, some of which may be driven by differences in ribosomal protein gene expression.

Response: Indeed, using the FCA dataset, we infer sex-specific differences in both cell size and cell number, which we validated with targeted experiments. We show that Drosophila cell types scale through distinct strategies-via cell size, cell number, or a mix of both-in an allometric rather than uniform fashion. We further propose that these scaling differences are driven, at least in part, by variation in translational activity, reflected in the expression of ribosomal proteins, translation elongation factors, and Myc.

-----------------------------------------------------------

I think the idea of mining this database is a clever one, however there a number of concerns about whether the existing data can really be used to draw the conclusions that are stated.

__Response: __We are pleased to see that the reviewer found the question and our approach interesting.

-----------------------------------------------------------

*One concern has to do with the assumption that RP (ribosome protein) expression is a proxy for cell size. It is well established that ribosome abundance scales with cell size, but is there reason to believe that ribosome nuclear gene EXPRESSION correlates with ribosome abundance? *

I'm not saying that this can't be true, but it seems like a big assumption that needs to be justified with some data. Maybe this is well known in the Drosophila literature, but in that case the relevant literature really needs to be cited.

__Response: __To avoid any misunderstanding: we use sex-biased RP expression as an indicator of sex differences in cell size only within the same cell type or subtype, as defined by expression-based clustering in the FCA-not as a general estimator of cell size. This measure is applied strictly within the same clusters, never between different ones. To prevent overinterpretation, we replaced the term 'proxy' with 'indicator,' since the earlier wording might have implied that ribosomal gene expression was being used to estimate cell size more broadly.

We should have begun by providing more background on the well-established link between ribosomal protein gene dosage and cell growth. This context was missing from the introduction, so we have now added a full paragraph outlining what is known about this connection:

*Added at line 85: *

"Cell growth, which supports both cell enlargement and cell division, demands elevated protein synthesis, accomplished by boosting translation rates. Indeed, ribosome abundance is known to scale with cell size in many organisms (Schmoller and Skotheim 2015; Cadart and Heald 2022; Serbanescu et al. 2022). Long before it was known that DNA was the carrier of genetic information, Drosophila researchers had identified a large class of mutations known as "Minutes" (Schultz 1929). These were universally haplo-insufficient. A single wild type copy resulted in a tiny slowly growing fly, and the homozygous loss-of-function alleles were lethal. In clones, the Minute cells are clearly smaller and compete poorly with surrounding wild type cells. We now know that most of the Minute loci encode ribosomal proteins (Marygold et al. 2007). Similarly, the Drosophila diminutive locus, also characterized by small flies almost a century ago, is now known to encode the Myc oncogene (Gallant 2013). This is significant as Myc is a regulator of ribosomal protein encoding genes in metazoans, including Drosophila (Grewal et al. 2005). The ribosome is assembled in a specialized nuclear structure called the nucleolus (Ponti 2025). Across species, including Drosophila (Diegmiller et al. 2021) and C. elegans (Ma et al. 2018), nucleolar size scales with cell size and is broadly correlated with growth in cell size and/or cell number, processes that are directly relevant to sex-specific allometry. Collectively, these and many other studies offer compelling evidence that ribosomal biogenesis is positively associated with cell size and growth, underscoring the value of measuring ribosome biogenesis as a metric."

We understand that the reviewer is asking whether reduced RP mRNA expression directly leads to reduced functional ribosome assembly. We do not have a definitive answer to that specific question. However, we directly measured translation in fat body cells (section: Female bias in ribosomal gene expression in fat body cells leads to sex-biased protein synthesis), and the results show a clear correlation between RP gene expression and biosynthetic activity; even though we did not track every step from transcription to ribosome assembly to polysome loading across all cell types. This would indeed be an excellent direction for future work, including polysome profiling and related assays. Importantly, we did examine the nucleolus (Figure 4), where ribosome assembly occurs, and showed that nucleolar volume scales with RP gene expression. This strongly supports the presence of sex-specific differences in ribosome biogenesis.

Added at line 115:

"Building on the earlier studies noted above, as well as our direct measurements of translation bias in the fat body, nucleolar size, and cell size, we used sex-biased expression of ribosomal proteins as an indicator of sex differences in per-nucleus cell size."

