Tiếp nối các công trình trước đó nghiên cứu bài toán tối ưu prompt liên tục và rời rạc, nhóm nghiên cứu quy ước rằng một tập huấn luyện là có sẵn để tính toán điểm accuracy trên tập train như một giá trị mục tiêu cho bài toán tối ưu hóa, và kết quả thực nghiệm cho thấy việc tối ưu hóa prompt dựa trên điểm accuracy trên 1 tập huấn luyện nhỏ là đủ để đạt được kết quả cao trên tập test.

các bố cục prompt tối ưu có thể được cụ thể hóa dựa trên mô hình và bài toán. Vì vậy, prompt engineering thường quan trọng với LLM để đạt được hiệu quả tốt nhất.

Các bài toán được tập trung đánh giá liên quan đến ngôn ngữ tự nhiên mà trong đó cả đầu vào và đầu ra đều là văn bản. LLM được chứng minh là nhạy cảm với bố cục của prompt

Khả năng của LLM trong việc hiểu ngôn ngữ tự nhiên đã đặt ra một khả năng mới cho bài toán tối ưu hóa: thay vì phải xác định bài toán tối ưu hóa và thu được bước cập nhật cùng với một phần mềm, bài toán tối ưu hóa được mô tả bằng ngôn ngữ tự nhiên, sau đó hướng dẫn LLM tạo sinh các giải pháp mới một cách lặp lại dựa trên mô tả của bài toán và các giải pháp trước đó. Việc tối ưu bằng LLM cho phép việc thích nghi nhanh với các bài toán khác nhau bằng cách thay đổi mô tả bài toán trong prompt, và quá trình tối ưu có thể được tùy chỉnh bằng việc thêm các chỉ dẫn để cụ thể hóa các tính chất cần có của các giải pháp.

Phương pháp OPRO được thử nghiệm với bài toán hồi quy tuyến tính (linear regression) và bái toán người giao hàng (traveling salesman), sau đó được đánh giá trên bài toán chính trong việc tối ưu hóa prompt mà trong đó, mục tiêu là tìm các prompt giúp tối đa hóa điểm accuracy.

Bài báo đề xuất phương pháp OPRO, một phương pháp đơn giản và hiệu quả tận dụng LLM làm trình tối ưu, trong đó bài toán tối ưu được mô tả dưới dạng ngôn ngữ tự nhiên. Ở mỗi bước tối ưu, LLM sẽ tạo ra các giải pháp mới từ prompt chứa các giải pháp được tạo trước đó với các giá trị của chúng, sau đó các giải pháp mới sẽ được đánh giá và thêm vào prompt cho bước tối ưu tiếp theo.

online editabe

Akkumulation und Ungleichheit hängen hier also eng zusammen. Eine Form der Akkumulation ist die Addition neuer Stoff-Flüsse zu den bestehenden. Bei Baumaterialien wie bei der Biomasse werden nicht vorhandene Typen durch neue ersetzt, sondern die neuen kommen zu den bestehenden dazu.

mapping

instead of mapping give construction rules and tools

Whispers of the Past: Ghostly Tales from Towns County, Georgia

flip the paradigm

information hiding is the root cause of the problem

leaky abstractions

information transparency visibility and malleability evergreen permanence slef-organization

eLife Assessment

This important study reports findings on the GnRH pulse generator's role in androgen-exposed mouse models, providing further insights into PCOS pathophysiology and advancing the field of reproductive endocrinology. The experimental data were collected using cutting-edge methodologies and are solid. The findings, while interesting, are primarily applicable to mouse models, and their translation to human physiology requires cautious interpretation and further validation. This work will be of interest to endocrinologists and reproductive biologists.

Reviewer #2 (Public review):

Summary:

The authors aimed to investigate the functionality of the GnRH (gonadotropin-releasing hormone) pulse generator in different mouse models to understand its role in reproductive physiology and its implications for conditions like polycystic ovary syndrome (PCOS). They compared the GnRH pulse generator activity in control mice, peripubertal androgen (PPA) treated mice, and prenatal androgen (PNA) exposed mice. The study sought to elucidate how androgen exposure affects the GnRH pulse generator and subsequent LH (luteinizing hormone) secretion, contributing to the pathophysiology of PCOS.

Strengths:

(1) Comprehensive Model Selection: The use of both PPA and PNA mouse models allows for a comparative analysis that can distinguish the effects of different timings of androgen exposure.

(2) Detailed Methodology: The methods employed, such as photometry recordings and serial blood sampling, are robust and allow for precise measurement of GnRH pulse generator activity and LH secretion.

(3) Clear Results Presentation: The experimental results are well-documented with appropriate statistical analyses, ensuring the findings are reliable and reproducible.

(4) Relevance to PCOS: The study addresses a significant gap in understanding the neuroendocrine mechanisms underlying PCOS, making the findings relevant to both basic science and potentially clinical research.

Weaknesses

(1) Model Limitations: While the PNA mouse model is suggested as the most appropriate for studying PCOS, the authors acknowledge that it does not completely replicate the human condition, particularly the elevated LH response seen in women with PCOS.

(2) Complex Data Interpretation: The reduced progesterone feedback and its effects on the GnRH pulse generator in PNA mice add complexity to data interpretation, making it challenging to draw straightforward conclusions.

(3) Machine Learning (ML) Selection and Validation: While k-means clustering is a useful tool for pattern recognition, the manuscript lacks detailed justification for choosing this specific algorithm over other potential methods. The robustness of clustering results has not been validated.

(4) Biological Interpretability: Although the machine learning approach identified cyclical patterns, the biological interpretation of these clusters in the context of PCOS is not thoroughly discussed. A deeper exploration of how these clusters correlate with physiological and pathological states could enhance the study's impact.

(5) Sample Size: The study uses a relatively small number of animals (n=4-7 per group), which may limit the generalisability of the findings. Larger sample sizes could provide more robust and statistically significant results.

(6) Scope of Application: The findings, while interesting, are primarily applicable to mouse models. The translation to human physiology requires cautious interpretation and further validation.

Comments on revised version:

I did not find the response to my main concerns regarding justification for the choice of the number of clusters (k) and providing evidence of cluster robustness satisfactory at all. It sounds contradictory to me to state that the authors have used unsupervised ML approach when at the same time had clear understanding of the data and the features they wanted to capture. Unsupervised approaches are meant to reveal features that are not apparent by eye... however in their response the authors state, "...our aim was to develop an unsupervised approach that would automatically detect the onset and existence of the key features of pulse generator cyclicity that were apparent by eye...". This sounds like a rather supervised ML approach to me.<br /> Furthermore, I am still unsure why did the authors choose k=5, i.e. assumed there are 5 clusters in the data, and did they explore other possible values for k?<br /> - If not why not? How does this fit with the claims that their ML approach is unsupervised, in other words purely data-driven without making any assumptions?<br /> - If yes did they compare the robustness of their clustering results obtained for different values of k?

Reviewer #3 (Public review):

Summary:

Zhou and colleagues elegantly used pre-clinical mouse models to understand the nature of abnormally high GnRH/LH pulse secretion in polycystic ovary syndrome (PCOS), a major endocrine disorder affecting female fertility worldwide. This work brings a fundamental question of how altered gonadotropin secretion takes place upstream within the GnRH pulse generator core, which is defined by arcuate nucleus kisspeptin neurons.

Strengths:

Authors use state-of-the-art in vivo calcium imaging with fiber photometry and important physiological manipulations and measurements to dissect the possible neuronal mechanisms underlying such neuroendocrine derangements in PCOS. The additional use of unsupervised k-means clustering analysis for the evaluation of calcium synchronous events greatly enhances the quality of their evidence. The authors nicely propose that neuroendocrine dysfunction in PCOS might involve different setpoints through the hypothalamic-pituitary-gonadal (HPG) axis, and beyond kisspeptin neurons, which importantly pushes our field forward toward future investigations.

Weaknesses:

The reviewer agrees that the authors provide important evidence and have improved the quality of the manuscript following first-round revisions. However, they seem resistant to show frequency and amplitude averages in Figure 1 or as supplemental data. Whether the amplitude is dependent on fiber position and its influences on the analysis should be a point of discussion and not data omission. A more detailed analysis of frequency data would enhance the quality of their manuscript.

Comments on revised version:

This comment is related to Reviewer 3's comment # 2 (major) response:

The response does not justify why authors could simply show frequency and amplitude averages in Figure 1 or as supplemental data. Whether the amplitude is dependent on fiber position and its influences on the analysis should be a point of discussion and not data omission.

Author response:

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

Reviewer #1 (Public Review):

The manuscript involves 11 research vignettes that interrogate key aspects of GnRH pulse generator in two established mouse models of PCOS (peripubertal and prenatal androgenisation; PPA and PNA) (9 of the vignettes focus on the latter model).

A key message of this paper is that the oft-quoted idea of rapid GnRH/LH pulses associated with PCOS is in fact not readily demonstrable in PNA and PPA mice. This is an important message to make known, but when established dogmas are being challenged, the experiments behind them need to be robust. In this case, underpowered experiments and one or two other issues greatly limit the overall robustness of the study.

General critiques

(1) My main concern is that many/most of the experiments were limited to 4-5 mice per group (PPA experiments 1 and 2, PNA experiments 3, 5, 6, 8, and 9). This seems very underpowered for trying to disprove established dogmas (sometimes falling back on "non-significant trends" - lines 105 and 239).

For the key characterization of GnRH pulse generator activity and LH pulsatility in intact PNA mice (Fig.3, 4, 6), we used 6-8 animals in each experiment which we believe to be sufficient. 

It is pertinent to explore the “established dogma”. While there is every expectation that the PNA model should have increased LH pulsatility, in fact there is only a single study (Moore, Prescott et al. 2015) that has shown this. The two other reports that have examined this issue find no change in LH pulse frequency (McCarthy, Dischino et al. 2021 and ours). Hence, we would suggest that expectations rather than evidence presently maintains the PNA “dogma”. For the PPA model, there is in fact not a single paper reporting increased LH pulse frequency.

(2) Page 133-142: it is concerning that the PNA mice didn't have elevated testosterone levels, and this clearly isn't the fault of the assay as this was re-tested in the laboratory of Prof Handelsman, an expert in the field, using LCMS. The point (clearly made in lines 315-336 of the Discussion) that elevated testosterone in PNA mice has been shown in some but not other publications is an important concern to describe for the field. However, the fact remains that it IS elevated in numerous studies, and in the current study it is not so, yet the authors go on to present GnRH pulse generator data as characteristic of the PNA model. Perhaps a demonstration of elevated testosterone levels (by LCMS?) should become a standard model validation prerequisite for publishing any PNA model data.

We provide a Table below showing the huge inconsistencies in testosterone levels reported in the PNA mouse model. If anything, these inconsistencies might be explained by age, although again this is very variable between studies. Much the same as the “dogma” related to LH pulsatility in the PNA model, we would question whether there is any robust increase in testosterone levels in this model. There is no question that women with PCOS have elevated testosterone but whether the PNA mouse is a good model for this is debatable. We have noted this caution and the need for further LC-MS studies in the Discussion.

Author response table 1.

*Same ELISA used in the current study.

(3) Line 191-196: the lack of a significant increase in LH pulse frequency in PNA mice is based on measurements using reasonable group sizes (7-8), although the sampling frequency is low for this type of analysis (10-minute intervals; 6-minute intervals would seem safer for not missing some pulses). The significance of the LH pulse frequency results is not stated (looks like about p=0.01). The authors note that LH concentration IS elevated (approximately doubled), and this clearly is not caused by an increase in amplitude (Figure 4 G, H, I). These things are worth commenting on in the discussion.

We have included the p-value of the LH pulse frequency results and included the relevant discussion.

(4) An interesting observation is that PNA mice appear to continue to have cyclical patterns of GnRH pulse generator activity despite reproductive acyclicity as determined by vaginal cytology (lines 209-241). This finding was used to analyse the frequency of GnRH pulse generator SEs in the machine-learning-identified diestrous-like stage of PNA mice and compare it to diestrous control mice (as identified by vaginal cytology?) (lines 245-254). The idea of a cycle stage-specific comparison is good, but surely the only valid comparison would be to use machine-learning to identify the diestrous-like stage in both groups of mice. Why use machine learning for one and vaginal cytology for the other?

As “machine learning-defined” diestrus is based on the control vaginal cytology information, the diestrous mice are in fact defined by the same machine learning parameters. We have now noted this.

Specific points

(5) With regard to point 2 above, it would be helpful to note the age at which the testosterone samples were taken.

We have included the age in the method.

(6) Lines 198-205 and 258-266: I think these are repeated measures of ANOVA data? If so, report the main relevant effect before the post hoc test result.

We have included the relevant main effect in the manuscript.

(7) Line 415: I don't think the word "although" works in this sentence.

We have changed the wording accordingly.

(8) Lines 514-518: what are the limits of hormone detection in the LCMS assay?

These were originally stated in the figure legend but have now been included in the Methods.

Reviewer #2 (Public Review):

Summary

The authors aimed to investigate the functionality of the GnRH (gonadotropin-releasing hormone) pulse generator in different mouse models to understand its role in reproductive physiology and its implications for conditions like polycystic ovary syndrome (PCOS). They compared the GnRH pulse generator activity in control mice, peripubertal androgen (PPA) treated mice, and prenatal androgen (PNA) exposed mice. The study sought to elucidate how androgen exposure affects the GnRH pulse generator and subsequent LH (luteinizing hormone) secretion, contributing to the pathophysiology of PCOS.

Strengths

(1) Comprehensive Model Selection: The use of both PPA and PNA mouse models allows for a comparative analysis that can distinguish the effects of different timings of androgen exposure.

(2) Detailed Methodology: The methods employed, such as photometry recordings and serial blood sampling, are robust and allow for precise measurement of GnRH pulse generator activity and LH secretion.

(3) Clear Results Presentation: The experimental results are well-documented with appropriate statistical analyses, ensuring the findings are reliable and reproducible.

(4) Relevance to PCOS: The study addresses a significant gap in understanding the neuroendocrine mechanisms underlying PCOS, making the findings relevant to both basic science and potentially clinical research.

Weaknesses

(1) Model Limitations: While the PNA mouse model is suggested as the most appropriate for studying PCOS, the authors acknowledge that it does not completely replicate the human condition, particularly the elevated LH response seen in women with PCOS.

We agree.

(2) Complex Data Interpretation: The reduced progesterone feedback and its effects on the GnRH pulse generator in PNA mice add complexity to data interpretation, making it challenging to draw straightforward conclusions.

We agree.

(3) Machine Learning (ML) Selection and Validation: While k-means clustering is a useful tool for pattern recognition, the manuscript lacks detailed justification for choosing this specific algorithm over other potential methods. The robustness of clustering results has not been validated.

Please see below.

(4) Biological Interpretability: Although the machine learning approach identified cyclical patterns, the biological interpretation of these clusters in the context of PCOS is not thoroughly discussed. A deeper exploration of how these clusters correlate with physiological and pathological states could enhance the study's impact.

It is presently difficult to ascribe specific functions of the various pulse generator states to physiological impact. While it is reasonable to suggest that Cluster_0 activity (representing very infrequent SEs) is responsible for the estrous/luteal-phase pause in pulsatility, we remain unclear on the physiological impact of multi-peak SEs on LH secretion, even in normal mice (see Vas et al., Endo 2024). Thus, for the moment, it is most appropriate to simply state that pulse generator activity remains cyclical in PNA mice without any unfounded speculation.

(5) Sample Size: The study uses a relatively small number of animals (n=4-7 per group), which may limit the generalisability of the findings. Larger sample sizes could provide more robust and statistically significant results.

For the key characterization of GnRH pulse generator activity and LH pulsatility in intact PNA mice (Fig.3, 4, 6), we used 6-8 animals in each experiment which we believe to be sufficient. Some of the subsequent experiments do have smaller N numbers and we are particularly aware of the progesterone treatment study that only has N=3 for the PNA group. However, as this was sufficient to show a statistical difference we did not generate more mice.

(6) Scope of Application: The findings, while interesting, are primarily applicable to mouse models. The translation to human physiology requires cautious interpretation and further validation.

We agree.

Reviewer #2 (Recommendations For The Authors):

(1) The validation of clustering results through additional metrics or comparison with other algorithms would strengthen the methodology. Specifically, the authors selected k=5 for k-means clustering without providing an explicit rationale or evidence of exploratory data analysis (EDA) to support this choice. They refer to their previous publication (Vas, Wall et al. 2024), which does not provide any EDA regarding the choice of a number of clusters nor their robustness. The arbitrary selection of "k" without justification can undermine confidence in the clustering results since clustering results heavily depend on "k". The authors also choose to use Euclidean distance as the "numerical measure" setting in the RapidMiner Studio's software without justification given the chosen features used for clustering and their properties. The lack of exploratory analysis to determine the optimal number of clusters, "k", to be considered means that the authors might have missed identifying the true structure of the data. Common cluster robustness methods, like the elbow method or silhouette analysis, are crucial for justifying the number of clusters. An inappropriate choice could lead to incorrect conclusions about the synchronisation patterns of ARN kisspeptin neurons and their implications for the study's hypotheses. Including EDA and other validation techniques (e.g., silhouette scores, elbow method) would have strengthened the manuscript by providing empirical support for the chosen algorithm and settings.

It is important to clarify that we did not start this exercise with an unknown or uncharacterised data set and that the objective of the clustering was not to provide any initial pattern to the data. Rather, our aim was to develop an unsupervised approach that would automatically detect the onset and existence of the key features of pulse generator cyclicity that were apparent by eye e.g. the estrous stage slowing and the presence of multi-peak SEs in metestrous. As such, our optimization was driven by the data as well as observation while retaining the unsupervised nature of k-means clustering. We started by assessed 10 variables describing all possible features of the recordings and through a process of elimination found that just 5 were sufficient to describe the key stages of the cycle. While we appreciate that the use of multiple different algorithms would progressively increase the robustness of the machine learning approach, it is evident that the current k-means approach with k=5 is already very effective at reporting the estrous cyclicity of the pulse generator in normal mice (Vas et al., Endo 2024). Having validated this approach, we have now used it here to compare the cyclical patterns of activity of PNA- and vehicle-treated mice.

(2) The data and methods presented in this study could be valuable for the research community studying reproductive endocrinology and neuroendocrine disorders provided the authors address my comments above regarding the application of ML methods. The insights gained from this work could potentially inform clinical research aiming to develop better diagnostic and therapeutic strategies for PCOS.

