Author response:

We would like to thank the editors and reviewers for taking the time to help improve our manuscript. We appreciate the feedback and will definitely increase the level of methodological detail in a revised submission.

Here is a brief summary of our plan to address the points raised by the reviewers. We will respond to the comments in a point-by-point manner when we resubmit a revised manuscript.

Reviewer 1

This reviewer raised a question about the 60 Hz frame rate for recording. We agree that increasing the number of cameras and frame rate would improve the tracking quality, but this would come at the cost of scalability. In the current study (and other concurrent studies in the lab), we recorded from 10-20 families simultaneously to try to sample the distribution of behavioral responses to stimuli observed in animals in our colony. This was only possible logistically because of the lightweight equipment design allowing us to record data from animals without large disruptions to their home-cage environment.

One strategy for acquiring higher-resolution data is to build a small number of enclosures that are fully surrounded by cameras, and to cycle animals through these enclosures (1). However, this strategy limits throughput by reducing the number of animals per day that can be studied. If the size and cost of cameras and computers decreases in the future, then this recording strategy will be scalable to the whole-colony level. For our current study and analysis, we are limited by the resolution of our dataset. We do believe that our data (although not a perfect 3d reconstruction or an extremely high frame rate) is sufficient to label behavioral states with high accuracy. We will add a figure to more clearly show that behavioral state data can be accurately inferred from this imperfect data, which has also been recently highlighted by other groups (2).

Additionally, with recent progress in the application of deep learning to animal pose tracking, new models can infer 3d pose dynamics from 2d data (3) and leverage spatiotemporal structure to clean up noisy data (4). We believe that other groups will be able to use these types of approaches to extract much more value from this dataset. So, in summary, we do understand the concern related to reconstruction quality and will 1) more clearly define the usefulness of our current models, 2) release our data and code so that others can build upon it or repurpose it, and 3) plan future experiments with higher camera count and frame rate as permitted by logistical constraints. 

Reviewer 2

This reviewer asked for an increased level of methodological detail. We will try to address this in a few ways:

(1) Code and data sharing. We believe that many of the questions related to the methodology will be best answered by sharing the data and code directly. Because there is a large amount of code associated with this manuscript, it is impractical to list every step and every parameter in the paper. Along with our revised manuscript, we will make our data and code publicly available. That said, we will improve our description of key parameters in the paper as the reviewer suggested.

(2) More detailed Methods section. The reviewer asked us to provide more methodological detail. We understand that this is currently a weakness of our manuscript, and we will focus on addressing it. For instance, the reviewer rightly points out that we did not describe the motion watches used to generate the data in Figure S7. We will address this.

(3) Simplify the manuscript. The paper currently has 22 figures, and further analysis could be done based on the results shown in any of them. For instance, this reviewer asked us to add a comparison across females and males (similar to our comparison of juveniles and adults). While we plan to add that analysis, we recognize that there are several figures/panels that are not closely related to our intended goal of describing the patterns we found in our large dataset. We will simplify the manuscript by removing some excess figures/panels and focus on describing the parts of the analysis that are crucial to our conclusions in greater detail.

(4) More careful language. This reviewer pointed out that there were some inaccuracies with our descriptive language. For instance, we used the term "natural" behavior to describe the behavior of animals in captivity, which may more accurately be described as their home-cage behavior. We will be more careful to align our language to the standard for the field. For instance, several studies refer to unrestrained behavior in a laboratory setting as "spontaneous" behavior rather than "natural" behavior (5). In our case, the data consists of both spontaneously occurring behavior and responses to a set of stimuli. We will make sure that the descriptions are more precise in the revised manuscript.

(1) Bala, P. C. et al. Automated markerless pose estimation in freely moving macaques with OpenMonkeyStudio. Nat Commun 11, (2020).

(2) Weinreb, C. et al. Keypoint-MoSeq: parsing behavior by linking point tracking to pose dynamics. bioRxiv (2023) doi:10.1101/2023.03.16.532307.

(3) Gosztolai, A. et al. LiftPose3D, a deep learning-based approach for transforming two-dimensional to three-dimensional poses in laboratory animals. Nat Methods 18, 975–981 (2021).

(4) Wu, A. et al. Deep Graph Pose: a semi-supervised deep graphical model for improved animal pose tracking. Adv Neural Inf Process Syst 33, 6040–6052 (2020).

(5) Levy, D. R. et al. Mouse spontaneous behavior reflects individual variation rather than estrous state. Curr Biol 33, 1358-1364.e4 (2023).

プログラムの動きが見えるツール

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

Reviewer #1 (Public review):

Summary:

The authors test the hypotheses, using an effort-exertion and an effort-based decision-making task, while recording brain dynamics with EEG, that the brain processes reward outcomes for effort differentially when they earned for themselves versus others.

Strengths:

The strengths of this experiment include what appears to be a novel finding of opposite signed effects of effort on the processing of reward outcomes when the recipient is self versus others. Also, the experiment is well-designed, the study seems sufficiently powered, and the data and code are publicly available.

Weaknesses:

Inferences rely heavily on the results of mixed effects models which may or may not be properly specified and are not supported by complementary analyses. Also, not all results hang together in a sensible way. For example, participants report feeling less subjective effort, but also more disliking of tasks when they were earning rewards for others versus self. Given that participants took longer to complete tasks when earning effort for others, it is conceivable that participants might have been working less hard for others versus themselves, and this may complicate the interpretation of results.

CONVERSATIONAL

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.ページ下のVisualize Executionボタンをクリック

Reviewer #2 (Public review):

Summary:

This work investigates transcriptional responses to varying levels of transcription factors (TFs). The authors aim for gradual up- and down-regulation of three transcription factors GFI1B, NFE2, and MYB in K562 cells, by using a CRISPRa- and a CRISPRi line, together with sgRNAs of varying potency. Targeted single-cell RNA sequencing is then used to measure gene expression of a set of 90 genes, which were previously shown to be downstream of GFI1B and NFE2 regulation. This is followed by an extensive computational analysis of the scRNA-seq dataset. By grouping cells with the same perturbations, the authors can obtain groups of cells with varying average TF expression levels. The achieved perturbations are generally subtle, not reaching half or double doses for most samples, and up-regulation is generally weak below 1.5-fold in most cases. Even in this small range, many target genes exhibit a non-linear response. Since this is rather unexpected, it is crucial to rule out technical reasons for these observations.

Strengths:

The work showcases how a single dataset of CRISPRi/a perturbations with scRNA-seq readout and an extended computational analysis can be used to estimate transcriptome dose responses, a general approach that likely can be built upon in the future.

Weaknesses:

(1) The experiment was only performed in a single replicate. In the absence of an independent validation of the main findings, the robustness of the observations remains unclear.

(2) The analysis is based on the calculation of log-fold changes between groups of single cells with non-targeting controls and those carrying a guide RNA driving a specific knockdown. How the fold changes were calculated exactly remains unclear, since it is only stated that the FindMarkers function from the Seurat package was used, which is likely not optimal for quantitative estimates. Furthermore, differential gene expression analysis of scRNA-seq data can suffer from data distortion and mis-estimations (Heumos et al. 2023 (https://doi.org/10.1038/s41576-023-00586-w), Nguyen et al. 2023 (https://doi.org/10.1038/s41467-023-37126-3)). In general, the pseudo-bulk approach used is suitable, but the correct treatment of drop-outs in the scRNA-seq analysis is essential.

(3) Two different cell lines are used to construct dose-response curves, where a CRISPRi line allows gene down-regulation and the CRISPRa line allows gene upregulation. Although both lines are derived from the same parental line (K562) the expression analysis of Tet2, which is absent in the CRISPRi line, but expressed in the CRISPRa line (Figure S3A) suggests substantial clonal differences between the two lines. Similarly, the PCA in S4A suggests strong batch effects between the two lines. These might confound this analysis.

(4) The study uses pseudo-bulk analysis to estimate the relationship between TF dose and target gene expression. This requires a system that allows quantitative changes in TF expression. The data provided does not convincingly show that this condition is met, which however is an essential prerequisite for the presented conclusions. Specifically, the data shown in Figure S3A shows that upon stronger knock-down, a subpopulation of cells appears, where the targeted TF is not detected anymore (drop-outs). Also Figure 3B (top) suggests that the knock-down is either subtle (similar to NTCs) or strong, but intermediate knock-down (log2-FC of 0.5-1) does not occur. Although the authors argue that this is a technical effect of the scRNA-seq protocol, it is also possible that this represents a binary behavior of the CRISPRi system. Previous work has shown that CRISPRi systems with the KRAB domain largely result in binary repression and not in gradual down-regulation as suggested in this study (Bintu et al. 2016 (https://doi.org/10.1126/science.aab2956), Noviello et al. 2023 (https://doi.org/10.1038/s41467-023-38909-4)).

(5) One of the major conclusions of the study is that non-linear behavior is common. This is not surprising for gene up-regulation, since gene expression will reach a plateau at some point, but it is surprising to be observed for many genes upon TF down-regulation. Specifically, here the target gene responds to a small reduction of TF dose but shows the same response to a stronger knock-down. It would be essential to show that his observation does not arise from the technical concerns described in the previous point and it would require independent experimental validations.

(6) One of the conclusions of the study is that guide tiling is superior to other methods such as sgRNA mismatches. However, the comparison is unfair, since different numbers of guides are used in the different approaches. Relatedly, the authors point out that tiling sometimes surpassed the effects of TSS-targeting sgRNAs, however, this was the least fair comparison (2 TSS vs 10 tiling guides) and additionally depends on the accurate annotation of TSS in the relevant cell line.

(7) Did the authors achieve their aims? Do the results support the conclusions?: Some of the most important conclusions are not well supported because they rely on accurately determining the quantitative responses of trans genes, which suffers from the previously mentioned concerns.

(8) Discussion of the likely impact of the work on the field, and the utility of the methods and data to the community:<br /> Together with other recent publications, this work emphasizes the need to study transcription factor function with quantitative perturbations. Missing documentation of the computational code repository reduces the utility of the methods and data significantly.

numerics = ['int16', 'int32', 'int64'] numerical_columns= adult_census.select_dtypes(include=numerics).columns.to_list() len(numerical_columns)

pd.read_csv("../datasets/penguins_classification.csv")

The primary virtue of this approach is that it is extremely simple. However, using return codes has a number of drawbacks which can quickly become apparent when used in non-trivial cases:

First, return values can be cryptic -- if a function returns -1, is it trying to indicate an error, or is that actually a valid return value? It’s often hard to tell without digging into the guts of the function or consulting documentation.

Second, functions can only return one value, so what happens when you need to return both a function result and a possible error code? Consider the following function:

double divide(int x, int y) { return static_cast<double>(x)/y; } This function is in desperate need of some error handling, because it will crash if the user passes in 0 for parameter y. However, it also needs to return the result of x/y. How can it do both? The most common answer is that either the result or the error handling will have to be passed back as a reference parameter, which makes for ugly code that is less convenient to use. For example:

include <iostream>

double divide(int x, int y, bool& outSuccess) { if (y == 0) { outSuccess = false; return 0.0; }

outSuccess = true;
return static_cast<double>(x)/y;

}

int main() { bool success {}; // we must now pass in a bool value to see if the call was successful double result { divide(5, 3, success) };

if (!success) // and check it before we use the result
    std::cerr << "An error occurred" << std::endl;
else
    std::cout << "The answer is " << result << '\n';

} Third, in sequences of code where many things can go wrong, error codes have to be checked constantly. Consider the following snippet of code that involves parsing a text file for values that are supposed to be there:

std::ifstream setupIni { "setup.ini" }; // open setup.ini for reading // If the file couldn't be opened (e.g. because it was missing) return some error enum if (!setupIni) return ERROR_OPENING_FILE;

// Now read a bunch of values from a file if (!readIntegerFromFile(setupIni, m_firstParameter)) // try to read an integer from the file return ERROR_READING_VALUE; // Return enum value indicating value couldn't be read

if (!readDoubleFromFile(setupIni, m_secondParameter)) // try to read a double from the file return ERROR_READING_VALUE;

if (!readFloatFromFile(setupIni, m_thirdParameter)) // try to read a float from the file return ERROR_READING_VALUE; We haven’t covered file access yet, so don’t worry if you don’t understand how the above works -- just note the fact that every call requires an error-check and return back to the caller. Now imagine if there were twenty parameters of differing types -- you’re essentially checking for an error and returning ERROR_READING_VALUE twenty times! All of this error checking and returning values makes determining what the function is trying to do much harder to discern.

Fourth, return codes do not mix with constructors very well. What happens if you’re creating an object and something inside the constructor goes catastrophically wrong? Constructors have no return type to pass back a status indicator, and passing one back via a reference parameter is messy and must be explicitly checked. Furthermore, even if you do this, the object will still be created and then has to be dealt with or disposed of.

Finally, when an error code is returned to the caller, the caller may not always be equipped to handle the error. If the caller doesn’t want to handle the error, it either has to ignore it (in which case it will be lost forever), or return the error up the stack to the function that called it. This can be messy and lead to many of the same issues noted above.

Using return codes is simpler but ends up linked to the control flow, constraining both how the code is laid out, and how errors can be reasonably handled.

For complex projects, build automation tools (such as make or build2) are often used to help automate the process of building programs and running automated tests. While such tools are powerful and flexible, because they are not part of the C++ core language, nor do you need to use them to proceed, we’ll not discuss them as part of this tutorial series.

Some programming fonts, such as Fira Code, use ligatures to combine such “art” back into a single character. For example, instead of displaying !=, Fira Code will display ≠ (using the same width as the two-character version). Some people find this easier to read, others prefer sticking with a more literal interpretation of the underlying characters.

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.ページ下のVisualize Executionをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.ページ下のVisualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.ページ下のVisualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.ページ下のVisualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.ページ下のVisualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

プログラムの動きが見えるツール

使用方法

1.ここをクリック

2.Visualize Executionボタンをクリック

How can they say "sofware devs" explicitly, wasn't the data anonymized? or to what extent?

This makes a lot of sense as it would reflect those first adopters being technical people (who code for instance), and or students (who where among the first appealed with the technology). However it is interesinf that there is an explicit focus on web and mobile app dev., is this an indicator that non-technical people are trying to leverage it for entreprenurial purposes for instace? again this notion of intent seems relevant in this context of evaluating use.

You need to remove the upload link to Miro.

Firstly, Miro is not open source, so the Miro component is not open and can't be part of the OER. Secondly, uploading the collage serves no stated purpose (why upload, to whom?) b) It creates the following ethical issues. Firstly, to collect data in this way you first must ask for and receive Informed Consent https://github.com/mrtayto/antart/wiki/Informed-Consent

You don't explain why you might be collecting the data, nor do you ask for consent. There's then the danger that the uploaded data might include Sensitive Data https://github.com/mrtayto/antart/wiki/Sensitive-Data

If you create or use such digital data, you need to manage that data according to the University's Research Data Management Policy (link). This means you have to inform the learner why you are collecting the data, how it will be securely stored, how it will be anonymised, how you will use it, when and how you will destroy the data. Since you haven't done any of these things, the link has to be removed.

Tumblr Blog (Account) Embeds

Relatively angry with myself that I somehow never made it to this page in the Tumblr docs, somehow.

To embed a blog, use this code:

https://blogname.tumblr.com/js

Replace “blogname” with your username. If you’re using a custom domain, replace “blogname.tumblr.com” with your domain.

For example, say you wanted to embed the Changes blog somewhere. You’d use this code:

<script type='text/javascript' src='https://blogname.tumblr.com/js'></script>

<footer>Tumblr Developer<cite> https://help.tumblr.com/embed-basics/</cite></footer>

Documentation

https://help.tumblr.com/embed-basics/

To embed a blog, use this code:

[https://blogname.tumblr.com/js](https://blogname.tumblr.com/js)

Replace “blogname” with your username. If you’re using a custom domain, replace “blogname.tumblr.com” with your domain.

For example, say you wanted to embed the Changes blog somewhere. You’d use this code:

```

```

<script type='text/javascript' src='https://asphaltapostle.tumblr.com/js'></script> `

I think the fact we control it brings about these concerns much more than AI becoming evil in a way similar to that described above in the article. Humans are imperfect creatures. We make mistakes and are prone to bias in our actions. AI as a result will be built on our actions, including the mistakes and biases we have. It can already be observed in some ways. The reading "The Unseen Black Faces of AI Algorithms" we did for class discusses this in a rather direct fashion. Due to the algorithms we create for facial recognition being built on mainly white faces, it struggles to pick up faces of darker skinned individuals due to the lack of data the algorithms were provided in regard to differing races.

Reviewer #1 (Public review):

Summary:

"Neural noise", here operationalized as an imbalance between excitatory and inhibitory neural activity, has been posited as a core cause of developmental dyslexia, a prevalent learning disability that impacts reading accuracy and fluency. This is study is the first to systematically evaluate the neural noise hypothesis of dyslexia. Neural noise was measured using neurophysiological (electroencephalography [EEG]) and neurochemical (magnetic resonance spectroscopy [MRS]) in adolescents and young adults with and without dyslexia. The authors did not find evidence of elevated neural noise in the dyslexia group from EEG or MRS measures, and Bayes factors generally informed against including the grouping factor in the models. Although the comparisons between groups with and without dyslexia did not support the neural noise hypothesis, a mediation model that quantified phonological processing and reading abilities continuously revealed that EEG beta power in the left superior temporal sulcus was positively associated with reading ability via phonological awareness. This finding lends support for analysis of associations between neural excitatory/inhibitory factors and reading ability along a continuum, rather than as with a case/control approach, and indicates the relevance of phonological awareness as an intermediate trait that may provide a more proximal link between neurobiology and reading ability. Further research is needed across developmental stages and over a broader set of brain regions to more comprehensively assess the neural noise hypothesis of dyslexia, and alternative neurobiological mechanisms of this disorder should be explored.

Strengths:

The inclusion of multiple methods of assessing neural noise (neurophysiological and neurochemical) is a major advantage of this paper. MRS at 7T confers an advantage of more accurately distinguishing and quantifying glutamate, which is a primary target of this study. In addition, the subject-specific functional localization of the MRS acquisition is an innovative approach. MRS acquisition and processing details are noted in the supplementary materials using according to the experts' consensus recommended checklist (https://doi.org/10.1002/nbm.4484). Commenting on rigor the EEG methods is beyond my expertise as a reviewer.<br /> Participants recruited for this study included those with a clinical diagnosis of dyslexia, which strengthens confidence in the accuracy of the diagnosis. The assessment of reading and language abilities during the study further confirms the persistently poorer performance of the dyslexia group compared to the control group.<br /> The correlational analysis and mediation analysis provide complementary information to the main case-control analyses, and the examination of associations between EEG and MRS measures of neural noise is novel and interesting.<br /> The authors follow good practice for open science, including data and code sharing. They also apply statistical rigor, using Bayes Factors to support conclusions of null evidence rather than relying only on non-significant findings. In the discussion, they acknowledge the limitations and generalizability of the evidence and provide directions for future research on this topic.

Weaknesses:

Though the methods employed in the paper are generally strong, the MRS acquisition was not optimized to quantify GABA, so the findings (or lack thereof) should be interpreted with caution. Specifically, while 7T MRS affords the benefit of quantifying metabolites, such as GABA, without spectral editing, this quantification is best achieved with echo times (TE) of 68 or 80 ms in order to minimize the spectral overlap between glutamate and GABA and reduce contamination from the macromolecular signal (Finkelman et al., 2022, https://doi.org/10.1016/j.neuroimage.2021.118810). The data in the present study were acquired at TE=28 ms, and are therefore likely affected by overlapping Glu and GABA peaks at 2.3 ppm that are much more difficult to resolve at this short TE, which could directly affect the measures that are meant to characterize the Glu/GABA+ ratio/imbalance. In future research, MRS acquisition schemes should be optimized for the acquisition of Glutamate, GABA, and their relative balance.

As the authors note in the discussion, additional factors such as MRS voxel location, participant age, and participant sex could influence associations between neural noise and reading abilities and should be considered in future studies.

Appraisal:

The authors present a thorough evaluation of the neural noise hypothesis of developmental dyslexia in a sample of adolescents and young adults using multiple methods of measuring excitatory/inhibitory imbalances as an indicator of neural noise. The authors concluded that there was not support for the neural noise hypothesis of dyslexia in their study based on null significance and Bayes factors. This conclusion is justified, and further research is called for to more broadly evaluate the neural noise hypothesis in developmental dyslexia.

Impact:

This study provides an exemplar foundation for the evaluation of the neural noise hypothesis of dyslexia. Other researcher may adopt the model applied in this paper to examine neural noise in various populations with/without dyslexia, or across a continuum of reading abilities, to more thoroughly examine evidence (or lack thereof) for this hypothesis. Notably, the lack of evidence here does not rule out the possibility for a role of neural noise in dyslexia, and the authors point out that presentation with co-occurring conditions, such as ADHD, may contribute to neural noise in dyslexia. Dyslexia remains a multi-faceted and heterogenous neurodevelopmental condition, and many genetic, neurobiological and environmental factors play a role. This study demonstrates one step toward evaluating neurobiological mechanisms that may contribute to reading difficulties.

Author response:

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

Reviewer #1 (Public Review):

Summary:

"Neural noise", here operationalized as an imbalance between excitatory and inhibitory neural activity, has been posited as a core cause of developmental dyslexia, a prevalent learning disability that impacts reading accuracy and fluency. This study is the first to systematically evaluate the neural noise hypothesis of dyslexia. Neural noise was measured using neurophysiological (electroencephalography [EEG]) and neurochemical (magnetic resonance spectroscopy [MRS]) in adolescents and young adults with and without dyslexia. The authors did not find evidence of elevated neural noise in the dyslexia group from EEG or MRS measures, and Bayes factors generally informed against including the grouping factor in the models. Although the comparisons between groups with and without dyslexia did not support the neural noise hypothesis, a mediation model that quantified phonological processing and reading abilities continuously revealed that EEG beta power in the left superior temporal sulcus was positively associated with reading ability via phonological awareness. This finding lends support for analysis of associations between neural excitatory/inhibitory factors and reading ability along a continuum, rather than as with a case/control approach, and indicates the relevance of phonological awareness as an intermediate trait that may provide a more proximal link between neurobiology and reading ability. Further research is needed across developmental stages and over a broader set of brain regions to more comprehensively assess the neural noise hypothesis of dyslexia, and alternative neurobiological mechanisms of this disorder should be explored.

Strengths:

The inclusion of multiple methods of assessing neural noise (neurophysiological and neurochemical) is a major advantage of this paper. MRS at 7T confers an advantage of more accurately distinguishing and quantifying glutamate, which is a primary target of this study. In addition, the subject-specific functional localization of the MRS acquisition is an innovative approach. MRS acquisition and processing details are noted in the supplementary materials according to the experts' consensus-recommended checklist (https://doi.org/10.1002/nbm.4484). Commenting on the rigor, the EEG methods is beyond my expertise as a reviewer.

Participants recruited for this study included those with a clinical diagnosis of dyslexia, which strengthens confidence in the accuracy of the diagnosis. The assessment of reading and language abilities during the study further confirms the persistently poorer performance of the dyslexia group compared to the control group.

The correlational analysis and mediation analysis provide complementary information to the main case-control analyses, and the examination of associations between EEG and MRS measures of neural noise is novel and interesting.

The authors follow good practice for open science, including data and code sharing. They also apply statistical rigor, using Bayes Factors to support conclusions of null evidence rather than relying only on non-significant findings. In the discussion, they acknowledge the limitations and generalizability of the evidence and provide directions for future research on this topic.

Weaknesses:

Though the methods employed in the paper are generally strong, there are certain aspects that are not clearly described in the Materials & Methods section, such as a description of the statistical analyses used for hypothesis testing.

Thank you for pointing this out. A description of the statistical models used in the analyses of EEG biomarkers has been added to the Materials and Methods:

“First, exponent and offset values were averaged across all electrodes and analyzed using a 2x2 repeated measures ANOVA with group (dyslexic, control) as a between-subjects factor and condition (resting state, language task) as a within-subjects factor. Age was included in the analyses as a covariate due to the correlation between variables. Next, exponent and offset values were averaged across electrodes corresponding to the left (F7, FT7, FC5) and right inferior frontal gyrus (F8, FT8, FC6), and to the left (T7, TP7, TP9) and right superior temporal sulcus (T8, TP8, TP10). The electrodes were selected based on the analyses outlined by Giacometti and colleagues (2014) and Scrivener and Reader (2022). For these analyses, a 2x2x2x2 repeated measures ANOVA with age as a covariate was conducted with group (dyslexic, control) as a between-subjects factor and condition (resting state, language task), hemisphere (left, right), and region (frontal, temporal) as within-subjects factors. Results for the alpha and beta bands were calculated for the same clusters of frontal and temporal electrodes and analyzed with a similar 2x2x2x2 repeated measures ANOVA; however, for these analyses, age was not included as a covariate due to a lack of significant correlations.”

We also expanded the description of the statistical models used in the analyses of MRS biomarkers:

“To analyze the metabolite results, separate univariate ANCOVAs were conducted for Glu, GABA+, Glu/GABA+ ratio and Glu/GABA+ imbalance measures with group (control, dyslexic) as a between-subjects factor and voxel gray matter volume (GMV) as a covariate. Additionally, for the Glu analysis, age was included as a covariate due to a correlation between variables. Both frequentist and Bayesian statistics were calculated. Glu/GABA+ imbalance measure was calculated as the square root of the absolute residual value of a linear relationship between Glu and GABA+ (McKeon et al., 2024).”

With regard to metabolite quantification, it is unclear why the authors chose to analyze and report metabolite values in terms of creatine ratios rather than quantification based on a water reference given that the MRS acquisition appears to support using a water reference.

We have decided to use the ratio of Glu and GABA to total creatine (tCr), as this is still a common practice in MRS studies at 7T (e.g., Nandi et al., 2022; Smith et al., 2021). This approach normalizes the signal, reducing the impact of intensity variations across different regions and tissue compositions. Additionally, total creatine concentration is considered relatively stable across different brain regions, which is particularly important in our study, where a functional localizer was used to establish the left STS region individually. Our decision was further influenced by previous studies on dyslexia (Del Tufo et al., 2018; Pugh et al., 2014) which have reported creatine ratios and included GM volume as a covariate in their models, thus providing comparability. It is now indicated in the Results:

“For comparability with previous studies in dyslexia (Del Tufo et al., 2018; Pugh et al., 2014) we report Glu and GABA as a ratio to total creatine (tCr).”

and in the Method sections:

“Glu and GABA+ concentrations were expressed as a ratio to total-creatine (tCr; Creatine + Phosphocreatine) following previous MRS studies in dyslexia (Del Tufo et al., 2018; Pugh et al., 2014).

We did not estimate absolute concentrations using water signals as a reference, as this would require accounting for water relaxation times, which may vary across our age range. Nevertheless, our dataset has been made publicly available for future researchers to calculate and compare absolute values.

Del Tufo, S. N., Frost, S. J., Hoeft, F., Cutting, L. E., Molfese, P. J., Mason, G. F., Rothman, D. L., Fulbright, R. K., & Pugh, K. R. (2018). Neurochemistry Predicts Convergence of Written and Spoken Language: A Proton Magnetic Resonance Spectroscopy Study of Cross-Modal Language Integration. Frontiers in Psychology, 9, 1507. https://doi.org/10.3389/fpsyg.2018.01507

Nandi, T., Puonti, O., Clarke, W. T., Nettekoven, C., Barron, H. C., Kolasinski, J., Hanayik, T., Hinson, E. L., Berrington, A., Bachtiar, V., Johnstone, A., Winkler, A. M., Thielscher, A., Johansen-Berg, H., & Stagg, C. J. (2022). tDCS induced GABA change is associated with the simulated electric field in M1, an effect mediated by grey matter volume in the MRS voxel. Brain Stimulation, 15(5), 1153–1162. https://doi.org/10.1016/j.brs.2022.07.049

Pugh, K. R., Frost, S. J., Rothman, D. L., Hoeft, F., Del Tufo, S. N., Mason, G. F., Molfese, P. J., Mencl, W. E., Grigorenko, E. L., Landi, N., Preston, J. L., Jacobsen, L., Seidenberg, M. S., & Fulbright, R. K. (2014). Glutamate and choline levels predict individual differences in reading ability in emergent readers. Journal of Neuroscience, 34(11), 4082–4089. https://doi.org/10.1523/JNEUROSCI.3907-13.2014

Smith, G. S., Oeltzschner, G., Gould, N. F., Leoutsakos, J. S., Nassery, N., Joo, J. H., Kraut, M. A., Edden, R. A. E., Barker, P. B., Wijtenburg, S. A., Rowland, L. M., & Workman, C. I. (2021). Neurotransmitters and Neurometabolites in Late-Life Depression: A Preliminary Magnetic Resonance Spectroscopy Study at 7T. Journal of Affective Disorders, 279, 417–425. https://doi.org/10.1016/j.jad.2020.10.011

GABA is typically quantified using J-editing sequences as lower field strengths (~3T), and there is some evidence that the GABA signal can be reliably measured at 7T without editing, however, the authors should discuss potential limitations, such as reliability of Glu and GABA measurements with short-TE semi-laser at 7T.

In addition, MRS measurements of GABA are known to be influenced by macromolecules, and GABA is often denoted as GABA+ to indicate that other compounds contribute to the measured signal, especially at a short TE and in the absence of symmetric spectral editing.

A general discussion of the strengths and limitations of unedited Glu and GABA quantification at 7T is warranted given the interest of this work to researchers who may not be experts in MRS.

While we agree with the Reviewer that at 3T, it is recommended to use J-edited MRS to measure GABA (Mullins et al., 2014), the better spectral resolution at 7T allows for more reliable results for both metabolites using moderate echo-time, non-edited MRS (Finkelman et al., 2022). In this study, we used a short echo time (TE), which is optimal for Glu but not ideal for GABA, as it interferes with other signals. We are grateful to the Reviewer for suggesting the addition of a short paragraph to the Discussion, describing the practicalities of 3T and 7T MRS and changing the abbreviation to GABA+ to inform readers of possible macromolecule contamination:

“We chose ultra-high-field MRS to improve data quality (Özütemiz et al., 2023), as the increased sensitivity and spectral resolution at 7T allows for better separation of overlapping metabolites compared to lower field strengths. Additionally, 7T provides a higher signal-to-noise ratio (SNR), improving the reliability of metabolite measurements and enabling the detection of small changes in Glu and GABA concentrations. Despite these theoretical advantages, several practical obstacles should be considered, such as susceptibility artifacts and inhomogeneities at higher field strengths that can impact data quality. Interestingly, actual methodological comparisons (Pradhan et al., 2015; Terpstra et al., 2016) show only a slight practical advantage of 7T single-voxel MRS compared to optimized 3T acquisition. For example, fitting quality yielded reduced estimates of variance in concentration of Glu in 7T (CRLB) and slightly improved reproducibility levels for Glu and GABA (at both fields below 5%). Choosing the appropriate MRS sequence involves a trade-off between the accuracy of Glu and GABA measurements, as different sequences are recommended for each metabolite. J-edited MRS is recommended for measuring GABA, particularly with 3T scanners (Mullins et al., 2014). However, at 7T, more reliable results can be obtained using moderate echo-time, non-edited MRS (Finkelman et al., 2022). We have opted for a short-echo-time sequence, which is optimal for measuring Glu. However, this approach results in macromolecule contamination of the GABA signal (referred to as GABA+).”

Finkelman, T., Furman-Haran, E., Paz, R., & Tal, A. (2022). Quantifying the excitatory-inhibitory balance: A comparison of SemiLASER and MEGA-SemiLASER for simultaneously measuring GABA and glutamate at 7T. NeuroImage, 247, 118810. https://doi.org/10.1016/j.neuroimage.2021.118810

Mullins, P. G., McGonigle, D. J., O'Gorman, R. L., Puts, N. A., Vidyasagar, R., Evans, C. J., Cardiff Symposium on MRS of GABA, & Edden, R. A. (2014). Current practice in the use of MEGA-PRESS spectroscopy for the detection of GABA. NeuroImage, 86, 43–52. https://doi.org/10.1016/j.neuroimage.2012.12.004

Özütemiz, C., White, M., Elvendahl, W., Eryaman, Y., Marjańska, M., Metzger, G. J., Patriat, R., Kulesa, J., Harel, N., Watanabe, Y., Grant, A., Genovese, G., & Cayci, Z. (2023). Use of a Commercial 7-T MRI Scanner for Clinical Brain Imaging: Indications, Protocols, Challenges, and Solutions-A Single-Center Experience. AJR. American Journal of Roentgenology, 221(6), 788–804. https://doi.org/10.2214/AJR.23.29342

Pradhan, S., Bonekamp, S., Gillen, J. S., Rowland, L. M., Wijtenburg, S. A., Edden, R. A., & Barker, P. B. (2015). Comparison of single voxel brain MRS AT 3T and 7T using 32-channel head coils. Magnetic Resonance Imaging, 33(8), 1013–1018. https://doi.org/10.1016/j.mri.2015.06.003

Terpstra, M., Cheong, I., Lyu, T., Deelchand, D. K., Emir, U. E., Bednařík, P., Eberly, L. E., & Öz, G. (2016). Test-retest reproducibility of neurochemical profiles with short-echo, single-voxel MR spectroscopy at 3T and 7T. Magnetic Resonance in Medicine, 76(4), 1083–1091. https://doi.org/10.1002/mrm.26022

Further, the single MRS voxel location is a limitation of the study as neurochemistry can vary regionally within individuals, and the putative excitatory/inhibitory imbalance in dyslexia may appear in regions outside the left temporal cortex (e.g., network-wide or in frontal regions involved in top-down executive processes). While the functional localization of the MRS voxel is a novelty and a potential advantage, it is unclear whether voxel placement based on left-lateralized reading-related neural activity may bias the experiment to be more sensitive to small, activity-related fluctuations in neurotransmitters in the CON group vs. the DYS group who may have developed an altered, compensatory reading strategy.