-----------------------------------------------------------

Second, the interpretation of RP expression as a proxy for cell size seems potentially at odds with the fact that some cells are multi-nucleate. Those cells are big because of multiple nuclei, and so they might not show any increase in ribosome expression per nucleus. presumably for multi-nucleate cells, RP expression if it reflects anything at all would be something to do with cell size PER nucleus.

Response: Yes, this is a very important point, and this is why we chose multinucleated indirect flight muscles for our direct experimental analysis. We show that in indirect flight muscle cells, adult cell size is greatly influenced by the sex-specific number of nuclei per cell. The female muscle cells are larger and have larger nuclei count per cell. Additionally, they also have higher expression of ribosomal protein coding genes. As the latter data are from the single nucleus sequencing atlas, this already demonstrates what this reviewer is asking for: per nucleus, female muscle cells express more ribosome protein coding mRNAs.

-----------------------------------------------------------

*Third, it is well known that many tissues in Drosophila are polyploid or polytene. I don't know enough about the methodology used to produce the FCA to know whether this is somehow normalized. Otherwise, my hypothesis would be that nuclei showing higher RP expression might just be polyploid or polytene. You might say that this could be controlled by asking if all genes are similary upregulated, but that isn't the case since at least in polytene chromosomes it is well known that only a small number of genes are expressed at a given time, while many are silent. *

Response: Yes, this is an excellent point. As noted above, our study does not distinguish among the different potential causes of sex differences in ribosomal mRNA copy number, as these may vary across cell types. We now explicitly acknowledge it in the discussion (line 327). Importantly, even in the cases when ribosomal gene expression bias primarily reflects differences in DNA content, this still represents a plausible mechanistic route linking ribosomal gene expression to increased nucleolar ribosome biogenesis and, ultimately, larger cell size. This possibility does not alter our main conclusions.

-----------------------------------------------------------

Overall, I think a lot more foundational work would need to be done in order to allow the inference of cell size from RP expression. In a way, it is a bit unfortunate that they chose to do this work in Drosophila where so many cells are polyploid, although I gather that even in humans some tissues have this issue, for example large neurons in the brain.

Response: We acknowledge that we did not clearly reference some of the foundational work in the literature. To address this, we have expanded the introduction to provide additional background and context. We also clarify that our fat body experiment offers independent support for the relationship between ribosomal gene expression bias, nuclear size bias, and corresponding biases in protein synthesis, thereby reinforcing the use of sex-specific ribosomal gene expression as an indicator of sex-specific cell size. Importantly, we assess this bias only within clusters, not between them. These clusters are derived from gene-expression-based clustering and are therefore relatively homogeneous. For example, as discussed in our response to Reviewer #3, the fat body contains several clusters that correspond to expression-defined subtypes of fat body cells. Our previous terminology may have inadvertently implied that we were using ribosomal gene expression to estimate cell size more broadly, which was not our intention.

As for the choice of the organism, most of the authors are Drosophila researchers and we benefit from the unique, highly replicated data from whole head and whole body of both sexes. Such data is necessary for a non-biased estimation of the differences in nuclear number.

-----------------------------------------------------------

*Reviewer #1 (Significance (Required)):

The idea that gene regulatory networks could "program" differences in scaling by changing levels of ribosomal protein gene expression is a tremendously important one if it can be established, because it would show a simple way for size scaling to be placed under control of developmental regulatory pathways. My original concern when I first looked at the abstract was going to be that yeah the results are interesting but a mechanism is not provided, but as I read it, that concern went away. showing that RP gene expression, which could be programmed by various driving pathways, can affect allometric scaling, would be extremely impactful and really change how we think about scaling, but putting it into the framework of gene expression networks that control other aspects of developmewnht. it would not be necessary to show which pathways actually drive these expression differences, the fact that they are different would be interesting enough to make everyone want to read this paper. But as discussed above I am not, however, convinced by the evidence presented here. So while I think it would be very significant if true, I am not convinced that the conclusion is well supported. This doesn't mean I have a reason to think it is false, just that its not well supported for the reasons I have given.*

Response: We are grateful to the reviewer for this positive assessment of our findings despite lack of a specific mechanism. We also regret that our initial writing did not clearly situate our work within the foundational literature on the relationship between ribosomal biogenesis and scaling. The key contribution of our study is to demonstrate that sex-biased ribosomal biogenesis plays a role in allometric scaling, providing a basis for future mechanistic exploration. We hope that the revised manuscript now offers clear and compelling support for the conclusion that RP gene expression bias can influence allometric scaling.