Reviewer #3 (Public Review):

Summary:

Zhou and colleagues elegantly used pre-clinical mouse models to understand the nature of abnormally high GnRH/LH pulse secretion in polycystic ovary syndrome (PCOS), a major endocrine disorder affecting female fertility worldwide. This work brings a fundamental question of how altered gonadotropin secretion takes place upstream within the GnRH pulse generator core, which is defined by arcuate nucleus kisspeptin neurons.

Strengths:

The authors use state-of-the-art in vivo calcium imaging with fiber photometry and important physiological manipulations and measurements to dissect the possible neuronal mechanisms underlying such neuroendocrine derangements in PCOS. The additional use of unsupervised k-means clustering analysis for the evaluation of calcium synchronous events greatly enhances the quality of their evidence. The authors nicely propose that neuroendocrine dysfunction in PCOS might involve different setpoints through the hypothalamic-pituitary-gonadal (HPG) axis, and beyond kisspeptin neurons, which importantly pushes our field forward toward future investigations.

Weaknesses:

Although the authors provide important evidence, additional efforts are required to improve the quality of the manuscript and back up their claims. For instance, animal experiments failed to detect high testosterone levels in PNA female mice, a well-established PCOS mouse model. Considering that androgen excess is a hallmark of PCOS, this highly influences the subsequent evaluation of calcium synchronous events in arcuate kisspeptin neurons and the implications for neuroendocrine derangements.

Please see our response to Reviewer 1. It will be important to establish a robust PCOS mouse model in the future that has elevated pulse generator activity in the presence of elevated testosterone concentrations.

Authors also may need to provide LH data from another mouse model used in their work, the peripubertal androgen (PPA) model. Their claims seem to fall short without the pairing evidence of calcium synchronous events in arcuate kisspeptin neurons and LH pulse secretion.

We have demonstrated that ARN-KISS neuron SEs are perfectly correlated with pulsatile LH secretion in intact and gonadectomized male and female mice on many occasions. Given that the pulse generator frequency slows by 50% in PPA mice, it is very hard to imagine how this could result in an elevated LH pulse frequency. While we were undertaking these studies the first paper (to our knowledge) looking at pulsatile LH secretion in the PPA model was published; no change was found.

Another aspect that requires reviewing, is further exploration of their calcium synchronous events data and the increase of animal numbers in some of their experiments.

Please see below.

Reviewer #3 (Recommendations For The Authors):

The reviewer believes that this work will greatly contribute to the field and, to provide better manuscript quality, there might be only a few minor and major revisions to be included in the future version.

Minor:

(1) Line 17: I would change the sentence to "One in ten women in their reproductive age suffer from PCOS" to adapt to more accurate prevalence studies.

We have revised the sentence as recommended.

(2) Line 18 and 19: Although the evidence indeed points to a high LH pulse secretion in PCOS, I would change it to "with increased LH secretion" as most studies show mean values and not LH pulse release data.

While we agree that most human studies show a mean increase in LH, when assessed with sufficient temporal resolution, this results from elevated LH pulse frequency. As such, and to keep the manuscript focussed on the pulse generator, we would like the retain the present wording.

(3) Line 47: Please correct "polycystic ovaries" to polycystic-like ovarian morphology to adapt to the current AEPCOS guidelines.

We have revised the sentence as recommended.

(4) Line 231: Authors stated that "These PNA mice exhibited a cyclical pattern of activity similar to that of control mice" (Figure 5C and D). Please, include the statistical tests here for this claim. Although they say there aren't differences, the colored fields do not reflect this and seem quite different. Could the authors re-evaluate these claims or provide better examples in the figure?

We used Sidak’s multiple comparisons tests for this analysis (as stated in Results). The key data for assessing overall cyclical activity in PNA and control mice is Fig 5B which suggest very little difference. We accept that the individual traces of activity (Fig.5D) do not look identical to controls and, indeed, they are representative of the data set. The key point is they remain cyclical in an acyclic mouse. We have made sure that this is clear in the text.

(5) Subheadings 6 and & of the result section: It sounds confusing to read the foremost claims of the absence of SE differences and next have a clear SE frequency difference in Figures 6 C and D. The reviewer suggests that authors could reorganize the text and figures to make their rationale flow better for future readers.

We have considered this point carefully but find that re-organization creates its own problems with having to use the machine learning algorithm before describing it. It will always be problematic to incorporate this type of data-reanalysis in an original paper but think this present sequence is the best that can be achieved.

(6) Discussion: If PNA female mice did not have elevated testosterone levels, how can the authors compare their results to the current literature? Could this be the case for lacking a more robust ARNKISS neuronal activity output in their experiments? The reviewer recommends a better discussion concerning these aspects.

Please refer to our response to Reviewer #1 comment (2).

(7) Discussion: the authors claim that diestrous PNA mice exhibited highly variable patterns of ARNKISS neuron activity. Would these differences be due to different circulating sex steroid levels or intrinsic properties? Would the inclusion of future in vitro calcium imaging (brain slices) studies contribute to their research question and conclusions? The reviewer recommends a better discussion concerning these aspects.

We have tried to clarify that the highly variable patterns of activity in “diestrous” PNA mice come from the fact that we are actually randomly recording from ARN-KISS neurons at metestrus, diestrus, proestrus and estrus.  The pulse generator is cycling but we only have the acyclic “diestrous” smear to go by. This also makes brain slice studies difficult as we would never know the actual cycle stage.

Major:

(1) Results section: The reviewer strongly recommends that the LH pulse secretion data for the PPA group be included in the manuscript. If the SEs represent the central mechanism of pulse generation, would the LH pulse frequency match those events? If not, could a mismatch be explained by androgen-mediated negative feedback at the pituitary level? What is the pituitary LH response to exogenous GnRH (i.p. injection) in the PPA group?

Our initial observation showed the frequency of ARNKISS neuron SEs was halved in PPA mice compared to controls. Additionally, one study reported pulsatile LH secretion to be unchanged in this animal model (Coyle, Prescott et al. 2022). Both pieces of evidence clearly indicate that the PPA mouse does not provide an appropriate PCOS model of elevated pulse generator activity. Therefore, we do not see the value of pursuing further experiments in this animal model.

(2) Although the evaluation of relative frequency and normalized amplitude indicate the dynamic over time, the authors should include the average amplitudes and frequencies of events within the recording session. For instance, looking at Figures 1 A and B and Figures 3 A and B, a reader can observe differences in the amplitude due to different scaling axes. Perhaps, using a Python toolbox such as GuPPy or any preferred analysis pipeline might help authors include these parameters.

The amplitude of recorded SEs for each mouse depends primarily on the fiber position. As such, it has only ever been possible to assess SE amplitude changes within the same mouse. It is not possible to assess differences in SE amplitude between mice.

(3) Line 144-156: (Immunoreactivity results): Authors should proceed with caution when describing these results and clearly state that results show a software-based measurement of immunoreactive signal intensity. In addition, the small sample size of the PNA group (N = 4) compared to controls (N = 6-7) seems to mask possible differences. Could the authors increase the N of the PNA group and re-evaluate these results?

We have clarified that the immunoreactive signal intensity is based on software-based measurement. The N number for PNA mice in these studies varies from 4 to 6 depending on brain section availability for the different immunohistochemistry runs. The scatter of data is such that any new data points would need to be at the extreme of the distributions to likely have any impact on statistical significance. As a minor part of the paper, we did not feel that the use of further mice was warranted.

(4) Considering the great variability of PNA's number of SE/hr, the review suggests increasing the N in this group, thus, authors can re-evaluate their findings and draw better analysis/ conclusion.

We have n=6 for the PNA group in the study. As noted above, the variability in SE/hr in Figure 3 comes from assessing the pulse generator at random times within the estrous cycle. Once we separate “diestrous-like” stage for the PNA animals, the variability is decreased as shown in Figure 6.

to - super browser paradigm web applications

os killer

eLife Assessment

This paper presents a new method called MINT that is effective at BCI-style decoding tasks. The authors show convincing evidence to support their claims regarding how MINT is a new method that produces excellent decoding performance relative to the state-of-the-art. This work is important and will be of broad interest to neuroscientists and neuroengineers.

Reviewer #1 (Public Review):

Summary:

This paper presents an innovative decoding approach for brain-computer interfaces (BCIs), introducing a new method named MINT. The authors develop a trajectory-centric approach to decode behaviors across several different datasets, including eight empirical datasets from the Neural Latents Benchmark. Overall, the paper is well written and their method shows impressive performance compared to more traditional decoding approaches that use a simpler approach. While there are some concerns (see below), the paper's strengths, particularly its emphasis on a trajectory-centric approach and the simplicity of MINT, provide a compelling contribution to the field.

Strengths:

The adoption of a trajectory-centric approach that utilizes statistical constraints presents a substantial shift in methodology, potentially revolutionizing the way BCIs interpret and predict neural behaviour. This is one of the strongest aspects of the paper.

The thorough evaluation of the method across various datasets serves as an assurance that the superior performance of MINT is not a result of overfitting. The comparative simplicity of the method in contrast to many neural network approaches is refreshing and should facilitate broader applicability.

Weaknesses:

Scope: Despite the impressive performance of MINT across multiple datasets, it seems predominantly applicable to M1/S1 data. Only one of the eight empirical datasets comes from an area outside the motor/somatosensory cortex. It would be beneficial if the authors could expand further on how the method might perform with other brain regions that do not exhibit low tangling or do not have a clear trial structure (e.g. decoding of position or head direction from hippocampus)

When comparing methods, the neural trajectories of MINT are based on averaged trials, while the comparison methods are trained on single trials. An additional analysis might help in disentangling the effect of the trial averaging. For this, the authors could average the input across trials for all decoders, establishing a baseline for averaged trials. Note that inference should still be done on single trials. Performance can then be visualized across different values of N, which denotes the number of averaged trials used for training.

Comments on revisions:

I have looked at the responses and they are thorough and answer all of my questions.

Reviewer #2 (Public Review):

Summary:

The goal of this paper is to present a new method, termed MINT, for decoding behavioral states from neural spiking data. MINT is a statistical method which, in addition to outputting a decoded behavioral state, also provides soft information regarding the likelihood of that behavioral state based on the neural data. The innovation in this approach is neural states are assumed to come from sparsely distributed neural trajectories with low tangling, meaning that neural trajectories (time sequences of neural states) are sparse in the high-dimensional space of neural spiking activity and that two dissimilar neural trajectories tend to correspond to dissimilar behavioral trajectories. The authors support these assumptions through analysis of previously collected data, and then validate the performance of their method by comparing it to a suite of alternative approaches. The authors attribute the typically improved decoding performance by MINT to its assumptions being more faithfully aligned to the properties of neural spiking data relative to assumptions made by the alternatives.

Strengths:

The paper did an excellent job critically evaluating common assumptions made by neural analytical methods, such as neural state being low-dimensional relative to the number of recorded neurons. The authors made strong arguments, supported by evidence and literature, for potentially high-dimensional neural states and thus the need for approaches that do not rely on an assumption of low dimensionality.

The paper was thorough in considering multiple datasets across a variety of behaviors, as well as existing decoding methods, to benchmark the MINT approach. This provided a valuable comparison to validate the method. The authors also provided nice intuition regarding why MINT may offer performance improvement in some cases and in which instances MINT may not perform as well.

In addition to providing a philosophical discussion as to the advantages of MINT and benchmarking against alternatives, the authors also provided a detailed description of practical considerations. This included training time, amount of training data, robustness to data loss or changes in the data, and interpretability. These considerations not only provided objective evaluation of practical aspects but also provided insights to the flexibility and robustness of the method as they relate back to the underlying assumptions and construction of the approach.

Impact:

This work is motivated by brain-computer interfaces applications, which it will surely impact in terms of neural decoder design. However, this work is also broadly impactful for neuroscientific analysis to relate neural spiking activity to observable behavioral features. Thus, MINT will likely impact neuroscience research generally. The methods are made publicly available, and the datasets used are all in public repositories, which facilitates adoption and validation of this method within the greater scientific community.

Author response:

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

Summary of reviewers’ comments and our revisions: 

We thank the reviewers for their thoughtful feedback. This feedback has motivated multiple revisions and additions that, in our view, have greatly improved the manuscript. This is especially true with regard to a major goal of this study: clearly defining existing scientific perspectives and delineating their decoding implications. In addition to building on this conceptual goal, we have expanded existing analyses and have added a new analysis of generalization using a newly collected dataset. We expect the manuscript will be of very broad interest, both to those interested in BCI development and to those interested in fundamental properties of neural population activity and its relationship with behavior.

Importantly, all reviewers were convinced that MINT provided excellent performance, when benchmarked against existing methods, across a broad range of standard tasks:

“their method shows impressive performance compared to more traditional decoding approaches” (R1) 

“The paper was thorough in considering multiple datasets across a variety of behaviors, as well as existing decoding methods, to benchmark the MINT approach. This provided a valuable comparison to validate the method.” (R2) 

“The fact that performance on stereotyped tasks is high is interesting and informative…” (R3)

This is important. It is challenging to design a decoder that performs consistently across multiple domains and across multiple situations (including both decoding and neural state estimation). MINT does so. MINT consistently outperformed existing lightweight ‘interpretable’ decoders, despite being a lightweight interpretable decoder itself. MINT was very competitive with expressive machine-learning methods, yet has advantages in flexibility and simplicity that more ‘brute force’ methods do not. We made a great many comparisons, and MINT was consistently a strong performer. Of the many comparisons we made, there was only one where MINT was at a modest disadvantage, and it was for a dataset where all methods performed poorly. No other method we tested was as consistent. For example, although the GRU and the feedforward network were often competitive with MINT (and better than MINT in the one case mentioned above), there were multiple other situations where they performed less well and a few situations where they performed poorly. Moreover, no other existing decoder naturally estimates the neural state while also readily decoding, without retraining, a broad range of behavioral variables.

R1 and R2 were very positive about the broader impacts of the study. They stressed its impact both on decoder design, and on how our field thinks, scientifically, about the population response in motor areas: 

“This paper presents an innovative decoding approach for brain-computer interfaces” (R1)

“presents a substantial shift in methodology, potentially revolutionizing the way BCIs interpret and predict neural behaviour” (R1)

“the paper's strengths, particularly its emphasis on a trajectory-centric approach and the simplicity of MINT, provide a compelling contribution to the field” (R1)

“The authors made strong arguments, supported by evidence and literature, for potentially high-dimensional neural states and thus the need for approaches that do not rely on an assumption of low dimensionality” (R2)

“This work is motivated by brain-computer interfaces applications, which it will surely impact in terms of neural decoder design.” (R2)

“this work is also broadly impactful for neuroscientific analysis... Thus, MINT will likely impact neuroscience research generally.” (R2)

We agree with these assessments, and have made multiple revisions to further play into these strengths. As one example, the addition of Figure 1b (and 6b) makes this the first study, to our knowledge, to fully and concretely illustrate this emerging scientific perspective and its decoding implications. This is important, because multiple observations convince us that the field is likely to move away from the traditional perspective in Figure 1a, and towards that in Figure 1b. We also agree with the handful of weaknesses R1 and R2 noted. The manuscript has been revised accordingly. The major weakness noted by R1 was the need to be explicit regarding when we suspect MINT would (and wouldn’t) work well in other brain areas. In non-motor areas, the structure of the data may be poorly matched with MINT’s assumptions. We agree that this is likely to be true, and thus agree with the importance of clarifying this topic for the reader. The revision now does so. R1 also wished to know whether existing methods might benefit from including trial-averaged data during training, something we now explore and document (see detailed responses below). R2 noted two weaknesses: 1) The need to better support (with expanded analysis) the statement that neural and behavioral trajectories are non-isometric, and 2) The need to more rigorously define the ‘mesh’. We agree entirely with both suggestions, and the revision has been strengthened by following them (see detailed responses below).

R3 also saw strengths to the work, stating that:

“This paper is well-structured and its main idea is clear.” 

“The fact that performance on stereotyped tasks is high is interesting and informative, showing that these stereotyped tasks create stereotyped neural trajectories.” 

“The task-specific comparisons include various measures and a variety of common decoding approaches, which is a strength.”

However, R3 also expressed two sizable concerns. The first is that MINT might have onerous memory requirements. The manuscript now clarifies that MINT has modest memory requirements. These do not scale unfavorably as the reviewer was concerned they might. The second concern is that MINT is: 

“essentially a table-lookup rather than a model.”

Although we don’t agree, the concern makes sense and may be shared by many readers, especially those who take a particular scientific perspective. Pondering this concern thus gave us the opportunity to modify the manuscript in ways that support its broader impact. Our revisions had two goals: 1) clarify the ways in which MINT is far more flexible than a lookup-table, and 2) better describe the dominant scientific perspectives and their decoding implications.

The heart of R3’s concern is the opinion that MINT is an effective but unprincipled hack suitable for situations where movements are reasonably stereotyped. Of course, many tasks involve stereotyped movements (e.g. handwriting characters), so MINT would still be useful. Nevertheless, if MINT is not principled, other decode methods would often be preferable because they could (unlike MINT in R3’s opinion) gain flexibility by leveraging an accurate model. Most of R3’s comments flow from this fundamental concern: 

“This is again due to MINT being a lookup table with a library of stereotyped trajectories rather than a model.”

“MINT models task-dependent neural trajectories, so the trained decoder is very task-dependent and cannot generalize to other tasks.”

“Unlike MINT, these works can achieve generalization because they model the neural subspace and its association to movement.”

“given that MINT tabulates task-specific trajectories, it will not generalize to tasks that are not seen in the training data even when these tasks cover the exact same space (e.g., the same 2D computer screen and associated neural space).”

“For proper training, the training data should explore the whole movement space and the associated neural space, but this does not mean all kinds of tasks performed in that space must be included in the training set (something MINT likely needs while modeling-based approaches do not).”

The manuscript has been revised to clarify that MINT is considerably more flexible than a lookup table, even though a lookup table is used as a first step. Yet, on its own, this does not fully address R3’s concern. The quotes above highlight that R3 is making a standard assumption in our field: that there exists a “movement space and associated neural space”. Under this perspective, one should, as R3 argues fully explore the movement space. This would perforce fully explore the associated neural subspace. One can then “model the neural subspace and its association to movement”. MINT does not use a model of this type, and thus (from R3’s perspective) does not appear to use a model at all. A major goal of our study is to question this traditional perspective. We have thus added a new figure to highlight the contrast between the traditional (Figure 1a) and new (Figure 1b) scientific perspectives, and to clarify their decoding implications.