We agree that including only one region of interest for the MRS measurements is a potential limitation of our study, and we have now added this information to the Discussion:

“Moreover, since the MRS data was collected only from the left STS, it is plausible that other areas might be associated with differences in Glu or GABA concentrations in dyslexia.”

However, differences in Glu and GABA concentrations in this region were directly predicted by the neural noise hypothesis of dyslexia. We acknowledge that this information was missing in the previous version of the manuscript. It is now included in the Results:

“Moreover, the neural noise hypothesis of dyslexia identifies perisylvian areas as being affected by increased glutamatergic signaling, and directly predicts associations between Glu and GABA levels in the superior temporal regions and phonological skills (Hancock et al., 2017).”

as well as in the Discussion:

“Nevertheless, the neural noise hypothesis predicted increased glutamatergic signaling in perisylvian regions, specifically in the left superior temporal cortex (Hancock et al., 2017).”

Figure 1 contains a lot of information, and it may be helpful to split it into 2 figures (EEG vs. MRS) so that the plots could be made larger and the reader could more easily digest the information.

(a) I would also recommend displaying separate metabolite fit plots for each group, since the current presentation in panel F makes it appear that the MRS data is examined by testing differences between groups across the full spectrum (where the lines diverge), which really isn't the case.

(b) The GABA peak is not visible in the spectrum, and Glutamate and GABA both have multiple peaks that should be shown on the spectrum. This may be best achieved by displaying the individual metabolite sub-spectra below the full spectrum

Thank you for these suggestions. We have split the information into two Figures following the Reviewer’s recommendations.

It is not clear why the 3T structural images were used for segmentation and calculation of tissue fraction if 7T structural images were also acquired (which would presumably have higher resolution).

Generally, T1-weighted images from the 7T scanner exhibit more artifacts than those from the 3T scanner due to higher magnetic field inhomogeneity. These artifacts are especially pronounced in regions near air-tissue interfaces, such as the temporal lobes. Therefore, we chose the 3T structural images for segmentation and tissue fraction calculations and clarified this in the Method section:

“Voxel segmentation was performed on structural images from a 3T scanner, coregistered to 7T structural images in SPM12, as the latter exhibited excessive artifacts and intensity bias in the temporal regions”.

The basis set includes a large number of metabolites (27), including many low-concentration metabolites/compounds (e.g., bHG, bHB, Citrate, Threonine, ethanol) that are typically only included in studies targeting specific metabolites in disease/pathology. Please justify the inclusion of this maximal set of metabolites in the basis set, given that the inclusion of overlapping low-concentration metabolites may influence metabolite measurements of interest (https://doi.org/10.1002/mrm.10246).

There is still no consensus in the MR community on which metabolites should be included in the model of human cerebral 1H-MR spectra. Typically, only major contributors such as NAA, Cr, Cho, Lac, mI, and possibly Glx are evaluated. Some studies also include additional metabolites like Ace, Ala, Asp, GABA, Glc, Gly, sI, NAAG, and Tau. In this study, as in a few others, further metabolites such as PCh, GPC, PCr, GSH, PE, and Thr were introduced and this approach seems suitable for high-field spectra (Hofmann et al., 2002).

Hofmann, L., Slotboom, J., Jung, B., Maloca, P., Boesch, C., & Kreis, R. (2002). Quantitative 1H-magnetic resonance spectroscopy of human brain: Influence of composition and parameterization of the basis set in linear combination model-fitting. Magnetic Resonance in Medicine, 48(3), 440–453. https://doi.org/10.1002/mrm.10246

Please provide a figure indicating the localization of the MRS voxel for a sample subject.

A figure indicating the localization of the MRS voxel for a sample subject was added to the MRS checklist.

It would be helpful to include Table S1 in the main article.

Table S1 from the Supplementary Material has now been added to the main manuscript as Table 1 in the Results section.

Please report descriptive statistics for EEG and MRS measures in Table S1.

We have added a new Table S1 in the Supplementary Material, providing descriptive statistics for EEG and MRS E/I balance measures, presented separately for the dyslexic and control groups.

I recommend avoiding using the terms "direct" and "indirect" to contrast MRS and EEG measures of E/I balance. Both of these measures are imperfect and it is misleading to say that MRS is a "direct" measure of neurotransmitters. There is also ambiguity in what is meant by "direct": in contrast to EEG, MRS does not measure neural activity and does not provide high-resolution temporal information, so in a sense, it is less direct.

Thank you for this suggestion. We have replaced the terms 'direct' and 'indirect' biomarkers with 'MRS' and 'EEG' biomarkers throughout the text.

There are many cases throughout the results in which Bayes and frequentist stats seem to contradict each other in terms of significance and what should be included in the models, especially with regard to the interaction effects (the Bayes factors appear to favor non-significant interactions). I think this is worth considering and describing to offer more clarity for the readers.

We agree that a discussion of the divergent results between Bayesian and frequentist models was missing in the previous version of the manuscript. To provide greater clarity for the readers, we have conducted follow-up Bayesian t-tests in every case where the results indicated the inclusion of non-significant interactions with the effect of group in the model. These additional analyses have been performed for the exponent, offset, as well as for beta bandwidth in the Supplementary Material. We have also added a paragraph addressing these discrepancies in the Discussion:

“Remarkably, in some models, results from Bayesian and frequentist statistics yielded divergent conclusions regarding the inclusion of non-significant effects. This was observed in more complex ANOVA models, whereas no such discrepancies appeared in t-tests or correlations. Given reports of high variability in Bayesian ANOVA estimates across repeated runs of the same analysis (Pfister, 2021), these results should be interpreted with caution. Therefore, following the recommendation to simplify complex models into Bayesian t-tests for more reliable estimates (Pfister, 2021), we conducted follow-up Bayesian t-tests in every case that favored the inclusion of non-significant interactions with the group factor. These analyses provided further evidence for the lack of differences between the dyslexic and control groups. Another source of discrepancy between the two methods may stem from the inclusion of interactions between covariates and within-subject effects in frequentist ANOVA, which were not included in Bayesian ANOVA to adhere to the recommendation for simpler Bayesian models (Pfister, 2021).”

Pfister, R. (2021). Variability of Bayes factor estimates in Bayesian analysis of variance. The Quantitative Methods for Psychology, 17(1), 40-45. doi:10.20982/tqmp.17.1.p040

It would be helpful to indicate whether participants in the DYS group had a history of reading intervention/remediation. In addition to showing that the DYS group performed lower than the CON group on reading assessments as a whole and given their age, was the performance on the reading assessments at an individual level considered for inclusion in the study? (i.e., were participants' persistent poor reading abilities confirmed with the research assessments?)

We were unable to assess individual reading skills due to the lack of standardized diagnostic norms for adult dyslexia in Poland. Therefore, participants in the dyslexic group were recruited based on a previous clinical diagnosis of dyslexia, and reading and reading-related tasks were used for group-level comparisons only. This information has been added to the Methods section:

“Since there are no standardized diagnostic norms for dyslexia in adults in Poland, individuals were assigned to the dyslexic group based on a past diagnosis of dyslexia.”

Unfortunately, we did not collect information about participants' history of reading intervention or remediation. In this context, we acknowledge that including a sample of adult participants is a potential limitation of our study, however, this was already mentioned in the Discussion.

Regarding the fMRI task, please indicate whether the participants whose threshold and/or contrast was changed for localization were from the DYS or CON group.

This information is now added to the Method section:

“For 6 participants (DYS n = 2, CON n = 4), the threshold was lowered to p < .05 uncorrected, while for another 6 participants (DYS n = 3, CON n = 3) the contrast from the auditory run was changed to auditory words versus fixation cross due to a lack of activation for other contrasts.”

Reviewer #2 (Public Review):

Summary:

This study utilized two complementary techniques (EEG and 7T MRI/MRS) to directly test a theory of dyslexia: the neural noise hypothesis. The authors report finding no evidence to support an excitatory/inhibitory balance, as quantified by beta in EEG and Glutamate/GABA ratio in MRS. This is important work and speaks to one potential mechanism by which increased neural noise may occur in dyslexia.

Strengths:

This is a well-conceived study with in-depth analyses and publicly available data for independent review. The authors provide transparency with their statistics and display the raw data points along with the averages in figures for review and interpretation. The data suggest that an E/I balance issue may not underlie deficits in dyslexia and is a meaningful and needed test of a possible mechanism for increased neural noise.

Weaknesses:

The researchers did not include a visual print task in the EEG task, which limits analysis of reading-specific regions such as the visual word form area, which is a commonly hypoactivated region in dyslexia. This region is a common one of interest in dyslexia, yet the researchers measured the I/E balance in only one region of interest, specific to the language network.

We agree with the Reviewer that including different tasks for the EEG biomarkers assessment would be valuable. However, this limitation was already addressed in the Discussion:

“Importantly, our study focused on adolescents and young adults, and the EEG recordings were conducted during rest and a spoken language task. These factors may limit the generalizability of our results. Future research should include younger populations and incorporate a broader array of tasks, such as reading and phonological processing, to provide a more comprehensive evaluation of the E/I balance hypothesis.”

Further, this work does not consider prior studies reporting neural inconsistency; a potential consequence of increased neural noise, which has been reported in several studies and linked with candidate-dyslexia gene variants (e.g., Centanni et al., 2018, 2022; Hornickel & Kraus, 2013; Neef et al., 2017). While E/I imbalance may not be a cause of increased neural noise, other potential mechanisms remain and should be discussed.

Thank you for referring us to other works reporting neural variability in dyslexia. We agree that a broader context regarding sources of reduced neural synchronization, beyond E/I imbalance, was missing in the previous version of the manuscript. We have now included these references in the Discussion:

“Furthermore, although our results do not support the idea of E/I balance alterations as a source of neural noise in dyslexia, they do not preclude other mechanisms leading to less synchronous neural firing posited by the hypothesis. In this context, there is evidence showing increased trial-to-trial inconsistency of neural responses in individuals with dyslexia (Centanni et al., 2022) or poor readers (Hornickel and Kraus, 2013) and its associations with specific dyslexia risk genes (Centanni et al., 2018; Neef et al., 2017). At the same time, the observed trial-to-trial inconsistency was either present only in a subset of participants (Centanni et al., 2018), limited to some experimental conditions (Centanni et al., 2022), or specific brain regions – e.g., brainstem in Hornickel and Kraus (2013), left auditory cortex in Centanni et al. (2018), or left supramarginal gyrus in Centanni et al. (2022).”

A better description of the exponent and offset components is needed at the beginning of the results, given that the methods are presented in detail at the end. I also do not see a clear description of these components in the methods.

A description of the aperiodic components is now included in the Results:

“In the initial step of the analysis, we analyzed the aperiodic (exponent and offset) components of the EEG spectrum. The exponent reflects the steepness of the EEG power spectrum, with a higher exponent indicating a steeper signal; while the offset represents a uniform shift in power across frequencies, with a higher offset indicating greater power across the entire EEG spectrum (Donoghue et al., 2020).”

as well as in the Materials and Methods:

“Two broadband aperiodic parameters were extracted: the exponent, which quantifies the steepness of the EEG power spectrum, and the offset, which indicates signal’s power across the entire frequency spectrum.”

Reviewer #3 (Public Review):

Summary:

This study by Glica and colleagues utilized EEG (i.e., Beta power, Gamma power, and aperiodic activity) and 7T MRS (i.e., MRS IE ratio, IE balance) to reevaluate the neural noise hypothesis in Dyslexia. Supported by Bayesian statistics, their results show solid 'no evidence' of EI balance differences between groups, challenging the neural noise hypothesis. The work will be of broad interest to neuroscientists, and educational and clinical psychologists.

Strengths:

Combining EEG and 7T MRS, this study utilized both the indirect (i.e., Beta power, Gamma power, and aperiodic activity) and direct (i.e., MRS IE ratio, IE balance) measures to reevaluate the neural noise hypothesis in Dyslexia.

Weaknesses:

The authors may need to provide more data to assess the quality of the MRS data.

We have addressed the following specific recommendations of the Reviewer providing more data about the quality of the MRS data.

The authors may need to explain how the number of subjects is determined in the MRS section.

We have clarified the MRS sample description in the Results section:

“Due to financial and logistical constraints, 59 out of the 120 recruited subjects, selected progressively as the study unfolded, were examined with MRS. Subjects were matched by age and sex between the dyslexic and control groups. Due to technical issues and to prevent delays and discomfort for the participants, we collected 54 complete sessions. Additionally, four datasets were excluded based on our quality control criteria, and three GABA+ estimates exceeded the selected CRLB threshold. Ultimately, we report 50 estimates for Glu (21 participants with dyslexia) and 47 for GABA+ and Glu/GABA+ ratios (20 participants with dyslexia).”

Is there a reason why theta and gamma peaks were not observed in the majority of participants? What are the possible reasons that likely caused the discrepancy between this study and previously reported relevant studies?

We have now added a discussion about the absence of oscillatory peaks in the theta and gamma bands to the Discussion section:

“We could not perform analyses for the gamma oscillations since in the majority of participants the gamma peak was not detected above the aperiodic component. Due to the 1/f properties of the EEG spectrum, both aperiodic and periodic components should be disentangled to analyze ‘true’ gamma oscillations; however, this approach is not typically recognized in electrophysiology research (Hudson and Jones, 2022). Indeed, previous studies that analyzed gamma activity in dyslexia (Babiloni et al., 2012; Lasnick et al., 2023; Rufener and Zaehle, 2021) did not separate the background aperiodic activity. For the same reason, we could not analyze results for the theta band, which often does not meet the criteria for an oscillatory component manifested as a peak in the power spectrum (Klimesch, 1999). Moreover, results from a study investigating developmental changes in both periodic and aperiodic components suggest that theta oscillations in older participants are mostly observed in frontal midline electrodes (Cellier et al., 2021), which were not analyzed in the current study.”

Hudson, M. R., & Jones, N. C. (2022). Deciphering the code: Identifying true gamma neural oscillations. Experimental Neurology, 357, 114205. https://doi.org/10.1016/j.expneurol.2022.114205

Klimesch, W. (1999). EEG alpha and theta oscillations reflect cognitive and memory performance: A review and analysis. Brain Research Reviews, 29(2-3), 169-195. https://doi.org/10.1016/S0165-0173(98)00056-3

Based on Figure 1F, the quality of the MRS data may be contaminated by the lipid signal, especially for the DYS group. To better evaluate the MRS data, especially the GABA measurements, the authors need to show:

(a) the placement of the MRS voxel on the anatomical images;

Averaged MRS voxel placement was already presented in Figure 1 (now Figure 2) in the manuscript. Now, we have also added exemplary single-subject images to the MRS checklist in the Supplement.

(b) Glu and GABA model functions

We have now provided more meaningful Glu and GABA indications in Figure 2.

(c) CRLB for GABA

We have added respective estimates to the Supplement:

%CRLB of Glu: mean 2.96, SD = 0.79

%CRLB of GABA: mean 10.59, SD = 2.76

%CRLB of NAA: 1.76 SD = 0.46

Further, the authors added voxel's gray matter volume as a covariate when performing separate ANCOVAs. The authors may need to use alpha correction or 1-fCSF correction to corroborate these results.

We chose to use the ratio of Glu and GABA to total creatine (tCr), as this remains a common practice in MRS studies at 7T (e.g., Nandi et al., 2022; Smith et al., 2021). This decision was also influenced by previous dyslexia studies (Del Tufo et al., 2018; Pugh et al., 2014) and is now clarified in the Results and Methods sections.

Regarding alpha correction, a recent paper (García-Pérez et al., 2023) recommends: 'In general, avoid corrections for multiple testing if statistical claims are to be made for each individual test, in the absence of an omnibus null hypothesis.' Since we report null findings, further alpha correction would not significantly impact the results.

García-Pérez, M. A. (2023). Use and misuse of corrections for multiple testing. Methods in Psychology, 8, 100120. https://doi.org/10.1016/j.metip.2023.100120

Reviewer #2 (Public review):

Summary:

It is generally believed that higher-order areas in the prefrontal cortex guide selection during working memory and attention through signals that selectively recruit neuronal populations in sensory areas that encode the relevant feature. In this work, Parto-Dezfouli and colleagues tested how these prefrontal signals influence activity in visual area V4 using a spatial working memory task. They recorded neuronal activity from visual area V4 and found that information about visual features at the behaviorally relevant part of space during the memory period is carried in a spatially selective manner in the timing of spikes relative to a beta oscillation (phase coding) rather than in the average firing rate (rate code). The authors further tested whether there is a causal link between prefrontal input and the phase encoding of visual information during the memory period. They found that indeed inactivation of the frontal eye fields, a prefrontal area known to send spatial signals to V4, decreased beta oscillatory activity in V4 and information about the visual features. The authors went one step further to develop a neural model that replicated the experimental findings and suggested that changes in the average firing rate of individual neurons might be a result of small changes in the exact beta oscillation frequency within V4. These data provide important new insights into the possible mechanisms through which top-down signals can influence activity in hierarchically lower sensory areas and can therefore have a significant impact on the Systems, Cognitive, and Computational Neuroscience fields.

Strengths:

This is a well-written paper with a well-thought-out experimental design. The authors used a smart variation of the memory-guided saccade task to assess how information about the visual features of stimuli is encoded during the memory period. By using a grating of various contrasts and orientations as the background the authors ensured that bottom-up visual input would drive responses in visual area V4 in the delay period, something that is not commonly done in experimental settings in the same task. Moreover, one of the major strengths of the study is the use of different approaches including analysis of electrophysiological data using advanced computational methods of analysis, manipulation of activity through inactivation of the prefrontal cortex to establish causality of top-down signals on local activity signatures (beta oscillations, spike locking and information carried) as well as computational neuronal modeling. This has helped extend an observation into a possible mechanism well supported by the results.

Weaknesses:

Although the authors provide support for their conclusions from different approaches, I found that the selection of some of the analyses and statistical assessments made it harder for the reader to follow the comparison between a rate code and a phase code. Specifically, the authors wish to assess whether stimulus information is carried selectively for the relevant position through a firing rate or a phase code. Results for the rate code are shown in Figures 1B-G and for the phase code are shown in Figure 2. Whereas an F-statistic is shown over time in Figure 1F (and Figure S1) no such analysis is shown for LFP power. Similarly, following FEF inactivation there is no data on how that influences V4 firing rates and information carried by firing rates in the two conditions (for positions inside and outside the V4 RF). In the same vein, no data are shown on how the inactivation affects beta phase coding in the OUT condition.

Moreover, some of the statistical assessments could be carried out differently including all conditions to provide more insight into mechanisms. For example, a two-way ANOVA followed by post hoc tests could be employed to include comparisons across both spatial (IN, OUT) and visual feature conditions (see results in Figures 2D, S4, etc.). Figure 2D suggests that the absence of selectivity in the OUT condition (no significant difference between high and low contrast stimuli) is mainly due to an increase in slope in the OUT condition for the low contrast stimulus compared to that for the same stimulus in the IN condition. If this turns out to be true it would provide important information that the authors should address.

There are also a few conceptual gaps that leave the reader wondering whether the results and conclusion are general enough. Specifically,

(1) the authors used microstimulation in the FEF to determine RFs. It is thus possible that the FEF sites that were inactivated were largely more motor-related. Given that beta oscillations and motor preparatory activity have been found to be correlated and motor sites show increased beta oscillatory activity in the delay period, it is possible that the effect of FEF inactivation on V4 beta oscillations is due to inactivation of the main source of beta activity. Had the authors inactivated sites with a preponderance of visual neurons in the FEF would the results be different?

(2) Somewhat related to this point and given the prominence of low-frequency activity in deeper layers of the visual cortex according to some previous studies, it is not clear where the authors' V4 recordings were located. The authors report that they do have data from linear arrays, so it should be possible to address this.

(3) The authors suggest that a change in the exact frequency of oscillation underlies the increase in firing rate for different stimulus features. However, the shift in frequency is prominent for contrast but not for orientation, something that raises questions about the general applicability of this observation for different visual features.

(4) One of the major points of the study is the primacy of the phase code over the rate code during the delay period. Specifically, here it is shown that information about the visual features of a stimulus carried by the rate code is similar for relevant and irrelevant locations during the delay period. This contrasts with what several studies have shown for attention in which case information carried in firing rates about stimuli in the attended location is enhanced relative to that for stimuli in the unattended location. If we are to understand how top-down signals work in cognitive functions it is inevitable to compare working memory with attention. The possible source of this difference is not clear and is not discussed. The reader is left wondering whether perhaps a different measure or analysis (e.g. a percent explained variance analysis) might reveal differences during the delay period for different visual features across the two spatial conditions.

The use of the memory-guided saccade task has certain disadvantages in the context of this study. Although delay activity is interpreted as memory activity by the authors, it is in principle possible that it reflects preparation for the upcoming saccade, spatial attention (particularly since there is a stimulus in the RF), etc. This could potentially change the conclusion and perspective.

For the position outside the V4 RF, there is a decrease in both beta oscillations and the clustering of spikes at a specific phase. It is therefore possible that the decrease in information about the stimuli features is a byproduct of the decrease in beta power and phase locking. Decreased oscillatory activity and phase locking can result in less reliable estimates of phase, which could decrease the mutual information estimates.

The authors propose that coherent oscillations could be the mechanism through which the prefrontal cortex influences beta activity in V4. I assume they mean coherent oscillations between the prefrontal cortex and V4. Given that they do have simultaneous recordings from the two areas they could test this hypothesis on their own data, however, they do not provide any results on that.

The authors make a strong point about the relevance of changes in the oscillation frequency and how this may result in an increase in firing rate although it could also be the reverse - an increase in firing rate leading to an increase in the frequency peak. It is not clear at all how these changes in frequency could come about. A more nuanced discussion based on both experimental and modeling data is necessary to appreciate the source and role (if any) of this observation.

Author response:

Reviewer #1 (Public review):

Summary:

This study investigates what happens to the stimulus-driven responses of V4 neurons when an item is held in working memory. Monkeys are trained to perform memory-guided saccades: they must remember the location of a visual cue and then, after a delay, make an eye movement to the remembered location. In addition, a background stimulus (a grating) is presented that varies in contrast and orientation across trials. This stimulus serves to probe the V4 responses, is present throughout the trial, and is task-irrelevant. Using this design, the authors report memory-driven changes in the LFP power spectrum, changes in synchronization between the V4 spikes and the ongoing LFP, and no significant changes in firing rate.

Strengths:

(1) The logic of the experiment is nicely laid out.

(2) The presentation is clear and concise.

(3) The analyses are thorough, careful, and yield unambiguous results.

(4) Together, the recording and inactivation data demonstrate quite convincingly that the signal stored in FEF is communicated to V4 and that, under the current experimental conditions, the impact from FEF manifests as variations in the timing of the stimulus-evoked V4 spikes and not in the intensity of the evoked activity (i.e., firing rate).

Weaknesses:

I think there are two limitations of the study that are important for evaluating the potential functional implications of the data. If these were acknowledged and discussed, it would be easier to situate these results in the broader context of the topic, and their importance would be conveyed more fairly and transparently.

(1) While it may be true that no firing rate modulations were observed in this case, this may have been because the probe stimuli in the task were behaviorally irrelevant; if anything, they might have served as distracters to the monkey's actual task (the MGS). From this perspective, the lack of rate modulation could simply mean that the monkeys were successful in attending the relevant cue and shielding their performance from the potentially distracting effect of the background gratings. Had the visual probes been in some way behaviorally relevant and/or spatially localized (instead of full field), the data might have looked very different.

Any task design involves tradeoffs; if the visual stimulus was behaviorally relevant, then any observed neurophysiological changes would be more confounded by possible attentional effects. We cannot exclude the possibility that a different task or different stimuli would produce different results; we ourselves have reported firing rate enhancements for other types of visual probes during an MGS task (Merrikhi et al. 2017). We have added an acknowledgement of these limitations in the discussion section (lines 311-319). At minimum, our results show a dissociation between the top-down modulation of phase coding, which is enhanced during WM even for these task-irrelevant stimuli, and rate coding. Establishing whether and how this phase coding is related to perception and behavior will be an important direction for future work.

With this in mind, it would be prudent to dial down the tone of the conclusions, which stretch well beyond the current experimental conditions (see recommendations).

We have edited the title (removing the word ‘primarily’) and key sentences throughout to tone down the conclusions, generally to state that the importance of a phase code in WM modulations is *possible* given the observed results, rather than certain (see abstract line 27, introduction lines 58-60, results line 215, conclusion lines 294-295).

(2) Another point worth discussing is that although the FEF delay-period activity corresponds to a remembered location, it can also be interpreted as an attended location, or as a motor plan for the upcoming eye movement. These are overlapping constructs that are difficult to disentangle, but it would be important to mention them given prior studies of attentional or saccade-related modulation in V4. The firing rate modulations reported in some of those cases provide a stark contrast with the findings here, and I again suspect that the differences may be due at least in part to the differing experimental conditions, rather than a drastically different encoding mode or functional linkage between FEF and V4.

We have added a paragraph to the discussion section addressing links to attention and motor planning (lines 301-322), and specifically acknowledging the inherent difficulties of fully dissociating these effects when interpreting our results (lines 311-319).

Reviewer #2 (Public review):

Summary:

It is generally believed that higher-order areas in the prefrontal cortex guide selection during working memory and attention through signals that selectively recruit neuronal populations in sensory areas that encode the relevant feature. In this work, Parto-Dezfouli and colleagues tested how these prefrontal signals influence activity in visual area V4 using a spatial working memory task. They recorded neuronal activity from visual area V4 and found that information about visual features at the behaviorally relevant part of space during the memory period is carried in a spatially selective manner in the timing of spikes relative to a beta oscillation (phase coding) rather than in the average firing rate (rate code). The authors further tested whether there is a causal link between prefrontal input and the phase encoding of visual information during the memory period. They found that indeed inactivation of the frontal eye fields, a prefrontal area known to send spatial signals to V4, decreased beta oscillatory activity in V4 and information about the visual features. The authors went one step further to develop a neural model that replicated the experimental findings and suggested that changes in the average firing rate of individual neurons might be a result of small changes in the exact beta oscillation frequency within V4. These data provide important new insights into the possible mechanisms through which top-down signals can influence activity in hierarchically lower sensory areas and can therefore have a significant impact on the Systems, Cognitive, and Computational Neuroscience fields.

Strengths:

This is a well-written paper with a well-thought-out experimental design. The authors used a smart variation of the memory-guided saccade task to assess how information about the visual features of stimuli is encoded during the memory period. By using a grating of various contrasts and orientations as the background the authors ensured that bottom-up visual input would drive responses in visual area V4 in the delay period, something that is not commonly done in experimental settings in the same task. Moreover, one of the major strengths of the study is the use of different approaches including analysis of electrophysiological data using advanced computational methods of analysis, manipulation of activity through inactivation of the prefrontal cortex to establish causality of top-down signals on local activity signatures (beta oscillations, spike locking and information carried) as well as computational neuronal modeling. This has helped extend an observation into a possible mechanism well supported by the results.

Weaknesses:

Although the authors provide support for their conclusions from different approaches, I found that the selection of some of the analyses and statistical assessments made it harder for the reader to follow the comparison between a rate code and a phase code. Specifically, the authors wish to assess whether stimulus information is carried selectively for the relevant position through a firing rate or a phase code. Results for the rate code are shown in Figures 1B-G and for the phase code are shown in Figure 2. Whereas an F-statistic is shown over time in Figure 1F (and Figure S1) no such analysis is shown for LFP power. Similarly, following FEF inactivation there is no data on how that influences V4 firing rates and information carried by firing rates in the two conditions (for positions inside and outside the V4 RF). In the same vein, no data are shown on how the inactivation affects beta phase coding in the OUT condition.

We plan to incorporate statistical analysis of this point in the revised version.

Moreover, some of the statistical assessments could be carried out differently including all conditions to provide more insight into mechanisms. For example, a two-way ANOVA followed by post hoc tests could be employed to include comparisons across both spatial (IN, OUT) and visual feature conditions (see results in Figures 2D, S4, etc.). Figure 2D suggests that the absence of selectivity in the OUT condition (no significant difference between high and low contrast stimuli) is mainly due to an increase in slope in the OUT condition for the low contrast stimulus compared to that for the same stimulus in the IN condition. If this turns out to be true it would provide important information that the authors should address.

We plan to incorporate statistical analysis of this point in the revised version.

There are also a few conceptual gaps that leave the reader wondering whether the results and conclusion are general enough. Specifically,

(1) the authors used microstimulation in the FEF to determine RFs. It is thus possible that the FEF sites that were inactivated were largely more motor-related. Given that beta oscillations and motor preparatory activity have been found to be correlated and motor sites show increased beta oscillatory activity in the delay period, it is possible that the effect of FEF inactivation on V4 beta oscillations is due to inactivation of the main source of beta activity. Had the authors inactivated sites with a preponderance of visual neurons in the FEF would the results be different?

We do not believe this to be likely based on what is known anatomically and functionally about this circuitry. Anatomically, the projections from FEF to V4 arise primarily from the supragranular layers, not layers which contain the highest proportion of motor activity (Barone et al. 2000, Pouget et al. 2009, Markov et al. 2013). Functionally, based on electrical identification of V4-projecting FEF neurons, we know that FEF to V4 projections are predominantly characterized by delay rather than motor activity (Merrikhi et al. 2017). We have now tried to emphasize these points when we introduce the inactivation experiments (lines 180-182).

Experimentally, the spread of the pharmacological effect with our infusion system is quite large relative to any clustering of visual vs. motor neurons within the FEF, with behavioral consequences of inactivation spreading to cover a substantial portion of the visual hemifield (e.g., Noudoost et al. 2014, Clark et al. 2014), and so our manipulation lacks the spatial resolution to selectively target motor vs. other FEF neurons.

(2) Somewhat related to this point and given the prominence of low-frequency activity in deeper layers of the visual cortex according to some previous studies, it is not clear where the authors' V4 recordings were located. The authors report that they do have data from linear arrays, so it should be possible to address this.

Unfortunately our chamber placement for V4 has produced linear array penetration angles which do not reliably allow identification of cortical layers. We are aware of previous results showing layer-specific effects of attention in V4 (e.g., Pettine et al. 2019, Buffalo et al. 2011), and it would indeed be interesting to determine whether our observed WM-driven changes follow similar patterns. We may be able to analyze a subset of the data with current source density analysis to look for layer-specific effects in the future, but are not able to provide any information at this time.

(3) The authors suggest that a change in the exact frequency of oscillation underlies the increase in firing rate for different stimulus features. However, the shift in frequency is prominent for contrast but not for orientation, something that raises questions about the general applicability of this observation for different visual features.

We plan to incorporate statistical analysis of this point in the revised version.

(4) One of the major points of the study is the primacy of the phase code over the rate code during the delay period. Specifically, here it is shown that information about the visual features of a stimulus carried by the rate code is similar for relevant and irrelevant locations during the delay period. This contrasts with what several studies have shown for attention in which case information carried in firing rates about stimuli in the attended location is enhanced relative to that for stimuli in the unattended location. If we are to understand how top-down signals work in cognitive functions it is inevitable to compare working memory with attention. The possible source of this difference is not clear and is not discussed. The reader is left wondering whether perhaps a different measure or analysis (e.g. a percent explained variance analysis) might reveal differences during the delay period for different visual features across the two spatial conditions.

We have added discussion regarding the relationship of these results to previous findings during attention in the discussion section (lines 301-322).

The use of the memory-guided saccade task has certain disadvantages in the context of this study. Although delay activity is interpreted as memory activity by the authors, it is in principle possible that it reflects preparation for the upcoming saccade, spatial attention (particularly since there is a stimulus in the RF), etc. This could potentially change the conclusion and perspective.