-----------------------------------------------------------

I hasten to point out that I could be entirely wrong, if the missing bits of logic (i.e. that RP expression matches ribosome abundance and that gene expression in the FCA dataset isn't influenced by ploidy of the nucleus). If suitable references can be provided to support these underlying assumptions, then in fact I think these concerns could be answered with very little effort. Otherwise, I think experiments would be needed to support these assumptions, and that might be non-trivial to do in a reasonable time frame. for that reason, in the next question I have put "cannot tell" for the time estimate.

Response: While gene expression in some FCA cell types may indeed be influenced by ploidy, our analysis does not depend on distinguishing among the possible sources of gene expression bias, which may vary across cell types. Rather, our key point is that-regardless of its origin-an increase in ribosomal gene expression is associated with enhanced ribosome biogenesis in the nucleolus and, ultimately, larger cell size. Thus, our main conclusions do not rely on any specific mechanism underlying RP gene expression upregulation. We now include additional references supporting the relationship between RP expression bias and cell size bias. We also strengthen the link between ribosomal gene expression and biosynthetic activity by clarifying its relationship with sex-biased Myc expression and the strong correlation with expression bias of EF1. We now include additional references supporting the relationship between RP expression bias and cell size bias. We also strengthen the link between ribosomal gene expression and biosynthetic activity by clarifying its relationship with sex-biased Myc expression and the strong correlation with expression bias of EF1.

We thank the reviewer for their thoughtful and constructive comments, which have prompted us to clarify both our reasoning and the relevant literature more fully.

-----------------------------------------------------------

*Reviewer #2 (Evidence, reproducibility and clarity (Required)):

The authors analyzed the FlyAtlas single-nucleus dataset to identify sex differences in gene expression and cell numbers. This led them to focus on muscles, cardiomyocytes, and fat body cells. They then measured cell and nucleolus size across different tissues and showed that reducing Myc function decreases sex differences in fat body cells. Overall, the manuscript provides a characterization of dimorphic differences in cell and organ size across three tissues.*

Response: This is a nice synopsis of the work.

-----------------------------------------------------------

Major Comments: The major claims of the manuscript are well supported by the reported experiments and analyses. While Reviewer #2 considered the major claims of the manuscript to be well supported, by the reported experiments and analysesStatistical analyses appear adequate.

Response: We agree, and we are glad that the reviewer found our work well supported.

-----------------------------------------------------------

*Minor Comments: The following minor issues should be addressed through textual edits:In the Introduction:

"Disruptions in proportionality, whether due to undergrowth or overgrowth, can lead to reduced fitness or diseases such as cancer." Could the authors provide a reference for this statement, particularly for the claim that disruptions in proportion*

Response: We apologize for this omission. The following explanation is now included starting at line 39:

"For example, scaled cell growth is a driver of symmetry in Myc-dependent scaling of bone growth in the skeleton by chondrocyte proliferation (Ota et al. 2007; Zhou et al. 2011). Increased nucleolus size is a well known marker of cancer progression in a histopathological setting (Pianese 1896; Derenzini et al. 1998; Elhamamsy et al. 2022)."

-----------------------------------------------------------

*The authors state:

"This study offers a comprehensive, cellular-resolution analysis of sexual size dimorphism in a model organism, uncovering how differences in cell number and size contribute to sex-specific body plans."*

The study cannot be considered comprehensive, as not all organs were examined.

Response: Indeed, "comprehensive" is a loaded word and in the revised manuscript we just omitted it.

-----------------------------------------------------------

*The following sentence from the abstract is unclear:

"By uncovering how a conserved developmental system produces sex-specific proportions through distinct cellular strategies..."*

* What do the authors mean by a conserved developmental system? Do they refer to a commonly used developmental model, or to a developmental system that is evolutionarily conserved?*

Response: We acknowledge that the use of the word 'conserved' was inappropriate, and we have therefore removed it from the statement.