While we favor the new perspective (Figure 1b), we concede that R3 may not share our view. This is fine. Part of the reason we believe this study is timely, and will be broadly read, is that it raises a topic of emerging interest where there is definitely room for debate. If we are misguided – i.e. if Figure 1a is the correct perspective – then many of R3’s concerns would be on target: MINT could still be useful, but traditional methods that make the traditional assumptions in Figure 1a would often be preferable. However, if the emerging perspective in Figure 1b is more accurate, then MINT’s assumptions would be better aligned with the data than those of traditional methods, making it a more (not less) principled choice.

Our study provides new evidence in support of Figure 1b, while also synthesizing existing evidence from other recent studies. In addition to Figure 2, the new analysis of generalization further supports Figure 1b. Also supporting Figure 1b is the analysis in which MINT’s decoding advantage, over a traditional decoder, disappears when simulated data approximate the traditional perspective in Figure 1a.

That said, we agree that the present study cannot fully resolve whether Figure 1a or 1b is more accurate. Doing so will take multiple studies with different approaches (indeed we are currently preparing other manuscripts on this topic). Yet we still have an informed scientific opinion, derived from past, present and yet-to-be-published observations. Our opinion is that Figure 1b is the more accurate perspective. This possibility makes it reasonable to explore the potential virtues of a decoding method whose assumptions are well-aligned with that perspective. MINT is such a method. As expected under Figure 1b, MINT outperforms traditional interpretable decoders in every single case we studied. 

As noted above, we have added a new generalization-focused analysis (Figure 6) based on a newly collected dataset. We did so because R3’s comments highlight a deep point: which scientific perspective one takes has strong implications regarding decoder generalization. These implications are now illustrated in the new Figure 6a and 6b. Under Figure 6a, it is possible, as R3 suggests, to explore “the whole movement space and associated neural space” during training. However, under Figure 6b, expectations are very different. Generalization will be ‘easy’ when new trajectories are near the training-set trajectories. In this case, MINT should generalize well as should other methods. In contrast, generalization will be ‘hard’ when new neural trajectories have novel shapes and occupy previously unseen regions / dimensions. In this case, all current methods, including MINT, are likely to fail. R3 points out that traditional decoders have sometimes generalized well to new tasks (e.g. from center-out to ‘pinball’) when cursor movements occur in the same physical workspace. These findings could be taken to support Figure 6a, but are equally consistent with ‘easy’ generalization in Figure 6b. To explore this topic, the new analysis in Figure 6c-g considers conditions that are intended to span the range from easy to hard. Results are consistent with the predictions of Figure 6b. 

We believe the manuscript has been significantly improved by these additions. The revisions help the manuscript achieve its twin goals: 1) introduce a novel class of decoder that performs very well despite being very simple, and 2) describe properties of motor-cortex activity that will matter for decoders of all varieties.

Reviewer #1: 

Summary: 

This paper presents an innovative decoding approach for brain-computer interfaces (BCIs), introducing a new method named MINT. The authors develop a trajectory-centric approach to decode behaviors across several different datasets, including eight empirical datasets from the Neural Latents Benchmark. Overall, the paper is well written and their method shows impressive performance compared to more traditional decoding approaches that use a simpler approach. While there are some concerns (see below), the paper's strengths, particularly its emphasis on a trajectory-centric approach and the simplicity of MINT, provide a compelling contribution to the field. 

We thank the reviewer for these comments. We share their enthusiasm for the trajectory-centric approach, and we are in complete agreement that this perspective has both scientific and decoding implications. The revision expands upon these strengths.

Strengths: 

The adoption of a trajectory-centric approach that utilizes statistical constraints presents a substantial shift in methodology, potentially revolutionizing the way BCIs interpret and predict neural behaviour. This is one of the strongest aspects of the paper. 

Again, thank you. We also expect the trajectory-centric perspective to have a broad impact, given its relevance to both decoding and to thinking about manifolds.

The thorough evaluation of the method across various datasets serves as an assurance that the superior performance of MINT is not a result of overfitting. The comparative simplicity of the method in contrast to many neural network approaches is refreshing and should facilitate broader applicability. 

Thank you. We were similarly pleased to see such a simple method perform so well. We also agree that, while neural-network approaches will always be important, it is desirable to also possess simple ‘interpretable’ alternatives.

Weaknesses:  

Comment 1) Scope: Despite the impressive performance of MINT across multiple datasets, it seems predominantly applicable to M1/S1 data. Only one of the eight empirical datasets comes from an area outside the motor/somatosensory cortex. It would be beneficial if the authors could expand further on how the method might perform with other brain regions that do not exhibit low tangling or do not have a clear trial structure (e.g. decoding of position or head direction from hippocampus) 

We agree entirely. Population activity in many brain areas (especially outside the motor system) presumably will often not have the properties upon which MINT’s assumptions are built. This doesn’t necessarily mean that MINT would perform badly. Using simulated data, we have found that MINT can perform surprisingly well even when some of its assumptions are violated. Yet at the same time, when MINT’s assumptions don’t apply, one would likely prefer to use other methods. This is, after all, one of the broader themes of the present study: it is beneficial to match decoding assumptions to empirical properties. We have thus added a section on this topic early in the Discussion: 

“In contrast, MINT and the Kalman filter performed comparably on simulated data that better approximated the assumptions in Figure 1a. Thus, MINT is not a ‘better’ algorithm – simply better aligned with the empirical properties of motor cortex data. This highlights an important caveat. Although MINT performs well when decoding from motor areas, its assumptions may be a poor match in other areas (e.g. the hippocampus). MINT performed well on two non-motor-cortex datasets – Area2_Bump (S1) and DMFC_RSG (dorsomedial frontal cortex) – yet there will presumably be other brain areas and/or contexts where one would prefer a different method that makes assumptions appropriate for that area.”

Comment 2) When comparing methods, the neural trajectories of MINT are based on averaged trials, while the comparison methods are trained on single trials. An additional analysis might help in disentangling the effect of the trial averaging. For this, the authors could average the input across trials for all decoders, establishing a baseline for averaged trials. Note that inference should still be done on single trials. Performance can then be visualized across different values of N, which denotes the number of averaged trials used for training. 

We explored this question and found that the non-MINT decoders are harmed, not helped, by the inclusion of trial-averaged responses in the training set. This is presumably because the statistics of trialaveraged responses don’t resemble what will be observed during decoding. This statistical mismatch, between training and decoding, hurts most methods. It doesn’t hurt MINT, because MINT doesn’t ‘train’ in the normal way. It simply needs to know rates, and trial-averaging is a natural way to obtain them. To describe the new analysis, we have added the following to the text.

“We also investigated the possibility that MINT gained its performance advantage simply by having access to trial-averaged neural trajectories during training, while all other methods were trained on single-trial data. This difference arises from the fundamental requirements of the decoder architectures: MINT needs to estimate typical trajectories while other methods don’t. Yet it might still be the case that other methods would benefit from including trial-averaged data in the training set, in addition to single-trial data. Alternatively, this might harm performance by creating a mismatch, between training and decoding, in the statistics of decoder inputs. We found that the latter was indeed the case: all non-MINT methods performed better when trained purely on single-trial data.”

Reviewer #2:

Summary: 

The goal of this paper is to present a new method, termed MINT, for decoding behavioral states from neural spiking data. MINT is a statistical method which, in addition to outputting a decoded behavioral state, also provides soft information regarding the likelihood of that behavioral state based on the neural data. The innovation in this approach is neural states are assumed to come from sparsely distributed neural trajectories with low tangling, meaning that neural trajectories (time sequences of neural states) are sparse in the high-dimensional space of neural spiking activity and that two dissimilar neural trajectories tend to correspond to dissimilar behavioral trajectories. The authors support these assumptions through analysis of previously collected data, and then validate the performance of their method by comparing it to a suite of alternative approaches. The authors attribute the typically improved decoding performance by MINT to its assumptions being more faithfully aligned to the properties of neural spiking data relative to assumptions made by the alternatives. 

We thank the reviewer for this accurate summary, and for highlighting the subtle but important fact that MINT provides information regarding likelihoods. The revision includes a new analysis (Figure 6e) illustrating one potential way to leverage knowledge of likelihoods.

Strengths:  

The paper did an excellent job critically evaluating common assumptions made by neural analytical methods, such as neural state being low-dimensional relative to the number of recorded neurons. The authors made strong arguments, supported by evidence and literature, for potentially high-dimensional neural states and thus the need for approaches that do not rely on an assumption of low dimensionality. 

Thank you. We also hope that the shift in perspective is the most important contribution of the study. This shift matters both scientifically and for decoder design. The revision expands on this strength. The scientific alternatives are now more clearly and concretely illustrated (especially see Figure 1a,b and Figure 6a,b). We also further explore their decoding implications with new data (Figure 6c-g).

The paper was thorough in considering multiple datasets across a variety of behaviors, as well as existing decoding methods, to benchmark the MINT approach. This provided a valuable comparison to validate the method. The authors also provided nice intuition regarding why MINT may offer performance improvement in some cases and in which instances MINT may not perform as well. 

Thank you. We were pleased to be able to provide comparisons across so many datasets (we are grateful to the Neural Latents Benchmark for making this possible).

In addition to providing a philosophical discussion as to the advantages of MINT and benchmarking against alternatives, the authors also provided a detailed description of practical considerations. This included training time, amount of training data, robustness to data loss or changes in the data, and interpretability. These considerations not only provided objective evaluation of practical aspects but also provided insights to the flexibility and robustness of the method as they relate back to the underlying assumptions and construction of the approach. 

Thank you. We are glad that these sections were appreciated. MINT’s simplicity and interpretability are indeed helpful in multiple ways, and afford opportunities for interesting future extensions. One potential benefit of interpretability is now explored in the newly added Figure 6e. 

Impact: 

This work is motivated by brain-computer interfaces applications, which it will surely impact in terms of neural decoder design. However, this work is also broadly impactful for neuroscientific analysis to relate neural spiking activity to observable behavioral features. Thus, MINT will likely impact neuroscience research generally. The methods are made publicly available, and the datasets used are all in public repositories, which facilitates adoption and validation of this method within the greater scientific community. 

Again, thank you. We have similar hopes for this study.

Weaknesses (1 & 2 are related, and we have switched their order in addressing them): 

Comment 2) With regards to the idea of neural and behavioral trajectories having different geometries, this is dependent on what behavioral variables are selected. In the example for Fig 2a, the behavior is reach position. The geometry of the behavioral trajectory of interest would look different if instead the behavior of interest was reach velocity. The paper would be strengthened by acknowledgement that geometries of trajectories are shaped by extrinsic choices rather than (or as much as they are) intrinsic properties of the data. 

We agree. Indeed, we almost added a section to the original manuscript on this exact topic. We have now done so:

“A potential concern regarding the analyses in Figure 2c,d is that they require explicit choices of behavioral variables: muscle population activity in Figure 2c and angular phase and velocity in Figure 2d. Perhaps these choices were misguided. Might neural and behavioral geometries become similar if one chooses ‘the right’ set of behavioral variables? This concern relates to the venerable search for movement parameters that are reliably encoded by motor cortex activity [69, 92–95]. If one chooses the wrong set of parameters (e.g. chooses muscle activity when one should have chosen joint angles) then of course neural and behavioral geometries will appear non-isometric. There are two reasons why this ‘wrong parameter choice’ explanation is unlikely to account for the results in Figure 2c,d. First, consider the implications of the left-hand side of Figure 2d. A small kinematic distance implies that angular position and velocity are nearly identical for the two moments being compared. Yet the corresponding pair of neural states can be quite distant. Under the concern above, this distance would be due to other encoded behavioral variables – perhaps joint angle and joint velocity – differing between those two moments. However, there are not enough degrees of freedom in this task to make this plausible. The shoulder remains at a fixed position (because the head is fixed) and the wrist has limited mobility due to the pedal design [60]. Thus, shoulder and elbow angles are almost completely determined by cycle phase. More generally, ‘external variables’ (positions, angles, and their derivatives) are unlikely to differ more than slightly when phase and angular velocity are matched. Muscle activity could be different because many muscles act on each joint, creating redundancy. However, as illustrated in Figure 2c, the key effect is just as clear when analyzing muscle activity. Thus, the above concern seems unlikely even if it can’t be ruled out entirely. A broader reason to doubt the ‘wrong parameter choice’ proposition is that it provides a vague explanation for a phenomenon that already has a straightforward explanation. A lack of isometry between the neural population response and behavior is expected when neural-trajectory tangling is low and output-null factors are plentiful [55, 60]. For example, in networks that generate muscle activity, neural and muscle-activity trajectories are far from isometric [52, 58, 60]. Given this straightforward explanation, and given repeated failures over decades to find the ‘correct’ parameters (muscle activity, movement direction, etc.) that create neural-behavior isometry, it seems reasonable to conclude that no such isometry exists.”

Comment 1) The authors posit that neural and behavioral trajectories are non-isometric. To support this point, they look at distances between neural states and distances between the corresponding behavioral states, in order to demonstrate that there are differences in these distances in each respective space. This supports the idea that neural states and behavioral states are non-isometric but does not directly address their point. In order to say the trajectories are non-isometric, it would be better to look at pairs of distances between corresponding trajectories in each space. 

We like this idea and have added such an analysis. To be clear, we like the original analysis too: isometry predicts that neural and behavioral distances (for corresponding pairs of points) should be strongly correlated, and that small behavioral distances should not be associated with large neural distances. These predictions are not true, providing a strong argument against isometry. However, we also like the reviewer’s suggestion, and have added such an analysis. It makes the same larger point, and also reveals some additional facts (e.g. it reveals that muscle-geometry is more related to neural-geometry than is kinematic-geometry). The new analysis is described in the following section:

“We further explored the topic of isometry by considering pairs of distances. To do so, we chose two random neural states and computed their distance, yielding dneural1. We repeated this process, yielding dneural2. We then computed the corresponding pair of distances in muscle space (dmuscle1 and dmuscle2) and kinematic space (dkin1 and dkin2). We considered cases where dneural1 was meaningfully larger than (or smaller than) dneural2, and asked whether the behavioral variables had the same relationship; e.g. was dmuscle1 also larger than dmuscle2? For kinematics, this relationship was weak: across 100,000 comparisons, the sign of dkin1 − dkin2 agreed with dneural1 − dneural2 only 67.3% of the time (with 50% being chance). The relationship was much stronger for muscles: the sign of dmuscle1 − dmuscle2 agreed with dneural1 − dneural2 79.2% of the time, which is far more than expected by chance yet also far from what is expected given isometry (e.g. the sign agrees 99.7% of the time for the truly isometric control data in Figure 2e). Indeed there were multiple moments during this task when dneural1 was much larger than dneural2, yet dmuscle1 was smaller than dmuscle2. These observations are consistent with the proposal that neural trajectories resemble muscle trajectories in some dimensions, but with additional output-null dimensions that break the isometry [60].”

Comment 3) The approach is built up on the idea of creating a "mesh" structure of possible states. In the body of the paper the definition of the mesh was not entirely clear and I could not find in the methods a more rigorous explicit definition. Since the mesh is integral to the approach, the paper would be improved with more description of this component. 

This is a fair criticism. Although MINTs actual operations were well-documented, how those operations mapped onto the term ‘mesh’ was, we agree, a bit vague. The definition of the mesh is a bit subtle because it only emerges during decoding rather than being precomputed. This is part of what gives MINT much more flexibility than a lookup table. We have added the following to the manuscript.

“We use the term ‘mesh’ to describe the scaffolding created by the training-set trajectories and the interpolated states that arise at runtime. The term mesh is apt because, if MINT’s assumptions are correct, interpolation will almost always be local. If so, the set of decodable states will resemble a mesh, created by line segments connecting nearby training-set trajectories. However, this mesh-like structure is not enforced by MINT’s operations.

Interpolation could, in principle, create state-distributions that depart from the assumption of a sparse manifold. For example, interpolation could fill in the center of the green tube in Figure 1b, resulting in a solid manifold rather than a mesh around its outer surface. However, this would occur only if spiking observations argued for it. As will be documented below, we find that essentially all interpolation is local”

We have also added Figure 4d. This new analysis documents the fact that decoded states are near trainingset trajectories, which is why the term ‘mesh’ is appropriate.

Reviewer #3:

Summary:  

This manuscript develops a new method termed MINT for decoding of behavior. The method is essentially a table-lookup rather than a model. Within a given stereotyped task, MINT tabulates averaged firing rate trajectories of neurons (neural states) and corresponding averaged behavioral trajectories as stereotypes to construct a library. For a test trial with a realized neural trajectory, it then finds the closest neural trajectory to it in the table and declares the associated behavior trajectory in the table as the decoded behavior. The method can also interpolate between these tabulated trajectories. The authors mention that the method is based on three key assumptions: (1) Neural states may not be embedded in a lowdimensional subspace, but rather in a high-dimensional space. (2) Neural trajectories are sparsely distributed under different behavioral conditions. (3) These neural states traverse trajectories in a stereotyped order.  

The authors conducted multiple analyses to validate MINT, demonstrating its decoding of behavioral trajectories in simulations and datasets (Figures 3, 4). The main behavior decoding comparison is shown in Figure 4. In stereotyped tasks, decoding performance is comparable (M_Cycle, MC_Maze) or better (Area 2_Bump) than other linear/nonlinear algorithms

(Figure 4). However, MINT underperforms for the MC_RTT task, which is less stereotyped (Figure 4).  

This paper is well-structured and its main idea is clear. The fact that performance on stereotyped tasks is high is interesting and informative, showing that these stereotyped tasks create stereotyped neural trajectories. The task-specific comparisons include various measures and a variety of common decoding approaches, which is a strength. However, I have several major concerns. I believe several of the conclusions in the paper, which are also emphasized in the abstract, are not accurate or supported, especially about generalization, computational scalability, and utility for BCIs. MINT is essentially a table-lookup algorithm based on stereotyped task-dependent trajectories and involves the tabulation of extensive data to build a vast library without modeling. These aspects will limit MINT's utility for real-world BCIs and tasks. These properties will also limit MINT's generalizability from task to task, which is important for BCIs and thus is commonly demonstrated in BCI experiments with other decoders without any retraining. Furthermore, MINT's computational and memory requirements can be prohibitive it seems. Finally, as MINT is based on tabulating data without learning models of data, I am unclear how it will be useful in basic investigations of neural computations. I expand on these concerns below.  