We have added a new discussion paragraph addressing the relationship to attention and motor planning (lines 301-322). We have also moderated the language used to describe our conclusions throughout the manuscript in light of this ambiguity.

For the position outside the V4 RF, there is a decrease in both beta oscillations and the clustering of spikes at a specific phase. It is therefore possible that the decrease in information about the stimuli features is a byproduct of the decrease in beta power and phase locking. Decreased oscillatory activity and phase locking can result in less reliable estimates of phase, which could decrease the mutual information estimates.

We plan to incorporate statistical analysis of this point in the revised version.

The authors propose that coherent oscillations could be the mechanism through which the prefrontal cortex influences beta activity in V4. I assume they mean coherent oscillations between the prefrontal cortex and V4. Given that they do have simultaneous recordings from the two areas they could test this hypothesis on their own data, however, they do not provide any results on that.

This paper only includes inactivation data. We are working on analyzing the simultaneous recording data for a future publication.

The authors make a strong point about the relevance of changes in the oscillation frequency and how this may result in an increase in firing rate although it could also be the reverse - an increase in firing rate leading to an increase in the frequency peak. It is not clear at all how these changes in frequency could come about. A more nuanced discussion based on both experimental and modeling data is necessary to appreciate the source and role (if any) of this observation.

As the reviewer notes, it is difficult to determine whether the frequency changes drive the rate changes, vice versa, or whether both are generated in parallel by a common source. We have adjusted our language to reflect this (lines 277-278). Future modeling work may be able to shed more light on the causal relationships between various neural signatures.

Reviewer #3 (Public review):

Summary:

In this report, the authors test the necessity of prefrontal cortex (specifically, FEF) activity in driving changes in oscillatory power, spike rate, and spike timing of extrastriate visual cortex neurons during a visual-spatial working memory (WM) task. The authors recorded LFP and spikes in V4 while macaques remembered a single spatial location over a delay period during which task-irrelevant background gratings were displayed on the screen with varying orientation and contrast. V4 oscillations (in the beta range) scaled with WM maintenance, and the information encoded by spike timing relative to beta band LFP about the task-irrelevant background orientation depended on remembered location. They also compared recorded signals in V4 with and without muscimol inactivation of FEF, demonstrating the importance of FEF input for WM-induced changes in oscillatory amplitude, phase coding, and information encoded about background orientations. Finally, they built a network model that can account for some of these results. Together, these results show that FEF provides meaningful input to the visual cortex that is used to alter neural activity and that these signals can impact information coding of task-irrelevant information during a WM delay.

Strengths:

(1) Elegant and robust experiment that allows for clear tests for the necessity of FEF activity in WM-induced changes in V4 activity.

(2) Comprehensive and broad analyses of interactions between LFP and spike timing provide compelling evidence for FEF-modulated phase coding of task-irrelevant stimuli at remembered location.

(3) Convincing modeling efforts.

Weaknesses:

(1) 0% contrast background data (standard memory-guided saccade task) are not reported in the manuscript. While these data cannot be used to consider information content of spike rate/time about task-irrelevant background stimuli, this condition is still informative as a 'baseline' (and a more typical example of a WM task).

We plan to incorporate statistical analysis of this point in the revised version.

(2) Throughout the manuscript, the primary measurements of neural coding pertain to task-irrelevant stimuli (the orientation/contrast of the background, which is unrelated to the animal's task to remember a spatial location). The remembered location impacts the coding of these stimulus variables, but it's unclear how this relates to WM representations themselves.

Indeed, here we have focused on how maintaining spatial WM impacts visual processing of incoming sensory information, rather than on how the spatial WM signal itself is represented and maintained. Behaviorally, this impact on visual signals could be related to the effects of the content of WM on perception and reaction times (e.g., Soto et al. 2008, Awh et al. 1998, Teng et al. 2019), but no such link to behavior is shown in our data.

eLife Assessment

This important study uses reinforcement learning to study how turbulent odor stimuli should be processed to yield successful navigation. They find that there is an optimal memory length over which an agent should ignore blanks in the odor to discriminate whether the agent is still inside the plume or outside of it, complementing recent studies using RNNs and finite state controllers to identify optimal strategies for navigating a turbulent plume. While the overall strength of evidence is convincing, better justification for using Brownian motion as a recovery strategy and the addition of accompanying code for reproducibility would add to this strength.

Reviewer #2 (Public review):

Summary:

The authors investigate the problem of olfactory search in turbulent environments using artificial agents trained using tabular Q-learning, a simple and interpretable reinforcement learning (RL) algorithm. The agents are trained solely on odor stimuli, without access to spatial information or prior knowledge about the odor plume's shape. This approach makes the emergent control strategy more biologically plausible for animals navigating exclusively using olfactory signals. The learned strategies show parallels to observed animal behaviors, such as upwind surging and crosswind casting. The approach generalizes well to different environments and effectively handles the intermittency of turbulent odors.

Strengths:

(1) The use of numerical simulations to generate realistic turbulent fluid dynamics sets this paper apart from studies that rely on idealized or static plumes.

(2) A key innovation is the introduction of a small set of interpretable olfactory states based on moving averages of odor intensity and sparsity, coupled with an adaptive temporal memory.

(3) The paper provides a thorough analysis of different recovery strategies when an agent loses the odor trail, offering insights into the trade-offs between various approaches.

(4) The authors provide a comprehensive performance analysis of their algorithm across a range of environments and recovery strategies, demonstrating the versatility of the approach.

(5) Finally, the authors list an interesting set of real-world experiments based on their findings, that might invite interest from experimentalists across multiple species.

Weaknesses:

(1) The inclusion of Brownian motion as a recovery strategy, seems odd since it doesn't closely match natural animal behavior, where circling (e.g. flies) or zigzagging (ants' "sector search") could have been more realistic.

(2) Using tabular Q-learning is both a strength and a limitation. It's simple and interpretable, making it easier to analyze the learned strategies, but the discrete action space seems somewhat unnatural. In real-world biological systems, actions (like movement) are continuous rather than discrete. Additionally, the ground-frame actions may not map naturally to how animals navigate odor plumes (e.g. insects often navigate based on their own egocentric frame).

(3) The lack of accompanying code is a major drawback since nowadays open access to data and code is becoming a standard in computational research. Given that the turbulent fluid simulation is a key element that differentiates this paper, the absence of simulation and analysis code limits the study's reproducibility.

Author response:

We thank the Editor and Reviewers for their work on our manuscript, and are happy to receive their positive comments, as well as their questions and suggestions. We are currently revising the manuscript and are planning to de-emphasize Brownian recovery as a simple yet biologically irrelevant benchmark and include comparisons with other biologically inspired strategies suggested by the reviewers. As for sharing the code and data: we completely agree: dataset 1 is already public and we will share the other dataset as well as the code. In a nutshell, we will be addressing the referee’s suggestions as follows:

(1)   As Referee 1 points out, even if the algorithm does not require a map of space, the agent is still required to tell apart North, East, South and West relative to the wind direction which is implicitly assumed known. We will better clarify the spatial encoding required to implement these strategies.

(2)   Referee 1 remarks that the learned recovery strategy works best and suggests to give it a more prominent role and better characterize it. We agree that what is done in the void state is definitely key and more work is needed to understand it. In the revised manuscript, we are planning to further substantiate the statistics of the learned recovery by repeating training several times and comparing several trajectories. Note that this strategy is much more flexible than the others and could potentially mix aspects of recovery to aspects of exploitation: we defer a more in-depth analysis that disentangles these two aspects elsewhere.

(3)   Referee 1 asks whether an optimal, minimal representation of the olfactory states exists. Q learning defines the olfactory states prior to training and does not allow to systematically optimize odor representation for the task. Given the odor features, we can however discretize them in more or less olfactory states. We expect that decreasing the number of olfactory states provides less positional information and potentially degrades performance, although loss in performance may be overshadowed by noise or by efficient recovery. We are planning to re-train our model with a smaller numer of non-void states and will provide the comparison. The number of void states does not need further testing: we chose 50 void states because it matches the time agents typically remain in the void and indeed achieves very high performance (less than 50 void states results in no convergence and more than 50 introduces states that are rarely visited)

(4)   Both reviewers correctly remark that Brownian motion is not biologically relevant. We will make sure to further clarify that this is a rather simple --but biologically irrelevant-- benchmark. We are planning to include results with both circling and zigzaging as biologically inspired recovery strategies.

(5)   We agree with reviewer 2 that animal locomotion does not look like a series of discrete displacements on a checkerboard. However, to overcome this limitation, one has to first focus on a specific system to define actions in a way that best adheres to a species’ motor controls. Second, these actions are likely continuous, which makes reinforcement learning notoriously more complex. While we agree that more realistic models are definitely needed for a comparison with real systems, this remains outside the scope of the current work.

(6)   We agree with the referees and editor that it is important to publish the code and data alongside with the manuscript. It was already planned and we will make sure to share the links within the revised version of the manuscript.

Reviewer #3 (Public review):

The author presents a novel theory and computational model suggesting that grid cells do not encode space, but rather encode non-spatial attributes. Place cells in turn encode memories of where those specific attributes occurred. The theory accounts for many experimental results and generates useful predictions for future studies. The model's simplicity and potential explanatory power will interest others in the field. There are, however, a few weaknesses outlined below which undermine the theory.

Main criticisms:

(1) A crucial assumption of the model is that grid cells express grid-like firing patterns if and only if the content of experience is constant in space. It is difficult to imagine a real world example that satisfies this assumption. Odors and sounds are used as examples. While they are often more spatially diffuse than an object on the ground, odors and sounds have sources that are readily detectable and thus are not constant in space. Animals can easily navigate to a food source or to a vocalizing conspecific. This assumption is especially problematic because it predicts that all grid cells should become silent when their preferred non-spatial attribute (e.g. a specific odor) is missing. I'm not aware of any experimental data showing that grid cells become silent. On the contrary, grid cells are known to remain active across all contexts that have been tested, including across sleep/wake states. Unlike place cells, grid cells have never been shown to turn off. Since grid cells are active in all contexts, their preferred attribute must also be present in all contexts, and therefore they would not convey any information about the specific content of an experience. The author lists many attributes that could in theory be constant in a laboratory setting, but there is no data I'm aware of that shows this is true in practice. As it stands, this crucial assumption of the model remains mere speculation.

(2) The proposed novelty of this theory is that other models all assume that grid cells encode space. This is not quite true of models based on continuous attractor networks, the discussion of which is essentially absent. More specifically, attractor models focus on the importance of intrinsic dynamics within entorhinal cortex in generating the grid pattern. While this firing pattern is aligned to space during navigation and therefore can be used a representation of that space, the neural dynamics are preserved even during sleep. Similarly, it is because the grid pattern does not strictly encode physical space that grid-like signals are also observed in relation to other two-dimensional continuous variables.

(3) The use of border cells or boundary vector cells as the main (or only) source of spatial information in the hippocampus is not well supported by experimental data. Border cells in entorhinal cortex are not active in the center of an environment. Boundary-vector cells can fire farther away from the walls, but are not found in entorhinal cortex. They are located in the subiculum, a major output of the hippocampus. While the entorhinal-hippocampal circuit is a loop, the route from boundary-vector cells to place cells is much less clear than from grid cells. Moreover, both border cells and boundary-vector cells (which are conflated in this paper) comprise a small population of neurons compared to grid cells.

Minor comments:

(1) There is substantial theoretical and experimental work supporting the idea that grid cell modules instantiate continuous attractor networks, yet this class of models is largely ignored:

p. 7 "In contrast, most grid cell models (Bellmund et al., 2016; Bush et al., 2015; Castro & Aguiar, 2014; Hasselmo, 2009; Mhatre et al., 2012; Solstad et al., 2006; Sorscher et al., 2023; Stepanyuk, 2015; Widloski & Fiete, 2014) are domain specific models of spatial navigation"

The following references should be added:

McNaughton, B. L., Battaglia, F. P., Jensen, O., Moser, E. I. & Moser, M.-B. Path integration and the neural basis of the 'cognitive map'. Nat. Rev. Neurosci. 7, 663-678 (2006).

Fuhs, M. C. & Touretzky, D. S. A spin glass model of path integration in rat medial entorhinal cortex. J. Neurosci. 26, 4266-4276 (2006).

Burak, Y. & Fiete, I. R. Accurate path integration in continuous attractor network models of grid cells. PLoS Comput. Biol. 5, e1000291 (2009).

Guanella, A., Kiper, D. & Verschure, P. A model of grid cells based on a twisted torus topology. Int. J. Neural Syst. 17, 231-240 (2007).

Couey, J. J. et al. Recurrent inhibitory circuitry as a mechanism for grid formation. Nat. Neurosci. 16, 318-324 (2013).

(Note: the Bellmund et al. (2016) citation is likely a mistake and was intended to be Bellmund et al. (2018).)

(2) The author claims in two places that this model is the first to explain that grid cell population activity lies on a torus. While it may be the first explicit computational account of why grid cell activity is mapped onto a torus, these claims should be moderated for clarity, for example by adding "but see McNaughton et al. (2006) and others".

Box 1. Results Uniquely Explained by this Memory Model - the population code of grid cells lies on a torus

p.11 "In addition, this simplifying assumption allows the model to capture the finding that the population of grid cells lies on a torus (Gardner et al., 2022), although I note that the model was developed before this result was known."

(3) Lateral entorhinal cortex is largely ignored in this memory model. It should be considered that the predominance of spatial representations reported in the literature is due to historical reasons. Namely, the discovery of hippocampal place cells spurred interest in looking upstream for the source of spatial information, which was later abundantly clear in medial entorhinal cortex. Lateral entorhinal cortex is relatively understudied, but is known to encode odors, objects, and time in a way that medial entorhinal cortex does not. It is therefore confusing to assume that these attributes are instead encoded by grid cells.

Author response:

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

Public Reviews

Reviewer #1 (Public Review): 

(1) Although the theory is based on memory, it also is based on spatially-selective cells.

Not all cells in the hippocampus fulfill the criteria of place/HD/border/grid cells, and place a role in memory. E.g., Tonegawa, Buszaki labs' work does not focus on only those cells, and there are certainly a lot of non-pure spatial cells in monkeys (Martinez-Trujillo) and humans (iEEG). Does the author mainly focus on saying that "spatial cells" are memory, but do not account for non-spatial memory cells? This seems to be an incomplete account of memory - which is fine, but the way the model is set up suggests that *all* memory is, place (what/where), and non-spatial attributes ("grid") - but cells that don't fulfil these criteria in MTL (Diehl et al., 2017, Neuron; non-grid cells; Schaeffer et al., 2022, ICML; Luo et al., 2024, bioRxiv) certainly contribute to memory, and even navigation. This is also related to the question of whether these cell definitions matter at all (Luo et al., 2024). The authors note "However, this memory conjunction view of the MTL must be reconciled with the rodent electrophysiology finding that most cells in MTL appear to have receptive fields related to some aspect of spatial navigation (Boccara et al., 2010; Grieves & Jeffery, 2017). The paucity of non-spatial cells in MTL could be explained if grid cells have been mischaracterized as spatial." Is the author mainly talking about rodent work?

There is a new section in the introduction that deals with these issues, titled ‘Why Model the Rodent Navigation Literature with a Memory Model?’ That section reads:

“Spatial navigation is inherently a memory problem – learning the spatial arrangement of a new enclosure requires memory for the conjunction of what and where. This has long been realized and in the introduction to ‘Hippocampus as a Cognitive Map’, O’Keefe and Nadel (1978) wrote “We shall argue that the hippocampus is the core of a neural memory system providing an objective spatial framework within which the items and events of an organism's experience are located and interrelated” (emphasis added). Furthermore, in the last chapter of their book, they extended cognitive map theory to human memory for non-spatial characteristics. However, in the decades since the development of cognitive map theory, the rodent spatial navigation and human memory literatures have progressed somewhat independently.

The ideas proposed in this model are an attempt to reunify these literatures by returning to the original claim that spatial navigation is inherently a memory problem. The goal of the current study is to explain the rodent spatial navigation literature using a memory model that has the potential to also explain the human memory literature. In contrast, most grid cell models (Bellmund et al., 2016; Bush et al., 2015; Castro & Aguiar, 2014; Hasselmo, 2009; Mhatre et al., 2012; Solstad et al., 2006; Sorscher et al., 2023; Stepanyuk, 2015; Widloski & Fiete, 2014) are domain specific models of spatial navigation and as such, they do not lend themselves to explanations of human memory. Thus, the reason to prefer this model is parsimony. Rather than needing to develop a theory of memory that is separate from a theory of spatial navigation, it might be possible to address both literatures with a unified account.

This study does not attempt to falsify other theories of grid cells. Instead, this model reaches a radically different interpretation regarding the function of grid cells; an interpretation that emerges from viewing spatial navigation as a memory problem. All other grid cell models assume that an entorhinal grid cell displaying a spatially arranged grid of firing fields serves the function of spatial coding (i.e., spatial grid cells exist to support a spatial metric). In contrast, the proposed memory model of grid cells assumes that the hexagonal tiling reflects the need to keep memories separate from each other to minimize confusion and confabulation – the grid pattern is the byproduct of pattern separation between memories rather than the basis of a spatial code. 

It is now understood that grid-like firing fields can occur for non-spatial twodimensional spaces. For instance, human entorhinal cortex exhibits grid-like responses to video morph trajectories in a two-dimensional bird neck-length versus bird leg-length space (Constantinescu et al., 2016). As a general theory of learning and memory, the proposed memory model of grid cells is easily extended to explain these results (e.g., relabeling the border cell inputs in the model as neck-length and leg-length inputs). However, there are other grid cell models that can explain both spatial grid cells as well as non-spatial grid-like responses (Mok & Love, 2019; Rodríguez-Domínguez & Caplan, 2019; Stachenfeld et al., 2017; Wei et al., 2015). Similar to this memory model of grid cells, these models are also positioned to explain both the rodent spatial navigation and human memory literatures. Nevertheless, there is a key difference between this model and other grid cell models that generalize to non-spatial representations. Specifically, these other models assume that grid cells exhibiting spatial receptive fields serve the function of identifying positions in the environment (i.e., their function is spatial). As such, these models do not explain why most of the input to rodent hippocampus appears to be spatial (Boccara et al., 2010; Diehl et al., 2017; Grieves & Jeffery, 2017). This memory model of grid cells provides an answer to the apparent paucity of nonspatial cell types in rodent MTL by proposing that grid cells with spatial receptive fields have been misclassified as spatial (they are what cells rather than where cells) and that place cells are fundamentally memory cells that conjoin what and where.”

(2) Related to the last point, how about non-grid multi-field mEC cells? In theory, these also should be the same; but the author only presents perfect-look grid cells. In empirical work, clearly, this is not the case, and many mEC cells are multi-field non-grid cells (Diehl et al., 2017). Does the model find these cells? Do they play a different role? As noted by the author "Because the non-spatial attributes are constant throughout the two-dimensional surface, this results in an array of discrete memory locations that are approximately hexagonal (as explained in the Model Methods, an "online" memory consolidation process employing pattern separation rapidly turns an approximately hexagonal array into one that is precisely hexagonal). " If they are indeed all precisely hexagonal, does that mean the model doesn't have non-grid spatial cells? 

Grid cells with irregular firing fields are now considered in the discussion with the following paragraphs

“According to this model, hexagonally arranged grid cells should be the exception rather than the rule when considering more naturalistic environments. In a more ecologically valid situation, such as with landmarks, varied sounds, food sources, threats, and interactions with conspecifics, there may still be remembered locations were events occurred or remembered properties can be found, but because the non-spatial properties are non-uniform in the environment, the arrangement of memory feedback will be irregular, reflecting the varied nature of the environment. This may explain the finding that even in a situation where there are regular hexagonal grid cells, there are often irregular non-grid cells that have a reliable multi-location firing field, but the arrangement of the firing fields is irregular (Diehl et al., 2017). For instance, even when navigating in an enclosure that has uniform properties as dictated by experimental procedures, they may be other properties that were not well-controlled (e.g., a view of exterior lighting in some locations but not others), and these uncontrolled properties may produce an irregular grid (i.e., because the uncontrolled properties are reliably associated with some locations but not others, hippocampal memory feedback triggers retrieval of those properties in the associations locations).

In this memory model, there are other situations in which an irregular but reliable multilocation grid may occur, even when everything is well controlled. In the reported simulations, when the hippocampal place cells were based on variation in X/Y (as defined by Border cells), nothing else changed as a function of location, and the model rapidly produced a precise hexagonal arrangement of hippocampal place cell memories. When head direction was included (i.e., real-world variation in X, Y, and head direction), the model still produced a hexagonal arrangement as per face-centered cubic packing of memories, but this precise arrangement was slower to emerge, with place cells continuing to shift their positions until the borders of the enclosure were sufficiently well learned from multiple viewpoints. If there is real-world variation in four or more dimensions, as is likely the case in a more ecologically valid situation, it will be even harder for place cell memories to settle on a precise regular lattice. Furthermore, in the case of four dimensions, mathematicians studying the “sphere packing problem” recently concluded that densest packing is irregular (Campos et al., 2023). This may explain why the multifield grid cells for freely flying bats have a systematic minimum distance between firing fields, but their arrangement is globally irregular (Ginosar et al., 2021). Assuming that the memories encoded by a bat include not just the three real-world dimensions of variation, but also head direction, the grid will likely be irregular even under optimal conditions of laboratory control.”

(3) Theoretical reasons for why the model is put together this way, and why grid cells must be coding a non-spatial attribute: Is this account more data-driven (fits the data so formulated this way), or is it theoretical - there is a reason why place, border, grid cells are formulated to be like this. For example, is it an efficient way to code these variables? It can be both, like how the BVC model makes theoretical sense that you can use boundaries to determine a specific location (and so place cell), but also works (creates realistic place cells). 

The motivation for this model is now articulated in the new section, quoted above, titled ‘Why Model the Rodent Navigation Literature with a Memory Model?’ Regarding the assumption that border cells provide a spatial metric, this assumption is made for the same reasons as in the BVC model. Regarding this, the text said: “These assumptions regarding border cells are based on the boundary vector cell (BVC) model of Barry et al. (2006). As in the BVC model, combinations of border cells encode where each memory occurred in the realworld X/Y plane.”. A new sentence is added to model methods, stating: “This assumption is made because border cells provide an efficient representation of Euclidean space (e.g., if the animal knows how far it is from different walls of the enclosure, this already available information can be used to calculate location).”

But in this case, the purpose of grid cell coding a non-spatial attribute, and having some kind of system where it doesn't fire at all locations seems a little arbitrary. If it's not encoding a spatial attribute, it doesn't have to have a spatial field. For example, it could fire in the whole arena - which some cells do (and don't pass the criteria of spatial cells as they are not spatially "selective" to another location, related to above).  

Some cells have a constant high firing rate, but they are the exception rather than the rule. More typically, cells habituate in the presence of ongoing excitatory drive and by doing so become sensitive to fluctuations in excitatory drive. Habituation is advantageous both in terms of metabolic cost and in terms of function (i.e., sensitivity to change). This is now explained in the following paragraph:

“In theory, a cell representing a non-spatial attribute found at all locations of an enclosure (aka, a grid cell in the context of this model), could fire constantly within the enclosure. However, in practice, cells habituate and rapidly reduce their firing rate by an order of magnitude when their preferred stimulus is presented without cessation (Abbott et al., 1997; Tsodyks & Markram, 1997). After habituation, the firing rate of the cell fluctuates with minor variation in the strength of the excitatory drive. In other words, habituation allows the cell to become sensitive to changes in the excitatory drive (Huber & O’Reilly, 2003). Thus, if there is stronger top-down memory feedback in some locations as compared to others, the cell will fire at a higher rate in those remembered locations rather than in all locations even though the attribute is found at all locations. In brief when faced with constant excitatory drive, the cell accommodates, and becomes sensitive to change in the magnitude of the excitatory drive. In the model simulation, this dynamic adaptation is captured by supposing that cells fire 5% of the time on-average across the simulation, regardless of their excitatory inputs.”

(4) Why are grid cells given such a large role for encoding non-spatial attributes? If anything, shouldn't it be lateral EC or perirhinal cortex? Of course, they both could, but there is less reason to think this, at least for rodent mEC.  

This is a good point and the following paragraph has been added to the introduction to explain that lateral EC is likely part of the explanation. But even when including lateral EC, it still appears that most of the input to hippocampus is spatial.

“One possible answer to the apparent lack of non-spatial cells in MTL is to highlight the role of the lateral entorhinal cortex (LEC) as the source of non-spatial what information for memory encoding (Deshmukh & Knierim, 2011). LEC can be contrasted with mEC, which appears to only provide where information (Boccara et al., 2010a; Diehl et al., 2017). Although it is generally true that LEC is involved in non-spatial processing, there is evidence that LEC provides some forms of spatial information (Knierim et al., 2014). The kind of non-spatial information provided by LEC appears to be in relation to objects (Connor & Knierim, 2017; Wilson et al., 2013). However, in a typical rodent spatial navigation study there are no objects within the enclosure. Thus, although the distinction between mEC and LEC is likely part of the explanation, it is still the case that rodent entorhinal input to hippocampus appears to heavily favor spatial information.”

(5) Clarification: why do place cells and grid cells differ in terms of stability in the model? Place cells are not stable initially but grid cells come out immediately. They seem directly connected so a bit unclear why; especially if place cell feedback leads to grid cell fields. There is an explanation in the text - based on grid cells coding the on-average memories, but these should be based on place cell inputs as well. So how is it that place fields are unstable then grid fields do not move at all? I wonder if a set of images or videos (gifs) showing the differences in spatial learning would be nice and clarify this point.  

In this revision, I provide a new video focused on learning of place cell memories that include head direction. This second video is in relation to the results reported in Figure 9. The short answer is that the grid fields for the non-spatial cell are based on the average across several view-dependent memories (i.e., across several place cells that have head direction sensitivity) and the average is reliable even if the place cells are unstable. The text of this explanation now reads:

“Why was the grid immediately apparent for the non-spatial attribute cell whereas the grid took considerable prior experience for the head direction cells? The answer relates to memory consolidation and the shifting nature of the hippocampal place cells. Head direction cells only produced a reliable grid once the hippocampal place cells (aka, memory cells) assumed stable locations. During the first few sessions, the hippocampal place cells were shifting their positions owing to pattern separation and consolidation. But once the place cells stabilized, they provided reliable top-down memory feedback to the head direction cells in some places but not others, thus producing a reliable grid arrangement to the firing maps of the head direction cells. In other words, for the head direction cells, the grid only appeared once the place cells stabilized. This slow stabilization of place fields is a known property (Bostock et al., 1991; Frank et al., 2004).

In the simulation, the place cells did not stabilize until a sufficient number of place cells were created (Figure 9C). Specifically, these additional memories were located immediately outside the enclosure, around all borders (Figure 9D). These “outside the box” memories served to constrain the interior place cells, locking them in position despite ongoing consolidation. This dynamic can be seen in a movie showing a representative simulation. The movie shows the positions of the head direction sensitive place cells during initial learning, and then during additional sessions of prior experience as the movie speeds up (see link in Figure 9 capture).

Why did the non-spatial grid cell (k) produce a grid immediately, before the place cells stabilized? As discussed in relation to Figure 8, the non-spatial grid cell is the projection through the 3D volume of real-world coordinates that includes X, Y, and head direction. Each grid field of a non-spatial grid cell reflects feedback from several place cells that each have a different head direction sensitivity (see for instance the allocentric pairs of memories illustrated in Figure 8C and 8D). Thus, each grid field is the average across several memories that entail different viewpoints and this averaging across memories provides stability even if the individual memories are not yet stable. This average of unstable memories produces a blurry sort of grid pattern without any prior experience.

A final piece of the puzzle relies on the same mechanism that caused the grid pattern to align with the borders as reported in the results of Figures 6 and 7. Specifically, there are some “sticky” locations with ongoing consolidation because the connection weights are bounded. Because weights cannot go below their minimum or above their maximum, it is slightly more difficult for consolidation to push or pull connection weights over the peak value or under the minimum value of the tuning curve. Thus, the place cells tend to linger in locations that correspond to the peak or trough of a border cell. There are multiple peak and trough locations but for the parameter values in this simulation, the grid pattern seen in Figure 9C shows the set of peak/trough locations that satisfy the desired spacing between memories. Thus, the average across memories shows a reliable grid field at these locations even though the memories are unstable.”

(6) Other predictions. Clearly, the model makes many interesting (and quite specific!) predictions. But does it make some known simple predictions? 

• More place cells at rewarded (or more visited) locations. Some empirical researchers seem to think this is not as obvious as it seems (e.g., Duvellle et al., 2019; JoN; Nyberg et al., 2021, Neuron Review).  

• Grid cell field moves toward reward (Butler et al., 2019; Boccera et al., 2019).  

• Grid cells deform in trapezoid (Krupic et al., 2015) and change in environments like mazes (Derikman et al., 2014).  

Thank you for these suggestions and I have added the following paragraph to the discussion:

“In terms of the animal’s internal state, all locations in the enclosure may be viewed as equally aversive and unrewarding, which is a memorable characteristic of the enclosure. Reward, or lack thereof, is arguably one of the most important nonspatial characteristics and application of this model to reward might explain the existence of goal-related activity in place cells (Hok et al., 2007; although see Duvelle et al., 2019), reflecting the need to remember rewarding locations for goal directed behavior. Furthermore, if place cell memories for a rewarding location activate entorhinal grid cells, this may explain the finding that grid cells remap in an enclosure with a rewarded location such that firing fields are attracted to that location (Boccara et al., 2019; Butler et al., 2019). Studies that introduce reward into the enclosure are an important first step in terms of examining what happens to grid cells when the animal is placed in a more varied environment.”

Regarding the changes in shape of the environment, this was discussed in the section of the paper that reads “As seen in Figure 12, because all but one of the place cells was exterior when the simulated animal was constrained to a narrow passage, the hippocampal place cell memories were no longer arranged in a hexagonal grid. This disruption of the grid array for narrow passages might explain the finding that the grid pattern (of grid cells) is disrupted in the thin corner of a trapezoid (Krupic et al., 2015) and disrupted when a previously open enclosure is converted to a hairpin maze by insertion of additional walls within the enclosure (Derdikman et al., 2009).” This particular section of the paper now appears in the Appendix and Figure 12 is now Appendix Figure 2.

Reviewer #2 (Public Review): 

The manuscript describes a new framework for thinking about the place and grid cell system in the hippocampus and entorhinal cortex in which these cells are fundamentally involved in supporting non-spatial information coding. If this framework were shown to be correct, it could have high impact because it would suggest a completely new way of thinking about the mammalian memory system in which this system is non-spatial. Although this idea is intriguing and thought-provoking, a very significant caveat is that the paper does not provide evidence that specifically supports its framework and rules out the alternate interpretations. Thus, although the work provides interesting new ideas, it leaves the reader with more questions than answers because it does not rule out any earlier ideas. 

Basically, the strongest claim in the paper, that grid cells are inherently non-spatial, cannot be specifically evaluated versus existing frameworks on the basis of the evidence that is shown here. If, for example, the author had provided behavioral experiments showing that human memory encoding/retrieval performance shifts in relation to the predictions of the model following changes in the environment, it would have been potentially exciting because it could potentially support the author's reconceptualization of this system. But in its current form, the paper merely shows that a new type of model is capable of explaining the existing findings. There is not adequate data or results to show that the new model is a significantly better fit to the data compared to earlier models, which limits the impact of the work. In fact, there are some key data points in which the earlier models seem to better fit the data.  

Overall, I would be more convinced that the findings from the paper are impactful if the author showed specific animal memory behavioral results that were only supported by their memory model but not by a purely spatial model. Perhaps the author could run new experiments to show that there are specific patterns of human or animal behavior that are only explained by their memory model and not by earlier models. But in its current form, I cannot rule out the existing frameworks and I believe some of the claims in this regard are overstated. 

As previously detailed in Box 1 and as explained in the text in several places, the model provides an explanation of several findings that remain unexplained by other theories (see “Results Uniquely Explained by the Memory Model”). But more generally this is a good point, and the initial draft failed to fully articulate why a researcher might choose this model to guide future empirical investigations. A new section in the introduction that deals with these issues, titled ‘Why Model the Rodent Navigation Literature with a Memory Model?’ That section reads:

“Spatial navigation is inherently a memory problem – learning the spatial arrangement of a new enclosure requires memory for the conjunction of what and where. This has long been realized and in the introduction to ‘Hippocampus as a Cognitive Map’, O’Keefe and Nadel (1978) wrote “We shall argue that the hippocampus is the core of a neural memory system providing an objective spatial framework within which the items and events of an organism's experience are located and interrelated” (emphasis added). Furthermore, in the last chapter of their book, they extended cognitive map theory to human memory for non-spatial characteristics. However, in the decades since the development of cognitive map theory, the rodent spatial navigation and human memory literatures have progressed somewhat independently.