-----------------------------------------------------------

*Reviewer #2 (Significance (Required)):

The manuscript presents a relevant exploration of sex-specific differences in cell size and cell number in Drosophila males and females. The limitations of the study are clearly acknowledged in the "Limitations" section. The work does not provide mechanistic insight into the causes or functional consequences of the observed differences. Nonetheless, the study extends our understanding of sexual dimorphism and establishes a foundation for future investigations into the autonomous and systemic mechanistic factors that regulate these differences.*

Response: Thank you.

-----------------------------------------------------------

*Reviewer #3 (Evidence, reproducibility and clarity (Required)):

The manuscript by Pal and colleagues addresses an important question: the cellular mechanisms underlying sex differences in organ size. By leveraging single-nucleus transcriptomic data from the adult Drosophila Cell Atlas, the authors show that different cell types adopt distinct strategies to achieve sex differences in organ size-either by increasing cell size or by altering cell number. They then focus on three organs-the indirect flight muscles, the heart, and the fat body-and provide supporting evidence for their transcriptomic analyses.*

Response: This is a nice summary of the study. Thank you.

-----------------------------------------------------------

This study tackles a highly relevant and often overlooked question, as our understanding of the molecular and cellular events driving sex differences remains incomplete. The work presents interesting observations; however, it is largely descriptive, establishing correlations without providing functional evidence or mechanistic insight.

Response: We agree that this is an often overlooked problem that has been difficult to address experimentally without single-cell genomics. Our work aims to help fill this gap. While the paper does contain descriptive elements, we believe such characterization is important at the early stages of developing a new area of inquiry. The study explores a unique dataset and includes experimental validation to support key observations. We also propose how allometry may be shaped by cell division and cell size, drawing on well-established molecular mechanisms. Thus, the reviewer's comment regarding a lack of mechanistic insight likely pertains to the absence of a direct connection to the sex-determination pathway, which is beyond the scope of the current study.

-----------------------------------------------------------

Below are four main points that should be addressed before publication: 1. Introduction and contextualisation of prior work The introduction does not adequately present the current state of knowledge. Several key studies are missing or insufficiently discussed. In particular, the following works should be included and integrated into the introduction: - PMID: 26710087 - shows that the sex determination gene transformer regulates male-female differences in Drosophila body size. - PMID: 28064166 - describes how differences in Myc gene dosage contribute to sex differences in body size. - PMID: 26887495 - demonstrates that the intrinsic sexual identity of adult stem cells can control sex-biased organ size through sex-biased proliferation. - PMID: 28976974 - reveals that Sxl modulates body growth through both tissue-autonomous and non-autonomous mechanisms. - PMID: 39138201 - shows that transformer drives sex differences in organ size and body weight. Incorporating and discussing these references would provide a more comprehensive and up-to-date framework for the study.

Response: We agree that the literature suggested by the reviewer strengthens the introduction and improves the contextualization of prior work relevant to our study. Although much of it was previously included in the discussion section on cell-autonomous and hormonal regulation, it has now been moved to the introduction, along with the discussion of the papers suggested by the reviewer (beginning at line 58).

"In Drosophila melanogaster, adult females are substantially larger than males (Fig. 1A1), yet both sexes develop from genetically similar zygotes and share most organs and cell types. In wild type flies, sex is determined by the number of X chromosomes in embryos, with XX flies developing as females and X(Y) flies developing as males due to the activation and stable expression of Sex-lethal only in XX flies (Erickson and Quintero 2007). While it is not entirely clear how sexually dimorphic size is regulated, the sex determination pathway is implicated in size regulation. Sex-reversed flies often show a size based on the X chromosome number rather than sexual morphology. Female Sex-lethal contributes to larger female size independently of sexual identity (Cline 1984), and Sex-lethal expression in insulin producing neurons in the brain also impacts body size (Sawala and Gould 2017). Female-specific Transformer protein is produced as a consequence of female-specific Sex-lethal and also contributes to increased female size (Rideout et al. 2015). This size scaling also applies to individual organs. For example, the Drosophila female gut is longer than the male gut due Transformer activity (Hudry et al. 2016). It has also been suggested that Myc dose (it is X-linked) is a regulator of body size (Mathews et al. 2017), although the failed dosage compensation model proposed has not been demonstrated."