We thank the reviewer for pointing out weaknesses in our framing and presentation. The comments above made us realize that we needed to 1) better document the ways in which MINT is far more flexible than a lookup-table, and 2) better explain the competing scientific perspectives at play. R3’s comments also motivated us to add an additional analysis of generalization. In our view the manuscript is greatly improved by these additions. Specifically, these additions directly support the broader impact that we hope the study will have.

For simplicity and readability, we first group and summarize R3’s main concerns in order to better address them. (These main concerns are all raised above, in addition to recurring in the specific comments below. Responses to each individual specific comment are provided after these summaries.)

(1) R3 raises concerns about ‘computational scalability.’ The concern is that “MINT's computational and memory requirements can be prohibitive.” This point was expanded upon in a specific comment, reproduced below:

I also find the statement in the abstract and paper that "computations are simple, scalable" to be inaccurate. The authors state that MINT's computational cost is O(NC) only, but it seems this is achieved at a high memory cost as well as computational cost in training. The process is described in section "Lookup table of log-likelihoods" on line [978-990]. The idea is to precompute the log-likelihoods for any combination of all neurons with discretization x all delay/history segments x all conditions and to build a large lookup table for decoding. Basically, the computational cost of precomputing this table is O(V^{Nτ} x TC) and the table requires a memory of O(V^{Nτ}), where V is the number of discretization points for the neural firing rates, N is the number of neurons, τ is the history length, T is the trial length, and C is the number of conditions. This is a very large burden, especially the V^{Nτ} term. This cost is currently not mentioned in the manuscript and should be clarified in the main text. Accordingly, computation claims should be modified including in the abstract.

The revised manuscript clarifies that our statement (that computations are simple and scalable) is absolutely accurate. There is no need to compute, or store, a massive lookup table. There are three tables: two of modest size and one that is tiny. This is now better explained:

“Thus, the log-likelihood of , for a particular current neural state, is simply the sum of many individual log-likelihoods (one per neuron and time-bin). Each individual log-likelihood depends on only two numbers: the firing rate at that moment and the spike count in that bin. To simplify online computation, one can precompute the log-likelihood, under a Poisson model, for every plausible combination of rate and spike-count. For example, a lookup table of size 2001 × 21 is sufficient when considering rates that span 0-200 spikes/s in increments of 0.1 spikes/s, and considering 20 ms bins that contain at most 20 spikes (only one lookup table is ever needed, so long as its firing-rate range exceeds that of the most-active neuron at the most active moment in Ω). Now suppose we are observing a population of 200 neurons, with a 200 ms history divided into ten 20 ms bins. For each library state, the log-likelihood of the observed spike-counts is simply the sum of 200 × 10 = 2000 individual loglikelihoods, each retrieved from the lookup table. In practice, computation is even simpler because many terms can be reused from the last time bin using a recursive solution (Methods). This procedure is lightweight and amenable to real-time applications.”

In summary, the first table simply needs to contain the firing rate of each neuron, for each condition, and each time in that condition. This table consumes relatively little memory. Assuming 100 one-second-long conditions (rates sampled every 20 ms) and 200 neurons, the table would contain 100 x 50 x 200 = 1,000,000 numbers. These numbers are typically stored as 16-bit integers (because rates are quantized), which amounts to about 2 MB. This is modest, given that most computers have (at least) tens of GB of RAM. A second table would contain the values for each behavioral variable, for each condition, and each time in that condition. This table might contain behavioral variables at a finer resolution (e.g. every millisecond) to enable decoding to update in between 20 ms bins (1 ms granularity is not needed for most BCI applications, but is the resolution used in this study). The number of behavioral variables of interest for a particular BCI application is likely to be small, often 1-2, but let’s assume for this example it is 10 (e.g. x-, y-, and z-position, velocity, and acceleration of a limb, plus one other variable). This table would thus contain 100 x 1000 x 10 = 1,000,000 floating point numbers, i.e. an 8 MB table. The third table is used to store the probability of s spikes being observed given a particular quantized firing rate (e.g. it may contain probabilities associated with firing rates ranging from 0 – 200 spikes/s in 0.1 spikes/s increments). This table is not necessary, but saves some computation time by precomputing numbers that will be used repeatedly. This is a very small table (typically ~2000 x 20, i.e. 320 KB). It does not need to be repeated for different neurons or conditions, because Poisson probabilities depend on only rate and count.

(2) R3 raises a concern that MINT “is essentially a table-lookup rather than a model.’ R3 states that MINT 

“is essentially a table-lookup algorithm based on stereotyped task-dependent trajectories and involves the tabulation of extensive data to build a vast library without modeling.”

and that,

“as MINT is based on tabulating data without learning models of data, I am unclear how it will be useful in basic investigations of neural computations.”

This concern is central to most subsequent concerns. The manuscript has been heavily revised to address it. The revisions clarify that MINT is much more flexible than a lookup table, even though MINT uses a lookup table as its first step. Because R3’s concern is intertwined with one’s scientific assumptions, we have also added the new Figure 1 to explicitly illustrate the two key scientific perspectives and their decoding implications. 

Under the perspective in Figure 1a, R3 would be correct in saying that there exist traditional interpretable decoders (e.g. a Kalman filter) whose assumptions better model the data. Under this perspective, MINT might still be an excellent choice in many cases, but other methods would be expected to gain the advantage when situations demand more flexibility. This is R3’s central concern, and essentially all other concerns flow from it. It makes sense that R3 has this concern, because their comments repeatedly stress a foundational assumption of the perspective in Figure 1a: the assumption of a fixed lowdimensional neural subspace where activity has a reliable relationship to behavior that can be modeled and leveraged during decoding. The phrases below accord with that view:

“Unlike MINT, these works can achieve generalization because they model the neural subspace and its association to movement.”

“it will not generalize… even when these tasks cover the exact same space (e.g., the same 2D computer screen and associated neural space).”

“For proper training, the training data should explore the whole movement space and the associated neural space”

“I also believe the authors should clarify the logic behind developing MINT better. From a scientific standpoint, we seek to gain insights into neural computations by making various assumptions and building models that parsimoniously describe the vast amount of neural data rather than simply tabulating the data. For instance, low-dimensional assumptions have led to the development of numerous dimensionality reduction algorithms and these models have led to important interpretations about the underlying dynamics”

Thus, R3 prefers a model that 1) assumes a low-dimensional subspace that is fixed across tasks and 2) assumes a consistent ‘association’ between neural activity and kinematics. Because R3 believes this is the correct model of the data, they believe that decoders should leverage it. Traditional interpretable method do, and MINT doesn’t, which is why they find MINT to be unprincipled. This is a reasonable view, but it is not our view. We have heavily revised the manuscript to clarify that a major goal of our study is to explore the implications of a different, less-traditional scientific perspective.

The new Figure 1a illustrates the traditional perspective. Under this perspective, one would agree with R3’s claim that other methods have the opportunity to model the data better. For example, suppose there exists a consistent neural subspace – conserved across tasks – where three neural dimensions encode 3D hand position and three additional neural dimensions encode 3D hand velocity. A traditional method such as a Kalman filter would be a very appropriate choice to model these aspects of the data.

Figure 1b illustrates the alternative scientific perspective. This perspective arises from recent, present, and to-be-published observations. MINT’s assumptions are well-aligned with this perspective. In contrast, the assumptions of traditional methods (e.g. the Kalman filter) are not well-aligned with the properties of the data under this perspective. This does not mean traditional methods are not useful. Yet under Figure 1b, it is traditional methods, such as the Kalman filter, that lack an accurate model of the data. Of course, the reviewer may disagree with our scientific perspective. We would certainly concede that there is room for debate. However, we find the evidence for Figure 1b to be sufficiently strong that it is worth exploring the utility of methods that align with this scientific perspective. MINT is such a method. As we document, it performs very well.

Thus, in our view, MINT is quite principled because its assumptions are well aligned with the data. It is true that the features of the data that MINT models are a bit different from those that are traditionally modeled. For example, R3 is quite correct that MINT does not attempt to use a biomimetic model of the true transformation from neural activity, to muscle activity, and thence to kinematics. We see this as a strength, and the manuscript has been revised accordingly (see paragraph beginning with “We leveraged this simulated data to compare MINT with a biomimetic decoder”).

(3) R3 raises concerns that MINT cannot generalize. This was a major concern of R3 and is intimately related to concern #2 above. The concern is that, if MINT is “essentially a lookup table” that simply selects pre-defined trajectories, then MINT will not be able to generalize. R3 is quite correct that MINT generalizes rather differently than existing methods. Whether this is good or bad depends on one’s scientific perspective. Under Figure 1a, MINT’s generalization would indeed be limiting because other methods could achieve greater flexibility. Under Figure 1b, all methods will have serious limits regarding generalization. Thus, MINT’s method for generalizing may approximate the best one can presently do. To address this concern, we have made three major changes, numbered i-iii below:

i) Large sections of the manuscript have been restructured to underscore the ways in which MINT can generalize. A major goal was to counter the impression, stated by R3 above, that: 

“for a test trial with a realized neural trajectory, [MINT] then finds the closest neural trajectory to it in the table and declares the associated behavior trajectory in the table as the decoded behavior”.

This description is a reasonable way to initially understand how MINT works, and we concede that we may have over-used this intuition. Unfortunately, it can leave the misimpression that MINT decodes by selecting whole trajectories, each corresponding to ‘a behavior’. This can happen, but it needn’t and typically doesn’t. As an example, consider the cycling task. Suppose that the library consists of stereotyped trajectories, each four cycles long, at five fixed speeds from 0.5-2.5 Hz. If the spiking observations argued for it, MINT could decode something close to one of these five stereotyped trajectories. Yet it needn’t. Decoded trajectories will typically resemble library trajectories locally, but may be very different globally. For example, a decoded trajectory could be thirty cycles long (or two, or five hundred) perhaps speeding up and slowing down multiple times across those cycles.

Thus, the library of trajectories shouldn’t be thought of as specifying a limited set of whole movements that can be ‘selected from’. Rather, trajectories define a scaffolding that outlines where the neural state is likely to live and how it is likely to be changing over time. When we introduce the idea of library trajectories, we are now careful to stress that they don’t function as a set from which one trajectory is ‘declared’ to be the right one:

“We thus designed MINT to approximate that manifold using the trajectories themselves, rather than their covariance matrix or corresponding subspace. Unlike a covariance matrix, neural trajectories indicate not only which states are likely, but also which state-derivatives are likely. If a neural state is near previously observed states, it should be moving in a similar direction. MINT leverages this directionality.

Training-set trajectories can take various forms, depending on what is convenient to collect. Most simply, training data might include one trajectory per condition, with each condition corresponding to a discrete movement. Alternatively, one might instead employ one long trajectory spanning many movements. Another option is to employ many sub-trajectories, each briefer than a whole movement. The goal is simply for training-set trajectories to act as a scaffolding, outlining the manifold that might be occupied during decoding and the directions in which decoded trajectories are likely to be traveling.”

Later in that same section we stress that decoded trajectories can move along the ‘mesh’ in nonstereotyped ways:

“Although the mesh is formed of stereotyped trajectories, decoded trajectories can move along the mesh in non-stereotyped ways as long as they generally obey the flow-field implied by the training data. This flexibility supports many types of generalization, including generalization that is compositional in nature. Other types of generalization – e.g. from the green trajectories to the orange trajectories in Figure 1b – are unavailable when using MINT and are expected to be challenging for any method (as will be documented in a later section).”

The section “Training and decoding using MINT” has been revised to clarify the ways in which interpolation is flexible, allowing decoded movements to be globally very different from any library trajectory.

“To decode stereotyped trajectories, one could simply obtain the maximum-likelihood neural state from the library, then render a behavioral decode based on the behavioral state with the same values of c and k. This would be appropriate for applications in which conditions are categorical, such as typing or handwriting. Yet in most cases we wish for the trajectory library to serve not as an exhaustive set of possible states, but as a scaffolding for the mesh of possible states. MINT’s operations are thus designed to estimate any neural trajectory – and any corresponding behavioral trajectory – that moves along the mesh in a manner generally consistent with the trajectories in Ω.”

“…interpolation allows considerable flexibility. Not only is one not ‘stuck’ on a trajectory from Φ, one is also not stuck on trajectories created by weighted averaging of trajectories in Φ. For example, if cycling speed increases, the decoded neural state could move steadily up a scaffolding like that illustrated in Figure 1b (green). In such cases, the decoded trajectory might be very different in duration from any of the library trajectories. Thus, one should not think of the library as a set of possible trajectories that are selected from, but rather as providing a mesh-like scaffolding that defines where future neural states are likely to live and the likely direction of their local motion. The decoded trajectory may differ considerably from any trajectory within Ω.”

This flexibility is indeed used during movement. One empirical example is described in detail:

“During movement… angular phase was decoded with effectively no net drift over time. This is noteworthy because angular velocity on test trials never perfectly matched any of the trajectories in Φ. Thus, if decoding were restricted to a library trajectory, one would expect growing phase discrepancies. Yet decoded trajectories only need to locally (and approximately) follow the flow-field defined by the library trajectories. Based on incoming spiking observations, decoded trajectories speed up or slow down (within limits).

This decoding flexibility presumably relates to the fact that the decoded neural state is allowed to differ from the nearest state in Ω. To explore… [the text goes on to describe the new analysis in Figure 4d, which shows that the decoded state is typically not on any trajectory, though it is typically close to a trajectory].”

Thus, MINT’s operations allow considerable flexibility, including generalization that is compositional in nature. Yet R3 is still correct that there are other forms of generalization that are unavailable to MINT. This is now stressed at multiple points in the revision. However, under the perspective in Figure 1b, these forms of generalization are unavailable to any current method. Hence we made a second major change in response to this concern…  ii) We explicitly illustrate how the structure of the data determines when generalization is or isn’t possible. The new Figure 1a,b introduces the two perspectives, and the new Figure 6a,b lays out their implications for generalization. Under the perspective in Figure 6a, the reviewer is quite right: other methods can generalize in ways that MINT cannot. Under the perspective in Figure 6b, expectations are very different. Those expectations make testable predictions. Hence the third major change… iii) We have added an analysis of generalization, using a newly collected dataset. This dataset was collected using Neuropixels Probes during our Pac-Man force-tracking task. This dataset was chosen because it is unusually well-suited to distinguishing the predictions in Figure 6a versus Figure 6b. Finding a dataset that can do so is not simple. Consider R3’s point that training data should “explore the whole movement space and the associated neural space”. The physical simplicity of the Pac-Man task makes it unusually easy to confirm that the behavioral workspace has been fully explored. Importantly, under Figure 6b, this does not mean that the neural workspace has been fully explored, which is exactly what we wish to test when testing generalization. We do so, and compare MINT with a Wiener filter. A Wiener filter is an ideal comparison because it is simple, performs very well on this task, and should be able to generalize well under Figure 1a. Additionally, the Wiener filter (unlike the Kalman Filter) doesn’t leverage the assumption that neural activity reflects the derivative of force. This matters because we find that neural activity does not reflect dforce/dt in this task. The Wiener filter is thus the most natural choice of the interpretable methods whose assumptions match Figure 1a.

The new analysis is described in Figure 6c-g and accompanying text. Results are consistent with the predictions of Figure 6b. We are pleased to have been motivated to add this analysis for two reasons. First, it provides an additional way of evaluating the predictions of the two competing scientific perspectives that are at the heart of our study. Second, this analysis illustrates an underappreciated way in which generalization is likely to be challenging for any decode method. It can be tempting to think that the main challenge regarding generalization is to fully explore the relevant behavioral space. This makes sense if a behavioral space has “an associated neural space”. However, we are increasingly of the opinion that it doesn’t. Different tasks often involve different neural subspaces, even when behavioral subspaces overlap. We have even seen situations where motor output is identical but neural subspaces are quite different. These facts are relevant to any decoder, something highlighted in the revised Introduction:

“MINT’s performance confirms that there are gains to be made by building decoders whose assumptions match a different, possibly more accurate view of population activity. At the same time, our results suggest fundamental limits on decoder generalization. Under the assumptions in Figure 1b, it will sometimes be difficult or impossible for decoders to generalize to not-yet-seen tasks. We found that this was true regardless of whether one uses MINT or a more traditional method. This finding has implications regarding when and how generalization should be attempted.”

We have also added an analysis (Figure 6e) illustrating how MINT’s ability to compute likelihoods can be useful in detecting situations that may strain generalization (for any method). MINT is unusual in being able to compute and use likelihoods in this way.

Detailed responses to R3: we reproduce each of R3’s specific concerns below, but concentrate our responses on issues not already covered above.

Main comments: 

Comment 1. MINT does not generalize to different tasks, which is a main limitation for BCI utility compared with prior BCI decoders that have shown this generalizability as I review below. Specifically, given that MINT tabulates task-specific trajectories, it will not generalize to tasks that are not seen in the training data even when these tasks cover the exact same space (e.g., the same 2D computer screen and associated neural space). 

First, the authors provide a section on generalization, which is inaccurate because it mixes up two fundamentally different concepts: 1) collecting informative training data and 2) generalizing from task to task. The former is critical for any algorithm, but it does not imply the latter. For example, removing one direction of cycling from the training set as the authors do here is an example of generating poor training data because the two behavioral (and neural) directions are non-overlapping and/or orthogonal while being in the same space. As such, it is fully expected that all methods will fail. For proper training, the training data should explore the whole movement space and the associated neural space, but this does not mean all kinds of tasks performed in that space must be included in the training set (something MINT likely needs while modeling-based approaches do not). Many BCI studies have indeed shown this generalization ability using a model. For example, in Weiss et al. 2019, center-out reaching tasks are used for training and then the same trained decoder is used for typing on a keyboard or drawing on the 2D screen. In Gilja et al. 2012, training is on a center-out task but the same trained decoder generalizes to a completely different pinball task (hit four consecutive targets) and tasks requiring the avoidance of obstacles and curved movements. There are many more BCI studies, such as Jarosiewicz et al. 2015 that also show generalization to complex realworld tasks not included in the training set. Unlike MINT, these works can achieve generalization because they model the neural subspace and its association to movement. On the contrary, MINT models task-dependent neural trajectories, so the trained decoder is very task-dependent and cannot generalize to other tasks. So, unlike these prior BCIs methods, MINT will likely actually need to include every task in its library, which is not practical. 