The ideas proposed in this model are an attempt to reunify these literatures by returning to the original claim that spatial navigation is inherently a memory problem. The goal of the current study is to explain the rodent spatial navigation literature using a memory model that has the potential to also explain the human memory literature. In contrast, most grid cell models (Bellmund et al., 2016; Bush et al., 2015; Castro & Aguiar, 2014; Hasselmo, 2009; Mhatre et al., 2012; Solstad et al., 2006; Sorscher et al., 2023; Stepanyuk, 2015; Widloski & Fiete, 2014) are domain specific models of spatial navigation and as such, they do not lend themselves to explanations of human memory. Thus, the reason to prefer this model is parsimony. Rather than needing to develop a theory of memory that is separate from a theory of spatial navigation, it might be possible to address both literatures with a unified account.

This study does not attempt to falsify other theories of grid cells. Instead, this model reaches a radically different interpretation regarding the function of grid cells; an interpretation that emerges from viewing spatial navigation as a memory problem. All other grid cell models assume that an entorhinal grid cell displaying a spatially arranged grid of firing fields serves the function of spatial coding (i.e., spatial grid cells exist to support a spatial metric). In contrast, the proposed memory model of grid cells assumes that the hexagonal tiling reflects the need to keep memories separate from each other to minimize confusion and confabulation – the grid pattern is the byproduct of pattern separation between memories rather than the basis of a spatial code. 

It is now understood that grid-like firing fields can occur for non-spatial twodimensional spaces. For instance, human entorhinal cortex exhibits grid-like responses to video morph trajectories in a two-dimensional bird neck-length versus bird leg-length space (Constantinescu et al., 2016). As a general theory of learning and memory, the proposed memory model of grid cells is easily extended to explain these results (e.g., relabeling the border cell inputs in the model as neck-length and leg-length inputs). However, there are other grid cell models that can explain both spatial grid cells as well as non-spatial grid-like responses (Mok & Love, 2019; Rodríguez-Domínguez & Caplan, 2019; Stachenfeld et al., 2017; Wei et al., 2015). Similar to this memory model of grid cells, these models are also positioned to explain both the rodent spatial navigation and human memory literatures. Nevertheless, there is a key difference between this model and other grid cell models that generalize to non-spatial representations. Specifically, these other models assume that grid cells exhibiting spatial receptive fields serve the function of identifying positions in the environment (i.e., their function is spatial). As such, these models do not explain why most of the input to rodent hippocampus appears to be spatial (Boccara et al., 2010; Diehl et al., 2017; Grieves & Jeffery, 2017). This memory model of grid cells provides an answer to the apparent paucity of nonspatial cell types in rodent MTL by proposing that grid cells with spatial receptive fields have been misclassified as spatial (they are what cells rather than where cells) and that place cells are fundamentally memory cells that conjoin what and where.”

- The paper does not fully take into account all the findings regarding grid cells, some of which very clearly show spatial processing in this system. For example, findings on grid-bydirection cells (e.g., Sargolini et al. 2006) would seem to suggest that the entorhinal grid system is very specifically spatial and related to path integration. Why would grid-bydirection cells be present and intertwined with grid cells in the author's memory-related reconceptualization? It seems to me that the existence of grid-by-direction cells is strong evidence that at least part of this network is specifically spatial.

Head by direction grid cells were a key part of the reported results. These grid cells naturally arise in the model as the animal forms memories (aka, hippocampal place cells) that conjoin location (as defined by border cells), head direction at the time of memory formation, and one or more non-spatial properties found at that location. In this revision, I have attempted to better explain how including head direction in hippocampal memories naturally gives rise to these cell types. The introduction to the head direction module simulations now reads:

“According to this memory model of spatial navigation, place cells are the conjunction of location, as defined by border cells, and one or more properties that are remembered to exist at that location. Such memories could, for instance, allow an animal to remember the location of a food cache (Payne et al., 2021). The next set of simulations investigates behavior of the model when one of the to-be-remembered properties is head direction at the time when the memory was formed (e.g., the direction of a pathway leading to a food cache). Indicating that head direction is an important part of place cell representations, early work on place cells in mazes found strong sensitivity to head direction, such that the place field is found in one direction of travel but not the other (McNaughton et al., 1983; Muller et al., 1994). Place cells can exhibit a less extreme version of head direction sensitivity in open field recordings (Rubin et al., 2014), but the nature of the sensitivity is more complicated, depending on location of the animal relative to the place field center (Jercog et al., 2019).

It is possible that some place cell memories do not receive head direction input, as was the case for the simulations reported in Figures 6/7 – in those simulations, place cells were entirely insensitive to head direction, owing to a lack of input from head direction cells. However, removal of head direction input to hippocampus affects place cell responses (Calton et al., 2003) and grid cell responses (Winter et al., 2015), suggesting that head direction is a key component of the circuit. Furthermore, if place cells represent episodic memories, it seems natural that they should include head direction (i.e., viewpoint at the time of memory formation).

In the simulations reported next, head direction is simply another property that is conjoined in a hippocampal place cell memory. In this case, a head direction cell should become a head direction conjunctive grid cell (i.e., a grid cell, but only when the animal is heading in a particular direction), owing to memory feedback from the hexagonal array of hippocampal place cell memories. When including head direction, the real-world dimensions of variation are across three dimensions (X, Y, and head direction) rather than two, and consolidation will cause the place cells to arrange in a three-dimensional volume. The simulation reported below demonstrates that this situation provides a “grid module”.”

- I am also concerned that the paper does not do enough to address findings regarding how the elliptical shape of grid fields shifts when boundaries of an environment compress in one direction or change shape/angles (Lever et al., & Krupic et al). Those studies show compression in grid fields based on boundary position, and I don't see how the authors' model would explain these findings.  

This finding was covered in the original submission: “For instance, perhaps one egocentric/allocentric pair of mEC grid modules is based on head direction (viewpoint) in remembered positions relative to the enclosure borders whereas a different egocentric/allocentric pair is based on head direction in remembered positions relative to landmarks exterior to the enclosure. This might explain why a deformation of the enclosure (moving in one of the walls to form a rectangle rather than a square) caused some of the grid modules but not others to undergo a deformation of the grid pattern in response to the deformation of the enclosure wall (see also Barry et al., 2007). More specifically, if there is one set of non-orthogonal dimensions for enclosure borders and the movement of one wall is too modest as to cause avoid global remapping, this would deform the grid modules based the enclosure border cells. At the same time, if other grid modules are based on exterior properties (e.g., perhaps border cells in relation to the experimental room rather than the enclosure), then those grid modules would be unperturbed by moving the enclosure wall.”

I apologize for being unclear in describing how the model might explain this result. The paragraph has been rewritten and now reads:

“Consider the possibility that one mEC grid modules is based on head direction (viewpoint) in remembered positions relative to the enclosure borders (e.g., learning the properties of the enclosure, such as the metal surface) while a different grid module is based on head direction in remembered positions relative to landmarks exterior to the enclosure (e.g., learning the properties of the experimental room, such as the sound of electronics that the animal is subject to at all locations). This might explain why a deformation of the enclosure (moving one of the walls to form a rectangle rather than a square) caused some of the grid modules but not others to undergo a deformation of the grid pattern in response to the deformation of the enclosure wall (see also Barry et al., 2007). More specifically, suppose that the movement of one wall is modest and after moving the wall, the animal views the enclosure as being the same enclosure, albeit slightly modified (e.g., when a home is partially renovated, it is still considered the same home). In this case, the set of non-orthogonal dimensions associated with enclosure borders would still be associated with the now-changed borders and any memories in reference to this border-determined space would adjust their positions accordingly in real-world coordinates (i.e., the place cells would subtly shift their positions owing to this deformation of the borders, producing a corresponding deformation of the grid). At the same time, there may be other sets of memories that are in relation to dimensions exterior to the enclosure. Because these exterior properties are unchanged, any place cells and grid cells associated with the exterior-oriented memories would be unchanged by moving the enclosure wall.”

- Are findings regarding speed modulation of grid cells problematic for the paper's memory results? 

- A further issue is that the paper does not seem to adequately address developmental findings related to the timecourses of the emergence of different cell types. In their simulation, researchers demonstrate the immediate emergence of grid fields in a novel environment, while noting that the stabilization of place cell positions takes time. However, these simulation findings contradict previous empirical developmental studies (Langston et al., 2010). Those studies showed that head direction cells show the earliest development of spatial response, followed by the appearance of place cells at a similar developmental stage. In contrast, grid cells emerge later in this developmental sequence. The gradual improvement in spatial stability in firing patterns likely plays a crucial role in the developmental trajectory of grid cells. Contrary to the model simulation, grid cells emerge later than place cells and head direction cells, yet they also hold significance in spatial mapping. 

- The model simulations suggest that certain grid patterns are acquired more gradually than others. For instance, egocentric grid cells require the stabilization of place cell memories amidst ongoing consolidation, while allocentric grid cells tend to reflect average place field positions. However, these findings seemingly conflict with empirical studies, particularly those on the conjunctive representation of distance and direction in the earliest grid cells. Previous studies show no significant differences were found in grid cells and grid cells with directional correlates across these age groups, relative to adults (Wills et al., 2012). This indicates that the combined representation of distance and direction in single mEC cells is present from the earliest ages at which grid cells emerge. 

These are good points and they have been addressed in a new section of the introduction titled ‘The Scope of the Proposed Model’. That section reads:

“The reported simulations explain why most mEC cell types in the rodent literature appear to be spatial (Boccara et al., 2010; Diehl et al., 2017; Grieves & Jeffery, 2017). Assuming that rodents can form non-spatial memories, rodent hippocampus must receive non-spatial input from entorhinal cortex. These simulations suggest that characterization of the rodent mEC cortex as primarily spatial might be incorrect if most grid cells (except perhaps head direction conjunctive grid cells) have been mischaracterized as spatial. Other literatures with other species find non-spatial representations in MTL (Gulli et al., 2020; Quiroga et al., 2005; Wixted et al., 2014) and non-spatial hippocampal memory encoding has been found in rodents (Liu et al., 2012; McEchron & Disterhoft, 1999). The proposed memory model is compatible with these results – the ideas contained in this model could be applied to nonspatial memory representations. However, surveys of cell types in rodent entorhinal cortex seem to indicate that most cells are spatial (Boccara et al., 2010; Diehl et al., 2017; Grieves & Jeffery, 2017). How can the rodent hippocampus encode nonspatial memories if most of its input is spatial? The goal of the reported simulations is to explain the apparent paucity of non-spatial cells in rodent entorhinal cortex by proposing that grid cells have been misclassified as spatial (see also Luo et al., 2024).

Given the simplicity of the proposed model, there are important findings that the model cannot address -- it is not that the model makes the wrong predictions but rather that it makes no predictions. The role of running speed (Kraus et al., 2015) is one such variable for which the model makes no predictions. Similarly, because the model is a rate-coded model rather than a model of oscillating spiking neurons, it makes no predictions regarding theta oscillations (Buzsáki & Moser, 2013). The model is an account of learning and memory for an adult animal, and it makes no predictions regarding the developmental (Langston et al., 2010; Muessig et al., 2015; Wills et al., 2012) or evolutionary (Rodrıguez et al., 2002) time course of different cell types. This model contains several purely spatial representations such as border cells, head direction cells, and head direction conjunctive grid cells and it may be that these purely spatial cell types emerged first, followed by the evolution and/or development of non-spatial cell types. However, this does not invalidate the model. Instead, this is a model for an adult animal that has both episodic memory capabilities and spatial navigation capabilities, irrespective of the order in which these capabilities emerged.

This model has the potential to explain context effects in memory (Godden & Baddeley, 1975; Gulli et al., 2020; Howard et al., 2005). According to this model, different grid cells represent different non-spatial characteristics and place cells represent the combination of these “context” factors and location. In the simulation, just one grid cell is simulated but the same results would emerge when simulating hundreds of different non-spatial inputs provided that all of the simulated non-spatial inputs exist throughout the recording session. However, there is evidence that hippocampus can explicitly represent the passage of time (Eichenbaum, 2014), and time is assuredly an important factor in defining episodic memory (Bright et al., 2020). Thus, although the current model addresses unique combinations of what and where, it is left to future work to incorporate representations of when in the memory model.”

Reviewer #3 (Public Review): 

A crucial assumption of the model is that the content of experience must be constant in space. It's difficult to imagine a real-world example that satisfies this assumption. Odors and sounds are used as examples. While they are often more spatially diffuse than an objects on the ground, odors and sounds have sources that are readily detectable. Animals can easily navigate to a food source or to a vocalizing conspecific. This assumption is especially problematic because it predicts that all grid cells should become silent when their preferred non-spatial attribute (e.g. a specific odor) is missing. I'm not aware of any experimental data showing that grid cells become silent. On the contrary, grid cells are known to remain active across all contexts that have been tested, including across sleep/wake states. Unlike place cells, grid cells do not seem to turn off. Since grid cells are active in all contexts, their preferred attribute must also be present in all contexts, and therefore they would not convey any information about the specific content of an experience.  

These are good points and in this revision I have attempted to explain that there is a great deal of contextual similarity across all recording sessions. One paragraph in the discussion now reads

“In a typical rodent spatial navigation study, the non-spatial attributes are wellcontrolled, existing at all locations regardless of the enclosure used during testing (hence, a grid cell in one enclosure will be a grid cell in a different enclosure). Because labs adopt standard procedures, the surfaces, odors (e.g., from cleaning), external lighting, time of day, human handler, electronic apparatus, hunger/thirst state, etc. might be the same for all recording sessions. Additionally, the animal is not allowed to interact with other animals during recording and this isolation may be an unusual and highly salient property of all recording sessions. Notably, the animal is always attached to wires during recording. The internal state of the animal (fear, aloneness, the noise of electronics, etc.) is likely similar across all recording situations and attributes of this internal state are likely represented in the hippocampus and entorhinal input to hippocampus. According to this model, hippocampal place cells are “marking” all locations in the enclosure as places where these things tend to happen.”

The proposed novelty of this theory is that other models all assume that grid cells encode space. This isn't quite true of models based on continuous attractor networks, the discussion of which is notably absent. More specifically, these models focus on the importance of intrinsic dynamics within the entorhinal cortex in generating the grid pattern. While this firing pattern is aligned to space during navigation and therefore can be used as a representation of that space, the neural dynamics are preserved even during sleep. Similarly, it is because the grid pattern does not strictly encode physical space that gridlike signals are also observed in relation to other two-dimensional continuous variables. 

These models were briefly discussed in the general discussion section and in this revision they are further discussed in the introduction in a new section, titled ‘Why Model the Rodent Navigation Literature with a Memory Model?’ That section reads:

“Spatial navigation is inherently a memory problem – learning the spatial arrangement of a new enclosure requires memory for the conjunction of what and where. This has long been realized and in the introduction to ‘Hippocampus as a Cognitive Map’, O’Keefe and Nadel (1978) wrote “We shall argue that the hippocampus is the core of a neural memory system providing an objective spatial framework within which the items and events of an organism's experience are located and interrelated” (emphasis added). Furthermore, in the last chapter of their book, they extended cognitive map theory to human memory for non-spatial characteristics. However, in the decades since the development of cognitive map theory, the rodent spatial navigation and human memory literatures have progressed somewhat independently.

The ideas proposed in this model are an attempt to reunify these literatures by returning to the original claim that spatial navigation is inherently a memory problem. The goal of the current study is to explain the rodent spatial navigation literature using a memory model that has the potential to also explain the human memory literature. In contrast, most grid cell models (Bellmund et al., 2016; Bush et al., 2015; Castro & Aguiar, 2014; Hasselmo, 2009; Mhatre et al., 2012; Solstad et al., 2006; Sorscher et al., 2023; Stepanyuk, 2015; Widloski & Fiete, 2014) are domain specific models of spatial navigation and as such, they do not lend themselves to explanations of human memory. Thus, the reason to prefer this model is parsimony. Rather than needing to develop a theory of memory that is separate from a theory of spatial navigation, it might be possible to address both literatures with a unified account.

This study does not attempt to falsify other theories of grid cells. Instead, this model reaches a radically different interpretation regarding the function of grid cells; an interpretation that emerges from viewing spatial navigation as a memory problem. All other grid cell models assume that an entorhinal grid cell displaying a spatially arranged grid of firing fields serves the function of spatial coding (i.e., spatial grid cells exist to support a spatial metric). In contrast, the proposed memory model of grid cells assumes that the hexagonal tiling reflects the need to keep memories separate from each other to minimize confusion and confabulation – the grid pattern is the byproduct of pattern separation between memories rather than the basis of a spatial code. 

It is now understood that grid-like firing fields can occur for non-spatial two dimensional spaces. For instance, human entorhinal cortex exhibits grid-like responses to video morph trajectories in a two-dimensional bird neck-length versus bird leg-length space (Constantinescu et al., 2016). As a general theory of learning and memory, the proposed memory model of grid cells is easily extended to explain these results (e.g., relabeling the border cell inputs in the model as neck-length and leg-length inputs). However, there are other grid cell models that can explain both spatial grid cells as well as non-spatial grid-like responses (Mok & Love, 2019; Rodríguez-Domínguez & Caplan, 2019; Stachenfeld et al., 2017; Wei et al., 2015). Similar to this memory model of grid cells, these models are also positioned to explain both the rodent spatial navigation and human memory literatures. Nevertheless, there is a key difference between this model and other grid cell models that generalize to non-spatial representations. Specifically, these other models assume that grid cells exhibiting spatial receptive fields serve the function of identifying positions in the environment (i.e., their function is spatial). As such, these models do not explain why most of the input to rodent hippocampus appears to be spatial (Boccara et al., 2010; Diehl et al., 2017; Grieves & Jeffery, 2017). This memory model of grid cells provides an answer to the apparent paucity of nonspatial cell types in rodent MTL by proposing that grid cells with spatial receptive fields have been misclassified as spatial (they are what cells rather than where cells) and that place cells are fundamentally memory cells that conjoin what and where.”

The use of border cells or boundary vector cells as the main (or only) source of spatial information in the hippocampus is not well supported by experimental data. Border cells in the entorhinal cortex are not active in the center of an environment. Boundary-vector cells can fire farther away from the walls but are not found in the entorhinal cortex. They are located in the subiculum, a major output of the hippocampus. While the entorhinalhippocampal circuit is a loop, the route from boundary-vector cells to place cells is much less clear than from grid cells. Moreover, both border cells and boundary-vector cells (which are conflated in this paper) comprise a small population of neurons compared to grid cells.

AUTHOR RESPONSE: The model can be built without assuming between-border cells (early simulations with the model did not make this assumption). Regarding this issue, the text reads “Unlike the BVC model, the boundary cell representation is sparsely populated using a basis set of three cells for each of the three dimensions (i.e., 9 cells in total), such that for each of the three non-orthogonal orientations, one cell captures one border, another the opposite border, and the third cell captures positions between the opposing borders (Solstad et al., 2008). However, this is not a core assumption, and it is possible to configure the model with border cell configurations that contain two opponent border cells per dimension, without needing to assume that any cells prefer positions between the borders (with the current parameters, the model predicts there will be two border cells for each between-border cell). Similarly, it is possible to configure the model with more than 3 cells for each dimension (i.e., multiple cells representing positions between the borders).” The Solstad paper found a few cells that responded in positions between borders, but perhaps not as many as 1 out of 3 cells, such as this particular model simulation predicts. If the paucity of between-border cells is a crucial data point, the model can be reconfigured with opponent-border cells without any between border cells. The reason that 3 border cells were used rather than 2 opponent border cells was for simplicity. Because 3 head direction cells were used to capture the face-centered cubic packing of memories, the simulation also used 3 border cells per dimensions to allow a common linear sum metric when conjoining dimensions to form memories. If the border dimensions used 2 cells while head direction used 3 cells, a dimensional weighting scheme would be needed to allow this mixing of “apples and oranges” in terms of distances in the 3D space that includes head direction.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors): 

Specific questions/clarifications:  

(1) Assumption of population-based vs single unit link to biological cells: At the start, the author assumes that each unit here can be associated with a population: "the simulated activation values can be thought of as proportional to the average firing rate of an ensemble of neurons with similar inputs and outputs (O'Reilly & Munakata, 2000)." But is a 'grid cell' found here a single cell or an average of many cells? Does this mean the model assumes many cells that have different fields that are averaged, which become a grid-like unit in the model? But in biology, these are single cells? Or does it mean a grid response is an average of the place cell inputs? 

I apologize for being unclear about this. The grid cells in the model are equivalent to real single cells except that the simulation uses a ratecoded cell rather than a spiking cell. The averaging that was mentioned in the paper is across identically behaving spiking cells rather than across cells with different grid field arrangements. To better explain this, I have added the following text:

“For instance, consider a set of several thousand spiking grid cells that are identical in terms of their firing fields. At any moment, some of these identically-behaving cells will produce an action potential while others do not (i.e., the cells are not perfectly synchronized), but a snapshot of their behavior can be extracted by calculating average firing rate across the ensemble. The simulated cells in the model represent this average firing rate of identically-behaving ensembles of spiking neurons.” 

This is a mathematical short-cut to avoid simulating many spiking neurons. Because this model was compared to real spike rate maps, this real-valued average firing rate is down-sampled to produce spikes by finding the locations that produced the top 5% of real-valued activation values across the simulation.

(2) It is not clear to me why they are circular border cells/basis sets.  

In the initial submission, there was a brief paragraph describing this assumption. In this revision, that paragraph has been expanded and modified for greater clarity. It now reads:

“Because head direction is necessarily a circular dimension, it was assumed that all dimensions are circular (a circular dimension is approximately linear for nearby locations). This assumption of circular dimensions was made to keep the model relatively simple, making it easier to combine dimensions and allowing application of the same processes for all dimensions. For instance, the model requires a weight normalization process to ensure that the pattern of weights for each dimension corresponds to a possible input value along that dimension. However, the normalization for a linear dimension is necessarily different than for a circular dimension. Because the neural tuning functions were assumed to be sine waves, normalization requires that the sum of squared weights add up to a constant value. For a linear dimension, this sum of squares rule only applies to the subset of cells that are relevant to a particular value along the dimension whereas for a circular dimension, this sum of squares rule is over the entire set of cells that represent the dimension (i.e., weight normalization is easier to implement with circular dimensions). Although all dimensions were assumed to be circular for reasons of mathematical convenience and parsimony, circular dimensions may relate to the finding that human observers have difficultly re-orienting themselves in a room depending on the degree of rotational symmetry of the room (Kelly et al., 2008). In addition, this simplifying assumption allows the model to capture the finding that the population of grid cells lies on a torus (Gardner et al., 2022), although I note that the model was developed before this result was known.”

(3) Why is it 3 components? I realise that the number doesn't matter too much, but I believe more is better, so is it just for simplicity? 

In this revision, additional text has been added to explain this assumption: “To keep the model simple, the same number of cells was assumed for all dimensions and all dimensions were assumed to be circular (head direction is necessarily circular and because one dimension needed to be circular, all dimensions were assumed to be circular). Three cells per dimensions was chosen because this provides a sparse population code of each dimension, with few border cells responding between borders, with few border cells responding between borders, while allowing three separate phases of grid cells within a grid cell module (in the model, a grid cell module arises from combination of a third dimension, such as head direction, with the real-world X/Y dimensions defined by border cells).”

As a reminder, the text explaining the sparse coding of border cells reads: “However, this is not a core assumption, and it is possible to configure the model with border cell configurations that contain two opponent border cells per dimension, without needing to assume that any cells prefer positions between the borders (with the current parameters, the model predicts there will be two border cells for each between-border cell). Similarly, it is possible to configure the model with more than 3 cells for each dimension (i.e., multiple cells representing positions between the borders).”

The model can work with just two opponent cells or with more than three cells per basis set. In different simulations, I have explored these possibilities. Three was chosen because it is a convenient way to highlight the face-centered cubic packing of memories that tends to occur (FCP produces 3 alternating layers of hexagonally arranged firing fields). Thus, each of the three head direction cells captures a different layer of the FCP arrangement. A more realistic simulation might combine 6 different head direction cells tiling the head direction dimension with opponent border cells (just 2 cells for each border dimensions). Such a combination would produce responses at borders, but no responses between borders and, at the same time, the head direction cells would still reveal the FCP arrangement. However, it is not easy to find the right parameters for such a mix-and-match simulation in which different dimensions have different numbers of tuning functions (e.g., some dimensions having 2 cells while others have 3 or 6 and some dimensions being linear while others are circular). When all of the dimensions are of the same type, the simple sum that arises from multiplying the input by the weight values gives rise to Euclidean distance (see Figure 3B). With a mix-and-match model of different dimension-types, it should be possible to adjust the sum to nevertheless produce a monotonic function with Euclidean distance although I leave this to future work. To keep things simple, I assumed that all dimensions are of the same type (circular, with 3 cells per dimension).  

(4) Confusion due to the border cells/box was unclear to me. "If the period of the circular border cells was the same as the width of the box, then a memory pushed outside the box on one side would appear on the opposite side of the box, in which case the partial grid field on one side should match up with its remainder on the other side. This would entail complete confusion between opposite sides of the box, and the representation of the box would be a torus (donut-shaped) rather than a flat two-dimensional surface. To reduce confusion ..." Is this confusion of the model? Of the animal?  

This would be confusion of the animal (e.g., a memory field overlapping with one border would also appear at the opposite border in the corresponding location). At one point in model development, I made the assumption that one side of the box wraps to the other side, and I asked Trygve Solstad to run some analyses of real data to see if cells actually wrap around in this manner. He did not find any evidence of this, and so I decided to include outsidethe-box representational area which, as it turned out, allowed the model to capture other behaviors as detailed in the paper.

This section of the paper now reads:

“The cosine tuning curves of the simulated border cells represent distance from the border on both sides of the border (i.e., firing rate increases as the animal approaches the border from either the inside or the outside of the enclosure). Experimental procedures do not allow the animal to experience locations immediately outside the enclosure, but these locations remain an important part of the hypothetic representation, particularly when considering the modification of memories through consolidation (i.e., a memory created inside the enclosure might be moved to a location outside the enclosure). This symmetry about the border cell’s preferred location is needed to maintain an unbiased representation, with a constant sum of squares for the border cell inputs (see methods section). Rather than using linear dimensions, all dimensions were assumed to be circular to keep the model relatively simple. This assumption was made because head direction is necessarily a circular dimension and by having all dimensions be circular, it is easy to combine dimensions in a consistent manner to produce multidimensional hippocampal place cell memories. Thus, the border cells define a torus (or more accurately a three-torus) of possible locations. This provides a hypothetical space of locations that could be represented.

In light of the assumption to represent border cells with a circular dimension, when a memory is pushed outside the East wall of the enclosure, it would necessarily be moved to the West wall of the enclosure if the period of the circular dimension was equal to the width of the enclosure. If this were true, then the partial grid field on one side of the enclosure would match up with its remainder on the other side. Such a situation would cause the animal to become completely confused regarding opposite sides of the enclosure (a location on the West wall would be indistinguishable from the corresponding location on the East wall). To reduce confusion between opposite sides of the enclosure, the width of the enclosure in which the animal navigated (Figure 5) was assumed to be half as wide as the full period of the border cells. In other words, although the space of possible representations was a three-torus, it was assumed that the real-world twodimensional enclosure encompassed a section of the torus (e.g., a square piece of tape stuck onto the surface of a donut). The torus is better thought of as “playing field” in which different sizes and shapes of enclosure can be represented (i.e., different sizes and shapes of tape placed on the donut). Furthermore, this assumption provides representational space that is outside the box without such locations wrapping around to the opposite side of the box.”

(5) Figure 3 - This result seems to be related to whether you use Euclidean or city-block distance. If you use Euclidean distances in two dimensions wouldn't this work out fine?  

Euclidean distance was the metric used in the analysis of the two-dimensional simulation, but this did not work out. To make this clear, I have changed the label on the x-axes to read “Euclidean distance” for both the two- and three-dimensional simulations. The two-dimensional simulation produced city block behavior rather than Euclidean behavior because memory retrieval is the sum of the two dimensions, as is standard in neural networks, rather than the Euclidian distance formula, which would require that memory retrieval be the square root of the sum of squares of the two dimensions. One way to address this problem with the two-dimensional simulation would be to use a specific Euclidean-mimicking activation function rather than a simple sum of dimensions. The very first model I developed used such an activation function as applied to opponent border cells with just two dimensions (so 4 cells in total – left/right and top/down). This produced Euclidean behavior, but the activation function was implausible and did not generalize to simulations that also included head direction. In contrast, with three non-orthogonal dimensions, the simple sum of dimensions is approximately Euclidean.

(6) Final sentence of the Discussion: "However, unlike the present model, these models still assume that entorhinal grid cells represent space rather than a non-spatial attribute." I am not sure if the authors of the cited papers will agree with this. They consider the spatial cases, but most argue they can treat non-spatial features as well. What the author might mean is that they assume non-spatial features are in some metric space that, in a way, is spatial. However, I am not sure if the author would argue that non-spatial features cannot be encoded metrically (e.g., Euclidean distance based on the similarity of odours). 

In this section, when referring to “entorhinal grid cells” I was specifically referring to traditional grid cells in a rodent spatial navigation experiment. I did not mean to imply that these other theories cannot explain nonspatial grid fields, such as in the two-dimensional bird space grid cells found with humans. The way in which the proposed memory model and these other models differ is in terms of what they assume regarding the function of grid cells that exhibit spatial grid fields. In this revision, I have changed this text to read:

“These models can capture some of the grid cell results presented in the current simulations, including extension to non-spatial grid-like responses (e.g., grid field that cover a two-dimensional neck/leg length bird space). Furthermore, these models may be able to explain memory phenomena similar to the model proposed in this study. However, unlike the proposed model, these models assume that the function of entorhinal grid cells that exhibit spatial X/Y grid fields during navigation is to represent space. In contrast, the memory model proposed in this study assume that the function of spatial X/Y grid cells is to represent a non-spatial attribute; the only reason they exhibit a spatial X/Y grid is because memories of that non-spatial attribute are arranged in a hexagonal grid owing to the uncluttered/unvarying nature of the enclosure. Thus, these model do not explain why most of the input to rodent hippocampus appears to be spatial (Boccara et al., 2010b; Diehl et al., 2017; Grieves & Jeffery, 2017) whereas the proposed model can explain this situation as reflecting the miss-classification of grid cells with a spatial arrangement as providing spatial input to hippocampus.”

(7) It would be interesting to see videos/gifs of the model learning, and an idea of how many steps of trials it takes (is it capturing real-time rodent cell firing whilst foraging, or is it more abstracted, taking more trials). 

The short answer is “yes”, the model is capturing real-time rodent cell firing while foraging. This is particularly true when simulating place cell memories in the absence of head direction information, as was shown in a video provided in the initial submission in relation to Figure 4. In this revision, I have provided a second video of learning when simulating place cell memories that include head direction. This second video is in relation to the results reported in Figure 9. This shows that even when learning a three-dimensional real-world space (X, Y, and head direction), the model rapidly produces an on-average hexagonal arrangement of place cells memories owing to the slight tendency of the place cell memories to linger in some locations as compared to others during consolidation. More specifically, they are more likely to linger in the locations that are the intersections of the peaks and/or troughs of the border cells and it is this tendency that supports the immediate appearance of grid cells. However, because the place cell memories are still shifting, head direction conjunctive grid cells are slower to emerge (the head direction conjunctive grid cells require stabilization of the place cells). The video then speeds up the learning process to so how place cells eventually stabilize after sufficient learning of the borders of the enclosure from different head/view directions.

(8) One question is whether all the results have to be presented in the main text. It was difficult to see which key predictions fit the data and do so better than a spatial/navigation account. 

Thank you for this suggestion. To make the paper more readable and easier for different readers with different interests to choose different aspects of the results to read, the second half of the results have been put in an appendix. More specifically, the second half of the results concerned place cells rather than grid cells. Thus, in this revision, the main text concerns grid cell results and the appendix concerns place cell results.

Reviewer #3 (Recommendations For The Authors):  

The title could usefully be shortened to focus on the main argument that observed firing patterns could be consistent with mapping memories instead of space. It's a stretch to argue that memory is the primary role when no such data is presented (i.e., there is no comparison of competing models). 

This is a good point (I do not present evidence that conclusively indicates the function of MTL). This original title was chosen to make clear how this account is a radical departure from other accounts of grid cells. The revised title highlights that: 1) a memory model can also explain rodent single cell recording data during navigation; and 2) grid cell may not be non-spatial. The revised title is: “A Memory Model of Rodent Spatial Navigation: Place Cells are Memories Arranged in a Grid and Grid Cells are Non-spatial”

When arguing that the main role of the hippocampus is memory, I strongly suggest engaging with the work of people like Howard Eichenbaum who spent the better part of their career arguing the same (e.g. DOI:10.1152/jn.00005.2017.)  