And again at line 74:

"These studies show that size is regulated, but they do not address whether scaling is uniform or non-uniform and the mechanism for sexual size differences (SSD). The origins of SSD can, in principle, arise from differences in (i) gene expression, (ii) the presence of sex-specific cell types, (iii) the number of cell-specific nuclei, or (iv) the size (per nucleus) of those cells. Previous research in Drosophila has largely focused on gene expression in sex-specific organs like the gonads (Arbeitman et al. 2002; Parisi et al. 2004; Graveley et al. 2011; Pal et al. 2023), which are governed by a well-characterized sex-determination pathway (Salz and Erickson 2010; Clough and Oliver 2012; Raz et al. 2023) However, whether and how scaling differences in shared, non-sex-specific tissues are achieved via changes in cell size and number remains largely unexamined (Fig. 1A2). These studies show that size is regulated, but they do not address whether scaling is uniform or non-uniform and the mechanism for size differences."

-----------------------------------------------------------

2. Use of ribosomal gene expression as a proxy for cell size The authors use ribosomal gene expression levels as a proxy for cell size, but this assumption is not adequately justified. The cited references (refs. 20-22) focus on unicellular organisms (bacteria and yeast) or cleavage divisions in frog embryos, which are fundamentally different from adult Drosophila tissues. The authors should provide evidence that ribosome abundance scales with cell size across the distinct adult Drosophila cell types. Given that most adult fly tissues are post-mitotic, it is more likely that ribosomal gene expression reflects protein synthesis activity rather than cell size, particularly in secretory cell types.

Response: Reviewer 1 raised a similar point, and we agree. We recognize that the term "proxy" may have been misleading. We use this measure only in the context of sex bias within homogeneous cell clusters, and not between clusters, even when such clusters share the same cell-type annotation. To avoid overinterpretation, we changed "poxy" to "indicator".

In response to the reviewer's concern, we have expanded our discussion of the relevant supporting literature (additional text starting line 75). We have also directly measured translation in the fat body cells (section: Female bias in ribosomal gene expression in fat body cells leads to sex biased protein synthesis), which clearly demonstrates a correlation between ribosomal protein gene expression and biosynthetic activity. Although, we have not traced the chain of events from expression to ribosome assembly to polysome loading in all cell types, we did examine the nucleolus (Figure 4), where ribosomes are assembled, and we make a strong point that the volume of the nucleolus scales like ribosome protein gene expression. This provides strong evidence for sex-specific ribosome biogenesis contributing to cell size.

Furthermore, the observation that ribosomal gene expression likely reflects protein synthesis activity is not at odds with increased cell size: biosynthesis increases in larger cells (Schmoller and Skotheim 2015). We have added a panel to Figure 4 showing the relationship between ribosomal gene expression bias and the average expression bias of Eukaryotic Elongation Factor 1 (eEF1).

-----------------------------------------------------------

3. Relationship between Myc and sex-biased Rp expression The proposed link between Myc and sex-biased Rp expression is unclear. Panels D and E of Figure 1 show no consistent relationship: some cell types with strong Rp sex bias exhibit either high or low female Myc bias, or even a male bias. The linear regression in Figure 4I (R = 0.07, p = 0.59) confirms the lack of correlation. The authors should clarify this point and adopt a more cautious interpretation regarding Myc as a potential regulator of sex-biased Rp expression and cell size differences. Experimentally, using Myc hypomorph or heterozygous conditions would be more appropriate than complete knockdown to test its role.

Response: Thank you for noting that the relationship between Myc expression bias and sex-biased RP expression required clarification. This response was prepared in consultation with Myc expert Dr. David Levens.

We demonstrate that both Myc and RP gene expression exhibit an overall female bias in the body. The absence of a strong correlation across cell clusters does not invalidate this conclusion. Myc is a well-established master regulator of ribosome biogenesis, but its quantitative effects are complex. According to recent models of Myc-mediated gene regulation (Nie et al. 2012; Lin et al. 2012), Myc upregulates all actively transcribed genes. Because this regulation is global, the relationship between changes in Myc expression and corresponding changes in ribosomal protein gene expression depends on cell type. Moreover, (Lorenzin et al. 2016) demonstrated that ribosomal protein genes saturate at relatively low levels of Myc, which helps explain why we observe a correlation in head cell clusters-where Myc expression is lower-but not in body clusters.