I suggest the authors remove claims of generalization and modify their arguments throughout the text and abstract. The generalization section needs to be substantially edited to clarify the above points. Please also provide the BCI citations and discuss the above limitation of MINT for BCIs. 

As discussed above, R3’s concerns are accurate under the view in Figure 1a (and the corresponding Figure 6a). Under this view, a method such as that in Gilja et al. or Jarosiewicz et al. can find the correct subspace, model the correct neuron-behavior correlations, and generalize to any task that uses “the same 2D computer screen and associated neural space”, just as the reviewer argues. Under Figure 1b things are quite different.

This topic – and the changes we have made to address it – is covered at length above. Here we simply want to highlight an empirical finding: sometimes two tasks use the same neural subspace and sometimes they don’t. We have seen both in recent data, and it is can be very non-obvious which will occur based just on behavior. It does not simply relate to whether one is using the same physical workspace. We have even seen situations where the patterns of muscle activity in two tasks are nearly identical, but the neural subspaces are fairly different. When a new task uses a new subspace, neither of the methods noted above (Gilja nor Jarosiewicz) will generalize (nor will MINT). Generalizing to a new subspace is basically impossible without some yet-to-be-invented approach. On the other hand, there are many other pairs of tasks (center-out-reaching versus some other 2D cursor control) where subspaces are likely to be similar, especially if the frequency content of the behavior is similar (in our recent experience this is often critical). When subspaces are shared, most methods will generalize, and that is presumably why generalization worked well in the studies noted above.

Although MINT can also generalize in such circumstances, R3 is correct that, under the perspective in Figure 1a, MINT will be more limited than other methods. This is now carefully illustrated in Figure 6a. In this traditional perspective, MINT will fail to generalize in cases where new trajectories are near previously observed states, yet move in very different ways from library trajectories. The reason we don’t view this is a shortcoming is that we expect it to occur rarely (else tangling would be high). We thus anticipate the scenario in Figure 6b.

This is worth stressing because R3 states that our discussion of generalization “is inaccurate because it mixes up two fundamentally different concepts: 1) collecting informative training data and 2) generalizing from task to task.” We have heavily revised this section and improved it. However, it was never inaccurate. Under Figure 6b, these two concepts absolutely are mixed up. If different tasks use different neural subspaces, then this requires collecting different “informative training data” for each. One cannot simply count on having explored the physical workspace.

Comment 2. MINT is shown to achieve competitive/high performance in highly stereotyped datasets with structured trials, but worse performance on MC_RTT, which is not based on repeated trials and is less stereotyped. This shows that MINT is valuable for decoding in repetitive stereotyped use-cases. However, it also highlights a limitation of MINT for BCIs, which is that MINT may not work well for real-world and/or less-constrained setups such as typing, moving a robotic arm in 3D space, etc. This is again due to MINT being a lookup table with a library of stereotyped trajectories rather than a model. Indeed, the authors acknowledge that the lower performance on MC_RTT (Figure 4) may be caused by the lack of repeated trials of the same type. However, real-world BCI decoding scenarios will also not have such stereotyped trial structure and will be less/un-constrained, in which MINT underperforms. Thus, the claim in the abstract or lines 480-481 that MINT is an "excellent" candidate for clinical BCI applications is not accurate and needs to be qualified. The authors should revise their statements according and discuss this issue. They should also make the use-case of MINT on BCI decoding clearer and more convincing. 

We discussed, above, multiple changes and additions to the revision that were made to address these concerns. Here we briefly expand on the comment that MINT achieves “worse performance on MC_RTT, which is not based on repeated trials and is less stereotyped”. All decoders performed poorly on this task. MINT still outperformed the two traditional methods, but this was the only dataset where MINT did not also perform better (overall) than the expressive GRU and feedforward network. There are probably multiple reasons why. We agree with R3 that one likely reason is that this dataset is straining generalization, and MINT may have felt this strain more than the two machine-learning-based methods. Another potential reason is the structure of the training data, which made it more challenging to obtain library trajectories in the first place. Importantly, these observations do not support the view in Figure 1a. MINT still outperformed the Kalman and Wiener filters (whose assumptions align with Fig. 1a). To make these points we have added the following:

“Decoding was acceptable, but noticeably worse, for the MC_RTT dataset… As will be discussed below, every decode method achieved its worst estimates of velocity for the MC_RTT dataset. In addition to the impact of slower reaches, MINT was likely impacted by training data that made it challenging to accurate estimate library trajectories. Due to the lack of repeated trials, MINT used AutoLFADS to estimate the neural state during training. In principle this should work well. In practice AutoLFADS may have been limited by having only 10 minutes of training data. Because the random-target task involved more variable reaches, it may also have stressed the ability of all methods to generalize, perhaps for the reasons illustrated in Figure 1b.

The only dataset where MINT did not perform the best overall was the MC_RTT dataset, where it was outperformed by the feedforward network and GRU. As noted above, this may relate to the need for MINT to learn neural trajectories from training data that lacked repeated trials of the same movement (a design choice one might wish to avoid). Alternatively, the less-structured MC_RTT dataset may strain the capacity to generalize; all methods experienced a drop in velocity-decoding R2 for this dataset compared to the others. MINT generalizes somewhat differently than other methods, and may have been at a modest disadvantage for this dataset. A strong version of this possibility is that perhaps the perspective in Figure 1a is correct, in which case MINT might struggle because it cannot use forms of generalization that are available to other methods (e.g. generalization based on neuron-velocity correlations). This strong version seems unlikely; MINT continued to significantly outperform the Wiener and Kalman filters, which make assumptions aligned with Figure 1a.”

Comment 3. Related to 2, it may also be that MINT achieves competitive performance in offline and trial-based stereotyped decoding by overfitting to the trial structure in a given task, and thus may not generalize well to online performance due to overfitting. For example, a recent work showed that offline decoding performance may be overfitted to the task structure and may not represent online performance (Deo et al. 2023). Please discuss. 

We agree that a limitation of our study is that we do not test online performance. There are sensible reasons for this decision:

“By necessity and desire, all comparisons were made offline, enabling benchmarked performance across a variety of tasks and decoded variables, where each decoder had access to the exact same data and recording conditions.”

We recently reported excellent online performance in the cycling task with a different algorithm

(Schroeder et al. 2022). In the course of that study, we consistently found that improvements in our offline decoding translated to improvements in our online decoding. We thus believe that MINT (which improves on the offline performance of our older algorithm) is a good candidate to work very well online. Yet we agree this still remains to be seen. We have added the following to the Discussion:

“With that goal in mind, there exist three important practical considerations. First, some decode algorithms experience a performance drop when used online. One presumed reason is that, when decoding is imperfect, the participant alters their strategy which in turn alters the neural responses upon which decoding is based. Because MINT produces particularly accurate decoding, this effect may be minimized, but this cannot be known in advance. If a performance drop does indeed occur, one could adapt the known solution of retraining using data collected during online decoding [13]. Another presumed reason (for a gap between offline and online decoding) is that offline decoders can overfit the temporal structure in training data [107]. This concern is somewhat mitigated by MINT’s use of a short spike-count history, but MINT may nevertheless benefit from data augmentation strategies such as including timedilated versions of learned trajectories in the libraries”

Comment 4. Related to 2, since MINT requires firing rates to generate the library and simple averaging does not work for this purpose in the MC_RTT dataset (that does not have repeated trials), the authors needed to use AutoLFADS to infer the underlying firing rates. The fact that MINT requires the usage of another model to be constructed first and that this model can be computationally complex, will also be a limiting factor and should be clarified. 

This concern relates to the computational complexity of computing firing-rate trajectories during training. Usually, rates are estimated via trial-averaging, which makes MINT very fast to train. This was quite noticeable during the Neural Latents Benchmark competition. As one example, for the “MC_Scaling 5 ms Phase”, MINT took 28 seconds to train while GPFA took 30 minutes, the transformer baseline (NDT) took 3.5 hours, and the switching nonlinear dynamical system took 4.5 hours.

However, the reviewer is quite correct that MINT’s efficiency depends on the method used to construct the library of trajectories. As we note, “MINT is a method for leveraging a trajectory library, not a method for constructing it”. One can use trial-averaging, which is very fast. One can also use fancier, slower methods to compute the trajectories. We don’t view this as a negative – it simply provides options. Usually one would choose trial-averaging, but one does not have to. In the case of MC_RTT, one has a choice between LFADS and grouping into pseudo-conditions and averaging (which is fast). LFADS produces higher performance at the cost of being slower. The operator can choose which they prefer. This is discussed in the following section:

“For MINT, ‘training’ simply means computation of standard quantities (e.g. firing rates) rather than parameter optimization. MINT is thus typically very fast to train (Table 1), on the order of seconds using generic hardware (no GPUs). This speed reflects the simple operations involved in constructing the library of neural-state trajectories: filtering of spikes and averaging across trials. At the same time we stress that MINT is a method for leveraging a trajectory library, not a method for constructing it. One may sometimes wish to use alternatives to trial-averaging, either of necessity or because they improve trajectory estimates. For example, for the MC_RTT task we used AutoLFADS to infer the library. Training was consequently much slower (hours rather than seconds) because of the time taken to estimate rates. Training time could be reduced back to seconds using a different approach – grouping into pseudo-conditions and averaging – but performance was reduced. Thus, training will typically be very fast, but one may choose time-consuming methods when appropriate.”

Comment 5. I also find the statement in the abstract and paper that "computations are simple, scalable" to be inaccurate. The authors state that MINT's computational cost is O(NC) only, but it seems this is achieved at a high memory cost as well as computational cost in training. The process is described in section "Lookup table of log-likelihoods" on line [978-990]. The idea is to precompute the log-likelihoods for any combination of all neurons with discretization x all delay/history segments x all conditions and to build a large lookup table for decoding. Basically, the computational cost of precomputing this table is O(V^{Nτ} x TC) and the table requires a memory of O(V^{Nτ}), where V is the number of discretization points for the neural firing rates, N is the number of neurons, τ is the history length, T is the trial length, and C is the number of conditions. This is a very large burden, especially the V^{Nτ} term. This cost is currently not mentioned in the manuscript and should be clarified in the main text. Accordingly, computation claims should be modified including in the abstract. 

As discussed above, the manuscript has been revised to clarify that our statement was accurate.

Comment 6. In addition to the above technical concerns, I also believe the authors should clarify the logic behind developing MINT better. From a scientific standpoint, we seek to gain insights into neural computations by making various assumptions and building models that parsimoniously describe the vast amount of neural data rather than simply tabulating the data. For instance, low-dimensional assumptions have led to the development of numerous dimensionality reduction algorithms and these models have led to important interpretations about the underlying dynamics (e.g., fixed points/limit cycles). While it is of course valid and even insightful to propose different assumptions from existing models as the authors do here, they do not actually translate these assumptions into a new model. Without a model and by just tabulating the data, I don't believe we can provide interpretation or advance the understanding of the fundamentals behind neural computations. As such, I am not clear as to how this library building approach can advance neuroscience or how these assumptions are useful. I think the authors should clarify and discuss this point. 

As requested, a major goal of the revision has been to clarify the scientific motivations underlying MINT’s design. In addition to many textual changes, we have added figures (Figures 1a,b and 6a,b) to outline the two competing scientific perspectives that presently exist. This topic is also addressed by extensions of existing analyses and by new analyses (e.g. Figure 6c-g). 

In our view these additions have dramatically improved the manuscript. This is especially true because we think R3’s concerns, expressed above, are reasonable. If the perspective in Figure 1a is correct, then R3 is right and MINT is essentially a hack that fails to model the data. MINT would still be effective in many circumstances (as we show), but it would be unprincipled. This would create limitations, just as the reviewer argues. On the other hand, if the perspective in Figure 1b is correct, then MINT is quite principled relative to traditional approaches. Traditional approaches make assumptions (a fixed subspace, consistent neuron-kinematic correlations) that are not correct under Figure 1b.

We don’t expect R3 to agree with our scientific perspective at this time (though we hope to eventually convince them). To us, the key is that we agree with R3 that the manuscript needs to lay out the different perspectives and their implications, so that readers have a good sense of the possibilities they should be considering. The revised manuscript is greatly improved in this regard.

Comment 7. Related to 6, there seems to be a logical inconsistency between the operations of MINT and one of its three assumptions, namely, sparsity. The authors state that neural states are sparsely distributed in some neural dimensions (Figure 1a, bottom). If this is the case, then why does MINT extend its decoding scope by interpolating known neural states (and behavior) in the training library? This interpolation suggests that the neural states are dense on the manifold rather than sparse, thus being contradictory to the assumption made. If interpolation-based dense meshes/manifolds underlie the data, then why not model the neural states through the subspace or manifold representations? I think the authors should address this logical inconsistency in MINT, especially since this sparsity assumption also questions the low-dimensional subspace/manifold assumption that is commonly made. 

We agree this is an important issue, and have added an analysis on this topic (Figure 4d). The key question is simple and empirical: during decoding, does interpolation cause MINT to violate the assumption of sparsity? R3 is quite right that in principle it could. If spiking observations argue for it, MINT’s interpolation could create a dense manifold during decoding rather than a sparse one. The short answer is that empirically this does not happen, in agreement with expectations under Figure 1b. Rather than interpolating between distant states and filling in large ‘voids’, interpolation is consistently local. This is a feature of the data, not of the decoder (MINT doesn’t insist upon sparsity, even though it is designed to work best in situations where the manifold is sparse).

In addition to adding Figure 4d, we added the following (in an earlier section):

“The term mesh is apt because, if MINT’s assumptions are correct, interpolation will almost always be local. If so, the set of decodable states will resemble a mesh, created by line segments connecting nearby training-set trajectories. However, this mesh-like structure is not enforced by MINT’s operations. Interpolation could, in principle, create state-distributions that depart from the assumption of a sparse manifold. For example, interpolation could fill in the center of the green tube in Figure 1b, resulting in a solid manifold rather than a mesh around its outer surface. However, this would occur only if spiking observations argued for it. As will be documented below, we find that essentially all interpolation is local.”

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors): 

I appreciate the detailed methods section, however, more specifics should be integrated into the main text. For example on Line 238, it should additionally be stated how many minutes were used for training and metrics like the MAE which is used later should be reported here.

Thank you for this suggestion. We now report the duration of training data in the main text:

“Decoding R^2 was .968 over ~7.1 minutes of test trials based on ~4.4 minutes of training data.”

We have also added similar specifics throughout the manuscript, e.g. in the Fig. 5 legend:

“Results are based on the following numbers of training / test trials: MC\_Cycle (174 train, 99 test), MC\_Maze (1721 train, 574 test), Area2\_Bump (272 train, 92 test), MC\_RTT (810 train, 268 test).”

Similar additions were made to the legends for Fig. 6 and 8. Regarding the request to add MAE for the multitask network, we did not do so for the simple reason that the decoded variable (muscle activity) has arbitrary units. The raw MAE is thus not meaningful. We could of course have normalized, but at this point the MAE is largely redundant with the correlation. In contrast, the MAE is useful when comparing across the MC_Maze, Area2_Bump, and MC_RTT datasets, because they all involve the same scale (cm/s).

Regarding the MC_RTT task, AutoLFADS was used to obtain robust spike rates, as reported in the methods. However, the rationale for splitting the neural trajectories after AutoLFADS is unclear. If the trajectories were split based on random recording gaps, this might lead to suboptimal performance? It might be advantageous to split them based on a common behavioural state? 

When learning neural trajectories via AutoLFADS, spiking data is broken into short (but overlapping) segments, rates are estimated for each segment via AutoLFADs, and these rates are then stitched together across segments into long neural trajectories. If there had been no recording gaps, these rates could have been stitched into a single neural trajectory for this dataset. However, the presence of recording gaps left us no choice but to stitch together these rates into more than one trajectory. Fortunately, recording gaps were rare: for the decoding analysis of MC_RTT there were only two recording gaps and therefore three neural trajectories, each ~2.7 minutes in duration. 

We agree that in general it is desirable to learn neural trajectories that begin and end at behaviorallyrelevant moments (e.g. in between movements). However, having these trajectories potentially end midmovement is not an issue in and of itself. During decoding, MINT is never stuck on a trajectory. Thus, if MINT were decoding states near the end of a trajectory that was cut short due to a training gap, it would simply begin decoding states from other trajectories or elsewhere along the same trajectory in subsequent moments. We could have further trimmed the three neural trajectories to begin and end at behaviorallyrelevant moments, but chose not to as this would have only removed a handful of potentially useful states from the library.

We now describe this in the Methods:

“Although one might prefer trajectory boundaries to begin and end at behaviorally relevant moments (e.g. a stationary state), rather than at recording gaps, the exact boundary points are unlikely to be consequential for trajectories of this length that span multiple movements. If MINT estimates a state near the end of a long trajectory, its estimate will simply jump to another likely state on a different trajectory (or earlier along the same trajectory) in subsequent moments. Clipping the end of each trajectory to an earlier behaviorally-relevant moment would only remove potentially useful states from the libraries.”

Are the training and execution times in Table 1 based on pure Matlab functions or Mex files? If it's Mex files as suggested by the code, it would be good to mention this in the Table caption.

They are based on a combination of MATLAB and MEX files. This is now clarified in the table caption:

“Timing measurements taken on a Macbook Pro (on CPU) with 32GB RAM and a 2.3 GHz 8-Core Intel Core i9 processor. Training and execution code used for measurements was written in MATLAB (with the core recursion implemented as a MEX file).”

As the method most closely resembles a Bayesian decoder it would be good to compare performance against a Naive Bayes decoder. 

We agree and have now done so. The following has been added to the text:

“A natural question is thus whether a simpler Bayesian decoder would have yielded similar results. We explored this possibility by testing a Naïve Bayes regression decoder [85] using the MC_Maze dataset. This decoder performed poorly, especially when decoding velocity (R2 = .688 and .093 for hand position and velocity, respectively), indicating that the specific modeling assumptions that differentiate MINT from a naive Bayesian decoder are important drivers of MINT’s performance.”