Thank you for pointing out this important oversight. Early in introduction, I now write: “The proposal that hippocampus represents the multimodal conjunctions that define an episode is not new (Marr et al., 1991; Sutherland & Rudy, 1989) and neither is the proposal that hippocampal memory supports spatial/navigation ability (Eichenbaum, 2017). This view of the hippocampus is consistent with “feature in place” results (O’Keefe & Krupic, 2021) in which hippocampal cells respond to the conjunction of a non-spatial attribute affixed to a specific location, rather than responding more generically to any instance of a non-spatial attribute. In other words, the what/where conjunction is unique. Furthermore, the uniqueness of the what/where conjunction may be the fundamental building block of spatial memory and navigation. In reviewing the hippocampal literature, Howard Eichenbaum (2017) concludes that ‘the hippocampal system is not dedicated to spatial cognition and navigation, but organizes experiences in memory, for which spatial mapping and navigation are both a metaphor for and a prominent application of relational memory organization.’”

With a focus on episodic memory, there should be a mention of the temporal component of memory. While it may rightfully be beyond the scope of this model, it's confusing to omit time completely from the discussion. 

This issue and several others are now addressed in a new section in the introduction titled ‘The Scope of the Proposed Model’. That section reads:

“The reported simulations explain why most mEC cell types in the rodent literature appear to be spatial (Boccara et al., 2010; Diehl et al., 2017; Grieves & Jeffery, 2017). Assuming that rodents can form non-spatial memories, rodent hippocampus must receive non-spatial input from entorhinal cortex. These simulations suggest that characterization of the rodent mEC cortex as primarily spatial might be incorrect if most grid cells (except perhaps head direction conjunctive grid cells) have been mischaracterized as spatial. Other literatures with other species find non-spatial representations in MTL (Gulli et al., 2020; Quiroga et al., 2005; Wixted et al., 2014) and non-spatial hippocampal memory encoding has been found in rodents (Liu et al., 2012; McEchron & Disterhoft, 1999). The proposed memory model is compatible with these results – the ideas contained in this model could be applied to nonspatial memory representations. However, surveys of cell types in rodent entorhinal cortex seem to indicate that most cells are spatial (Boccara et al., 2010; Diehl et al., 2017; Grieves & Jeffery, 2017). How can the rodent hippocampus encode nonspatial memories if most of its input is spatial? The goal of the reported simulations is to explain the apparent paucity of non-spatial cells in rodent entorhinal cortex by proposing that grid cells have been misclassified as spatial (see also Luo et al., 2024).

Given the simplicity of the proposed model, there are important findings that the model cannot address -- it is not that the model makes the wrong predictions but rather that it makes no predictions. The role of running speed (Kraus et al., 2015) is one such variable for which the model makes no predictions. Similarly, because the model is a rate-coded model rather than a model of oscillating spiking neurons, it makes no predictions regarding theta oscillations (Buzsáki & Moser, 2013). The model is an account of learning and memory for an adult animal, and it makes no predictions regarding the developmental (Langston et al., 2010; Muessig et al., 2015; Wills et al., 2012) or evolutionary (Rodrıguez et al., 2002) time course of different cell types. This model contains several purely spatial representations such as border cells, head direction cells, and head direction conjunctive grid cells and it may be that these purely spatial cell types emerged first, followed by the evolution and/or development of non-spatial cell types. However, this does not invalidate the model. Instead, this is a model for an adult animal that has both episodic memory capabilities and spatial navigation capabilities, irrespective of the order in which these capabilities emerged.

This model has the potential to explain context effects in memory (Godden & Baddeley, 1975; Gulli et al., 2020; Howard et al., 2005). According to this model, different grid cells represent different non-spatial characteristics and place cells represent the combination of these “context” factors and location. In the simulation, just one grid cell is simulated but the same results would emerge when simulating hundreds of different non-spatial inputs provided that all of the simulated non-spatial inputs exist throughout the recording session. However, there is evidence that hippocampus can explicitly represent the passage of time (Eichenbaum, 2014), and time is assuredly an important factor in defining episodic memory (Bright et al., 2020). Thus, although the current model addresses unique combinations of what and where, it is left to future work to incorporate representations of when in the memory model.”

I recommend explaining the motivation of the theory in more detail in the introduction. It reads as "what if it's like this?" It would be helpful to instead highlight the limitations of current theories and argue why this theory is either a better fit for the data or is logically simpler. 

This issue and several others are now addressed in the new section in the introduction titled ‘Why Model the Rodent Navigation Literature with a Memory Model?’, which I quoted above in response to the public reviews.

It's worth considering shortening the results section to include only those that most convincingly support the main claim. The manuscript is quite long and appears to lack focus at times. 

Thank you for this suggestion. To make the paper more readable and easier for different readers with different interests to choose different aspects of the results to read, the second half of the results have been put in an appendix. More specifically, the second half of the results concerned place cells rather than grid cells. Thus, in this revision, the main text concerns grid cell results and the appendix concerns place cell results.

The discussion of path dependence on the formation of the grid pattern is important but only briefly discussed. It may be useful to add simulations testing whether different paths (not random walks) produce distorted grid patterns. 

The short answer is that the path doesn’t affect things in general. The consolidation rule ensures equally spaced memories even if, for instance, one side of the enclosure is explored much more than the other side. As just one example, I have run simulations with a radial arm maze and even though the animal is constrained to only run on the maze arms. The memories still arrange hexagonally as memories become pushed outside the arms. Rather than adding additional simulations to study, I now briefly describe this in the model methods:

“Of note, the ability of the model to produce grid cell responses does not depend on this decision to simulate an animal taking a random walk – the same results emerge if the animal is more systematic in its path. All that matters for producing grid cell responses is that the animal visits all locations and that the animal takes on different head directions for the same location in the case of simulations that also include head direction as an input to hippocampal place cells.”

I struggle to understand in Figure 3 why retrieval strength ought to scale monotonically with Euclidean distance, and why that justifies a more complex model (three non-orthogonal dimensions). 

The introduction to this section now reads: “Animals can plan novel straight line paths to reach a known position and evidence suggests they do so by learning Euclidean representations of space (Cheng & Gallistel, 2014; Normand & Boesch, 2009; Wilkie, 1989). Thus, it was assumed that hippocampal place cells represent positions in Euclidean space (as opposed to non-Euclidean space, such a occurs with a city-block metric).”

p.17 "although the representational space is a torus (or more specifically a three-torus), it is assumed that the real-world two-dimensional surface is only a section of the torus (e.g., a square piece of tape stuck onto the surface of a donut)." I fail to understand how the realworld surface is only a part of the torus. In the existing theoretical and experimental work on toroidal topology of grid cell activity, the torus represents a very small fraction of the real world, and repeating activity on the toroidal manifold is a crucial feature of how it maps 2D space in a regular manner. Why then here do you want the torus to be larger than the realworld? 

This section has been rewritten to better explain these assumptions. The relevant paragraphs now read:

“The cosine tuning curves of the simulated border cells represent distance from the border on both sides of the border (i.e., firing rate increases as the animal approaches the border from either the inside or the outside of the enclosure). Experimental procedures do not allow the animal to experience locations immediately outside the enclosure, but these locations remain an important part of the hypothetic representation, particularly when considering the modification of memories through consolidation (i.e., a memory created inside the enclosure might be moved to a location outside the enclosure). This symmetry about the border cell’s preferred location is needed to maintain an unbiased representation, with a constant sum of squares for the border cell inputs (see methods section). Rather than using linear dimensions, all dimensions were assumed to be circular to keep the model relatively simple. This assumption was made because head direction is necessarily a circular dimension and by having all dimensions be circular, it is easy to combine dimensions in a consistent manner to produce multidimensional hippocampal place cell memories. Thus, the border cells define a torus (or more accurately a three-torus) of possible locations. This provides a hypothetical space of locations that could be represented.

In light of the assumption to represent border cells with a circular dimension, when a memory is pushed outside the East wall of the enclosure, it would necessarily be moved to the West wall of the enclosure if the period of the circular dimension was equal to the width of the enclosure. If this were true, then the partial grid field on one side of the enclosure would match up with its remainder on the other side. Such a situation would cause the animal to become completely confused regarding opposite sides of the enclosure (a location on the West wall would be indistinguishable from the corresponding location on the East wall). To reduce confusion between opposite sides of the enclosure, the width of the enclosure in which the animal navigated (Figure 5) was assumed to be half as wide as the full period of the border cells. In other words, although the space of possible representations was a three-torus, it was assumed that the real-world twodimensional enclosure encompassed a section of the torus (e.g., a square piece of tape stuck onto the surface of a donut). The torus is better thought of as “playing field” in which different sizes and shapes of enclosure can be represented (i.e., different sizes and shapes of tape placed on the donut). Furthermore, this assumption provides representational space that is outside the box without such locations wrapping around to the opposite side of the box.”

p.28 "More specifically, egocentric grid cells (e.g., head direction conjunctive grid cells) require stabilization of the place cell memories in the face of ongoing consolidation whereas allocentric grid cells reflect on-average place field positions." and p.32 "if place cells represent episodic memories, it seems natural that they should include head direction (an egocentric viewpoint)." But the head direction signal is not egocentric, it is allocentric. I'm unsure whether this is a typo or a potentially more serious conceptual misunderstanding. 

Any reference to egocentric has been removed in this revision. In the initial submission, when I used egocentric, I was referring to memories that depended on the head direction of the animal at the time of memory formation. I was using “egocentric” in relation to whether the memory was related to the animal’s personal bodily experience at the time of memory formation. But I concede that this is confusing since the ego/allo distinction is typically used to differentiate angular directions that are relative to the person (left/right) versus earth (East/West). Instead, throughout the manuscript I now refer to these as view-dependent memories since head direction would entail having a different view of the environment at the time of memory formation. I still refer to the stacking of multiple view-dependent memories on the same X/Y location as being the development of an allocentric representation however, since this can be thought of as one way to learn a cognitive map of the enclosure that is view independent.

p.37 "But if the border cells had changed their alignment with the new enclosure (e.g., if the E border dimension aligned with the North-South borders), then the place cells would have appeared to undergo global remapping as their positions rotated by 90 degrees and the grid pattern would have also rotated." But this would not be interpreted as global remapping by standard analyses of place and grid cell responses. A coherent rotation of firing patterns is not interpreted as remapping. 

This sentence now reads: “But if the border cells had changed their alignment with the new enclosure (e.g., if the E border dimension aligned with the North-South borders), then the place cells would remain in their same positions relative to the now-rotated borders (i.e., no remapping relative to the enclosure) and the corresponding grid cells would also retain their same alignment relative to the enclosure.”

p.37 "this is more accurately described as partial remapping (nearly all place fields were unaffected)." If nearly all place fields were unaffected, this should be interpreted as a stable map. Partial remapping is a mix of stability, rate remapping, and global remapping within a population of place cells. 

This sentence has been removed.

p.40 "The dependence of grid cell responses on memory may help explain why grid cells have been found for bats crawling on a two-dimensional surface (Yartsev et al., 2011), but three-dimensional grid cells have never been observed for flying bats." This is not true. Ginosar et al. (2021) observed 3D grid cells in flying bats.  

Thank you for highlighting this issue. In the initial submission I was using “grid cell” to mean a cell that produced a precise hexagonal grid, which is not the case for the 3D grid cells in bats. In this revision, I now discuss grid cell that produce irregular grid fields, writing:

“According to this model, hexagonally arranged grid cells should be the exception rather than the rule when considering more naturalistic environments. In a more ecologically valid situation, such as with landmarks, varied sounds, food sources, threats, and interactions with conspecifics, there may still be remembered locations were events occurred or remembered properties can be found, but because the non-spatial properties are non-uniform in the environment, the arrangement of memory feedback will be irregular, reflecting the varied nature of the environment. This may explain the finding that even in a situation where there are regular hexagonal grid cells, there are often irregular non-grid cells that have a reliable multi-location firing field, but the arrangement of the firing fields is irregular (Diehl et al., 2017). For instance, even when navigating in an enclosure that has uniform properties as dictated by experimental procedures, they may be other properties that were not well-controlled (e.g., a view of exterior lighting in some locations but not others), and these uncontrolled properties may produce an irregular grid (i.e., because the uncontrolled properties are reliably associated with some locations but not others, hippocampal memory feedback triggers retrieval of those properties in the associations locations).

In this memory model, there are other situations in which an irregular but reliable multi-location grid may occur, even when everything is well controlled. In the reported simulations, when the hippocampal place cells were based on variation in X/Y (as defined by Border cells), nothing else changed as a function of location, and the model rapidly produced a precise hexagonal arrangement of hippocampal place cell memories. When head direction was included (i.e., real-world variation in X, Y, and head direction), the model still produced a hexagonal arrangement as per face centered cubic packing of memories, but this precise arrangement was slower to emerge, with place cells continuing to shift their positions until the borders of the enclosure were sufficiently well learned from multiple viewpoints. If there is realworld variation in four or more dimensions, as is likely the case in a more ecologically valid situation, it will be even harder for place cell memories to settle on a precise regular lattice. Furthermore, in the case of four dimensions, mathematicians studying the “sphere packing problem” recently concluded that densest packing is irregular (Campos et al., 2023). This may explain why the multifield grid cells for freely flying bats have a systematic minimum distance between firing fields, but their arrangement is globally irregular (Ginosar et al., 2021). Assuming that the memories encoded by a bat include not just the three realworld dimensions of variation, but also head direction, the grid will likely be irregular even under optimal conditions of laboratory control.”

Multiple typos are found on page 25, end of paragraph 3: "More specifically, if there is one set of non-orthogonal dimensions for enclosure borders and the movement of one wall is too modest as to cause avoid global remapping, this would deform the grid modules based the enclosure border cells."

As detailed above in the response the public reviews, this paragraph has been rewritten.

eLife Assessment

The authors studied the development of mesentery borders in the rice coral Montipora, a new experimental system, to complement existing data from the sea anemone Nematostella. They make a solid case that in Montipora, there is a sequence of Hox-Gbx genes whose staggered expression in the unsegmented larva is suggestive of their role in subdividing the gastric cavity into repeated units bordered by mesenteries, as in the sea anemone Nematostella. Pharmacological experiments also point to the involvement of the BMP pathway in this process, but additional experiments validating this are necessary. This is a valuable contribution to the field of cnidarian evolution, suggesting that BMP- and "Hox-Gbx code"-dependent patterning of the directive axis was ancestral for Anthozoa.

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:

(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.)

(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.

(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.

(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. 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?

(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.

(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).

Virtuous project! I love the effect.

ern fait je ne vois pas comment un cadre légal peut contraindre des pratiques qui par définition sont illicte. On peut mettre tous les cadres légaux en place, il n'auront aucun effet si un Etat décide de passer outre

Reviewer #3 (Public review):

In this study, O'Brien et al. address the need for scalable and cost-effective approaches to finding lead compounds for the treatment of the growing number of Mendelian diseases. They used state-of-the-art phenotypic screening based on an established high-dimensional phenotypic analysis pipeline in the nematode C. elegans.

First, a panel of 25 C. elegans models was created by generating CRISPR/Cas9 knock-out lines for conserved human disease genes. These mutant strains underwent behavioral analysis using the group's published methodology. Clustering analysis revealed common features for genes likely operating in similar genetic pathways or biological functions. The study also presents results from a more focused examination of ciliopathy disease models.

Subsequently, the study focuses on the NALCN channel gene family, comparing the phenotypes of mutants of nca-1, unc-77, and unc-80. This initial characterization identifies three behavioral parameters that exhibit significant differences from the wild type and could serve as indicators for pharmacological modulation.

As a proof-of-concept, O'Brien et al. present a drug repurposing screen using an FDA-approved compound library, identifying two compounds capable of rescuing the behavioral phenotype in a model with UNC80 deficiency. The relatively short time and low cost associated with creating and phenotyping these strains suggest that high-throughput worm tracking could serve as a scalable approach for drug repurposing, addressing the multitude of Mendelian diseases. Interestingly, by measuring a wide range of behavioural parameters, this strategy also simultaneously reveals deleterious side effects of tested drugs that may confound the analysis.

Considering the wealth of data generated in this study regarding important human disease genes, it is regrettable that the data is not made accessible to researchers less versed in data analysis methods. This diminishes the study's utility. It would have a far greater impact if an accessible and user-friendly online interface were established to facilitate data querying and feature extraction for specific mutants. This would empower researchers to compare their findings with the extensive dataset created here.

Another technical limitation of the study is the use of single alleles. Large deletion alleles were generated by CRISPR/Cas9 gene editing. At first glance, this seems like a good idea because it limits the risk that background mutations, present in chemically-generated alleles, will affect behavioral parameters. However, these large deletions can also remove non-coding RNAs or other regulatory genetic elements, as found, for example, in introns. Therefore, it would be prudent to validate the behavioral effects by testing additional loss-of-function alleles produced through early stop codons or targeted deletion of key functional domains.

Comments on revisions:

In this final round of revisions, the authors have improved their manuscript and provide useful information about analysis procedures and code and updated figures.

Author response:

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

This important study provides proof of principle that C. elegans models can be used to accelerate the discovery of candidate treatments for human Mendelian diseases by detailed high-throughput phenotyping of strains harboring mutations in orthologs of human disease genes. The data are compelling and support an approach that enables the potential rapid repurposing of FDA-approved drugs to treat rare diseases for which there are currently no effective treatments. The authors should provide a clearer explanation of how the statistical analyses were performed, as well as a link to a GitHub repository to clarify how figures and tables in the manuscript were generated from the phenotypic data.

We have amended our description of the statistical analysis in the materials and methods section of the manuscript. We have also updated the GitHub repository link to a dedicated repository for this study, this contains all of the code needed to generated all the figures made from the phenotypic data provided. Additionally, we have updated the Zenodo repository to contain both the code and datasets within the same file.

We have also updated the GitHub repository link to a dedicated repository for this manuscript, that contains all of the code needed to generate all figures from the phenotypic data provided. Additionally, we have updated the Zenodo repository link to contain both the code and datasets within the same folder structure. 

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

The authors have responded to previous review to improve the presentation of the work. The paper more than meets publication standards.

No response required.

Reviewer #2 (Recommendations for the authors):

The authors have addressed all of my questions and concerns. I'm happy to see this updated paper of record.

No response required.

Reviewer #3 (Recommendations for the authors):

Regarding the interactive heatmap

The html version and the panel in Figure 2C appear not to coincide visually. Maybe the features are ordered in a different way?

The html version of Figure 2C is for the entire feature set extract per strain and not the condensed Tierpsy256 set shown in the panel figure. We have now remade this figure to show this reduced feature set (aligning with what is shown in Figure 2C) and included both versions of the interactive heatmaps as static html files within the same repository.

Regarding data accessibility overall

More generally, the html file does not address my initial concern about the accessibility of the data to non-experts. Making the full dataset available was a necessary first step, but the hermetic nature of its format and the lack of a simple way to query the data remains an issue for me that limits the usefulness of this data to the broadest audience.

We agree, but unfortunately do not currently have the resources to build a public-facing database to facilitate this.

Enshittification (Cory Doctorow)

In essence, the bad actors pushed toward centralization and reliance on large providers to cut the noise from their "signaling/code" surface.

We gotta bring back color coding for these prestige TV shows with thirty billion characters.

lien vers l'article

Lien bleu avec -> devant

manque la puce

enlever espacement au-dessus

Lien DSFR avec "->" devant

Puce

Lien : https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000042813464/

Manque le lien https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000044467582

manque l'ancre dans le code source

manque le lien "https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000044467582"

remonter l'article

Manque le lien https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000044467582

mettre un espace (retour à la ligne) avant la 2e puce

"In the world of The Road , there is a simple rule for distinguishing the good guys from the bad guys. Bad guys eat people; good guys don't. This is what remains of the Categorical Imperative: don't treat people as mere food. While this is the most obvious principle to which good guys are committed, it is not the only one. It is possible to discern in The Road a Code of the Good Guys, a set of principles to which good guys are committed. That Code includes the following rules:

1. Don't eat people.
2. Don't steal.
3. Don't lie.
4. Keep your promises.
5. Help others.
6. Never give up.

The man tries to teach these principles to the child and he tries to follow them himself. Throughout the novel we witness the man's struggle to be a good guy, to do what is right in a world in which most people seem to have abandoned morality altogether." (Wielenberg 5-6).

https://www.jstor.org/stable/42909407

"Every social institution and convention that could serve as a hallmark of civilization has passed so far into oblivion that, as Ashley Kunsa argues, the names of places, road, and people have passed into meaninglessness, leaving only the deeds of individuals to providing meaning and morality to the world (61–63). The most important dividing line for the boy is the assurance from his father that they will not eat people." (Dominy 146).

https://www.jstor.org/stable/10.5325/cormmccaj.13.1.0143

More on Worldview here: https://ontheroadtotheroad7.wordpress.com/2024/11/26/worldview/

Author response:

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

Reviewer #1 (Public Review):

In this revision, the authors significantly improved the manuscript. They now address some of my concerns. Specifically, they show the contribution of end-effects on spreading the inputs between dendrites. This analysis reveals greater applicability of their findings to cortical cells, with long, unbranching dendrites than other neuronal types, such as Purkinje cells in the cerebellum.

They now explain better the interactions between calcium and voltage signals, which I believe improve the take-away message of their manuscript. They modified and added new figures that helped to provide more information about their simulations.

However, some of my points remain valid. Figure 6 shows depolarization of ~5mV from -75. This weak depolarization would not effectively recruit nonlinear activation of NMDARs. In their paper, Branco and Hausser (2010) showed depolarizations of ~10-15mV.

More importantly, the signature of NMDAR activation is the prolonged plateau potential and activation at more depolarized resting membrane potentials (their Figure 4). Thus, despite including NMDARs in the simulation, the authors do not model functional recruitment of these channels. Their simulation is thus equivalent to AMPA only drive, which can indeed summate somewhat nonlinearly.

In the current study, we used short sequences of 5 inputs, since the convergence of longer sequences is extremely unlikely in the network configurations we have examined. This resulted in smaller EPSP amplitudes of ~5mV (Figure 6 - Supplement 2A, B). Longer sequences containing 9 inputs resulted in larger somatic depolarizations of ~10mV (Figure 6 - Supplement 2E, F). Although we had modified the (Branco, Clark, and Häusser 2010) model to remove the jitter in the timing of arrival of inputs and made slight modifications to the location of stimulus delivery on the dendrite, we saw similar amplitudes when we tested a 9-length sequence using (Branco, Clark, and Häusser 2010)’s published code (Figure 6 - Supplement 2I, J). In all the cases we tested (5 input sequence, 9 input sequence, 9 input sequence with (Branco, Clark, and Häusser 2010) code repository), removal of NMDA synapses lowered both the somatic EPSPs (Figure 6 - Supplement 2C,D,G,H,K,L) as well as the selectivity (measured as the difference between the EPSPs generated for inward and outward stimulus delivery) (Figure 6 Supplement 2M,N,O). Further, monitoring the voltage along the dendrite for a sequence of 5 inputs showed dendritic EPSPs in the range of 20-45 mV (Figure 6 - Supplement 2P, Q), which came down notably (10-25mV) when NMDA synapses were abolished (Figure 6 - Supplement 2R, S). Thus, even sequences containing as few as 5 inputs were capable of engaging the NMDA-mediated nonlinearity to show sequence selectivity, although the selectivity was not as strong as in the case of 9 inputs.

Reviewer #1 (Recommendations for the authors):

Minor points:

Figure 8, what does the scale in A represent? I assume it is voltage, but there are no units. Figure 8, C, E, G, these are unconventional units for synaptic weights, usually, these are given in nS / per input.

We have corrected these. The scalebar in 8A represents membrane potential in mV. The units of 8C,E,G are now in nS.

Reviewer #2 (Public Review):

Summary:

If synaptic input is functionally clustered on dendrites, nonlinear integration could increase the computational power of neural networks. But this requires the right synapses to be located in the right places. This paper aims to address the question of whether such synaptic arrangements could arise by chance (i.e. without special rules for axon guidance or structural plasticity), and could therefore be exploited even in randomly connected networks. This is important, particularly for the dendrites and biological computation communities, where there is a pressing need to integrate decades of work at the single-neuron level with contemporary ideas about network function.

Using an abstract model where ensembles of neurons project randomly to a postsynaptic population, back-of-envelope calculations are presented that predict the probability of finding clustered synapses and spatiotemporal sequences. Using data-constrained parameters, the authors conclude that clustering and sequences are indeed likely to occur by chance (for large enough ensembles), but require strong dendritic nonlinearities and low background noise to be useful.

Strengths:

(1) The back-of-envelope reasoning presented can provide fast and valuable intuition. The authors have also made the effort to connect the model parameters with measured values. Even an approximate understanding of cluster probability can direct theory and experiments towards promising directions, or away from lost causes.

(2) I found the general approach to be refreshingly transparent and objective. Assumptions are stated clearly about the model and statistics of different circuits. Along with some positive results, many of the computed cluster probabilities are vanishingly small, and noise is found to be quite detrimental in several cases. This is important to know, and I was happy to see the authors take a balanced look at conditions that help/hinder clustering, rather than to just focus on a particular regime that works.

(3) This paper is also a timely reminder that synaptic clusters and sequences can exist on multiple spatial and temporal scales. The authors present results pertaining to the standard `electrical' regime (~50-100 µm, <50 ms), as well as two modes of chemical signaling (~10 µm, 100-1000 ms). The senior author is indeed an authority on the latter, and the simulations in Figure 5, extending those from Bhalla (2017), are unique in this area. In my view, the role of chemical signaling in neural computation is understudied theoretically, but research will be increasingly important as experimental technologies continue to develop.

Weaknesses:

(1) The paper is mostly let down by the presentation. In the current form, some patience is needed to grasp the main questions and results, and it is hard to keep track of the many abbreviations and definitions. A paper like this can be impactful, but the writing needs to be crisp, and the logic of the derivation accessible to non-experts. See, for instance, Stepanyants, Hof & Chklovskii (2002) for a relevant example.

It would be good to see a restructure that communicates the main points clearly and concisely, perhaps leaving other observations to an optional appendix. For the interested but time-pressed reader, I recommend starting with the last paragraph of the introduction, working through the main derivation on page 7, and writing out the full expression with key parameters exposed. Next, look at Table 1 and Figure 2J to see where different circuits and mechanisms fit in this scheme. Beyond this, the sequence derivation on page 15 and biophysical simulations in Figures 5 and 6 are also highlights.

We appreciate the reviewers' suggestions. We have tightened the flow of the introduction. We understand that the abbreviations and definitions are challenging and have therefore provided intuitions and summaries of the equations discussed in the main text.

Clusters calculations

Our approach is to ask how likely it is that a given set of inputs lands on a short segment of dendrite, and then scale it up to all segments on the entire dendritic length of the cell.

Thus, the probability of occurrence of groups that receive connections from each of the M ensembles (PcFMG) is a function of the connection probability (p) between the two layers, the number of neurons in an ensemble (N), the relative zone-length with respect to the total dendritic arbor (Z/L) and the number of ensembles (M).

Sequence calculations

Here we estimate the likelihood of the first ensemble input arriving anywhere on the dendrite, and ask how likely it is that succeeding inputs of the sequence would arrive within a set spacing.

Thus, the probability of occurrence of sequences that receive sequential connections (PcPOSS) from each of the M ensembles is a function of the connection probability (p) between the two layers, the number of neurons in an ensemble (N), the relative window size with respect to the total dendritic arbor (Δ/L) and the number of ensembles (M).

(2) I wonder if the authors are being overly conservative at times. The result highlighted in the abstract is that 10/100000 postsynaptic neurons are expected to exhibit synaptic clustering. This seems like a very small number, especially if circuits are to rely on such a mechanism. However, this figure assumes the convergence of 3-5 distinct ensembles. Convergence of inputs from just 2 ense mbles would be much more prevalent, but still advantageous computationally. There has been excitement in the field about experiments showing the clustering of synapses encoding even a single feature.

We agree that short clusters of two inputs would be far more likely. We focused our analysis on clusters with three of more ensembles because of the following reasons:

(1) The signal to noise in these clusters was very poor as the likelihood of noise clusters is high.

(2) It is difficult to trigger nonlinearities with very few synaptic inputs.

(3) At the ensemble sizes we considered (100 for clusters, 1000 for sequences), clusters arising from just two ensembles would result in high probability of occurrence on all neurons in a network (~50% in cortex, see p_CMFG in figures below.). These dense neural representations make it difficult for downstream networks to decode (Foldiak 2003).

However, in the presence of ensembles containing fewer neurons or when the connection probability between the layers is low, short clusters can result in sparse representations (Figure 2 - Supplement 2). Arguments 1 and 2 hold for short sequences as well.

(3) The analysis supporting the claim that strong nonlinearities are needed for cluster/sequence detection is unconvincing. In the analysis, different synapse distributions on a single long dendrite are convolved with a sigmoid function and then the sum is taken to reflect the somatic response. In reality, dendritic nonlinearities influence the soma in a complex and dynamic manner. It may be that the abstract approach the authors use captures some of this, but it needs to be validated with simulations to be trusted (in line with previous work, e.g. Poirazi, Brannon & Mel, (2003)).

We agree that multiple factors might affect the influence of nonlinearities on the soma. The key goal of our study was to understand the role played by random connectivity in giving rise to clustered computation. Since simulating a wide range of connectivity and activity patterns in a detailed biophysical model was computationally expensive, we analyzed the exemplar detailed models for nonlinearity separately (Figures 5, 6, and new figure 8), and then used our abstract models as a proxy for understanding population dynamics. A complete analysis of the role played by morphology, channel kinetics and the effect of branching requires an in-depth study of its own, and some of these questions have already been tackled by (Poirazi, Brannon, and Mel 2003; Branco, Clark, and Häusser 2010; Bhalla 2017). However, in the revision, we have implemented a single model which incorporates the range of ion-channel, synaptic and biochemical signaling nonlinearities which we discuss in the paper (Figure 8, and Figure 8 Supplement 1, 2,3). We use this to demonstrate all three forms of sequence and grouped computation we use in the study, where the only difference is in the stimulus pattern and the separation of time-scales inherent in the stimuli.

(4) It is unclear whether some of the conclusions would hold in the presence of learning. In the signal-to-noise analysis, all synaptic strengths are assumed equal. But if synapses involved in salient clusters or sequences were potentiated, presumably detection would become easier? Similarly, if presynaptic tuning and/or timing were reorganized through learning, the conditions for synaptic arrangements to be useful could be relaxed. Answering these questions is beyond the scope of the study, but there is a caveat there nonetheless.

We agree with the reviewer. If synapses receiving connectivity from ensembles had stronger weights, this would make detection easier. Dendritic spikes arising from clustered inputs have been implicated in local cooperative plasticity (Golding, Staff, and Spruston 2002; Losonczy, Makara, and Magee 2008). Further, plasticity related proteins synthesized at a synapse undergoing L-LTP can diffuse to neighboring weakly co-active synapses, and thereby mediate cooperative plasticity (Harvey et al. 2008; Govindarajan, Kelleher, and Tonegawa 2006; Govindarajan et al. 2011). Thus if clusters of synapses were likely to be co-active, they could further engage these local plasticity mechanisms which could potentiate them while not potentiating synapses that are activated by background activity. This would depend on the activity correlation between synapses receiving ensemble inputs within a cluster vs those activated by background activity. We have mentioned some of these ideas in a published opinion paper (Pulikkottil, Somashekar, and Bhalla 2021). In the current study, we wanted to understand whether even in the absence of specialized connection rules, interesting computations could still emerge. Thus, we focused on asking whether clustered or sequential convergence could arise even in a purely randomly connected network, with the most basic set of assumptions. We agree that an analysis of how selectivity evolves with learning would be an interesting topic for further work.

References

• Bhalla, Upinder S. 2017. “Synaptic Input Sequence Discrimination on Behavioral Timescales Mediated by Reaction-Diffusion Chemistry in Dendrites.” Edited by Frances K Skinner. eLife 6 (April):e25827. https://doi.org/10.7554/eLife.25827.

• Branco, Tiago, Beverley A. Clark, and Michael Häusser. 2010. “Dendritic Discrimination of Temporal Input Sequences in Cortical Neurons.” Science (New York, N.Y.) 329 (5999): 1671–75. https://doi.org/10.1126/science.1189664.

• Foldiak, Peter. 2003. “Sparse Coding in the Primate Cortex.” The Handbook of Brain Theory and Neural Networks. https://research-repository.st-andrews.ac.uk/bitstream/handle/10023/2994/FoldiakSparse HBTNN2e02.pdf?sequence=1.