Importantly, on average, the female-specific Myc expression bias is stronger in body cell clusters than in head cell clusters, consistent with the stronger female bias in ribosomal protein gene expression observed in the head relative to the body.

To make this relationship more transparent, we combined the head and body clusters, which yielded a strong overall correlation (Fig. 4J, replacing the previous Fig. 4H).

To further strengthen the evidence linking ribosomal gene expression to cell size, we also examined the relationship between ribosomal gene expression bias and Elongation Factor 1 (eEF1) expression bias, a key component of protein biosynthesis during the elongation step of translation. The resulting correlation exceeds 0.9 (new Fig. 4H, added as an additional panel in Fig. 4).

-----------------------------------------------------------

4. Conclusions about fat body cell number I have concerns about drawing conclusions on sex differences in fat body cell number from single-nucleus transcriptomic data for two reasons:

1- Drosophila fat body tissue is heterogeneous, comprising distinct subpopulations (e.g., visceral fat cells, subcuticular fat cells), some of which are sex-specific-such as fat cells associated with the spermathecae in females.

Response: Thank you for giving us the opportunity to clarify our analysis of the FCA data. Our approach does account for subpopulations within the fat body as well as within other cell types. Based on gene expression profiles, we identify three fat body clusters, all of which are reported in Table S3. One small female-specific cluster (

When all fat body clusters are combined into a single supercluster, this supercluster still shows a male bias. We have now clarified this point in the manuscript (line 113). Note that both subclusters of fat body are already shown in Fig. 1C and 1D.

-----------------------------------------------------------

2- Adult fat body cells can be multinucleated (PMID: 13723227). Apparent sex differences in nucleus number may reflect differences in specific subpopulations or degrees of multinucleation rather than true differences in cell number. To strengthen the conclusions, the analysis should be performed at the level of fat body subpopulations, distinguishing clusters where possible. Additionally, quantifying nuclei relative to actual cell number-as done for muscle tissue-would clarify whether observed sex differences reflect true variation in cell number or differences in nuclear content per cell.

Response: Yes, some cells can be multinucleate. We specifically address this in the context of muscle cells, where multinucleation is prominent, and we also conducted experimental validation in this tissue. As noted above, our analysis is performed at the subpopulation level, since clusters are defined by expression similarity (Leiden resolution 4.0) rather than by annotation.

Because our work relies on single-nucleus data, each nucleus is treated as an individual unit of analysis. Nevertheless, we observe genuine nuclear differences within each cluster. Importantly, the presence of multinucleated cells does not alter our conclusions; it simply represents one form of variation in cell number that can be thought of as a subcomponent of cell/nuclei number.

-----------------------------------------------------------

Minor corrections/points: 1-The term body size in the title does not accurately reflect the content of the paper. I recommend replacing it with organ size to better align with the study's focus.

Response: Thank you for the suggestion.

----------------------------------------------------------- 2-The term sexual size dimorphism is somewhat inaccurate in this context. Sex differences in size would be more appropriate. The term sexual dimorphism typically refers to traits that exhibit two distinct forms in males and females-such as primary or secondary sexual characteristics like sex organs or sex combs. In contrast, size is a quantitative trait that follows a normal distribution. Although the average female may be larger than the average male, the distributions overlap, making the term dimorphism imprecise.

Response: Thank you for the suggestion.

-----------------------------------------------------------

3-In Figure 2E, there appears to be an inconsistency between the text, figure legend, and the data presented. The text and legend state that the total volume of dorsal longitudinal flight muscle cells was quantified, whereas the graph indicates measurements of nuclear size. This discrepancy should be clarified.

Response: Thank you for pointing this out. We figured out that Y-axis label in the graph was incorrect and it is now fixed.

-----------------------------------------------------------

4-The authors proposed: "This increased biosynthetic activity in fat body cells may contribute to cell size differences, but also to the regulation of body size via production of factors that mediate body growth via interorgan communication". Please note that this hypothesis has already been tested functionally in PMID: 39138201 and was shown to be incorrect. Sex differences in body size are completely independent of fat body sexual identity or any intrinsic sex differences within fat cells.