Line 199 Typo: The assumption of stereotypy trajectory also enables neural states (and decoded behaviors) to be updated in between time bins. 

Fixed

Table 3: It's unclear why the Gaussian binning varies significantly across different datasets. Could the authors explain why this is the case and what its implications might be? 

We have added the following description in the “Filtering, extracting, and warping data on each trial” subsection of the Methods to discuss how 𝜎 may vary due to the number of trials available for training and how noisy the neural data for those trials is:

“First, spiking activity for each neuron on each trial was temporally filtered with a Gaussian to yield single-trial rates. Table 3 reports the Gaussian standard deviations σ (in milliseconds) used for each dataset. Larger values of σ utilize broader windows of spiking activity when estimating rates and therefore reduce variability in those rate estimates. However, large σ values also yield neural trajectories with less fine-grained temporal structure. Thus, the optimal σ for a dataset depends on how variable the rate estimates otherwise are.”

An implementation of the method in an open-source programming language could further enhance the widespread use of the tool. 

We agree this would be useful, but have yet not implemented the method in any other programming languages. Implementation in Python is still a future goal.

Reviewer #2 (Recommendations For The Authors): 

- Figures 4 and 5 should show the error bars on the horizontal axis rather than portraying them vertically. 

[Note that these are now Figures 5 and 6]

The figure legend of Figure 5 now clarifies that the vertical ticks are simply to aid visibility when symbols have very similar means and thus overlap visually. We don’t include error bars (for this analysis) because they are very small and would mostly be smaller than the symbol sizes. Instead, to indicate certainty regarding MINT’s performance measurements, the revised text now gives error ranges for the correlations and MAE values in the context of Figure 4c. These error ranges were computed as the standard deviation of the sampling distribution (computed via resampling of trials) and are thus equivalent to SEMs. The error ranges are all very small; e.g. for the MC_Maze dataset the MAE for x-velocity is 4.5 +/- 0.1 cm/s. (error bars on the correlations are smaller still).

Thus, for a given dataset, we can be quite certain of how well MINT performs (within ~2% in the above case). This is reassuring, but we also don’t want to overemphasize this accuracy. The main sources of variability one should be concerned about are: 1) different methods can perform differentially well for different brain areas and tasks, 2) methods can decode some behavioral variables better than others, and 3) performance depends on factors like neuron-count and the number of training trials, in ways that can differ across decode methods. For this reason, the study examines multiple datasets, across tasks and brain areas, and measures performance for a range of decoded variables. We also examine the impact of training-set-size (Figure 8a) and population size (solid traces in Fig. 8b, see R2’s next comment below). 

There is one other source of variance one might be concerned about, but it is specific to the neuralnetwork approaches: different weight initializations might result in different performance. For this reason, each neural-network approach was trained ten times, with the average performance computed. The variability around this average was very small, and this is now stated in the Methods.

“For the neural networks, the training/testing procedure was repeated 10 times with different random seeds. For most behavioral variables, there was very little variability in performance across repetitions. However, there were a few outliers for which variability was larger. Reported performance for each behavioral group is the average performance across the 10 repetitions to ensure results were not sensitive to any specific random initialization of each network.”

- For Figure 6, it is unclear whether the neuron-dropping process was repeated multiple times. If not, it should be since the results will be sensitive to which particular subsets of neurons were "dropped". In this case, the results presented in Figure 6 should include error bars to describe the variability in the model performance for each decoder considered. 

A good point. The results in Figure 8 (previously Figure 6) were computed by averaging over the removal of different random subsets of neurons (50 subsets per neuron count), just as the reviewer requests. The figure has been modified to include the standard deviation of performance across these 50 subsets. The legend clarifies how this was done.

Reviewer #3 (Recommendations For The Authors): 

Other comments: 

(1) [Line 185-188] The authors argue that in a 100-dimensional space with 10 possible discretized values, 10^100 potential neural states need to be computed. But I am not clear on this. This argument seems to hold only in the absence of a model (as in MINT). For a model, e.g., Kalman filter or AutoLFADS, information is encoded in the latent state. For example, a simple Kalman filter for a linear model can be used for efficient inference. This 10^100 computation isn't a general problem but seems MINT-specific, please clarify. 

We agree this section was potentially confusing. It has been rewritten. We were simply attempting to illustrate why maximum likelihood computations are challenging without constraints. MINT simplifies this problem by adding constraints, which is why it can readily provide data likelihoods (and can do so using a Poisson model). The rewritten section is below:

“Even with 1000 samples for each of the neural trajectories in Figure 3, there are only 4000 possible neural states for which log-likelihoods must be computed (in practice it is fewer still, see Methods). This is far fewer than if one were to naively consider all possible neural states in a typical rate- or factor-based subspace. It thus becomes tractable to compute log-likelihoods using a Poisson observation model. A Poisson observation model is usually considered desirable, yet can pose tractability challenges for methods that utilize a continuous model of neural states. For example, when using a Kalman filter, one is often restricted to assuming a Gaussian observation model to maintain computational tractability “

(2) [Figure 6b] Why do the authors set the dropped neurons to zero in the "zeroed" results of the robustness analysis? Why not disregard the dropped neurons during the decoding process? 

We agree the terminology we had used in this section was confusing. We have altered the figure and rewritten the text. The following, now at the beginning of that section, addresses the reviewer’s query: 

“It is desirable for a decoder to be robust to the unexpected loss of the ability to detect spikes from some neurons. Such loss might occur while decoding, without being immediately detected. Additionally, one desires robustness to a known loss of neurons / recording channels. For example, there may have been channels that were active one morning but are no longer active that afternoon. At least in principle, MINT makes it very easy to handle this second situation: there is no need to retrain the decoder, one simply ignores the lost neurons when computing likelihoods. This is in contrast to nearly all other methods, which require retraining because the loss of one neuron alters the optimal parameters associated with every other neuron.”

The figure has been relabeled accordingly; instead of the label ‘zeroed’, we use the label ‘undetected neuron loss’.

(3) Authors should provide statistical significance on their results, which they already did for Fig. S3a,b,c but missing on some other figures/places. 

We have added error bars in some key places, including in the text when quantifying MINT’s performance in the context of Figure 4. Importantly, error bars are only as meaningful as the source of error they assess, and there are reasons to be careful given this. The standard method for putting error bars on performance is to resample trials, which is indeed what we now report. These error bars are very small. For example, when decoding horizontal velocity for the MC_Maze dataset, the correlation between MINT’s decode and the true velocity had a mean and SD of the sampling distribution of 0.963 +/- 0.001. This means that, for a given dataset and target variable, we have enough trials/data that we can be quite certain of how well MINT performs. However, we want to be careful not to overstate this certainty. What one really wants to know is how well MINT performs across a variety of datasets, brain areas, target variables, neuron counts, etc. It is for this reason that we make multiple such comparisons, which provides a more valuable view of performance variability.

For Figure 7, error bars are unavailable. Because this was a benchmark, there was exactly one test-set that was never seen before. This is thus not something that could be resampled many times (that would have revealed the test data and thus invalidated the benchmark, not to mention that some of these methods take days to train). We could, in principle, have added resampling to Figure 5. In our view it would not be helpful and could be misleading for the reasons noted above. If we computed standard errors using different train/test partitions, they would be very tight (mostly smaller than the symbol sizes), which would give the impression that one can be quite certain of a given R^2 value. Yet variability in the train/test partition is not the variability one is concerned about in practice. In practice, one is concerned about whether one would get a similar R^2 for a different dataset, or brain area, or task, or choice of decoded variable. Our analysis thus concentrated on showing results across a broad range of situations. In our view this is a far more relevant way of illustrating the degree of meaningful variability (which is quite large) than resampling, which produces reassuringly small but (mostly) irrelevant standard errors.

Error bars are supplied in Figure 8b. These error bars give a sense of variability across re-samplings of the neural population. While this is not typically the source of variability one is most concerned about, for this analysis it becomes appropriate to show resampling-based standard errors because a natural concern is that results may depend on which neurons were dropped. So here it is both straightforward, and desirable, to compute standard errors. (The fact that MINT and the Wiener filter can be retrained many times swiftly was also key – this isn’t true of the more expressive methods). Figure S1 also uses resampling-based confidence intervals for similar reasons.

(4) [Line 431-437] Authors state that MINT outperforms other methods with the PSTH R^2 metric (trial-averaged smoothed spikes for each condition). However, I think this measure may not provide a fair comparison and is confounded because MINT's library is built using PSTH (i.e., averaged firing rate) but other methods do not use the PSTH. The author should clarify this. 

The PSTH R^2 metric was not created by us; it was part of the Neural Latents Benchmark. They chose it because it ensures that a method cannot ‘cheat’ (on the Bits/Spike measure) by reproducing fine features of spiking while estimating rates badly. We agree with the reviewer’s point: MINT’s design does give it a potential advantage in this particular performance metric. This isn’t a confound though, just a feature. Importantly, MINT will score well on this metric only if MINT’s neural state estimate is accurate (including accuracy in time). Without accurate estimation of the neural state at each time, it wouldn’t matter that the library trajectory is based on PSTHs. This is now explicitly stated:

“This is in some ways unsurprising: MINT estimates neural states that tend to resemble (at least locally) trajectories ‘built’ from training-set-derived rates, which presumably resemble test-set rates. Yet strong performance is not a trivial consequence of MINT’s design. MINT does not ‘select’ whole library trajectories; PSTH R2 will be high only if condition (c), index (k), and the interpolation parameter (α) are accurately estimated for most moments.”

are

remove

Which - ? (question mark is missing? maybe start 'You can...' on a new line as well

sensitive

You will also use this catchment for the final assignment.

do we want to include a deadline in this book?

non-negligible

instead

PDF

Refer to the help function to understand which arguments this function requires and what output it produces

switch between the PDFs for comparison

For easy comparison, open all three generated PDF figures, make them fullscreen, and switch between them using ALT+TAB.

PDF

It may be necessary

Zoom in on this period

November

entire

August

you don’t need to complete all of them

the

exercises

corresponding line

, or

assign

Ensure that no errors occur. Warnings can be ignored.

complete

To make the note more user-friendly and accessible while clearly indicating what the link redirects to, you could rephrase it like this:

"Note: The answers for this practical are available in this download containing the Lowlands scripts and answers."

You want to prevent using a link text like 'here'.

PC practical: Catchment and climate hydrology

Author response:

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

Public Reviews:

Reviewer #1 (Public Review): 

In the presented manuscript, the authors investigate how neural networks can learn to replay presented sequences of activity. Their focus lies on the stochastic replay according to learned transition probabilities. They show that based on error-based excitatory and balance-based inhibitory plasticity networks can selforganize towards this goal. Finally, they demonstrate that these learning rules can recover experimental observations from song-bird song learning experiments. 

Overall, the study appears well-executed and coherent, and the presentation is very clear and helpful. However, it remains somewhat vague regarding the novelty. The authors could elaborate on the experimental and theoretical impact of the study, and also discuss how their results relate to those of Kappel et al, and others (e.g., Kappel et al (doi.org/10.1371/journal.pcbi.1003511))). 

We agree with the reviewer that our previous manuscript lacked comparison with previously published similar works. While Kappel et al. demonstrated that STDP in winner-take-all circuits can approximate online learning of hidden Markov models (HMMs), a key distinction from our model is that their neural representations acquire deterministic sequential activations, rather than exhibiting stochastic transitions governing Markovian dynamics. Specifically, in their model, the neural representation of state B would be different in the sequences ABC and CBA, resulting in distinct deterministic representations like ABC and C'B'A', where ‘A’ and ‘A'’ are represented by different neural states (e.g., activations of different cell assemblies). In contrast, our network learns to generate stochastically transitioning cell assemblies which replay Markovian trajectories of spontaneous activity obeying the learned transition probabilities between neural representations of states. For example, starting from reactivation from assembly ‘A’, there may be an 80% probability to transition to assembly ‘B’ and 20% to ‘C’. Although Kappel et al.'s model successfully solves HMMs, their neural representations do not themselves stochastically transition between states according to the learned model. Similar to the Kappel et al.'s model, while the models proposed in Barber (2002) and Barber and Agakov (2002) learn the Markovian statistics, these models learned a static spatiotemporal input patterns only and how assemblies of neurons show stochastic transition in spontaneous activity has been still unclear. In contrast with these models, our model captures the probabilistic neural state trajectories, allowing spontaneous replay of experienced sequences with stochastic dynamics matching the learned environmental statistics.

We have included new sentences for explain these in ll. 509-533 in the revised manuscript.

Overall, the work could benefit if there was either (A) a formal analysis or derivation of the plasticity rules involved and a formal justification of the usefulness of the resulting (learned) neural dynamics; 

We have included a derivation of our plasticity rules in ll. 630-670 in the revised manuscript. Consistent with our claim that excitatory plasticity updates the excitatory synapse to predict output firing rates, we have shown that the corresponding cost function measures the discrepancy between the recurrent prediction and the output firing rate. Similarly, for inhibitory plasticity, we defined the cost function that evaluates the difference between the excitatory and inhibitory potential within each neuron. We showed that the resulting inhibitory plasticity rule updates the inhibitory synapses to maintain the excitation-inhibition balance.

and/or (B) a clear connection of the employed plasticity rules to biological plasticity and clear testable experimental predictions. Thus, overall, this is a good work with some room for improvement. 

Our proposed plasticity mechanism could be implemented through somatodendritic interactions. Analogous to previous computational works (Urbanczik and Senn., 2014; Asabuki and Fukai., 2020; Asabuki et al., 2022), our model suggests that somatic responses may encode the stimulus-evoked neural activity states, while dendrites encode predictions based on recurrent dynamics that aim to minimize the discrepancy between somatic and dendritic activity. To directly test this hypothesis, future experimental studies could simultaneously record from both somatic and dendritic compartments to investigate how they encode evoked responses and predictive signals during learning (Francioni et al., 2022).

We have included new sentences for explain these in ll. 476-484 in the revised manuscript.

Reviewer #2 (Public Review): 

Summary: 

This work proposes a synaptic plasticity rule that explains the generation of learned stochastic dynamics during spontaneous activity. The proposed plasticity rule assumes that excitatory synapses seek to minimize the difference between the internal predicted activity and stimulus-evoked activity, and inhibitory synapses try to maintain the E-I balance by matching the excitatory activity. By implementing this plasticity rule in a spiking recurrent neural network, the authors show that the state-transition statistics of spontaneous excitatory activity agree with that of the learned stimulus patterns, which are reflected in the learned excitatory synaptic weights. The authors further demonstrate that inhibitory connections contribute to well-defined state transitions matching the transition patterns evoked by the stimulus. Finally, they show that this mechanism can be expanded to more complex state-transition structures including songbird neural data. 

Strengths: 

This study makes an important contribution to computational neuroscience, by proposing a possible synaptic plasticity mechanism underlying spontaneous generations of learned stochastic state-switching dynamics that are experimentally observed in the visual cortex and hippocampus. This work is also very clearly presented and well-written, and the authors conducted comprehensive simulations testing multiple hypotheses. Overall, I believe this is a well-conducted study providing interesting and novel aspects of the capacity of recurrent spiking neural networks with local synaptic plasticity. 

Weaknesses: 

This study is very well-thought-out and theoretically valuable to the neuroscience community, and I think the main weaknesses are in regard to how much biological realism is taken into account. For example, the proposed model assumes that only synapses targeting excitatory neurons are plastic, and uses an equal number of excitatory and inhibitory neurons. 

We agree with the reviewer. The network shown in the previous manuscript consists of an equal number of excitatory and inhibitory neurons, which seems to lack biological plausibility. Therefore, we first tested whether a biologically plausible scenario would affect learning performance by setting the ratio of excitatory to inhibitory neurons to 80% and 20% (Supplementary Figure 7a; left). Even in such a scenario, the network still showed structured spontaneous activity (Supplementary Figure 7a; center), with transition statistics of replayed events matching the true transition probabilities (Supplementary Figure 7a; right). We then asked whether the model with our plasticity rule applied to all synapses would reproduce the corresponding stochastic transitions. We found that the network can learn transition statistics but only under certain conditions. The network showed only weak replay and failed to reproduce the appropriate transition (Supplementary Fig. 7b) if the inhibitory neurons were no longer driven by the synaptic currents reflecting the stimulus, due to a tight balance of excitatory and inhibitory currents on the inhibitory neurons. We then tested whether the network with all synapses plastic can learn transition statistics if the external inputs project to the inhibitory neurons as well. We found that, when each stimulus pattern activates a non-overlapping subset of neurons, the network does not exhibit the correct stochastic transition of assembly reactivation (Supplementary Fig. 7c). Interestingly, when each neuron's activity is triggered by multiple stimuli and has mixed selectivity, the reactivation reproduced the appropriate stochastic transitions (Supplementary Fig. 7d).

We have included these new results as new Supplementary Figure 7 and they are explained in ll.215-230 in the revised manuscript.

The model also assumes Markovian state dynamics while biological systems can depend more on history. This limitation, however, is acknowledged in the Discussion. 

We have included the following sentence to provide a possible solution to this limitation: “Therefore, to learn higher-order stochastic transitions, recurrent neural networks like ours may need to integrate higher-order inputs with longer time scales.” in ll.557-559 in the revised manuscript. 

Finally, to simulate spontaneous activity, the authors use a constant input of 0.3 throughout the study. Different amplitudes of constant input may correspond to different internal states, so it will be more convincing if the authors test the model with varying amplitudes of constant inputs. 

We thank the reviewer for pointing this out. In the revised manuscript, we have tested constant input with three different strengths. If the strength is moderate, the network showed accurate encoding of transition statistics in the spontaneous activity as we have seen in Fig.2. We have additionally shown that the weaker background input causes spontaneous activity with lower replay rate, which in turn leads to high variance of encoded transition, while stronger inputs make assembly replay transitions more uniform. We have included these new results as new Supplementary Figure 6 and they are explained in ll.211214 in the revised manuscript.

Reviewer #3 (Public Review): 

Summary: 

Asabuki and Clopath study stochastic sequence learning in recurrent networks of Poisson spiking neurons that obey Dale's law. Inspired by previous modeling studies, they introduce two distinct learning rules, to adapt excitatory-to-excitatory and inhibitory-to-excitatory synaptic connections. Through a series of computer experiments, the authors demonstrate that their networks can learn to generate stochastic sequential patterns, where states correspond to non-overlapping sets of neurons (cell assemblies) and the state-transition conditional probabilities are first-order Markov, i.e., the transition to a given next state only depends on the current state. Finally, the authors use their model to reproduce certain experimental songbird data involving highly-predictable and highly-uncertain transitions between song syllables. 