• Golding, Nace L., Nathan P. Staff, and Nelson Spruston. 2002. “Dendritic Spikes as a Mechanism for Cooperative Long-Term Potentiation.” Nature 418 (6895): 326–31. https://doi.org/10.1038/nature00854.

• Govindarajan, Arvind, Inbal Israely, Shu-Ying Huang, and Susumu Tonegawa. 2011. “The Dendritic Branch Is the Preferred Integrative Unit for Protein Synthesis-Dependent LTP.” Neuron 69 (1): 132–46. https://doi.org/10.1016/j.neuron.2010.12.008.

• Govindarajan, Arvind, Raymond J. Kelleher, and Susumu Tonegawa. 2006. “A Clustered Plasticity Model of Long-Term Memory Engrams.” Nature Reviews Neuroscience 7 (7): 575–83. https://doi.org/10.1038/nrn1937.

• Harvey, Christopher D., Ryohei Yasuda, Haining Zhong, and Karel Svoboda. 2008. “The Spread of Ras Activity Triggered by Activation of a Single Dendritic Spine.” Science (New York, N.Y.) 321 (5885): 136–40. https://doi.org/10.1126/science.1159675.

• Losonczy, Attila, Judit K. Makara, and Jeffrey C. Magee. 2008. “Compartmentalized Dendritic Plasticity and Input Feature Storage in Neurons.” Nature 452 (7186): 436–41. https://doi.org/10.1038/nature06725.

• Poirazi, Panayiota, Terrence Brannon, and Bartlett W. Mel. 2003. “Pyramidal Neuron as Two-Layer Neural Network.” Neuron 37 (6): 989–99. https://doi.org/10.1016/S0896-6273(03)00149-1.

• Pulikkottil, Vinu Varghese, Bhanu Priya Somashekar, and Upinder S. Bhalla. 2021. “Computation, Wiring, and Plasticity in Synaptic Clusters.” Current Opinion in Neurobiology, Computational Neuroscience, 70 (October):101–12. https://doi.org/10.1016/j.conb.2021.08.001.

Author response:

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

Reviewer #1 (Public Review):

In this revision, the authors significantly improved the manuscript. They now address some of my concerns. Specifically, they show the contribution of end-effects on spreading the inputs between dendrites. This analysis reveals greater applicability of their findings to cortical cells, with long, unbranching dendrites than other neuronal types, such as Purkinje cells in the cerebellum.

They now explain better the interactions between calcium and voltage signals, which I believe improve the take-away message of their manuscript. They modified and added new figures that helped to provide more information about their simulations.

However, some of my points remain valid. Figure 6 shows depolarization of ~5mV from -75. This weak depolarization would not effectively recruit nonlinear activation of NMDARs. In their paper, Branco and Hausser (2010) showed depolarizations of ~10-15mV.

More importantly, the signature of NMDAR activation is the prolonged plateau potential and activation at more depolarized resting membrane potentials (their Figure 4). Thus, despite including NMDARs in the simulation, the authors do not model functional recruitment of these channels. Their simulation is thus equivalent to AMPA only drive, which can indeed summate somewhat nonlinearly.

In the current study, we used short sequences of 5 inputs, since the convergence of longer sequences is extremely unlikely in the network configurations we have examined. This resulted in smaller EPSP amplitudes of ~5mV (Figure 6 - Supplement 2A, B). Longer sequences containing 9 inputs resulted in larger somatic depolarizations of ~10mV (Figure 6 - Supplement 2E, F). Although we had modified the (Branco, Clark, and Häusser 2010) model to remove the jitter in the timing of arrival of inputs and made slight modifications to the location of stimulus delivery on the dendrite, we saw similar amplitudes when we tested a 9-length sequence using (Branco, Clark, and Häusser 2010)’s published code (Figure 6 - Supplement 2I, J). In all the cases we tested (5 input sequence, 9 input sequence, 9 input sequence with (Branco, Clark, and Häusser 2010) code repository), removal of NMDA synapses lowered both the somatic EPSPs (Figure 6 - Supplement 2C,D,G,H,K,L) as well as the selectivity (measured as the difference between the EPSPs generated for inward and outward stimulus delivery) (Figure 6 Supplement 2M,N,O). Further, monitoring the voltage along the dendrite for a sequence of 5 inputs showed dendritic EPSPs in the range of 20-45 mV (Figure 6 - Supplement 2P, Q), which came down notably (10-25mV) when NMDA synapses were abolished (Figure 6 - Supplement 2R, S). Thus, even sequences containing as few as 5 inputs were capable of engaging the NMDA-mediated nonlinearity to show sequence selectivity, although the selectivity was not as strong as in the case of 9 inputs.

Reviewer #1 (Recommendations for the authors):

Minor points:

Figure 8, what does the scale in A represent? I assume it is voltage, but there are no units. Figure 8, C, E, G, these are unconventional units for synaptic weights, usually, these are given in nS / per input.

We have corrected these. The scalebar in 8A represents membrane potential in mV. The units of 8C,E,G are now in nS.

Reviewer #2 (Public Review):

Summary:

If synaptic input is functionally clustered on dendrites, nonlinear integration could increase the computational power of neural networks. But this requires the right synapses to be located in the right places. This paper aims to address the question of whether such synaptic arrangements could arise by chance (i.e. without special rules for axon guidance or structural plasticity), and could therefore be exploited even in randomly connected networks. This is important, particularly for the dendrites and biological computation communities, where there is a pressing need to integrate decades of work at the single-neuron level with contemporary ideas about network function.

Using an abstract model where ensembles of neurons project randomly to a postsynaptic population, back-of-envelope calculations are presented that predict the probability of finding clustered synapses and spatiotemporal sequences. Using data-constrained parameters, the authors conclude that clustering and sequences are indeed likely to occur by chance (for large enough ensembles), but require strong dendritic nonlinearities and low background noise to be useful.

Strengths:

(1) The back-of-envelope reasoning presented can provide fast and valuable intuition. The authors have also made the effort to connect the model parameters with measured values. Even an approximate understanding of cluster probability can direct theory and experiments towards promising directions, or away from lost causes.

(2) I found the general approach to be refreshingly transparent and objective. Assumptions are stated clearly about the model and statistics of different circuits. Along with some positive results, many of the computed cluster probabilities are vanishingly small, and noise is found to be quite detrimental in several cases. This is important to know, and I was happy to see the authors take a balanced look at conditions that help/hinder clustering, rather than to just focus on a particular regime that works.

(3) This paper is also a timely reminder that synaptic clusters and sequences can exist on multiple spatial and temporal scales. The authors present results pertaining to the standard `electrical' regime (~50-100 µm, <50 ms), as well as two modes of chemical signaling (~10 µm, 100-1000 ms). The senior author is indeed an authority on the latter, and the simulations in Figure 5, extending those from Bhalla (2017), are unique in this area. In my view, the role of chemical signaling in neural computation is understudied theoretically, but research will be increasingly important as experimental technologies continue to develop.

Weaknesses:

(1) The paper is mostly let down by the presentation. In the current form, some patience is needed to grasp the main questions and results, and it is hard to keep track of the many abbreviations and definitions. A paper like this can be impactful, but the writing needs to be crisp, and the logic of the derivation accessible to non-experts. See, for instance, Stepanyants, Hof & Chklovskii (2002) for a relevant example.

It would be good to see a restructure that communicates the main points clearly and concisely, perhaps leaving other observations to an optional appendix. For the interested but time-pressed reader, I recommend starting with the last paragraph of the introduction, working through the main derivation on page 7, and writing out the full expression with key parameters exposed. Next, look at Table 1 and Figure 2J to see where different circuits and mechanisms fit in this scheme. Beyond this, the sequence derivation on page 15 and biophysical simulations in Figures 5 and 6 are also highlights.

We appreciate the reviewers' suggestions. We have tightened the flow of the introduction. We understand that the abbreviations and definitions are challenging and have therefore provided intuitions and summaries of the equations discussed in the main text.

Clusters calculations

Our approach is to ask how likely it is that a given set of inputs lands on a short segment of dendrite, and then scale it up to all segments on the entire dendritic length of the cell.

Thus, the probability of occurrence of groups that receive connections from each of the M ensembles (PcFMG) is a function of the connection probability (p) between the two layers, the number of neurons in an ensemble (N), the relative zone-length with respect to the total dendritic arbor (Z/L) and the number of ensembles (M).

Sequence calculations

Here we estimate the likelihood of the first ensemble input arriving anywhere on the dendrite, and ask how likely it is that succeeding inputs of the sequence would arrive within a set spacing.

Thus, the probability of occurrence of sequences that receive sequential connections (PcPOSS) from each of the M ensembles is a function of the connection probability (p) between the two layers, the number of neurons in an ensemble (N), the relative window size with respect to the total dendritic arbor (Δ/L) and the number of ensembles (M).

(2) I wonder if the authors are being overly conservative at times. The result highlighted in the abstract is that 10/100000 postsynaptic neurons are expected to exhibit synaptic clustering. This seems like a very small number, especially if circuits are to rely on such a mechanism. However, this figure assumes the convergence of 3-5 distinct ensembles. Convergence of inputs from just 2 ense mbles would be much more prevalent, but still advantageous computationally. There has been excitement in the field about experiments showing the clustering of synapses encoding even a single feature.

We agree that short clusters of two inputs would be far more likely. We focused our analysis on clusters with three of more ensembles because of the following reasons:

(1) The signal to noise in these clusters was very poor as the likelihood of noise clusters is high.

(2) It is difficult to trigger nonlinearities with very few synaptic inputs.

(3) At the ensemble sizes we considered (100 for clusters, 1000 for sequences), clusters arising from just two ensembles would result in high probability of occurrence on all neurons in a network (~50% in cortex, see p_CMFG in figures below.). These dense neural representations make it difficult for downstream networks to decode (Foldiak 2003).

However, in the presence of ensembles containing fewer neurons or when the connection probability between the layers is low, short clusters can result in sparse representations (Figure 2 - Supplement 2). Arguments 1 and 2 hold for short sequences as well.

(3) The analysis supporting the claim that strong nonlinearities are needed for cluster/sequence detection is unconvincing. In the analysis, different synapse distributions on a single long dendrite are convolved with a sigmoid function and then the sum is taken to reflect the somatic response. In reality, dendritic nonlinearities influence the soma in a complex and dynamic manner. It may be that the abstract approach the authors use captures some of this, but it needs to be validated with simulations to be trusted (in line with previous work, e.g. Poirazi, Brannon & Mel, (2003)).

We agree that multiple factors might affect the influence of nonlinearities on the soma. The key goal of our study was to understand the role played by random connectivity in giving rise to clustered computation. Since simulating a wide range of connectivity and activity patterns in a detailed biophysical model was computationally expensive, we analyzed the exemplar detailed models for nonlinearity separately (Figures 5, 6, and new figure 8), and then used our abstract models as a proxy for understanding population dynamics. A complete analysis of the role played by morphology, channel kinetics and the effect of branching requires an in-depth study of its own, and some of these questions have already been tackled by (Poirazi, Brannon, and Mel 2003; Branco, Clark, and Häusser 2010; Bhalla 2017). However, in the revision, we have implemented a single model which incorporates the range of ion-channel, synaptic and biochemical signaling nonlinearities which we discuss in the paper (Figure 8, and Figure 8 Supplement 1, 2,3). We use this to demonstrate all three forms of sequence and grouped computation we use in the study, where the only difference is in the stimulus pattern and the separation of time-scales inherent in the stimuli.

(4) It is unclear whether some of the conclusions would hold in the presence of learning. In the signal-to-noise analysis, all synaptic strengths are assumed equal. But if synapses involved in salient clusters or sequences were potentiated, presumably detection would become easier? Similarly, if presynaptic tuning and/or timing were reorganized through learning, the conditions for synaptic arrangements to be useful could be relaxed. Answering these questions is beyond the scope of the study, but there is a caveat there nonetheless.

We agree with the reviewer. If synapses receiving connectivity from ensembles had stronger weights, this would make detection easier. Dendritic spikes arising from clustered inputs have been implicated in local cooperative plasticity (Golding, Staff, and Spruston 2002; Losonczy, Makara, and Magee 2008). Further, plasticity related proteins synthesized at a synapse undergoing L-LTP can diffuse to neighboring weakly co-active synapses, and thereby mediate cooperative plasticity (Harvey et al. 2008; Govindarajan, Kelleher, and Tonegawa 2006; Govindarajan et al. 2011). Thus if clusters of synapses were likely to be co-active, they could further engage these local plasticity mechanisms which could potentiate them while not potentiating synapses that are activated by background activity. This would depend on the activity correlation between synapses receiving ensemble inputs within a cluster vs those activated by background activity. We have mentioned some of these ideas in a published opinion paper (Pulikkottil, Somashekar, and Bhalla 2021). In the current study, we wanted to understand whether even in the absence of specialized connection rules, interesting computations could still emerge. Thus, we focused on asking whether clustered or sequential convergence could arise even in a purely randomly connected network, with the most basic set of assumptions. We agree that an analysis of how selectivity evolves with learning would be an interesting topic for further work.

References

Bhalla, Upinder S. 2017. “Synaptic Input Sequence Discrimination on Behavioral Timescales Mediated by Reaction-Diffusion Chemistry in Dendrites.” Edited by Frances K Skinner. eLife 6 (April):e25827. https://doi.org/10.7554/eLife.25827.

Branco, Tiago, Beverley A. Clark, and Michael Häusser. 2010. “Dendritic Discrimination of Temporal Input Sequences in Cortical Neurons.” Science (New York, N.Y.) 329 (5999): 1671–75. https://doi.org/10.1126/science.1189664.

Foldiak, Peter. 2003. “Sparse Coding in the Primate Cortex.” The Handbook of Brain Theory and Neural Networks. https://research-repository.st-andrews.ac.uk/bitstream/handle/10023/2994/FoldiakSparse HBTNN2e02.pdf?sequence=1.

Golding, Nace L., Nathan P. Staff, and Nelson Spruston. 2002. “Dendritic Spikes as a Mechanism for Cooperative Long-Term Potentiation.” Nature 418 (6895): 326–31. https://doi.org/10.1038/nature00854.

Govindarajan, Arvind, Inbal Israely, Shu-Ying Huang, and Susumu Tonegawa. 2011. “The Dendritic Branch Is the Preferred Integrative Unit for Protein Synthesis-Dependent LTP.” Neuron 69 (1): 132–46. https://doi.org/10.1016/j.neuron.2010.12.008.

Govindarajan, Arvind, Raymond J. Kelleher, and Susumu Tonegawa. 2006. “A Clustered Plasticity Model of Long-Term Memory Engrams.” Nature Reviews Neuroscience 7 (7): 575–83. https://doi.org/10.1038/nrn1937.

Harvey, Christopher D., Ryohei Yasuda, Haining Zhong, and Karel Svoboda. 2008. “The Spread of Ras Activity Triggered by Activation of a Single Dendritic Spine.” Science (New York, N.Y.) 321 (5885): 136–40. https://doi.org/10.1126/science.1159675.

Losonczy, Attila, Judit K. Makara, and Jeffrey C. Magee. 2008. “Compartmentalized Dendritic Plasticity and Input Feature Storage in Neurons.” Nature 452 (7186): 436–41. https://doi.org/10.1038/nature06725.

Poirazi, Panayiota, Terrence Brannon, and Bartlett W. Mel. 2003. “Pyramidal Neuron as Two-Layer Neural Network.” Neuron 37 (6): 989–99. https://doi.org/10.1016/S0896-6273(03)00149-1.

Pulikkottil, Vinu Varghese, Bhanu Priya Somashekar, and Upinder S. Bhalla. 2021. “Computation, Wiring, and Plasticity in Synaptic Clusters.” Current Opinion in Neurobiology, Computational Neuroscience, 70 (October):101–12. https://doi.org/10.1016/j.conb.2021.08.001.

Surely the first thing to do is to install the ocaml extension in vs code.

Thanks for putting your code on github! I noticed on your github page that you masked low confidence ends of protein structures prior to alignment. This is an interesting consideration and I think is worth mentioning in the methods here.

Note: it's okay to use some # pragma: no cover's to get make test coverage to pass if there are some parts of the code that are too awkward or not worthwhile to write tests for. Our approach to unittests is that we enforce 100% coverage but are pragmatic about using # pragma: no cover to get there if necessary.

Note: the code in this project was hacked together without a lot of care. You might find that you have to refactor it in order to make it unit-testable. You might find it helpful to write functests first (there's a separate issue for those: #7) as these will enable large-scale refactorings with confidence.

là on passe de la cognition distribuée au protocole... c'est pas la même chose... est-ce que tu critiques l'idée de H. et tu proposes une alternative? ou une manière de l'interpréter (ce qu'on dirait plus bas... tu dis: cogn distr -> nécessité du protocole. Mais est-ce vrai par ex pour des oiseaux en formation? Ont-ils un protocole??

I see from your code page that you used other stats like Kolmogorov-Smirnov Test and Levene’s Test, for example. I think, these should be described somewhere here in this section as well as assumptions and potential violations to assumptions.

Add versions and relevant references to tese packages, if available. Add references to reference list.

I assume that you will provide the full code as an appendix to your thesis? If so, please add link to repository here.

I see from your code page that you used other stats like Kolmogorov-Smirnov Test and Levene’s Test, for example. I think, these should be described somewhere here in this section as well as assumptions and potential violations to assumptions.

Add versions and relevant references to tese packages, if available. Add references to reference list.

I assume that you will provide the full code as an appendix to your thesis? If so, please add link to repository here.

The 'Crypto' Underwood Typewriter by [[Robert Messenger]]

This citation should be provided for the first reference to this code base.

benefit

I see from your code page that you used other stats like Kolmogorov-Smirnov Test and Levene’s Test, for example. I think, these should be described somewhere here in this section as well as assumptions and potential violations to assumptions.

Add versions and relevant references to tese packages, if available. Add references to reference list.

I assume that you will provide the full code as an appendix to your thesis? If so, please add link to repository here.

Reviewer #1 (Public review):

Summary:

Cesar, Santos & Cogni use a meta-analysis to report on the direction and magnitude of three fundamental fitness components in defensive symbioses. Specifically, the work focuses on interactions between three arthropod host families (Aphididae, Culicidae, Drosophilidae, and others) and common bacterial endosymbionts (Wolbachia, Serratia, Hamiltonella, Spiroplasma, Rickettsia, Regiella X-type and Arsenophonus). The results of the overall analysis confirm common assumptions and previous work on such fitness components, showing that defensive symbionts provide strong protection to hosts and cause detectable costs to both hosts and the enemy. The analysis provides insight into the extent of the cost/benefit tradeoff for hosts, reporting that the cost is six times lower than the protective effect. The confirmation that natural enemies attacking hosts infected with symbionts have a reduction in their fitness is also an interesting one, as this shows that the majority of defensive symbionts provide protection by resisting enemy infection, as opposed to tolerating it. This finding has important consequences for evolutionary counter-responses in the enemy species. Of course, this result has less relevance for certain types of enemies (such as parasitoids) where successful infection is dependent upon host killing.

Interesting results also emerge from the subgroup analysis. For the full dataset, both natural and introduced symbionts were similarly effective in positively influencing the fitness of hosts. However, in the Wolbachia-specific analysis, the artificially introduced symbionts caused costs to the hosts where the natural strain did not. These findings have potentially important ramifications for schemes that use endosymbionts for biocontrol or vector competence, suggesting that (in some cases) natural strains may be the more stable choice for deploying (as they are associated with lower costs).

The analysis draws from an impressively large dataset, but the interpretation of the full impact of the results would be helped by greater detail on the species/strain level systems included, the data extraction approach, and inclusion criteria. Accounting for phylogenetic nonindependence and alternative coding of one of the moderator variables could also strengthen the biological relevance of the models. Suggestions and thoughts are outlined below.

Strengths & Potential Improvements:

An impressively large number of effect sizes (3000) from only 226 studies is collected, robustly confirming common assumptions on the magnitude of fundamental fitness components. However the paper would benefit from a clear breakdown in the main text of the specificities of each system included (e.g. a table at the host species/symbiont strain level, where it is possible). Currently, there is not enough detail for those who want a deep dive to understand what data was extracted for the analysis from these 226 studies, or those who want to understand the underlying diversity in the dataset.

Currently, when the 'natural enemy group' is tested as a moderator it is coded broadly by type of organism (e.g. virus, bacterium, fungi, parasitoid). But this doesn't adequately capture the mode of killing/fitness reduction by the enemy, which would be the much more biologically relevant categorisation for your questions. For example, parasitoid infection is dependent upon host death (thus host fecundity is not relevant, because the host either survived or did not). Among bacterial and viral pathogens antagonists there is scope for both fecundity and survival to be affected. This in turn may be a very influential factor for the outcome. You could consider recoding this enemy moderator.

The analysis is restricted to arthropod hosts and defensive symbionts that are also classed as endosymbionts. This focus should be made clear early on in the paper, as there are many systems (that are classed by many as defensive symbioses) that are not part of the analysis.

There is fairly minimalistic testing of moderators/sub-groups (which probably has its statistical strengths) but perhaps there are also some missed opportunities for testing other ecological contributors to variance, including coinfection (although perhaps limited by power) and other approaches to coding enemy group (as detail above).

Looking at the overview of systems included, there's likely a high degree of phylogenetic non-independence in the dataset. Where it is possible, using phylogenetically controlled models could strengthen this analysis.

Looking at your included systems (Table S5), you might be able to test the effect of coinfection on the 3 variables of interest. For example, it would be particularly important to see if the effects of two symbionts are additive or not.

No code for the analysis is provided for review at this stage and full details of the dataset are also not available. This slightly limits the ability to assess the full scope and robustness of the study. It would be helpful to have an extensive table in the supplementary detailing (minimum) the reference, study, experiment, host species, symbiont strain, and a description of the exact data extraction source (e.g.table/figure/in text), and method of extraction.

Author response:

Reviewer #1 (Public review):

Summary:

Cesar, Santos & Cogni use a meta-analysis to report on the direction and magnitude of three fundamental fitness components in defensive symbioses. Specifically, the work focuses on interactions between three arthropod host families (Aphididae, Culicidae, Drosophilidae, and others) and common bacterial endosymbionts (Wolbachia, Serratia, Hamiltonella, Spiroplasma, Rickettsia, Regiella X-type and Arsenophonus). The results of the overall analysis confirm common assumptions and previous work on such fitness components, showing that defensive symbionts provide strong protection to hosts and cause detectable costs to both hosts and the enemy. The analysis provides insight into the extent of the cost/benefit tradeoff for hosts, reporting that the cost is six times lower than the protective effect. The confirmation that natural enemies attacking hosts infected with symbionts have a reduction in their fitness is also an interesting one, as this shows that the majority of defensive symbionts provide protection by resisting enemy infection, as opposed to tolerating it. This finding has important consequences for evolutionary counter-responses in the enemy species. Of course, this result has less relevance for certain types of enemies (such as parasitoids) where successful infection is dependent upon host killing.

Interesting results also emerge from the subgroup analysis. For the full dataset, both natural and introduced symbionts were similarly effective in positively influencing the fitness of hosts. However, in the Wolbachia-specific analysis, the artificially introduced symbionts caused costs to the hosts where the natural strain did not. These findings have potentially important ramifications for schemes that use endosymbionts for biocontrol or vector competence, suggesting that (in some cases) natural strains may be the more stable choice for deploying (as they are associated with lower costs).

The analysis draws from an impressively large dataset, but the interpretation of the full impact of the results would be helped by greater detail on the species/strain level systems included, the data extraction approach, and inclusion criteria. Accounting for phylogenetic nonindependence and alternative coding of one of the moderator variables could also strengthen the biological relevance of the models. Suggestions and thoughts are outlined below.

We sincerely thank Reviewer #1 for the time and effort dedicated to reviewing our manuscript. The suggestions provided are highly constructive and will greatly assist us in improving both our analyses and the manuscript overall.

Strengths & Potential Improvements:

An impressively large number of effect sizes (3000) from only 226 studies is collected, robustly confirming common assumptions on the magnitude of fundamental fitness components. However the paper would benefit from a clear breakdown in the main text of the specificities of each system included (e.g. a table at the host species/symbiont strain level, where it is possible). Currently, there is not enough detail for those who want a deep dive to understand what data was extracted for the analysis from these 226 studies, or those who want to understand the underlying diversity in the dataset.

We thank the reviewer for the suggestion, and we will add this information to our revised manuscript.

Currently, when the 'natural enemy group' is tested as a moderator it is coded broadly by type of organism (e.g. virus, bacterium, fungi, parasitoid). But this doesn't adequately capture the mode of killing/fitness reduction by the enemy, which would be the much more biologically relevant categorisation for your questions. For example, parasitoid infection is dependent upon host death (thus host fecundity is not relevant, because the host either survived or did not). Among bacterial and viral pathogens antagonists there is scope for both fecundity and survival to be affected. This in turn may be a very influential factor for the outcome. You could consider recoding this enemy moderator.

We agree, and we will implement this in the analysis to our revised manuscript.

The analysis is restricted to arthropod hosts and defensive symbionts that are also classed as endosymbionts. This focus should be made clear early on in the paper, as there are many systems (that are classed by many as defensive symbioses) that are not part of the analysis.

We agree, and we will implement this to our revised manuscript.

There is fairly minimalistic testing of moderators/sub-groups (which probably has its statistical strengths) but perhaps there are also some missed opportunities for testing other ecological contributors to variance, including coinfection (although perhaps limited by power) and other approaches to coding enemy group (as detail above).

We agree, and we will implement this in the analysis to our revised manuscript.

Looking at the overview of systems included, there's likely a high degree of phylogenetic non-independence in the dataset. Where it is possible, using phylogenetically controlled models could strengthen this analysis.

We thank the reviewer for the suggestion. We will explore the possibility of using phylogenetically controlled models in our analyses, although we recognize the challenges associated with their implementation, particularly in the case of the natural enemies, given the great diversity of distant related groups included in our study - viruses, bacteria, fungi, protozoans, nematodes and parasitoids wasps.

Looking at your included systems (Table S5), you might be able to test the effect of coinfection on the 3 variables of interest. For example, it would be particularly important to see if the effects of two symbionts are additive or not.

We agree, and we will implement this in the analysis to our revised manuscript.

No code for the analysis is provided for review at this stage and full details of the dataset are also not available. This slightly limits the ability to assess the full scope and robustness of the study. It would be helpful to have an extensive table in the supplementary detailing (minimum) the reference, study, experiment, host species, symbiont strain, and a description of the exact data extraction source (e.g.table/figure/in text), and method of extraction.

The code for the analysis and the full raw data with the suggested information are available at https://github.com/cassiasqr/MetaSymbiont (The link is available at the end of the manuscript).

Reviewer #2 (Public review):

Summary:

In this exciting study, Cesar and co-authors perform a meta-analysis on the influence of arthropod symbionts on the fitness of their hosts when they are exposed or not to natural enemies. These so-called defensive symbionts are increasingly recognized as key elements in arthropod survival against natural enemies, with effects that ripple through entire terrestrial ecosystems. The topic is timely, the approach is sound, and the manuscript is well-written. I believe this manuscript will attract the attention of entomologists and of microbiologists interested in symbiosis. This study builds on a previous meta-analysis that I was involved in, which was based on phloem-feeding insects. This novel data set is much larger and includes flies (including the model system Drosophila) and mosquitoes (a group of high medical interest). While the previous metaanalysis considered only parasitoids as natural enemies, this study also includes fungi, bacteria, and viruses.

Strengths:

The authors compile a very large dataset and provide a broad quantitative overview of the effects of defensive symbionts in insects. By measuring symbiont effects in the presence and absence of natural enemies, the authors are able to infer whether a trade-off between defense and the costs of mutualism in the absence of enemy pressure exists. Defensive symbioses are an important research topic that had its initial "momentum" a decade ago, so the timing for such a systematic review is very appropriate.

We sincerely thank Reviewer #2 for dedicating their time and effort to reviewing our manuscript. The suggestions are very insightful and will significantly contribute to improving our manuscript.

Weaknesses:

I think the manuscript could be improved by clarifying several sections, particularly the introduction and methods. The introduction section is too specific and heavily reliant on particular examples. In my view, the theoretical background of the study could be made clearer, and the knowledge gap identified more explicitly. A focus on how widespread defensive symbioses are, along with a brief, up-to-date review of the groups possessing such symbionts, would help. This lack of focus is also observed in the methods section, where more details are needed in many instances to better understand how data was collected and analyzed. Regarding the analyses, the multi-level analysis contains many moderators, but it's unclear why these moderators were included. While this may seem a minor issue, it highlights a disconnection between the analyses, the conceptual background, and the hypotheses tested. 

We thank the reviewer for the suggestions, and we will try to make the introduction and the methods section clearer. 

Another important weakness is that the analyses are too general, and much-hidden information is not immediately apparent. For instance, readers cannot easily identify which species of symbionts are studied (and the effects they have), or which natural enemies are involved. Although this information is found in the supplementary material, including it in the main body would significantly improve the manuscript.

We agree, and we will implement this to our   revised manuscript.

assuming this happens in order for the individuals behind the process of collecting data do this in order to see where the common affects that the crisis caused?

In my experience this is also the main benefit of using node.js as your backend. Being able to write your front and backend code in the same language (javascript) removes a switching cost I didn't fully realize existed until I tried node the first time.

I see from your code page that you used other stats like Kolmogorov-Smirnov Test and Levene’s Test, for example. I think, these should be described somewhere here in this section as well as assumptions and potential violations to assumptions.

Add versions and relevant references to tese packages, if available. Add references to reference list.

I assume that you will provide the full code as an appendix to your thesis? If so, please add link to repository here.

measures delayed recall for visual stimuli

tests visual search, sustained attention and encoding

Reviewer #1 (Public review):

Summary:

This paper presents a data processing pipeline to discover causal interactions from time-lapse imaging data and convincingly illustrates it on a challenging application for the analysis of tumor-on-chip ecosystem data.

The core of the discovery module is the original tMIIC method of the authors, which is shown in supplementary material to compare favourably to two state-of-the-art methods on synthetic temporal data on a 15 nodes network.

Strengths:

This paper tackles the problem of learning causal interactions from temporal data which is an open problem in presence of latent variables.

The core of the method tMIIC of the authors is nicely presented in connection to Granger-Schreiber causality and to the novel graphical conditions used to infer latent variables and based on a theorem about transfer entropy.

tMIIC compares favourably to PC and PCMCI+ methods using different kernels on synthetic datasets generated from a network of 15 nodes.

A full application to tumor-on-chip cellular ecosystems data including cancer cells, immune cells, cancer-associated fibroblasts, endothelial cells and anti cancer drugs, with convincing inference results with respect to both known and novel effects between those components and their contact.

The code and dataset are available online for the reproducibility of the results.

Weaknesses:

The references to "state-of-the-art methods" concerning the inference of causal networks should be more precise by giving citations in the main text, and better discussed in general terms, both in the first section and in the section of presentation of CausalXtract. It is only in the legend of the figures of the supplementary material that we get information.

Of course, comparison on our own synthetic datasets can always be criticized but this is rather due to the absence of a common benchmark in this domain compared to other domains. I recommend the authors to explicitly propose their datasets made accessible in supplementary material as benchmark for the community.

Comments on revisions:

This is a very nice paper.

Reviewer #2 (Public review):

Summary:

The authors propose a methodology to perform causal (temporal) discovery. The approach appears to be robust and is tested in the different scenarios: one related to live-cell imaging data, and another one using synthetic (mathematically defined) time series data. They compare the performance of their findings against another well-known method by using metrics like F-score, precision and recall,

Strengths:

--Performance, robustness, the text is clear and concise, The authors provide the code to review.

Comments on revisions:

The authors have addressed my concerns properly providing the needed explanations.

Author response:

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

Reviewer #1 (Public review):

Summary:

This paper presents a data processing pipeline to discover causal interactions from time-lapse imaging data, and convicingly illustrates it on a challenging application for the analysis of tumor-on-chip ecosystem data. The core of the discovery module is the original tMIIC method of the authors, which is shown in supplementary material to compare favourably to two state-of-the-art methods on synthetic temporal data on a 15 nodes network.

Strengths:

This paper tackles the problem of learning causal interactions from temporal data which is an open problem in presence of latent variables. The core of the method tMIIC of the authors is nicely presented in connection to Granger- Schreiber causality and to the novel graphical conditions used to infer latent variables and based on a theorem about transfer entropy. tMIIC compares favourably to PC and PCMCI+ methods using different kernels on synthetic datasets generated from a network of 15 nodes. A full application to tumor-onchip cellular ecosystems data including cancer cells, immune cells, cancer-associated fibroblasts, endothelial cells and anti cancer drugs, with convincing inference results with respect to both known and novel effects between those components and their contact.

The code and dataset are available online for the reproducibility of the results.