__Response: __We thank the reviewer for the opportunity to discuss why the data shown in PMID 39138201 (Hérault et al. 2024) do not rule out a model in which the fat body contributes to the sex-specific regulation of body size via interorgan communication. The main reason data in Herault et al cannot rule out such a model is that they use wing size as a proxy for body size. This is in contrast to prior studies, such as (Rideout et al. 2015), in which pupal volume was used to directly measure body size and show a non-autonomous effect of sex determination gene transformer on body size. Measuring body size directly is a more precise readout of growth during the larval stages of development, as opposed to using adult wing area which reflects the growth of a single organ. It is also important to note that the diets used to rear flies in Herault and Rideout differ, which is an important consideration as females do not achieve their maximal size without high dietary protein levels (Millington et al. 2021). To ensure all these points are communicated to readers, we added text to this effect in the revised version of our manuscript.

Added at line 254:

"This increased biosynthetic activity in fat body cells may contribute to cell size differences, but also to the regulation of body size via production of factors that mediate body growth via interorgan communication (Colombani et al. 2003; Géminard et al. 2009; Rajan and Perrimon 2012; Sano et al. 2015; Koyama and Mirth 2016). Indeed, one study showed the sexual identity of the fat body influenced pupal volume, which is an accurate readout of larval growth (Rideout et al. 2015; Delanoue et al. 2010). While a recent study suggests that male-female differences in body size were regulated independently of fat body sexual identity (Hérault et al. 2024), this study measured the growth of a single organ, the wing, as a proxy for body size. Additional studies are therefore needed to resolve whether fat body protein synthesis plays an important role in regulating sex differences in body size."

-----------------------------------------------------------

*5-The authors state: "This demonstrate that Myc plays a key role in regulating the sex difference in nucleolar size." This is an overstatement given the functional data presented. The claim should be toned down to reflect the limited evidence.

**Referee cross-commenting**

I completely agree with the main comments of Reviewer 1, as they address the paper's core.*

Response: We have addressed the comments of Reviewer 1 in the response to reviewer's comments above.

-----------------------------------------------------------

*Reviewer #3 (Significance (Required)):

The main novelty and strongest aspect of this study is its use of single-nucleus transcriptomic data from the adult Drosophila Cell Atlas to investigate how different cell types adopt distinct strategies to generate sex differences in organ size-either by increasing cell size or by altering cell number. Previous studies have largely focused on specific tissues, whereas this work provides a comprehensive, organism-wide view that encompasses all tissues, enabling direct cross-comparison between organs. This represents a clear advance in the field, primarily from a technical perspective, by leveraging organism-wide single-cell transcriptomics. The main limitations lie in the lack of functional experiments and mechanistic insights. Moreover, the proposed mechanism-differences in Myc gene dosage or expression levels-is not entirely novel, as Myc dosage has previously been implicated in contributing to sex differences in body size (PMID: 28064166).*

Response: We do have some functional testing in the 3 tissues, flight muscle, heart and fat body, however, providing mechanistic insights is beyond the scope of this paper. The paper suggested by the reviewer is an example of one attempt to provide such a mechanism, probably not the only one. We hope that our rich data that we have assembled in this paper provide resources for generating hypotheses and stimulate further research.

-----------------------------------------------------------

References

Cadart, Clotilde, and Rebecca Heald. 2022. "Scaling of Biosynthesis and Metabolism with Cell Size." Molecular Biology of the Cell 33 (9): pe5. https://doi.org/10.1091/mbc.E21-12-0627.

Diegmiller, Rocky, Caroline A. Doherty, Tomer Stern, Jasmin Imran Alsous, and Stanislav Y. Shvartsman. 2021. "Size Scaling in Collective Cell Growth." Development (Cambridge, England) 148 (18): dev199663. https://doi.org/10.1242/dev.199663.

Gallant, Peter. 2013. "Myc Function in Drosophila." Cold Spring Harbor Perspectives in Medicine 3 (10): a014324. https://doi.org/10.1101/cshperspect.a014324.

Grewal, Savraj S., Ling Li, Amir Orian, Robert N. Eisenman, and Bruce A. Edgar. 2005. "Myc-Dependent Regulation of Ribosomal RNA Synthesis during Drosophila Development." Nature Cell Biology 7 (3): 295-302. https://doi.org/10.1038/ncb1223.