Strengths: 

This is an easy-to-follow, well-written paper, whose results are likely easy to reproduce. The experiments are clear and well-explained. The study of songbird experimental data is a good feature of this paper; finches are classical model animals for understanding sequence learning in the brain. I also liked the study of rapid task-switching, it's a good-to-know type of result that is not very common in sequence learning papers. 

Weaknesses: 

While the general subject of this paper is very interesting, I missed a clear main result. The paper focuses on a simple family of sequence learning problems that are well-understood, namely first-order Markov sequences and fully visible (nohidden-neuron) networks, studied extensively in prior work, including with spiking neurons. Thus, because the main results can be roughly summarized as examples of success, it is not entirely clear what the main point of the authors is. 

We apologize the reviewer that our main claim was not clear. While various computational studies have suggested possible plasticity mechanisms for embedding evoked activity patterns or their probability structures into spontaneous activity (Litwin-Kumar et al., Nat. Commun. 2014, Asabuki and Fukai., Biorxiv 2023), how transition statistics of the environment are learned in spontaneous activity is still elusive and poorly understood. Furthermore, while several network models have been proposed to learn Markovian dynamics via synaptic plasticity (Brea, et al. (2013); Pfister et al. (2004); Kappel et al. (2014)), they have been limited in a sense that the learned network does not show stochastic transition in a neural state space. For instance, while Kappel et al. demonstrated that STDP in winner-take-all circuits can approximate online learning of hidden Markov models (HMMs), a key distinction from our model is that their neural representations acquire deterministic sequential activations, rather than exhibiting stochastic transitions governing Markovian dynamics. Specifically, in their model, the neural representation of state B would be different in the sequences ABC and CBA, resulting in distinct deterministic representations like ABC and C'B'A', where ‘A’ and ‘A'’ are represented by different neural states (e.g., activations of different cell assemblies). In contrast, our network learns to generate stochastically transitioning cell assemblies that replay Markovian trajectories of spontaneous activity obeying the learned transition probabilities between neural representations of states. For example, starting from reactivation from assembly ‘A’, there may be an 80% probability to transition to assembly ‘B’ and 20% to ‘C’. Although Kappel et al.'s model successfully solves HMMs, their neural representations do not themselves stochastically transition between states according to the learned model. Similar to the Kappel et al.'s model, while the models proposed in Barber (2002) and Barber and Agakov (2002) learn the Markovian statistics, these models learned a static spatiotemporal input patterns only and how assemblies of neurons show stochastic transition in spontaneous activity has been still unclear. In contrast with these models, our model captures the probabilistic neural state trajectories, allowing spontaneous replay of experienced sequences with stochastic dynamics matching the learned environmental statistics.

We have explained this point in ll.509-533 in the revised manuscript.

Going into more detail, the first major weakness I see in this paper is the heuristic choice of learning rules. The paper studies Poisson spiking neurons (I return to this point below), for which learning rules can be derived from a statistical objective, typically maximum likelihood. For fully-visible networks, these rules take a simple form, similar in many ways to the E-to-E rule introduced by the authors. This more principled route provides quite a lot of additional understanding on what is to be expected from the learning process. 

We thank the reviewer for pointing this out. To better demonstrate the function of our plasticity rules, we have included the derivation of the rules of synaptic plasticity in ll. 630-670 in the revised manuscript. Consistent with our claim that excitatory plasticity updates the excitatory synapse to predict output firing rates, we have shown that the corresponding cost function measures the discrepancy between the recurrent prediction and the output firing rate. Similarly, for inhibitory plasticity, we defined the cost function that evaluates the difference between the excitatory and inhibitory potential within each neuron. We showed that the resulting inhibitory plasticity rule updates the inhibitory synapses to maintain the excitation-inhibition balance.

For instance, should maximum likelihood learning succeed, it is not surprising that the statistics of the training sequence distribution are reproduced. Moreover, given that the networks are fully visible, I think that the maximum likelihood objective is a convex function of the weights, which then gives hope that the learning rule does succeed. And so on. This sort of learning rule has been studied in a series of papers by David Barber and colleagues [refs. 1, 2 below], who applied them to essentially the same problem of reproducing sequence statistics in recurrent fully-visible nets. It seems to me that one key difference is that the authors consider separate E and I populations, and find the need to introduce a balancing I-to-E learning rule. 

The reviewer’s understanding that inhibitory plasticity to maintain EI balance is one of a critical difference from previous works is correct. However, we believe that the most striking point of our study is that we have shown numerically that predictive plasticity rules enable recurrent networks to learn and replay the assembly activations whose transition statistics match those of the evoked activity. Please see our reply above.

Because the rules here are heuristic, a number of questions come to mind. Why these rules and not others - especially, as the authors do not discuss in detail how they could be implemented through biophysical mechanisms? When does learning succeed or fail? What is the main point being conveyed, and what is the contribution on top of the work of e.g. Barber, Brea, et al. (2013), or Pfister et al. (2004)? 

Our proposed plasticity mechanism could be implemented through somatodendritic interactions. Analogous to previous computational works (Senn, Asabuki), our model suggests that somatic responses may encode the stimulusevoked neural activity states, while dendrites encode predictions based on recurrent dynamics that aim to minimize the discrepancy between somatic and dendritic activity. To directly test this hypothesis, future experimental studies could simultaneously record from both somatic and dendritic compartments to investigate how they encode evoked responses and predictive signals during learning.

To address the point of the reviewer, we conducted addionnal simulations to test where the model fails. We found that the model with our plasticity rule applied to all synapses only showed faint replays and failed to replay the appropriate transition (Supplementary Fig. 7b). This result is reasonable because the inhibitory neurons were no longer driven by the synaptic currents reflecting the stimulus, due to a tight balance of excitatory and inhibitory currents on the inhibitory neurons. Our model predicts that mixed selectivity in the inhibitory population is crucial to learn an appropriate transition statistics (Supplementary Fig. 7d). Future work should clarify the role of synaptic plasticity on inhibitory neurons, especially plasticity at I to I synapses. We have explained this result as new supplementary Figure7 in the revised manuscript.

The use of a Poisson spiking neuron model is the second major weakness of the study. A chief challenge in much of the cited work is to generate stochastic transitions from recurrent networks of deterministic neurons. The task the authors set out to do is much easier with stochastic neurons; it is reasonable that the network succeeds in reproducing Markovian sequences, given an appropriate learning rule. I believe that the main point comes from mapping abstract Markov states to assemblies of neurons. If I am right, I missed more analyses on this point, for instance on the impact that varying cell assembly size would have on the findings reported by the authors.

The reviewer’s understanding is correct. Our main point comes from mapping Markov statistics to replays of cell assemblies. In the revised manuscript, we performed additional simulations to ask whether varying the size of the cell assemblies would affect learning. We ran simulations with two different configurations in the task shown in Figure 2. The first configuration used three assemblies with a size ratio of 1:1.5:2. After training, these assemblies exhibited transition statistics that closely matched those of the evoked activity (Supplementary Fig.4a,b). In contrast, the second configuration, which used a size ratio of 1:2:3, showed worse performance compared to the 1:1.5:2 case (Supplementary Fig.4c,d). These results suggest that the model can learn appropriate transition statistics as long as the size ratio of the assemblies is not drastically varied.

Finally, it was not entirely clear to me what the main fundamental point in the HVC data section was. Can the findings be roughly explained as follows: if we map syllables to cell assemblies, for high-uncertainty syllable-to-syllable transitions, it becomes harder to predict future neural activity? In other words, is the main point that the HVC encodes syllables by cell assemblies? 

The reviewer's understanding is correct. We wanted to show that if the HVC learns transition statistics as a replay of cell assemblies, a high-uncertainty syllable-to-syllable transition would make predicting future reactivations more difficult, since trial-averaged activities (i.e., poststimulus activities; PSAs) marginalized all possible transitions in the transition diagram.

(1) Learning in Spiking Neural Assemblies, David Barber, 2002. URL: https://proceedings.neurips.cc/paper/2002/file/619205da514e83f869515c782a328d3c-Paper.pdf  

(2) Correlated sequence learning in a network of spiking neurons usingmaximum likelihood, David Barber, Felix Agakov, 2002. URL: http://web4.cs.ucl.ac.uk/staff/D.Barber/publications/barber-agakovTR0149.pdf  

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors): 

In more detail: 

A) Theoretical analysis 

The plasticity rules in the study are introduced with a vague reference to previous theoretical studies of others. Doing this, one does not provide any formal insight as to why these plasticity rules should enable one to learn to solve the intended task, and whether they are optimal in some respect. This becomes noticeable, especially in the discussion of the importance of inhibitory balance, which does not go into any detail, but rather only states that its required, both in the results and discussion sections. Another unclarity appears when error-based learning is discussed and compared to Hebbian plasticity, which, as you state, "alone is insufficient to learn transition probabilities". It is not evident how this claim is warranted, nor why error-based plasticity in comparison should be able to perform this (other than referring to the simulation results). Please either clarify formally (or at least intuitively) how plasticity rules result in the mentioned behavior, or alternatively acknowledge explicitly the (current) lack of intuition. 

The lack of formal discussion is a relevant shortcoming compared to previous research that showed very similar results with formally more rigorous and principled approaches. In particular, Kappel et al derived explicitly how neural networks can learn to sample from HMMs using STDP and winner-take-all dynamics. Even though this study has limitations, the relation with respect to that work should be made very clear; potentially the claims of novelty of some results (sampling) should be adjusted accordingly. See also Yanping Huang, Rajesh PN Rao (NIPS 2014), and possibly other publications. While it might be difficult to formally justify the learning rules post-hoc, it would be very helpful to the field if you very clearly related your work to that of others, where learning rules have been formally justified, and elaborate on the intuition of how the employed rules operate and interact (especially for inhibition). 

Lastly, while the importance of sampling learned transition probabilities is discussed, the discussion again remains on a vague level, characterized by the lack of references in the relevant paragraphs. Ideally, there should be a proof of concept or a formal understanding of how the learned behaviour enables to solve a problem that is not solved by deterministic networks. Please incorporate also the relation to the literature on neural sampling/planning/RL etc. and substantiate the claims with citations. 

We have included sentences in ll. 691-696 in the revised manuscript to explain that for Poisson spiking neurons, the derived learning rule is equivalent to the one that minimizes the Kullback-Leibler divergence between the distributions of output firing and the dendritic prediction, in our case, the recurrent prediction (Asabuki and Fukai; 2020). Thus, the rule suggests that the recurrent prediction learns the statistical model of the evoked activity, which in turn allows the network to reproduce the learned transition statistics.

We have also added a paragraph to discuss the differences between previously published similar models (e.g., Kappel et al.). Please see our response above.

B) Connection to biology 

The plasticity rules in the study are introduced with a vague reference to previous theoretical studies of others. Please discuss in more detail if these rules (especially the error-based learning rule) could be implemented biologically and how this could be achieved. Are there connections to biologically observed plasticity? E.g. for error-based plasticity has been discussed in the original publication by Urbanzcik and Senn, or more recently by Mikulasch et al (TINS 2023). The biological plausibility of inhibitory balance has been discussed many times before, e.g. by Vogels and others, and a citation would acknowledge that earlier work. This also leaves the question of how neurons in the songbird experiment could adapt and if the model does capture this well (i.e., do they exhibit E-I balance? etc), which might be discussed as well. 

Last, please provide some testable experimental predictions. By proposing an interesting experimental prediction, the model could become considerably more relevant to experimentalists. Also, are there potentially alternative models of stochastic sequence learning (e.g., Kappel et al)? How could they be distinguished? (especially, again, why not Hebbian/STDP learning?) 

We have cited the Vogels paper to acknowledge the earlier work. We have also included additional paragraphs to discuss a possible biologically plausible implementation of our model and how our model differs from similar models proposed previously (e.g., Kappel et al.). Please see our response above.

Other comments 

As mentioned, a derivation of recurrent plasticity rules is missing, and parameters are chosen ad-hoc. This leaves the question of how much the results rely on the specific choice of parameters, and how robust they are to perturbations. As a robustness check, please clarify how the duration of the Markov states influences performance. It can be expected that this interacts with the timescale of recurrent connections, so having longer or shorter Markov states, as it would be in reality, should make a difference in learning that should be tested and discussed.

We thank the reviewer for pointing this out. To address this point, we performed new simulations and asked to what extent the duration of Markov states affect performance. Interestingly, even when the network was trained with input states of half the duration, the distributions of the durations of assembly reactivations remain almost identical to those in the original case (Supplementary Figure 3a). Furthermore, the transition probabilities in the replay were still consistent with the true transition probabilities (Supplementary Figure 3b). We have also included the derivation of our plasticity rule in ll. 630-670 in the revised manuscript. 

Similarly, inhibitory plasticity operates with the same plasticity timescale parameter as excitatory plasticity, but, as the authors discuss, lags behind excitatory plasticity in simulation as in experiment. Is this required or was the parameter chosen such that this behaviour emerges? Please clarify this in the methods section; moreover, it would be good to test if the same results appear with fast inhibitory plasticity. 

We have performed a new simulation and showed that even when the learning rate of inhibitory plasticity was larger than that of excitatory plasticity, inhibitory plasticity still occurred on a slower timescale than excitatory plasticity. We have included this result in a new Supplementary Figure 2 in the revised manuscript.

What is the justification (biologically and theoretically) for the memory trace h and its impact on neural spiking? Is it required for the results or can it be left away? Since this seems to be an important and unconventional component of the model, please discuss it in more detail. 

In the model, it is assumed that each stimulus presentation drives a specific subset of network neurons with a fixed input strength, which avoids convergence to trivial solutions. Nevertheless, we choose to add this dynamic sigmoid function to facilitate stable replay by regulating neuron activity to prevent saturation. We have explained this point in ll.605-611 in the revised manuscript.

Reviewer #2 (Recommendations For The Authors): 

I noticed a couple of minor typos: 

Page 3 "underly"->"underlie" 

Page 7 "assemblies decreased settled"->"assemblies decreased and settled"

We have modified the text. We thank the reviewer for their careful review.

I think Figure 1C is rather confusing and not intuitive. 

We apologize that the Figure 1C was confusing. In the revised figure, we have emphasized the flow of excitatory and inhibitory error for updating synapses.

Reviewer #3 (Recommendations For The Authors): 

One possible path to improve the paper would be to establish a relationship between the proposed learning rules and e.g. the ones derived by Barber. 

When reading the paper, I was left with a number of more detailed questions I omitted from the public review: 

(1) The authors introduce a dynamic sigmoidal function for excitatory neurons, Eq. 3. This point requires more discussion and analysis. How does this impact the results? 

In the model, it is assumed that each stimulus presentation drives a specific subset of network neurons with a fixed input strength, which avoids convergence to trivial solutions. Nevertheless, we choose to add this dynamic sigmoid function to facilitate stable replay by regulating neuron activity to prevent saturation. We have explained this point in ll.605-611 in the revised manuscript.

(2) For Poisson spiking neurons, it would be great to understand what cell assemblies bring (apart from biological realism, i.e., reproducing data where assemblies can be found), compared to self-connected single neurons. For example, how do the results shown in Figure 2 depend on assembly size? 

We have changed the cell assembly size ratio and how it affects learning performance in a new Supplementary Figure 4. Please see our reply above.

(3) The authors focus on modeling spontaneous transitions, corresponding to a highly stochastic generative model (with most transition probabilities far from 1). A complementary question is that of learning to produce a set of stereotypical sequences, with probabilities close to 1. I wondered whether the learning rules and architecture of the model (in particular under the I-to-E rule) would also work in such a scenario. 

We thank the reviewer for pointing this out. In fact, we had the same question, so we considered a situation in which the setting in Figure 2 includes both cases where the transition matrix is very stochastic (prob=0.5) and near deterministic (prob=0.9).

(4) An analysis of what controls the time so that the network stays in a certain state would be welcome. 

We trained the network model in two cases, one with a fast speed of plasticity and one with a slow speed of plasticity. As a result, we found that the duration of assembly becomes longer in the slow learning case than in the fast case. We have included these results as Supplementary Figure 5 in the revised manuscript.

Regarding the presentation, given that this is a computational modeling paper, I wonder whether *all* the formulas belong in the Methods section. I found myself skipping back and forth to understand what the main text meant, mainly because I missed a few key equations. I understand that this is a style issue that is very much community-dependent, but I think readability would improve drastically if the main model and learning rule equations could be introduced in the main text, as they start being discussed. 

We thank the reviewer for the suggestion. To cater to a wider audience, we try to explain the principle of the paper without using mathematical formulas as much as possible in the main text.

eLife Assessment

This is an important study that investigates how neural networks can learn to stochastically replay presented sequences of activity according to learned transition probabilities. The authors use error-based excitatory plasticity to minimize the difference between internally predicted activity and stimulus-driven activity, and inhibitory plasticity to maintain E-I balance. The approach is solid but the choice of learning rules and parameters is not always always justified, with some unclear aspects to the formal derivation.

Reviewer #2 (Public review):

Summary:

This work proposes a synaptic plasticity rule which explains the generation of learned stochastic dynamics during spontaneous activity. The proposed plasticity rule assumes that excitatory synapses seek to minimize the difference between the internal predicted activity and stimulus-evoked activity, and inhibitory synapses tries to maintain the E-I balance by matching the excitatory activity. By implementing this plasticity rule in a spiking recurrent neural network, the authors show that the state-transition statistics of spontaneous excitatory activity agrees with that of the learned stimulus patterns, which is reflected in the learned excitatory synaptic weights. The authors further demonstrate that inhibitory connections contribute to well-defined state-transitions matching the transition patterns evoked by the stimulus. Finally, they show that this mechanism can be expanded to more complex state-transition structures including songbird neural data.

Strengths:

This study makes an important contribution to computational neuroscience, by proposing a possible synaptic plasticity mechanism underlying spontaneous generations of learned stochastic state-switching dynamics that are experimentally observed in the visual cortex and hippocampus. This work is also very clearly presented and well-written, and the authors conducted comprehensive simulations testing multiple hypotheses. Overall, I believe this is a well-conducted study providing interesting and novel aspects on the capacity of recurrent spiking neural networks with local synaptic plasticity.