We thank Reviewer #1 for highlighting the main results and strengths of our paper, as well as, for his/her recommendations below to further improve the manuscript.

Weaknesses:

The references to ”state-of-the-art methods” concerning the inference of causal networks should be more precise by giving citations in the main text, and better discussed in general terms, both in the first section and in the section of presentation of CausalXtract. It is only in the legend of the figures of the supplementary material that we get information. Of course, comparison on our own synthetic datasets can always be criticized but this is rather due to the absence of common benchmark and I would recommend the authors to explicitly propose their datasets as benchmark to the community.

Following Reviewer #1’s suggestion, we now compare tMIIC’s performance to other state-of-the-art causal discovery methods for time series data in the main text and in a new Figure 2. This Figure 2 also highlights the relation between graph-based causal discovery methods for time series data and Granger-Schreiber temporal causality, as discussed in more details in Methods (Theorem 1).

We also agree about the importance of sharing benchmark datasets with the community. This is the reason why we provide the dynamical equations of the 15-node benchmarks in Supplementary Tables 1 & 2, so that anyone can generate equivalent time series datasets of any desired length.

Reviewer #2 (Public review):

Summary:

The authors propose a methodology to perform causal (temporal) discovery. The approach appears to be robust and is tested in the different scenarios: one related with live-cell imaging data, and another one using synthetic (mathematically defined) time series data. They compare the performance of their findings against another well-know method by using metrics like F-score, precision and recall,

Strengths:

Performance, robustness, the text is clear and concise, The authors provide the code to review.

We thank Reviewer #2 for his/her positive assessment of our work and the suggestions below to improve the manuscript.

Weaknesses:

One concern could be the applicability of the method in other areas like climate, economy. For those areas, public data are available and might be interesting to test how the method performs with this kind of data.

While our main expertise concerns the analysis of biological and biomedical data, we agree that tMIIC (which is included in MIIC R package) could in principle be applied to other areas, like climate, economy.

We have not included benchmarks on such diverse types of datasets in the present manuscript, which focuses on CausalXtract’s pipeline for the analysis and causal interpretation of live-cell time-lapse imaging data from complex cellular systems.

Reviewer #3 (Public review):

This work brings a computational approach to the study of promoters and transcription. The paper is improved but there are still factual errors and implausible explanations. I am not convinced by the response from the authors, concerning the promoter -35 element, in their rebuttal.

Comments on author rebuttal:

- We respectfully but strongly disagree that our analysis has misrepresented the true nature of -35 boxes. First, accounting for more A's at position 5 in the PWM is not going to lead to a "critical error." This is because positions 4-6 of the motif barely have any information content (bits) compared to positions 1-3 (see Fig 1A).

The analysis does misrepresent the consensus -35 element, which is, unequivocally, TTGACA. I agree that positions 4-6 of the element are less well-conserved.

- This assertion is not just based on our own PWM, but based on ample precedent in the literature. In PMID 14529615, TTG is present in 38% of all -35 boxes, but ACA only in 8%.

This does not mean that TTGACA is not the consensus, or that "ACA" is not important at promoters where it's present.

- In PMID 29388765, with the -10 instance TATAAT, the -35 instance TTGCAA yields stronger promoters compared to the -35 instance TTGACA (See their Figure 3B).

This is a known phenomenon and results from "perfect" promoters being limited at the point of RNA polymerase promoter escape (because the RNAP struggles to "let go" of perfect promoters). This does not mean the TTGACA is not the consensus. Indeed, and this is a key point, it is evident in the figure the authors refer to that TTGACA stimulates more transcription than alternative -35 sequences when -10 elements are not perfect.

- In PMID 29745856 (Figure 2), the most information content lies in positions 1-3, with the A and C at position 5 both nearly equally represented, as in our PWM.

The motif shown in this paper suffers from exactly the same issue as the paper under review; the variable spacing between the -35 hexamer and -10 element isn't taken into account by MEME.

- In PMID 33958766 (Figure 1) an experimentally-derived -35 box is even reduced to a "partial" -35 box which only includes positions 1 and 2, with consensus: TTnnnn.

This paper does not show an "experimentally-derived -35 box" in Figure 1 (or anywhere else, as far as I can see).

- In addition, we did not derive the PWMs as the reviewer describes. The PWMs we use are based on computational predictions that are in excellent agreement with experimental results. Specifically, the PWMs we use are from PMID 29728462, which acquired 145 -10 and -35 box sequences from the top 3.3% of computationally predicted boxes from Regulon DB.

The paper mentioned states "for the genomic RNAP logo, sequences were taken from computationally predicted RNAP binding sites on RegulonDB" so these are not experimentally defined promoters? It's not obvious from the paper, or regulon DB, which sequences these are or how they were predicted.

- Thank you for pointing out that our original submission was incomplete in this regard. We address these concerns by new analyses, including some new experiments. First, Rho dependent termination is associated with the RUT motif, which is very rich in Cytosines (PMID: 30845912). Given that our sequences confer between 65%-78% of AT-content, canonical rho dependent termination is unlikely. However, we computationally searched for rho-dependent terminators using the available code from PMID: 30845912, but the algorithm did not identify any putative RUTs. Because this analysis was not informative, we did not include it in the paper.

I don't believe it is the case that Rho absolutely requires a RUT sequence. My understanding is that, if an RNA is not translated, Rho will intervene (e.g. see PMID: 18487194).

- We respectfully disagree that the reviewer's point is pertinent because what the reviewer is referring to is the likelihood that the sequence is a promoter, which indeed increases with AT content, but we are focused on the likelihood that a sequence becomes a promoter through DNA mutation

I disagree that this distinction is relevant. An AT-rich sequence will much more closely resemble a promoter by chance than a GC rich sequence. As an extreme example, the sequence TTTTTT can be converted into a reasonable -10 element by one change (to TATTTT) but the sequence GGGGGG can't.

This explanation underscores the importance of using functions to promote efficiency and maintainability in programming. By reusing code, we reduce redundancy, make our programs easier to debug, and set a foundation for more scalable software development practices.

You can tell how capitalism was inspired or infnluenced by colonialism which was older. This is shown in how one group or person has the msot pwoer and sets the rules and laws in a place. In colonialism, the country imposes laws, culture, and language on the group and in capitalism, the same can be said in how there might be a language policy or code of conduct in the company that everyone has to follow

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

Summary:

The work from Petazzi et al. aimed at identifying novel factors supporting the differentiation of human hematopoietic progenitors from induced pluripotent stem cells (iPSCs). The authors developed an inducible CRISPR-mediated activation strategy (iCRISPRa) to test the impact of newly identified candidate factors on the generation of hematopoietic progenitors in vitro. They first compared previously published transcriptomic data of iPSCderived hemato-endothelial populations with cells isolated ex vivo from the aorta-gonadmesonephros (AGM) region of the human embryo and they identified 9 transcription factors expressed in the aortic hemogenic endothelium that were poorly expressed in the in vitro differentiated cells. They then tested the activation of these candidate factors in an iPSCbased culture system supporting the differentiation of hematopoietic progenitors in vitro. They found that the IGF binding protein 2 (IGFBP2) was the most upregulated gene in arterial endothelium after activation and they demonstrated that IGFBP2 promotes the generation of functional hematopoietic progenitors in vitro.

Strengths:

The authors developed an extremely useful doxycycline-inducible system to activate the expression of specific candidate genes in human iPSC. This approach allows us to simultaneously test the impact of 9 different transcription factors on in vitro differentiation of hematopoietic cells, and the system appears to be very versatile and applicable to a broad variety of studies.

The system was extensively validated for the expression of 1 transcription factor (RUNX1) in both HeLa cells and human iPSC, and a detailed characterization of this test experiment was provided.

The authors exhaustively demonstrated the role of IGFBP2 in promoting the generation of functional hematopoietic progenitors in vitro from iPSCs. Even though the use of IGFBP2interacting proteins IGF1 and IGF2 have been previously reported in human iPSC-derived hematopoietic differentiation in vitro (Ditadi and Sturgeon, Methods 2016; Ng et al., Nature Biotechnology 2016), and IGFBP-2 itself has been shown to promote adult HSC expansion ex vivo (Zhang et al., Blood 2008), its role on supporting in vitro hematopoiesis was demonstrated here for the first time.

Weaknesses:

Although the authors performed a very thorough characterization of the system in proof-ofprinciple experiments activating a single transcription factor, the data provided when 9 independent factors were used is not sufficient to fully validate the experimental strategy. Indeed, in the current version of the manuscript, it is not clear whether the results presented in both the scRNAseq analysis and the functional assays are the consequence of the simultaneous activation of all 9 TF or just a subset of them. This is essential to establish whether all the proposed factors play a role during embryonic hematopoiesis, and a more complete analysis of the scRNAseq dataset could help clarify this aspect.

Similarly, the data presented in the manuscript are not sufficient to clarify at what stage of the endothelial-to-hematopoietic transition (EHT) the TF activation has an impact. Indeed, even though the overall increase of functional hematopoietic progenitors is fully demonstrated, the assays proposed in the manuscript do not clarify whether this is due to a specific effect at the endothelial level or to an increased proliferation rate of the generated hematopoietic progenitors. Similar conclusions can be applied to the functional validation of IGFBP2 in vitro.

The overall conclusions are sometimes vague and not always supported by the data. For instance, the authors state that the CRISPR activation strategy resulted in transcriptional remodeling and a steer in cell identity, but they do not specify which cell types are involved and at what level of the EHT process this is happening. In the discussion, the authors also claim that they provided evidence to support that RUNX1T1 could regulate IGFBP2 expression. However, this is exclusively based on the enrichment of RUNX1T1 gRNA in cells expressing higher levels of IGFBP2 and it does not demonstrate any direct or indirect association of the two factors.

We thank the reviewer for the positive comments about the importance of our work and have now addressed the points raised as weaknesses by performing additional analysis and experiments, adding a new schematic of the mechanism, and rewording our claims.

We have clarified the different effects mediated by the activation and the IGFBP2 addition in a summary section at the end of the results and added Figure 6, showing this in visual form. We have also clearly stated the limitations related to the correlation between RUNX1T1 and IGFBP2 in the discussion and toned down our claims regarding this throughout the entire paper. We have also reworded the text to clarify the specific cell types identified in the sequencing data that we refer to.

Reviewer #2 (Public Review):

To enable robust production of hematopoietic progenitors in-vitro, Petazzi et al examined the role of transcription factors in the arterial hemogenic endothelium. They use IGFBP2 as a candidate gene to increase the directed differentiation of iPSCs into hematopoietic progenitors. They have established a novel induced-CRISPR mediated activation strategy to drive the expression of multiple endogenous transcription factors and show enhanced production of hematopoietic progenitors through expansion of the arterial endothelial cells. Further, upregulation of IGFBP2 in the arterial cells facilitates the metabolic switch from glycolysis to oxidative phosphorylation, inducing hematopoietic differentiation. While the overall study and resources generated are good, assertions in the manuscript are not entirely supported by the experimental data and some claims need further experimental validation.

We thank the reviewer for the positive comments, and we have provided new data and analysis to make sure that all our assertations are clearly supported and also reworded those where limitations were identified by the reviewers.

Recommendations for the authors:

Reviewing Editor (Recommendations For The Authors):

The assessment could change from "incomplete" to "solid" if the authors: i) improve data analysis (for both scRNAseq and functional assays) by providing additional information that could strengthen their conclusions, as suggested in the specific comments by both reviewers; ii) either provide new functional evidence supporting their mechanistic conclusion or alternatively tone down the claims that are not fully supported by data and acknowledge the limitations raised by reviewers in the discussion; (iii) the issue of paracrine signaling to expand only hematopoietic progenitors needs to be addressed.

We have now improved the data analysis and provided additional functional tests to strengthen our conclusions and toned down those that were identified by the reviewers as not supported enough and included a discussion on these limitations. We have also reworded the section about the paracrine signaling throughout the paper.

Reviewer #1 (Recommendations For The Authors):

Figure 1 contains exclusively published data. It might be more appropriate to use it as a supplementary figure or as part of a more exhaustive figure (maybe combining Figures 1 and 2 together?).

Figure 1 contained novel bioinformatic analyses that represent the base of our research and it has a different content and focus to figure 2, which is already a large figure. We therefore believe it is better to keep it as a separate figure, containing a new panel now too. 

It seems there is an issue with Figure S3 labelling:

• In line 112, Figure S2A-B does not display genomic PCR and sequencing results;

• In line 123, Figure S3D-E does not show viability and proliferation data;

• In line 127, Figure S3G does not show mCherry expression in response to DOX;

We apologies for the confusion with the numbers, we have now correctly labelled the figures.

It would be more informative to include gates and frequency on flow cytometry plots in Figure S3, to be able to evaluate the extent of the reduction in mCherry expression.

We have now included the gating and frequency of mCherry-expressing cells in Supplementary Figure 3D.

It is not clear from the text and figures whether the SB treatment was maintained throughout the hematopoietic differentiation protocol (line 122):

• If so, it would be important to confirm that HDAC treatment does not affect EHT cultures

• If not, can the authors provide some evidence that transgene silencing is not occurring during hematopoietic differentiation?

We have clarified that we decided to treat the cells with SB exclusively in maintenance condihons because HDACs have been shown to be essenhal for the EHT (lines 138-142). We have now also included addihonal data showing the high expression of the mCherry tag reporhng the iSAM expression on day 8 (Supplementary Figure 4F).

Can the authors provide a simple diagram summarizing the experimental strategy for each differentiation experiment in the respective supplementary figure? For instance, at what stage of the protocol was DOX added in Figure 3? Or at what stage IGFBP2 was added in Figure 5? It would be a very useful addition to the interpretation of the results.

We have now included three schemahcs for all the experiments in the manuscript in supplementary figure 4 A-C.

In Figure 3, the authors should provide more detailed information about the data filtering of the scRNAseq experiment, and more specifically:

• How many cells were included in the analysis for each library after QC and filtering?

• How "cells in which the gRNAs expression was detected" were selected? Do they include only cells showing expression of gRNAs for all 9 TF?

This informahon is now included in the method sechon lines 773-781; the detailed code is available on the GitHub link provided in the same sechon. We have filtered the cells expressing one gRNA for the non-targehng gRNA (iSAM_NT) control and more than one for the iSAM_AGM sample. 

In Figure 3A, it is not clear whether the expression of the 9 factors is consistently detected in all cells or just a subset of them, and the heatmap in Figure 3A does not provide this information. It would be more accurate to provide expression on a per-cell basis, for instance, as a violin plot displaying single dots representing each cell. 

We have now included this violin plot in Supplementary Figure 4G as requested. However, this visualisation is difficult to interpret because some of the target genes’ expression seems variable in both experimental and control conditions. We had envisaged that this could have been the case and so this is why we had included the three different controls.  For this reason we chose to show the normalised expression which takes all the different variables into account (Figure 3A). 

In Figure 3B-C, it seems that clusters EHT1 and EHT2 do not express endothelial markers anymore. Are these fully differentiated hematopoietic cells rather than cells undergoing EHT? In general, it would be quite important to provide evidence of expressed marker genes characterizing each cluster (eg. heatmap summarizing top DEG in the supplementary figure?). 

We have now provided a spreadsheet containing the clusters’ markers that we used in

Supplementary Table 1) a heatmap in Figure 3E. Furthermor,e we have now edited Figure 3C to include Pan Endothelial markers (PECAM1 and CDH5). These data show that the EHT1 and EHT2 cluster both express endothelial markers but are progressively downregulated as expected during endothelial to hematopoietic transition. We have also included and discussed this in the manuscript lines 192-195 and a schematic for the mechanism in Figure 6.

In Figure 3E, displaying the proportion of clusters within each sample/library would be a more accurate way of comparing the cell types present in each library (removing potential bias introduced by loading different numbers of cells in each sample).

We have now included the requested data in Supplementary Figure 4I and it confirms again the expansion of arterial cells in the activated cells.    

In Figure 3G, by plating 20,000 total CD34+, the assay does not account for potential differences in sample composition. It is then hard to discriminate between the increased number of progenitors in the input or an enhanced ability of HE to undergo EHT. This is an important aspect to consider to precisely identify at what level the activation of the 9 factors is acting. A proper quantification of flow cytometry data summarizing the % of progenitors, arterial cells, etc. would be useful to interpret these results.

Lines 204-205 reworded. We are very much aware of the fact that the CD34+ cell population consists of a range of cells across the EHT process and this is precisely why we carried out this single cell sequencing analyses.  We purposely tested the effect of the observed changes in composition by colony assays

In Figure 3G, it seems that NT cells w/o DOX have very little CFU potential (if any). Can the authors provide an explanation for this?

We think that the limited CFU potential is due to the extensive genetic manipulation and selection that the cells underwent for the derivation of all the iSAM lines but this did not impede us from observing an effect of gene activation on CFU numbers. This is one of the primary reasons that we then validated our overall findings using the parental iPSC line in control condition and with the addition of IGFBP2. We show that the parental iPSC line gives rise to hematopoietic progenitor, both immunophenotypically (Figure 4D) and functionally, at expected levels (Figure 4B left column).

Figure 4A shows an upregulation of IGFBP2 in arterial cells as a result of TF activation. However, from the data presented here, it is not possible to evaluate whether this is specific to the arterial cluster, or it is a common effect shared by all cell types regardless of their identity. 

Data has now been included in Supplementary Figure 4H, which shows that all the cells show an increase in IGFBP2, but arterial cells show the highest increase. We have now edited the text to reflect this, in lines 228-230.

In Figure 5A-B only a minority of arterial cells express RUNX1 in response to IGFBP2 treatment. Is this sufficient to explain the very significant increase in the generation of functional hematopoietic progenitors described in Figure 4? Quantification and statistical analysis of RUNX1 upregulation would strengthen this conclusion.

We have now provided the statistical analysis showing significant upregulation of RUNX1 upon IGFBP2 addition. The p values are now provided in the figure 5 legend.

In Figure 5 the authors conclude that IGFBP2 remodels the metabolic profile of endothelial cells. However, it is not clear which cell types and clusters were included in the analysis of Figure 5C-G. Is the switch from Glycolysis to Oxidative Phosphorylation specific to endothelial cells? Or it is a more general effect on the entire culture, including hematopoietic cells? 

We based this conclusion on the fact that the single-cell RNAseq allows to verify that the metabolic differences are obtained in the endothelial cells. Given that we sorted the adherent cells, the majority of these are endothelial cells as shown in Figure 5A. The Seahorse pipeline includes a number of washing steps resulting in the analyses being performed on the adherent compartment which we know consists primarily of endothelial cells. We cannot exclude some contamination from non-endothelial cells but we highlight to this reviewer that the initial observation of the metabolic changes was identified in endothelial cells in the single cell sequencing data. Taken together, we believe that this implies that metabolic changes are specific to this population. We have clarified this in the line 317.

In the discussion, the authors conclude that they "provide evidence to support the hypothesis that RUNX1T1 could regulate IGFBP2 expression". To further support this conclusion, the authors could provide a correlation analysis of the expression of the two genes in the cell type of interest. 

Following the observation of the IGFBP2 high expression across clusters, we have now reworded this sentence in lines 382-385  We have tried to perform the correlation analysis but we believe this not to be appropriate due to the detection level of the gRNA, we have now included this as a limitation point in the discussion lines 416-427, and also toned down the conclusion we did draw about RUNX1T1 throughout the whole manuscript.

As mentioned by the authors, IGFBP2 binds IGF1 and IGF2 modulating their function. Both IGF1 (http://dx.doi.org/10.1016/j.ymeth.2015.10.001) and IGF2 (doi:10.1038/nbt.3702) have been used in iPSC differentiation into definitive hematopoietic cells. It would be relevant to discuss/reference this in the discussion.

We have now included the suggested reference in the section where we discuss the role of IGFBP2 in binding IGF1 and IGF2.

Reviewer #2 (Recommendations For The Authors):

(1) Figure 1 compares the transcriptome of human AGM and in-vitro derived hemogenic endothelial cells (HECs). It is not clear why only the genes downregulated in the latter were chosen. Are there any significantly upregulated genes, knockdown/knockout which could also serve a similar purpose? Single-cell transcriptome database analysis is very preliminary. A detailed panel with differences in cluster properties of HECs between the two systems should be provided. A heatmap of all differentially expressed genes between the two samples must be generated, along with a logical explanation for choosing the given set of genes. 

We have now included another panel in figure 1 to better clarify the logic behind the strategy used to identify our target genes (Figure 1A).

(2) Figure 2 - a panel describing the workflow of gRNA design and targeting for the 9 candidate genes, along with lentiviral packaging and transduction would make it easier to follow. 

We have now included three schematics for all the experiments in the manuscript in supplementary figure 4 A-C. 

(3) Figure 3- to assess the effect of arterial cell expansion on the emergence of hematopoietic progenitors, CD34+ Dll4+ cells should be sorted for OP9 co-culture assay.

Using only CD34+ cells does not answer the question raised. Also, the CFU assay performed does not fully support the claim of enhanced hematopoietic differentiation since only CFU-E and CFU-GM colonies are increased in Dox-treated samples, with no effect on other colony types. OP9 co-culture assay with these cells would be required to strengthen this claim. 

We wanted to clarify that the effect on the methylcellulose coming from the activated cells was not limited to CFU-E, as the reviewer reported; instead, it also affected CFU-GM and CFU-M. 

We have now performed additional experiments where we sorted the CD34+ compartment into DLL4- and DLL4+ in Supplementary Figure 5D-E, which we discussed in lines 250-258. 

(4) In Figure 3F, there appears to be a lot of variation in the DLL4% fold change values for

DOX treated iSAM_AGM sample, which weakens the claim of increased arterial expansion.

Can the authors explain the probable reason? It is suggested that the two other controls (iSAM_+DOX and iSAM_-DOX) should be included in this analysis. It is imperative to also show % populations rather than just fold change to gain confidence.

We agree that there is a lot of variability. That is because differentiation happens in 3D in embryoid bodies, which contain many different cell types that differentiate in different proportions across independent experiments. We have now included the raw data in Supplementary Figure 4 D, with additional statistical analysis to show the expansion of arterial cells including also the suggested additional controls.

(5) How does activation of these target genes cause increased arterialization? Is the emergence of non-HE populations suppressed? Or is it specific to the HE? The data on this should be clarified and also discussed. ANTO/Lesley text

We have provided additional data clarifying the connection between increased arterialisation and hemogenic potential. We showed that the activation induces increased arterialisation and that IGFBP2 acts by supporting the acquisition of hemogenic potential. We have discussed this in lines 326-348 and provided a new figure to explain this in detail (figure 6)

(6) Considering that IGFBP2 was chosen from the activated target gene(s) cluster, can the authors explain why the reduced CFU-M phenomenon observed in Figure 3G does not appear in the MethoCult assay for IGFBP2 treated cells (Figure 4B)?

The difference could be explained by the fact that in Figure 3G, the cells underwent activation of multiple genes, while in Figure 4B, they were only exposed to IGFBP2. Our results show that IGFBP2 could at least partially explain the phenotype that we see with the activation, but we believe that during the activation experiments, there might be other signals available that might not be induced by IGFBP2 alone. We have also added a summary section and a figure to clarify the different mechanisms of action of the gene activation and IGFBP2.

(7) Figure 4- while the experiments conducted support the role of IGFBP2 in increasing hematopoietic output, there is no experimental evidence to prove its function through paracrine signalling in HECs. The authors need to provide some evidence of how IGFBP2 supplementation specifically expands only the hematopoietic progenitors. Experimental strategies involving specifically targeting IGFBP2 in hemogenic/arterial endothelial cells are required to prove its cell type specific function. Additionally, assessing the in vivo functional potential of the hematopoietic cells generated in the presence of IGFBP2, by bone-marrow transplantation of CD34+ CD43+ cells, is essential. 

The role of IGFBP2 in the context of HSC production and expansion was not the topic of our research, and we have not claimed that IGFBP2  affects the long-term repopulating capacity of HSPCs. Therefore, we believe that the requested experiments are not required to support the specific claims that we do make. We have now provided more experiments and bioinformatic analysis that support the role of IGFBP2 in inducing the progression of EHT from arterial cells to hemogenic endothelium, and to avoid misunderstandings, we have toned down our claims by editing the text regarding its paracrine effect s. 

(8) Figure 4C-D -It is recommended to plot % populations along with fold change value. As this is a key finding, it is important to perform flow cytometry for additional hematopoietic markers- CD144, CD235a and CD41a to demonstrate whether this strategy can also expand erythroid-megakaryocyte progenitors. Telma

Figure 4C already shows the percentage values; we have now added the percentage for Figure 4D in SF5C. We have also performed additional analysis as requested and added the data obtained to Supplementary Figure 5D.

(9) In Figure 5, analysis showing the frequency of cells constituting different clusters, between untreated and IGFBP2-treated samples in the single-cell transcriptome analysis is essential. Additional experiments are required to validate the function of IGFBP2 through modulation of metabolic activity. Inhibition of oxidative phosphorylation in the IGFBP2treated cells should reduce the hematopoietic output. Authors should consider doing these experiments to provide a stronger mechanistic insight into IGFBP2-mediated regulation of hematopoietic emergence.

We have now included the requested cluster composition in Supplementary Figure 5F. We decided not to include further tests on the metabolic profile of IGFBP2 as we already discussed in other papers that showed, using selective inhibitors, that the EHT coincides with a glycol to OxPhos switch. 

(10) It is very striking to see that IGFBP2 supplementation changes the transcriptional profile of developing hematopoietic cells by increasing transcription of OXPHOS-related genes with concomitant reduction of glycolytic signatures, particularly at Day 13. However, the mitochondrial ATP rate measurements do not seem convincing. The bioenergetic profiles show that when mitochondrial inhibitors are added, both groups exhibit decreased OCR values and, on the other hand, higher ECAR. This indicates that both groups have the capability to utilize OXPHOS or glycolysis and may only differ in their basal respiration rates.

Differences in proliferation rate can cause basal respiration to change. There is no information on how the bioenergetic profile was normalized (cell no./protein amount). Given that IGFBP2 has been shown to increase proliferation, it is very likely that the cells treated with IGFBP2 proliferated faster and therefore have higher OCR. The data needs to be normalized appropriately to negate this possibility.

We have previously tested whether IGFBP2 causes an increase in proliferation by analysing the cell cycle of cells treated with it, as we initially thought this could be a mechanism of action. We have now provided the quantification of the cell cycle in the cells treated with IGFBP2, showing no effect was observed in cell cycle Supplementary Figure 4E. Following this analysis, we decided to plate the same number of cells and test their density under the microscope before running the experiment; each experiment was done in triplicate for each condition. We have now added this info to the method sections lines 806-813.  We did not comment on the basal difference, which we agree might be due to several factors, but we only compared the difference in response to the inhibitors, which isn’t affected by the basal level but exclusively by their D values. We have also included the formulas used to calculate the ATP production rate.

Overall, it appears that IGFBP2 does not seem to primarily cause metabolic changes, but simply accelerates the metabolic dependency on OXPHOS. Hence, the term 'metabolic remodelling' must be avoided unless IGFBP2 depletion/loss of function analysis is shown.

We thank the reviewer for suggesting how to interpret the data about the dependency on OXPHOS. We have now changed the conclusions and claims about the effect of IGFBP2. We have also included a cell cycle analysis of the hematopoietic cells derived upon IGFBP2 addition to show that they don’t show differences in proliferation that could cause the increase in colony formation we observed. Regarding the assay, we have plated the same number of cells for each group to make sure we were comparing the same number of cells, which we also assessed in the microscope before the test, and we eliminated the suspension cells during the washes that preceded the measurement. The review is correct in indicating that there is a basal difference in the value of OCR and ECAR where the IGFBP2 is lower at the start and not higher, which would not conceal higher proliferation. Finally, the ATP production rate is calculated on the variation of OCR and ECAR upon the addition of inhibitors, which normalizes for the basal differences.

Explanation:

The annotated text "Reference #18.aa2d3e17.1733239827.ea4e98df" appears to be a reference code or link to an external resource, likely pertaining to the Spanish bill mentioned in the user question. The significance of this annotation lies in its function as a placeholder or identifier for additional information that is not directly included in the provided text. This reference could potentially lead to a detailed document, database, or error page that contains the key provisions of the bill or outlines its relationship to other bills.

The implications of this annotation are multifaceted:

1. Identification and Retrieval: The reference code is crucial for locating the specific document or webpage that contains the relevant details about the Spanish bill. This ensures that users can access comprehensive information that may not be immediately visible in the main text.

2. Contextual Linking: By referencing an external source, the annotation suggests that the full understanding of the bill's provisions and related legislation requires consulting the linked material. This highlights the interconnected nature of legal documents and the importance of cross-referencing for thorough legal analysis.

3. Potential Error: The URL provided in the reference indicates an "errors.edgesuite.net" domain, which might imply that the link is broken or leads to an error page. This could signify issues with accessing the necessary information, suggesting the need for alternative methods to obtain the details about the bill.

In summary, the annotated text serves as a critical reference point for accessing detailed information regarding the Spanish bill. Its significance is rooted in its role as a connector to external resources, which are essential for a full understanding of the bill's key provisions and its relationship to other legislation.

Author response:

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

Reviewer 1:

- The manuscript needs comprehensive proofreading for language and formatting. In many instances, spaces are missing or not required.

Thank you for your comments. The manuscript has been thoroughly proofread for errors in language and formatting.

- Could the authors explore correlation network analyses to get additional insights into the structure of different clusters? 

We have added a co-occurrence analysis (at species taxonomic level) based on SparCC to the manuscript (Figure 2).

This is described on Page 9 line 141-148

- The GitHub link is not correct. 

The github repository has now been made public.

- It is not possible to access the dataset on ENA. 

We have changed the ENA study PRJEB57401 status to open.

- Add the graphs obtained with decontam analysis as a supplementary figure. 

We have added the outputs of decontam (.csv files with feature lists of ASVs that were filtered based on the prevalence and frequency tests) to the github repository.

- There is nothing about the RPL group in the results section, while the authors discuss this issue in the introduction. What about the controls with proven fertility? 

Thank you. We have amended the manuscript to compare characteristics between the RPL, unexplained subfertility and controls groups.

Line 1279-130 page 8:  

“The study group represented 85% of samples with high sperm DNA fragmentation, 85% of samples with elevated ROS and 79% of samples with oligospermia. Rates of abnormal seminal parameters including low sperm concentration, reduced progressive motility and ROS concentrations were found to be highest in the MFI group (Supplementary Figure 1). Baseline characteristics between the RPL, unexplained subfertility and controls groups were similar.

Line 150-154 Page 9: 

“Bacterial richness, diversity and load were similar between all patient groups examined in the study (Supplementary Figure 4).

- While correctly stated in the title, the term microbiota should be used throughout the manuscript instead of "microbiome" 

Thank you. This misnomer has been amended throughout the manuscript.

Minor corrections:

Line 25: provoke is not a good term here. 

Thank you. The term ‘provoke’ has been removed

Line 26: why does semen culture have a limited scope? 

Thank you. Line 40-41 Page 3 has been amended:

“It is therefore plausible that asymptomatic seminal infections may be associated with impaired reproductive function in some men. Since semen culture has a limited scope for studying the seminal microbiota due to its inability to identify all present microbiota next generation sequencing (NGS) approaches have been reported recently by a growing number of investigators (13, 14, 15, 16, 17, 18, 19)”.

Line 68: write μl correctly

Thank you. This has been corrected

Line 131: several organisms at the genus level. 

Thank you. This has been corrected

Line 136: what are the relative abundances of these genera? Is this relevant? 

The mean relative abundances for the key taxa mention in each cluster are all above 20%. This information has been added to the manuscript text on page 9, line 153.

Line 173: Molina et al. 

Thank you. This has been corrected

Line 173: the contaminations are referred to the low biomass nature of testicular samples. If present, bacteria of accessory gland secretions are an integral part of the seminal microbiota itself. Please review these sentences. 

Thank you. This had been reworked to highlight the important of urethral contamination, which you later allude to as a limitation of our study is the failure to provide paired urine and semen samples.

Page 11 line 194-196

“Molina et al report that 50%-70% of detected bacterial reads may be environmental contaminants in a sample from extracted testicular spermatozoa (35); with the addition of passage along the urethra it is likely that contamination of ejaculated semen would be much higher.”

Table 1: remove results interpretation from table caption. 

Thank you this has been acted upon.

Table 1: why in some cases, like in DNA fragmentation index, the total is not equal to n=223? 

This is due to missing data/ analysis not possible for some men due to the requirement of a minimum number of sperm in the ejaculate to perform DNA fragmentation testing.

Table 1: "frag" is not defined. 

Thank you, this has been amended

Tables 2, 3 & 4: bacterial genera in italics. 

Thank you, this has been amended

Figure 1A: add the fertility status information above the cluster colors. 