Hérault, Chloé, Thomas Pihl, and Bruno Hudry. 2024. "Cellular Sex throughout the Organism Underlies Somatic Sexual Differentiation." Nature Communications 15 (1): 6925. https://doi.org/10.1038/s41467-024-51228-6.

Lin, Charles Y., Jakob Lovén, Peter B. Rahl, et al. 2012. "Transcriptional Amplification in Tumor Cells with Elevated C-Myc." Cell 151 (1): 56-67. https://doi.org/10.1016/j.cell.2012.08.026.

Lorenzin, Francesca, Uwe Benary, Apoorva Baluapuri, et al. 2016. "Different Promoter Affinities Account for Specificity in MYC-Dependent Gene Regulation." eLife 5 (July): e15161. https://doi.org/10.7554/eLife.15161.

Ma, Tian-Hsiang, Po-Hsiang Chen, Bertrand Chin-Ming Tan, and Szecheng J. Lo. 2018. "Size Scaling of Nucleolus in Caenorhabditis Elegans Embryos." Biomedical Journal 41 (5): 333-36. https://doi.org/10.1016/j.bj.2018.07.003.

Marygold, Steven J., John Roote, Gunter Reuter, et al. 2007. "The Ribosomal Protein Genes and Minute Loci of Drosophila Melanogaster." Genome Biology 8 (10): R216. https://doi.org/10.1186/gb-2007-8-10-r216.

Millington, Jason W., George P. Brownrigg, Charlotte Chao, et al. 2021. "Female-Biased Upregulation of Insulin Pathway Activity Mediates the Sex Difference in Drosophila Body Size Plasticity." eLife 10 (January): e58341. https://doi.org/10.7554/eLife.58341.

Nie, Zuqin, Gangqing Hu, Gang Wei, et al. 2012. "C-Myc Is a Universal Amplifier of Expressed Genes in Lymphocytes and Embryonic Stem Cells." Cell 151 (1): 68-79. https://doi.org/10.1016/j.cell.2012.08.033.

Ponti, Donatella. 2025. "The Nucleolus: A Central Hub for Ribosome Biogenesis and Cellular Regulatory Signals." International Journal of Molecular Sciences 26 (9): 4174. https://doi.org/10.3390/ijms26094174.

Rideout, Elizabeth J., Marcus S. Narsaiya, and Savraj S. Grewal. 2015. "The Sex Determination Gene Transformer Regulates Male-Female Differences in Drosophila Body Size." PLOS Genetics 11 (12): e1005683. https://doi.org/10.1371/journal.pgen.1005683.

Schmoller, Kurt M., and Jan M. Skotheim. 2015. "The Biosynthetic Basis of Cell Size Control." Trends in Cell Biology 25 (12): 793-802. https://doi.org/10.1016/j.tcb.2015.10.006.

Schultz, J. 1929. "The Minute Reaction in the Development of DROSOPHILA MELANOGASTER." Genetics 14 (4): 366-419. https://doi.org/10.1093/genetics/14.4.366.

Serbanescu, Diana, Nikola Ojkic, and Shiladitya Banerjee. 2022. "Cellular Resource Allocation Strategies for Cell Size and Shape Control in Bacteria." The FEBS Journal 289 (24): 7891-906. https://doi.org/10.1111/febs.16234.

DOI: 10.64898/2025.12.05.692606

Resource: (BD Biosciences Cat# 550538, RRID:AB_393732)

Curator: @scibot

SciCrunch record: RRID:AB_393732

What is this?

DOI: 10.6084/m9.figshare.30661148

Resource: (DSHB Cat# BA-D5, RRID:AB_2235587)

Curator: @scibot

SciCrunch record: RRID:AB_2235587

What is this?

DOI: 10.1200/OA-25-00109

Resource: FlowJo (RRID:SCR_008520)

Curator: @scibot

SciCrunch record: RRID:SCR_008520

What is this?

DOI: 10.1200/OA-25-00109

Resource: BD LSRFortessa X-20 Cell Analyzer (RRID:SCR_025285)

Curator: @scibot

SciCrunch record: RRID:SCR_025285

What is this?

DOI: 10.1200/OA-25-00109

Resource: Miltenyi gentleMACS Octo Dissociator with Heaters Tissue Dissociator (RRID:SCR_020271)

Curator: @scibot

SciCrunch record: RRID:SCR_020271

What is this?