Thank you, this has been amended in Figure 1.

Figure 1C: the color code is confusing. Use different colors for each cluster. 

Figure 1 legend: bacterial genera in italics. 

Figures 1 & 2: the authors should use similar chart formatting in the two tables. 

Thank you, this has been amended

Reviewer 2:

(1) The patient groups have different diagnoses and should be handled as different groups, and not fused into one 'patient' group in analyses. <br /> Why are the data in tables presented as controls and cases? I would consider men from couples with recurrent pregnancy loss, unexplained infertility, and male factor infertility to have different seminal parameters (not to fuse them into one group). This means, that the statistical analyses should be performed considering each group separately, and not to fuse 3 different infertility diagnoses into one patient group. 

We have conducted detailed analyses, requested by the reviewer, comparing seminal DNA, ROS and microbiota characteristics between each individual patient groups (Supplimental figures 1 and 4). No specific taxa (at either genera or species-level) were found to differ in relative abundance between the diagnostic groups. However, we expect associations between parameters such as reactive oxygen species, or DNA fragmentation, and relative abundance of bacterial species, to be general and not restricted to or specific to each diagnostic group. Therefore, we also conducted further analyses aggregating data from all patient groups to investigate relationships common to these different forms of male reproductive dysfunction.

(2) Were any covariables included in the statistical analyses, e.g. age, BMI, smoking, time of sexual abstinence, etc? 

Covariates were not included in the statistical analyses. This has been added in the manuscript to the limitations.

Page 14 line 267-268

“Additionally, we did not have other covariables such as smoking status with which to include in further analyses”.

(3) Furthermore, it is known that 16S rRNA gene analysis does not provide sensitive enough detection of bacteria on the species level. How much do the authors trust their results on the species level? 

The limitations of taxonomic assignment using 16S rRNA gene metataxonomics are well documented. However, the capacity to assign sequence amplicons at species level depends on the sequence variability of the 16S rRNA gene for each of the taxa reported and the specific gene region chosen. In this study, amplification of the V1-V2 region was performed using a mixed 28f primer set (see methods for details) that enables resolution and assignment of several bacterial species highly relevant to the reproductive tract including Lactobacillus spp., such as L. crispatus and L. iners, (e.g. https://doi.org/10.3389/fcell.2021.641921, https://doi.org/10.1128/msystems.01039-23, https://doi.org/10.1186/s12915-023-01702-2). In this study, we report the presence of L. iners, but not L. crispatus in semen samples, and we have also identified a specific association/co-occurrence between Gardnerella vaginalis and Lactobacillus iners, similar to that observed in vaginal bacterial communities.

(4) Were the analyses of bacterial genera and species abundances with seminal quality parameters controlled for diagnosis and other confounders? 

As stated in point 2, no adjustment was made for co-variates. No differences in microbiome composition were observed among the three diagnostic groups, so no adjustments were made to our analysis.

(5) The authors stress that their study is the biggest on the microbiome in semen. However, when considering that the study consists of 4 groups (with n=46-63), it does not stand out from previous studies. 

Our study is overall the largest investigating interactions between the seminal microbiome and male reproductive dysfunction. Other studies have included greater numbers of men with infertility.

(6) Weaknesses: There is a lack of paired seminal/urinal samples. 

Thank you. This limitation has been added.

Page 14 line 266-267

“A further limitation of this study, and others, is the lack of reciprocal genital tract microbiota testing of the female partners, or paired seminal and urinary samples from male participants”.

Recommendation for authors to consider:

Including previous classical reviews in the introduction: DOI:10.1097/MOU.0000000000000742 <br /> DOI: 10.1038/s41585-019-0250-y 

Thank you. This has been added.

Mentioning in the M&M section that there is a supplementary text with a more detailed M&M part. 

Thank you. This has been added. Further methodological detail can be found in supplementary text.

Revising the use of 'microbiota' and 'microbiome', they are not synonyms. When talking of 16S rRNA gene analysis, we consider 'microbiome' analysis. 

Thank you. This misnomer has been amended throughout the manuscript.

Revising the text, there are several erratas (e.g. verb missing, etc). 

Thank you for your comments. The manuscript has been thoroughly proofread for errors in language and formatting.

Wave glimmer in enginners' eye

Total: 75/100

Original Content 30/30

Technical Understanding 25/30: * Use of Required Tools (6/6) * The project effectively integrates tools such as the image classifier, video input, and audio playback. * Integration of Technology (4/5) * The use of classification, video display, and label-based progression is seamless and functional. * The classification results are used effectively to trigger specific actions and audio playback. However, there is no fallback logic if classification fails or if the environment lacks recognized objects. * Functionality of Classifier Models (3/5) * The classifier operates with sufficient functionality, identifying objects and triggering associated instructions. * The model is not retrained, as indicated in the code comments, which could reduce its accuracy in recognizing project-specific elements. * Interactivity and Feedback (4/5) * The user is guided through the experience with dynamic audio and text feedback based on classified labels. * Feedback is engaging and humorous but could be enhanced with more visual or interactive cues for the detected elements. * Complexity of Decision Tree or Flowchart (3/4) * The flow is linear but has some branching logic tied to specific labels. This creates a guided but somewhat limited exploration experience. * Increasing the number of paths or conditions could enrich the overall structure and align more closely with the dérive concept. * Creative Application of Rules (5/5) * The rules are really creative, with witty and engaging text prompts that enhance the psychogeographic aspect of the project.

Timely Submission 20/40

DOI: 10.1016/j.celrep.2024.115025

Resource: (IMSR Cat# CRL_022,RRID:IMSR_CRL:022)

Curator: @vtello

SciCrunch record: RRID:IMSR_CRL:022

What is this?

This seems interesting. Do I still consume the feeds on the Bluesky application?

via mikael

I wish that MySpace profiles hadn't been lost. I wasn't technical back then (n.b. this is nostalgia about when I was a literal child) but I used all kinds of tools and generators to try to get the visual look I wanted, and the look of translucency above a large photo background was a big one. Why did I like it? What was the influence? For the life of me I can't identify one.

State management is a complex topic in Flutter with various approaches to choose from.

"State management is a complex topic."

Provider: A commonly used state management solution in Flutter.

"Provider package"

Riverpod offers compile safety and testing without depending on the Flutter SDK, similar to Provider.

"Riverpod works in a similar fashion to Provider. It offers compile safety and testing without depending on the Flutter SDK."

setState is the low-level approach for widget-specific, ephemeral state.

"The low-level approach to use for widget-specific, ephemeral state."

ValueNotifier & InheritedNotifier use Flutter's built-in tools to update state and notify the UI of changes.

"An approach using only Flutter provided tooling to update state and notify the UI of changes."

InheritedWidget & InheritedModel facilitate communication between ancestors and children in the widget tree and are the foundation for other state management solutions.

"The low-level approach used to communicate between ancestors and children in the widget tree. This is what provider and many other approaches use under the hood."

June is a lightweight and modern library focusing on a pattern similar to Flutter's built-in state management.

"A lightweight and modern state management library that focuses on providing a pattern similar to Flutter's built-in state management."

Redux is a state container approach familiar to web developers, suitable for managing application state.

"A state container approach familiar to many web developers."

Fish Redux is an assembled Flutter application framework based on Redux, suitable for medium and large applications.

"Fish Redux is an assembled flutter application framework based on Redux state management. It is suitable for building medium and large applications."

BLoC / Rx comprise a family of stream/observable-based patterns for state management.

"A family of stream/observable based patterns."

GetIt is a service locator approach that doesn't require a BuildContext for state management.

"A service locator based state management approach that doesn't need a BuildContext."

MobX is a popular library based on observables and reactions for state management.

"A popular library based on observables and reactions."

Dart Board is a modular feature management framework designed to encapsulate and isolate features in Flutter applications.

"A modular feature management framework for Flutter. Dart Board is designed to help encapsulate and isolate features, including examples/frameworks, small kernel, and many ready-to-use decoupled features such as debugging, logging, auth, redux, locator, particle system and more."

Flutter Commands uses the Command Pattern and ValueNotifiers for reactive state management.

"Reactive state management that uses the Command Pattern and is based on ValueNotifiers."

Binder is a state management package using InheritedWidget at its core, promoting separation of concerns.

"A state management package that uses InheritedWidget at its core. Inspired in part by recoil. This package promotes the separation of concerns."

GetX is a simplified and powerful reactive state management solution.

"A simplified reactive state management solution."

states_rebuilder combines state management with dependency injection and an integrated router.

"An approach that combines state management with a dependency injection solution and an integrated router."

Triple Pattern (Segmented State Pattern) uses Streams or ValueNotifier, focusing on three core values: Error, Loading, and State.

"Triple is a pattern for state management that uses Streams or ValueNotifier. This mechanism (nicknamed triple because the stream always uses three values: Error, Loading, and State), is based on the Segmented State pattern."

solidart is a simple yet powerful state management solution inspired by SolidJS.

"A simple but powerful state management solution inspired by SolidJS."

flutter_reactive_value offers a minimalistic solution, allowing newcomers to add reactivity without complexity.

"The flutter_reactive_value library might offer the least complex solution for state management in Flutter. It might help Flutter newcomers add reactivity to their UI, without the complexity of the mechanisms described before."

Elementary provides a straightforward way to build Flutter applications with MVVM, enhancing productivity and testability.

"Elementary is a simple and reliable way to build applications with MVVM in Flutter. It offers a pure Flutter experience with clear code separation by responsibilities, efficient rebuilds, easy testability, and enhancing team productivity."

Developers are encouraged to review these options to select an approach that best fits their use case.

"If you feel that some of your questions haven't been answered, or that the approach described on these pages is not viable for your use cases, you are probably right."

The annotated text, "Bayesian neural network on the supervised regression task," is highlighted to emphasize a key methodological innovation in the research presented in the paper. Here’s a comprehensive explanation of its significance:

Explanation

The highlighted phrase "Bayesian neural network on the supervised regression task" is significant because it underscores a novel approach in the study of cosmological density fields and warm dark matter (WDM) models. By leveraging a Bayesian neural network, the authors introduce a machine learning technique that can perform a supervised regression task to predict the optical depth-weighted density field (Δτ) from the Lyman-α forest flux field. This method offers several transformational benefits:

1. Enhanced Predictive Accuracy: The Bayesian neural network is trained to predict not only the density field but also the uncertainty in its reconstruction. This dual prediction capability allows for more robust and reliable inferences about the underlying cosmological structures.

2. Efficient Data Utilization: The machine learning approach enables the extraction of meaningful constraints on WDM particle masses using significantly less observational data compared to traditional methods like Markov Chain Monte Carlo (MCMC) techniques. This efficiency is crucial for advancing our understanding of dark matter with limited data resources.

3. Innovative Statistical Inference: The implementation of a Bayesian framework allows the researchers to quantify and propagate uncertainties through the statistical analysis pipeline. This results in more precise and credible constraints on the properties of dark matter particles.

4. Validation on Diverse Datasets: The neural network's performance is validated against both the Sherwood-Relics simulation suite and the Nyx simulation code, demonstrating its generalization capability across different hydrodynamical solvers and physical prescriptions.

Overall, the use of a Bayesian neural network for supervised regression in this context represents a significant methodological advancement in astrophysics and cosmology. It enables more accurate reconstructions of the intergalactic medium density field and provides tighter constraints on warm dark matter models, thus contributing to our broader understanding of the universe's structure and composition.

So clearly this is meant to let us admire the monogram and wordmark, but actually what's grabbing me is the use of the repeated QR code. QR codes vibrate with potential in bringing hypertextuality to printed (or knitted or etc. etc.) materials, but something about how they look stamped onto posters and stickers and such has always turned me off. (If someone were out there getting good results with a vegan diffusion model, this kind of nonsense wouldn't be entirely unappealing to me, pursued with a bit more taste...) But the simple repeat as a border changes how this reads, so profoundly I feel stupid for not having ever tried it myself.

Sometimes it's fun to live in the time you live in!

Author response:

Reviewer #1 (Recommendations for the authors):

(1) Storyline and Narrative Flow:

Consider revising the manuscript to create a more coherent and consistent narrative. Clarify how each section of the study-particularly the transition from multi-omics data integration to single-cell RNA-seq validation-contributes to the overall research question. This will help readers better understand the logical flow of the study.

In the upcoming revisions, we will optimize the logical connections between sections of the manuscript to clarify the role each part plays in the overall research question, making it easier for readers to follow.

(2) Immune Cell Activity Analysis:

Reevaluate the methods used to assess immune cell activities within the context of the tumor microenvironment. Consider providing additional justification for the relevance of using the cancer cell model for this analysis. If necessary, explore alternative methods or models that might offer more meaningful insights into immune-tumor interactions.

We fully recognize the importance of using tumor models to analyze and validate immune activity results, and we are considering experimental research in this area in future projects.

(3) Single-Cell RNA-Seq Validation:

Expand the validation of your findings using single-cell RNA-seq data. This could include more in-depth analyses that explore the heterogeneity within the subtypes and confirm the robustness of your classification method at the single-cell level. This would strengthen the support for your claims about the relevance of the identified subtypes.

In the current study, we have applied the obtained multi-omics profiling features to single-cell sequencing data to classify malignant cells. We analyzed the metabolic and cell communication differences between different subtypes of malignant cells and explored potential reasons for these differences. Next, we plan to conduct further analysis of the differences between malignant cell subtypes to identify additional clues and mechanisms underlying these variations.

(4) Methodological Justification:

Provide a more detailed rationale for the selection of machine learning algorithms and integration strategies used in the study. Explain why the chosen methods are particularly well-suited for this research, and discuss any potential limitations they might have.

In the revised manuscript, we will include descriptions of the principles of these analytical methods, as well as examples of their application in other studies, to discuss the rationale and limitations of applying these methods in this research.

(5) Figures and Visualizations:

Improve the clarity of your figures by addressing the following:

a) Figure 3A: Cluster the pathways to make the comparisons clearer and more meaningful.

b) Figure 4A: Clearly explain the significance of the blue bar.

c) Figure 4B: Ensure this figure is discussed in the main text to justify its inclusion.

d) Figure 7C: Enhance the figure legend to provide more informative details.

Additionally, ensure that figure descriptions go beyond the captions and provide detailed explanations that help the reader understand the significance of each figure.

We fully agree with the reviewer’s suggestions regarding these figures, and we will make the necessary revisions in the revised manuscript.

(6) Supplementary Materials:

Consider including more detailed supplementary materials that provide additional validation data, extended methodological descriptions, and any other information that would support the robustness of your findings.

When we submission the revised manuscript, we will include supplementary materials such as figures or tables that may enhance the presentation of the manuscript's completeness.

(7) Recent Literature:

a) Incorporate more recent studies in your discussion, especially those related to HCC subtypes and the application of machine learning in oncology. This will provide a more current context for your work and help position your findings within the broader field.

We appreciate the reviewer's suggestion. We will incorporate more recent studies into the discussion section and optimize its content.

(8) Data and Code Availability:

Ensure that all data, code, and materials used in your study are made available in line with eLife's policies. Provide clear links to repositories where readers can access the data and code used in your analyses.

We have indicated the sources of the data and tools used in the analysis process within the text, and these data and tools can be accessed through the websites or literature we have cited.

Reviewer #2 (Recommendations for the authors):

(1) While the computational findings are robust, further experimental validation of the two subtypes, particularly the role of the MIF signaling pathway, would strengthen the biological relevance of the findings. In vitro or in vivo validation could confirm the proposed mechanisms and their influence on patient prognosis.

We fully recognize the importance of using tumor models to analyze and validate immune activity results, and we are considering experimental research in this area in future projects.

(2) Consider testing the model on additional independent cohorts beyond the TCGA and ICGC datasets to further demonstrate its generalizability and applicability across different patient populations.

We are considering looking for independent external datasets in the GEO database or other databases to validate our model.

(3) Review the manuscript for long or complex sentences, which can be broken down into shorter, more readable parts.

In the revised manuscript, we will address any grammatical issues present in the manuscript and modify long and complex sentences that may hinder reader comprehension.

Author response:

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

Thank you for your assessment and constructive critique, which helped us to improve the manuscript and its clarity. Upon carefully reading through the comments, we noticed that, based on the Reviewer's questions, some of our answers were already available but “hidden” as supplementary data. Thus, we changed the following two figures and text accordingly to showcase our results to the reader better:

A) To highlight how mobile service data can indicate the spread of highly prevalent variants, we added a high-prevalence subcluster to Figure 2 (previously shown in Supplementary Figures S4 and S5) and, in exchange, moved one low-prevalence subcluster from Figure 2 back into the supplement. The figure is now showing a low and a high prevalent subcluster instead of two low prevalent subclusters.

B) Based on Reviewer 1’s question about where samples were taken in regards to the mobility data from the community of the first identification (negative controls), we now highlight all the mobility data that was available to us in Figure 3 (as triangles) instead of just a few top mobility hits for both - mobility guided and random surveillance (serving as a negative control for the former). This way, we think, it is clearer how random sampling was also performed in some regions where mobility was coming from the community of origin (as asked by Reviewer 1) - the detailed trips and sampling are now part of the supplement for data transparency reasons. We also noticed a typo in the GPS coordinates, aligning one of the arrows falsely, which is corrected in the improved Figure 3.

We have also included the R-Scripts used to generate all the figures in the manuscript in an OSF repository (we updated the “Data sharing statement”). We also updated Figure 1 slightly and extended the supplemental material. The remaining comments to reviewers are addressed point-by-point below.

Reviewer 1 (Public Review):

In "1 Exploring the Spatial Distribution of Persistent SARS-CoV-2 Mutations -Leveraging mobility data for targeted sampling" Spott et al. combine SARS-CoV-2 genomic data alongside granular mobility data to retrospectively evaluate the spread of SARS-CoV-2 alpha lineages throughout Germany and specifically Thuringia. They further prospectively identified districts with strong mobility links to the first district in which BQ.1.1 was observed to direct additional surveillance efforts to these districts. The additional surveillance effort resulted in the earlier identification of BQ.1.1 in districts with strong links to the district in which BQ.1.1 was first observed.

Thank you for taking the time to review our work.

(1) It seems the mobility-guided increased surveillance included only districts with significant mobility links to the origin district and did not include any "control" districts (those without strong mobility links). As such, you can only conclude that increasing sampling depth increased the rate of detection for BQ.1.1., not necessarily that doing so in a mobility-guided fashion provided an additional benefit. I absolutely understand the challenges of doing this in a real-world setting and think that the work remains valuable even with this limitation, but I would like the lack of control districts to be more explicitly discussed.

Thank you for the critical assessment of our work. We agree that a control is essential for interpreting the results. In our case, randomized surveillance (“the gold standard”) served as a control with a total sampling depth seven times higher than the mobility-guided sampling. To better reflect the sampling in regards to the available mobility data, we revisited Figure 3 and added all the mobility information from the origin that was available to us. We also added this information to the random surveillance to provide a clearer picture to the reader. This now clearly shows how randomized surveillance covered communities with varying degrees of incoming mobility from the community of first occurrences, thereby underlining its role as a negative control. We updated the manuscript to reflect these changes and included the October 2020 and June 2021 mobility datasets in Supplementary Table S6. We agree that the sampling depth increases the detection, which is the point of guided sampling to increase sampling, specifically in areas where mobility points towards a possible spread. In regards to the negative control: Random surveillance (not Mobility-guided) in October covered 40 samples in the northwest region of Thuringia (Mobility-guided covered 19 samples). Thus, random surveillance also contained 31 out of 132 samples with a mobility link towards the first occurrence of BQ1.1 but with varying amounts of mobility (low to high).

We added this information to the main text:

Line 270 to 293:

Following its first Thuringian identification, we utilized the latest available dataset of the past two years of mobile service data (October 2020 and June 2021) to investigate the residential movements for the community of first detection. Considering the highest incoming mobility from both datasets, we identified 18 communities with high (> 10,000), 34 with medium (2,001-10,000), and 82 with low (30-2,000) number of incoming one-way trips from the originating community (purple triangles in Figure 3a). As a result, we specifically requested all the available samples from the eight communities with the highest incoming mobility. Still, we were restricted to the submission of third parties over whom we had no influence. This led to the inclusion of the following eight communities with the most residential movement from the originating community: four in central and three in NW of Thuringia, one in NW-neighboring state Saxony-Anhalt. The samples requested from central Thuringia were also due to their geographic arrangement as a “belt” in central Thuringia, linking three major cities (see Supplementary Figure S1). Subsequently, we collected 19 additional samples (isolated between the 17th and 25th of October 2022; see “Guided Sampling” for October 2022, Figure 3a) besides the randomized sampling strategy. Thus, the sampling depth was increased in communities with high incoming mobility from the first origin.

As part of the general Thuringian surveillance, we collected 132 samples for October (covering dates between the 5th and 31st) and 69 samples in November (covering dates between the 1st and 25th; see Figure 3b and c). Randomized sampling was not influenced or adjusted based on the mobility-guided sample collection. Thus, it also contains samples from communities with a mobility link towards the first occurrence of BQ.1.1, as they were part of the regular random collection (see gray triangles in Figure 3b). A complete overview of all samples is provided in Supplementary Table S5. The mobility datasets from October 2020 and June 2021 for all sampled communities are provided in Supplementary Table S6.

Line 305 to 313:

Among the 19 samples specifically collected based on mobile service data, we identified one additional sample of the specific Omicron sublineage BQ.1.1 in a community with high incoming mobility (n = 14, number of trips = 37,499) with a distance of approximately 16 km between both towns. Our randomly sampled routine surveillance strategy did not detect another sample during the same period. This was despite a seven times higher overall sample rate, which included 31 samples from communities with an identified incoming mobility from the community of the first occurrence (October 2022, Figure 3b). Only in the one-month follow-up were four other samples identified across Thuringia through routine surveillance (November 2022, Figure 3c).

Line 325 to 333:

In summary, increasing the sampling depth in the suspected regions successfully identified the specified lineage using only a fraction of the samples from the randomized sampling. Conversely, randomized surveillance, the “gold standard” acting as our negative control, did not identify additional samples with similar sampling depths in regions with no or low incoming mobility or even in high mobility regions with less sampling depth. Implementing such an approach effectively under pandemic conditions poses difficult challenges due to the fluctuating sampling sizes. Although the finding of the sample may have been coincidental, our proof of concept demonstrated how we can leverage the potential of mobile service data for targeted surveillance sampling.

(2) Line 313: While this work has reliably shown that the spread of Alpha was slower in Thuringia, I don't think there have been sufficient analyses to conclude that this is due to the lack of transportation hubs. My understanding is that only mobility within Thuringia has been evaluated here and not between Thuringia and other parts of Germany.

Thank you for pointing this out. We noticed that the original sentence lacked the necessary clarity. The statement in line 313 was based on the observation that Alpha first occurred in federal states with major transport hubs, such as international airports and ports, which Thuringia lacks, as demonstrated in the Microreact dataset. For clarification, we adjusted the sentence as follows:

Line 340 and following:

A plausible explanation for the delayed spread of the Alpha lineage in Thuringia is the lack of major transport hubs, as Alpha first occurred in federal states with such hubs. Previous studies have already highlighted the impact of major transportation hubs in the spread of Sars-CoV-2.

(3) Line 333 (and elsewhere): I'm not convinced, based on the results presented in Figure 2, that the authors have reliably identified a sampling bias here. This is only true if you assume (as in line 235) that the variant was in these districts, but that hasn't actually been demonstrated here. While I recognize that for high-prevalence variants, there is a strong correlation between inflow and variant prevalence, low-prevalence variants by definition spread less and may genuinely be missing from some districts. To support this conclusion that they identified a bias, I'd like to see some type of statistical model that is based e.g. on the number of sequences, prevalence of a given variant in other districts, etc. Alternatively, the language can be softened ("putative sampling bias").

Thank you for addressing this legitimate point of criticism in our interpretation. Due to the retrospective nature of the analysis and the fact that we found no additional samples of the clusters after the specified timeframes, we were limited to the samples in our dataset. Therefore, it is impossible to demonstrate if a variant was present in the relevant districts afterward. We agree that the variant’s low prevalence means they may genuinely not have spread to some districts. For clarification, we added the following statements and changed the wording accordingly:

Additional statement in line 248:

However, due to their low prevalence, it is also possible that these subclusters have not spread to the indicated districts.

Adjusted wording in line 361:

We exemplified this approach with the Alpha lineage, where mobile service data indicated a putative sampling bias and partially predicted the spread of our Thuringian subclusters.

Recommendations:

(1) I applaud the use of the microreact page to make the data public, however, I don't see any reference to a GitHub or Zenodo repository with the analysis code. The NextStrain code is certainly appreciated but there is presumably additional code used to identify the clusters, generate figures, etc. I generally prefer this code be made public and it is recommended by eLife.

Thank you for your appreciation. We have now included the R-scripts in the manuscript’s OSF repository. These were used to create the figures in the manuscript and supplement utilizing the supplementary tables 1-6, which are also stored in the repository. To clearly communicate which data is provided, we changed lines 513 and 514 of the “Data sharing statement” as follows:

Line 513 and following:

Supplementary tables and the R-scripts used to generate all figures are also provided in the repository under https://osf.io/n5qj6/. These include the mobile service data used in this study, which is available in processed and anonymized form.

The subcluster identification was performed manually. By adding each sample's mutation profile to the Microreact metadata file, we visually screened the phylogenetic time tree for all non-Alpha specific mutations present in at least 20 Thuringian genomes. We then applied the criteria described in the Methods section to identify the nine Alpha subclusters. For clarification, we changed line 436:

Line 436:

We then manually screened for mutations present in at least 20 genomes with a small phylogenetic distance and a time occurrence of at least two months.

Reviewer 2 (Public Review):

In the manuscript, the authors combine SARS-CoV-2 sequence data from a state in Germany and mobility data to help in understanding the movement of the virus and the potential to help decide where to focus sequencing. The global expansion in sequencing capability is a key outcome of the public health response. However, there remains uncertainty about how to maximise the insights the sequence data can give. Improved ability to predict the movement of emergent variants would be a useful public health outcome. Also knowing where to focus sequencing to maximising insights is also key. The presented case study from one State in Germany is therefore a useful addition to the literature. Nevertheless, I have a few comments.

Thank you for taking the time to review our work.

(1) One of the key goals of the paper is to explore whether mobile phone data can help predict the spread of lineages. However, it appears unclear whether this was actually addressed in the analyses. To do this, the authors could hold out data from a period of time, and see whether they can predict where the variants end up being found.

Based on your feedback, we noticed that the results of the other seven clusters presented in the supplement were not appropriately highlighted, causing them to be overlooked. We indeed demonstrated that predicting viral spread based on mobility data is possible, as shown for the high-prevalence subcluster 7 (Cluster “ORF1b:A520V”, 811 samples). This was briefly mentioned in lines 240-242, but the cluster was only shown in Supplementary Figures S4 and S5. Instead, we focused more on the putative sampling bias that the mobility for low-prevalence subclusters could indicate as an interesting use case of mobility data. This addresses a concrete problem of every surveillance: successfully identifying low-prevalence targets. However, based on your feedback, we revisited Figure 2, adding the plots of the high-prevalence subcluster: “ORF1b:A520V” from Supplementary Figures S4 and S5 while moving the low-prevalence subcluster “S:N185D” from Figure 2 into the Supplementary Figures S4 and S5. Additionally, we changed line 229 to highlight this result properly.

line 229 and following:

The mobile service data-based prediction of a subcluster’s spread aligned well with the subsequent regional coverage of fast-spreading, highly prevalent subclusters, such as subcluster 7, which covered 811 samples (see Figure 2). In contrast, the predicted spread for the low-prevalence subclusters did not correspond well with the actual occurrence.

(2) The abstract presents the mobility-guided sampling as a success, however, the results provide a much more mixed result. Ultimately, it's unclear what having this strategy really achieved. In a quickly moving pandemic, it is unclear what hunting for extra sequences of a specific, already identified, variant really does. I'm not sure what public health action would result, especially given the variant has already been identified.

Thank you for your critical assessment of the presented results and their interpretation.

Here, we aimed to provide an alternative to the standard randomized surveillance strategy. Through mobility-guided sampling, we sought to increase identification chances while necessitating fewer samples and decreasing costs, ultimately enhancing surveillance efficiency. The Omicron-lineage BQ.1.1 was the perfect example to prove this concept under actual pandemic conditions. Yet, the strategy is not limited to low-prevalence sublineages but can be applied to virtually any surveillance case. However, from your question, we recognize that this conclusion was unclear from the text. Therefore, we adapted the conclusion to better communicate the real implications of our proof of concept. Additionally, we altered line 42 in the abstract for clarification.

However, we did not assess the benefits of surveillance itself, as the German Robert Koch Institute (RKI) already had outlined its importance for tracking different viral variants. This tracking served several reasons, like monitoring vaccine escapism, mutational progress, and assessing available antibodies for treatment.

Line 42:

The latter concept was successfully implemented as a proof-of-concept for a mobility-guided sampling strategy in response to the surveillance of Omicron sublineage BQ.1.1.

Line 364 to 374:

Another approach is actively guiding the sampling process through mobile service data, which we demonstrated with our proof of principle focusing on the Omicron-lineage BQ.1.1 as a real-life example. This approach could allow for a flexible allocation of surveillance resources, enabling adaptation to specific circumstances and increasing sampling depth in regions where a variant is anticipated. By incorporating guided sampling, much fewer resources may be needed for unguided or random sampling, thereby reducing overall surveillance costs.

Additionally, while this approach is particularly useful for identifying low-prevalence variants, it is not limited to such variants. Still, it can provide a guided, more cost-efficient, low-sampling alternative to general randomized surveillance that can also be applied to other viruses or lineages.

(3) Relatedly, it is unclear to me whether simply relying on spatial distance would not be an alternative simpler approach than mobile phone data. From Figure 2, it seems clear that a simple proximity matrix would work well at reconstructing viral flow. The authors could compare the correlation of spatial, spatial proximity, and CDR data.

Thank you for pointing this out. While proximity data might appear to be an obvious choice, it has significant limitations compared to mobility data, especially in the context of our study. Proximity data assumes that spatial distance alone can accurately represent movement patterns, which would only be true in a normally distributed traffic network. Geographic features such as mountains, cities, and highways affect traffic flows, leading to variability over distance and time, which are beyond the scope of spatial proximity but efficiently captured by mobility data. In Figure 2, we presented a simplified view of the mobility data. Hence, proximity and mobility data appear to provide the same insights. However, as shown in the updated Figure 3, a detailed overview of the available mobility data reveals obvious and non-obvious spatial connections that proximity data can not capture. Incorporating such a level of detail in Figure 2 would have cluttered the figure and reduced its clarity (e.g., adding triangles for each Thuringian community).

While a comparison between proximity data and mobility data would indeed be informative, it is beyond the scope of our current study, as our primary focus was to examine the useability of mobility data in explaining our subcluster’s spread in the first place. However, we agree it would be a valuable direction for future research. We summarized our thoughts from above in the following additional sentence:

Line 374:

Pre-generated mobility networks automatically tailored to each state's unique infrastructure and population dynamics could provide better-targeted sampling guidance rather than simple geographical proximity.

Recommendations:

(1) Line 128: What do these percentages mean - the proportion of States with at least one Alpha variant? Please clarify.

We clarified the values at their first appearance in the text:

Line 127:

By March, Alpha had spread to nearly all states and districts (districts are similar to counties or provinces) in Germany (Median: 76·47 % Alpha samples among a federal states total sequenced samples compared to 36·03 % in February, excluding Thuringia) and Thuringia (Median: 85·29 %, up from 50·00 % in February).

(2) Line 134: It's a little strange to compare the dynamics of a state with that of the whole country. For it lagged as compared to all other States?

Line 134: “In summary, the spread of the Alpha lineage in Thuringia lagged roughly two weeks behind the general spread in the rest of Germany but showed similar proportions.”

Thank you for the feedback. The statement refers to the comparison of Alpha-lineage proportions across federal states, excluding Thuringia, in lines 118 to 130. To simplify, we collectively referred to these federal states as “Germany” in the text. However, we recognize that this formulation is misleading, so we adjusted line 135 for clarification:

Line 135:

In summary, the spread of the Alpha lineage in Thuringia lagged roughly two weeks behind the general spread of other German federal states but showed similar proportions.

It is forbidden to re-use a project directory, which is especially annoying if you are testing the code. Would it be possible to allow this with a warning that you are replacing the last version?

Not sure I follow this sentence

Le code présenté ci-dessous n'est pas optimal. La pointe de la flèche se superpose sur le haut de la lettre B (ceci est corrigé en chargeant amsmath, mais ça fait doublon avec le point 1.2 qui justement propose une solution en chargeant amsmath). De plus, il faudrait éviter l'emploi simultané de \vec et de \overrightarrow car le dessin de leurs pointes diffère sensiblement.