Weaknesses:

This study is very well-thought out and theoretically valuable to the neuroscience community, and I think the main weaknesses are in regard to how much biological realism is taken into account. For example, the proposed model assumes that only synapses targeting excitatory neurons are plastic, and uses an equal number of excitatory and inhibitory neurons.<br /> The model also assumes Markovian state dynamics while biological systems can depend more on history. This limitation, however, is acknowledged in the Discussion.<br /> Finally, to simulate spontaneous activity, the authors use a constant input of 0.3 throughout the study. Different amplitudes of constant input may correspond to different internal states, so it will be more convincing if the authors test the model with varying amplitudes of constant inputs.

Comments on revisions:

The authors have addressed all of the previously raised concerns satisfactorily, by running extra simulations with a biologically plausible composition of excitatory and inhibitory neurons, plasticity assumed for all synapses, and varied amounts of constant inputs representing internal states or background activities. While in some of these cases the stochastic dynamics during spontaneous activity change or do not replicate those of the learned stimulus patterns as well as before, these extended studies provide thorough evaluations of the strengths and limitations of the proposed plasticity rule as the underlying mechanism of stochastic dynamics during spontaneous activity. Overall, the revision has strengthened the paper significantly.

Reviewer #3 (Public review):

Summary:

Asabuki and Clopath study stochastic sequence learning in recurrent networks of Poisson spiking neurons that obey Dale's law. Inspired by previous modeling studies, they introduce two distinct learning rules, to adapt excitatory-to-excitatory and inhibitory-to-excitatory synaptic connections. Through a series of computer experiments, the authors demonstrate that their networks can learn to generate stochastic sequential patterns, where states correspond to non-overlapping sets of neurons (cell assemblies) and the state-transition conditional probabilities are first-order Markov, i.e., the transition to a given next state only depends on the current state. Finally, the authors use their model to reproduce certain experimental songbird data involving highly-predictable and highly-uncertain transitions between song syllables. While the findings are only moderately surprising, this is a well-written and welcome detailed study that may be of interest to experts of plasticity and learning in recurrent neural networks that respect Dale's law.

Strengths:

This is an easy-to-follow, well-written paper, whose results are likely easy to reproduce. The experiments are clear and well-explained. In particular, the study of the interplay between excitation and inhibition (and their different plasticity rules) is a highlight of the study. The study of songbird experimental data is another good feature of this paper; finches are classical model animals for understanding sequence learning in the brain. I also liked the study of rapid task-switching, it's a good-to-know type of result that is not very common in sequence learning papers.

Weaknesses:

One weakness I see in this paper is the derivation of the learning rules, which is semi-heuristic. The paper studies Poisson spiking neurons, for which learning rules can be derived from a statistical objective, typically maximum likelihood, as previously done in the cited literature. The authors provide a brief section connecting the learning rules to gradient descent on objective functions, but the link is only heuristic or at least not entirely presented. The reason is that the neural network state is not fully determined by (or "clamped to") the target during learning (for instance, inhibitory neurons do not even have a target assigned). So, the (total) gradient should take into account the recurrent contributions from other neurons, and equation 13 does not appear to be complete/correct to me. Moreover, the target firing rate is a mixture of external currents with currents arising from other neurons in the recurrent network. The authors ideally should start from an actual distribution matching objective (e.g., KL divergence, and not such a squared error), so that their main claims immediately follow from the mathematical derivations. Along the same line, it would be excellent to get some additional insights on the interaction of the two distinct plasticity rules, one of the highlights of the study. This could be naturally achieved by relating their distinct rules to a common principled objective.

The other major weakness (albeit one that is clearly discussed by the authors) is that the study assumes that every excitatory neuron is directly given its target state when learning. In machine learning language, there are no 'hidden' excitatory neurons. While this assumption greatly simplifies the derivation of efficient and biologically-plausible learning rules that can be mapped to synaptic plasticity, it also limits considerably the distributions that can be learned by the network, more precisely to those that satisfy the Markov property.

"Razorous" is a made-up yet intuitively word used by McCarthy to mean "like a razor".

Exploration des Intolérances Alimentaires : Réalité ou Illusion ?

Source : Extrait du documentaire "Les intolérances alimentaires : mal du siècle ou illusion ? | Les questions qui fâchent | ARTE"

I. Introduction : Le Boom des Intolérances Alimentaires (0:00 - 4:00)

L'extrait débute en présentant le narrateur, Bertold Meer, professeur de psychologie, qui se soumet à un test sanguin pour détecter d'éventuelles intolérances alimentaires, illustrant la popularité croissante de ce phénomène.

Le documentaire questionne ensuite les causes de cet engouement, se demandant s'il s'agit d'une simple mode ou d'un symptôme d'une alimentation moderne néfaste.

L'interview de Natalia Avlon, une femme souffrant d'intolérances alimentaires depuis l'enfance, met en lumière les difficultés et les restrictions liées à ces conditions.

II. Le Rôle de l'Alimentation Moderne et de la Flore Intestinale (4:00 - 9:00)

Le nutritionniste Mathias Ridle explique que la consommation excessive d'aliments transformés, pauvre en fibres, perturbe la flore intestinale et engendre divers maux, souvent confondus avec des intolérances alimentaires.

Il critique l'industrie agroalimentaire qui, selon lui, profite de l'ignorance des consommateurs pour promouvoir des produits de substitution inutiles et potentiellement néfastes.

Ridle souligne l'importance d'une alimentation équilibrée et riche en aliments frais pour préserver la santé intestinale.

III. Influences Socio-Culturelles et Intérêts Economiques (9:00 - 14:00)

Le sociologue Daniel C. analyse l'influence de la politique, de l'industrie et de la science sur nos habitudes alimentaires.

Il démontre que les recommandations nutritionnelles varient selon les régions du monde, reflétant les intérêts économiques locaux, comme l'exemple du poisson dans les pays nordiques.

Daniel C. met en lumière la dimension identitaire de l'alimentation, expliquant que nos choix alimentaires reflètent désormais ce que nous refusons de manger.

Il souligne les différences culturelles dans la perception de l'alimentation, contrastant la préoccupation allemande pour la santé avec l'importance du plaisir gustatif en France et en Italie.

IV. Le Rôle des Produits Chimiques et la Complexité des Diagnostics (14:00 - 19:00)

L'immunologiste Catherine Negler met en cause la présence de produits chimiques dans les aliments, suspectés de provoquer des réactions indésirables et des diagnostics erronés d'allergies.

Elle explique que ces substances chimiques altèrent le microbiote intestinal, augmentant le risque de maladies chroniques comme le diabète, l'obésité et les MICI.

Negler déplore l'impact négatif de l'alimentation moderne sur notre santé, soulignant la multiplication des allergies et des maladies chroniques dans les sociétés occidentales.

V. Le Mythe du "Sans Gluten" et la Fiabilité des Tests (19:00 - 24:00)

Le documentaire explore ensuite le phénomène du "sans gluten," soulignant son adoption massive malgré l'absence de bénéfices scientifiques pour la majorité de la population.

Le narrateur, Bertold Meer, confronté aux résultats de son test sanguin, découvre une liste d'aliments à éviter, suscitant son étonnement et son questionnement.

Christina Hel, docteur en nutrition, met en garde contre la fiabilité des tests IgG, expliquant qu'ils ne révèlent pas d'intolérances mais simplement la présence d'anticorps après la consommation d'un aliment, un processus naturel et bénéfique.

VI. Conclusion : Intolérances Réelles, Incomforts et Alimentation Industrielle (24:00 - 28:00)

Le documentaire s'achève sur une réflexion nuancée, reconnaissant l'existence d'intolérances alimentaires réelles tout en soulignant la tendance à l'autodiagnostic erroné.

L'impact négatif de l'alimentation industrielle est réaffirmé, pointant du doigt la consommation excessive d'aliments transformés comme cause principale des problèmes digestifs.

L'extrait se conclut sur un appel à la prudence, encourageant une alimentation équilibrée et riche en aliments frais, tout en invitant à savourer sa nourriture sans se laisser envahir par l'inquiétude.

Remove

climate

Jupyter Book

has

provided

Jupyter Book

slides

Information

test -> test3

使用图片学习 使用影像学习 给别人讲解 来加深记忆

eLife Assessment

This fundamental study combines Global Positioning System tracking and the analysis of social interactions among feral pigs, to provide insights into the likelihood of disease transmission based on contact rates both within and between sounders. The method used for data collection is compelling, but the varying sample sizes across populations could be a potential source of bias. With the potential biases from varying sample sizes strengthened this paper would be of interest to the fields of Veterinary Medicine, Public Health, and Epidemiology.

Reviewer #1 (Public review):

Summary:

The authors aimed to quantify feral pig interactions in eastern Australia to inform disease transmission networks. They used GPS tracking data from 146 feral pigs across multiple locations to construct proximity-based social networks and analyze contact rates within and between pig social units.

Strengths:

(1) Addresses a critical knowledge gap in feral pig social dynamics in Australia.

(2) Uses robust methodology combining GPS tracking and network analysis.

(3) Provides valuable insights into sex-based and seasonal variations in contact rates.

(4) Effectively contextualizes findings for disease transmission modeling and management.

(5) Includes comprehensive ethical approval for animal research.

(6) Utilizes data from multiple locations across eastern Australia, enhancing generalizability.

Weaknesses:

(1) Limited discussion of potential biases from varying sample sizes across populations

(2) Some key figures are in supplementary materials rather than the main text.

(3) Economic impact figures are from the US rather than Australia-specific data.

(4) Rationale for spatial and temporal thresholds for defining contacts could be clearer.

(5) Limited discussion of ethical considerations beyond basic animal ethics approval.

The authors largely achieved their aims, with the results supporting their conclusions about the importance of sex and seasonality in feral pig contact networks. This work is likely to have a significant impact on feral pig management and disease control strategies in Australia, providing crucial data for refining disease transmission models.

Reviewer #2 (Public review):

Summary:

The paper attempts to elucidate how feral (wild) pigs cause distortion of the environment in over 54 countries of the world, particularly Australia.

The paper displays proof that over $120 billion worth of facilities were destroyed annually in the United States of America.

The authors have tried to infer that the findings of their work were important and possess a convincing strength of evidence.

Strengths:

(1) Clearly stating feral (wild) pigs as a problem in the environment.

(2) Stating how 54 countries were affected by the feral pigs.

(3) Mentioning how $120 billion was lost in the US, annually, as a result of the activities of the feral pigs.

(4) Amplifying the fact that 14 species of animals were being driven into extinction by the feral pigs.

(5) Feral pigs possessing zoonotic abilities.

(6) Feral pigs acting as reservoirs for endemic diseases like brucellosis and leptospirosis.

(7) Understanding disease patterns by the social dynamics of feral pig interactions.

(8) The use of 146 GPS-monitored feral pigs to establish their social interaction among themselves.

Weaknesses:

(1) Unclear explanation of the association of either the female or male feral pigs with each other, seasonally.

(2) The "abstract paragraph" was not justified.

(3) Typographical errors in the abstract.

Reviewer #3 (Public review):

Summary:

The authors sought to understand social interactions both within and between groups of feral pigs, with the intent of applying their findings to models of disease transmission. The authors analyzed GPS tracking data from across various populations to determine patterns of contact that could support the transmission of a range of zoonotic and livestock diseases. The analysis then focused on the effects of sex, group dynamics, and seasonal changes on contact rates that could be used to base targeted disease control strategies that would prioritize the removal of adult males for reducing intergroup disease transmission.

Strengths:

It utilized GPS tracking data from 146 feral pigs over several years, effectively capturing seasonal and spatial variation in the social behaviors of interest. Using proximity-based social network analysis, this work provides a highly resolved snapshot of contact rates and interactions both within and between groups, substantially improving research in wildlife disease transmission. Results were highly useful and provided practical guidance for disease management, showing that control targeted at adult males could reduce intergroup disease transmission, hence providing an approach for the control of zoonotic and livestock diseases.

Weaknesses:

Despite their reliability, populations can be skewed by small sample sizes and limited generalizability due to specific environmental and demographic characteristics. Further validation is needed to account for additional environmental factors influencing social dynamics and contact rates

Author response:

Public Reviews:

Reviewer #1 (Public review):

Summary:

The authors aimed to quantify feral pig interactions in eastern Australia to inform disease transmission networks. They used GPS tracking data from 146 feral pigs across multiple locations to construct proximity-based social networks and analyze contact rates within and between pig social units.

Strengths:

(1) Addresses a critical knowledge gap in feral pig social dynamics in Australia.

(2) Uses robust methodology combining GPS tracking and network analysis.

(3) Provides valuable insights into sex-based and seasonal variations in contact rates.

(4) Effectively contextualizes findings for disease transmission modeling and management.

(5) Includes comprehensive ethical approval for animal research.

(6) Utilizes data from multiple locations across eastern Australia, enhancing generalizability.

Weaknesses:

(1) Limited discussion of potential biases from varying sample sizes across populations

This is a really good comment, and we will address this in the discussion as one of the limitations of the study.

(2) Some key figures are in supplementary materials rather than the main text.

We will move some of our supplementary material to the main text as suggested.

(3) Economic impact figures are from the US rather than Australia-specific data.

We included the impact figures that are available for Australia (for FDM), and we will include the estimated impact of ASF in Australia in the introduction.

(4) Rationale for spatial and temporal thresholds for defining contacts could be clearer.

We will improve the explanation of why we chose the spatial and temporal thresholds based on literature, the size of animals and GPS errors.

(5) Limited discussion of ethical considerations beyond basic animal ethics approval.

This research was conducted under an ethics committee's approval for collaring the feral pigs. This research is part of an ongoing pest management activity, and all the ethics approvals have been highlighted in the main manuscript.

The authors largely achieved their aims, with the results supporting their conclusions about the importance of sex and seasonality in feral pig contact networks. This work is likely to have a significant impact on feral pig management and disease control strategies in Australia, providing crucial data for refining disease transmission models.

Reviewer #2 (Public review):

Summary:

The paper attempts to elucidate how feral (wild) pigs cause distortion of the environment in over 54 countries of the world, particularly Australia.

The paper displays proof that over $120 billion worth of facilities were destroyed annually in the United States of America.

The authors have tried to infer that the findings of their work were important and possess a convincing strength of evidence.

Strengths:

(1) Clearly stating feral (wild) pigs as a problem in the environment.

(2) Stating how 54 countries were affected by the feral pigs.

(3) Mentioning how $120 billion was lost in the US, annually, as a result of the activities of the feral pigs.

(4) Amplifying the fact that 14 species of animals were being driven into extinction by the feral pigs.

(5) Feral pigs possessing zoonotic abilities.

(6) Feral pigs acting as reservoirs for endemic diseases like brucellosis and leptospirosis.

(7) Understanding disease patterns by the social dynamics of feral pig interactions.

(8) The use of 146 GPS-monitored feral pigs to establish their social interaction among themselves.

Weaknesses:

(1) Unclear explanation of the association of either the female or male feral pigs with each other, seasonally.

This will be better explain in the methods.

(2) The "abstract paragraph" was not justified.

We have justified the abstract paragraph as requested by the reviewer.

(3) Typographical errors in the abstract.

Typographical errors have been corrected in the Abstract.

Reviewer #3 (Public review):

Summary:

The authors sought to understand social interactions both within and between groups of feral pigs, with the intent of applying their findings to models of disease transmission. The authors analyzed GPS tracking data from across various populations to determine patterns of contact that could support the transmission of a range of zoonotic and livestock diseases. The analysis then focused on the effects of sex, group dynamics, and seasonal changes on contact rates that could be used to base targeted disease control strategies that would prioritize the removal of adult males for reducing intergroup disease transmission.

Strengths:

It utilized GPS tracking data from 146 feral pigs over several years, effectively capturing seasonal and spatial variation in the social behaviors of interest. Using proximity-based social network analysis, this work provides a highly resolved snapshot of contact rates and interactions both within and between groups, substantially improving research in wildlife disease transmission. Results were highly useful and provided practical guidance for disease management, showing that control targeted at adult males could reduce intergroup disease transmission, hence providing an approach for the control of zoonotic and livestock diseases.

Weaknesses:

Despite their reliability, populations can be skewed by small sample sizes and limited generalizability due to specific environmental and demographic characteristics. Further validation is needed to account for additional environmental factors influencing social dynamics and contact rates

This is a good point, and we thank the reviewer for pointing out this issue. We will discuss the potential biases due to sample size in our discussion. We agree that environmental factors need to be incorporated and tested for their influence on social dynamics, and this will be added to the discussion as we have plans to expand this research and conduct, the analysis to determine if environmental factors are influencing social dynamics.

eLife Assessment

This valuable study uses extensive comparative analysis to examine the relationship between plasma glucose levels, albumin glycation levels, and diet and life history, within the framework of the "pace of life syndrome" hypothesis. The evidence that glucose and glycation levels are broadly correlated is convincing. However, concerns about the consistency of the data quality across species and some aspects of data analysis make the key conclusion about higher glycation resistance in species with higher glucose levels currently incomplete. Still, as the first extensive comparative analysis of glycation rates, life history, and glucose levels in birds, the study has potential to be of interest to evolutionary ecologists and the aging research community more broadly.

Reviewer #1 (Public review):

The paper explored cross-species variance in albumin glycation and blood glucose levels in the function of various life-history traits. Their results show that<br /> (1) blood glucose levels predict albumin gylcation rates<br /> (2) larger species have lower blood glucose levels<br /> (3) lifespan positively correlates with blood glucose levels and<br /> (4) diet predicts albumin glycation rates.

The data presented is interesting, especially due to the relevance of glycation to the ageing process and the interesting life-history and physiological traits of birds. Most importantly, the results suggest that some mechanisms might exist that limit the level of glycation in species with the highest blood glucose levels.

While the questions raised are interesting and the amount of data the authors collected is impressive, I have some major concerns about this study:

(1) The authors combine many databases and samples of various sources. This is understandable when access to data is limited, but I expected more caution when combining these. E.g. glucose is measured in all samples without any description of how handling stress was controlled for. E.g glucose levels can easily double in a few minutes in birds, potentially introducing variation in the data generated. The authors report no caution of this effect, or any statistical approaches aiming to check whether handling stress had an effect here, either on glucose or on glycation levels.

(2) The database with the predictors is similarly problematic. There is information pulled from captivity and wild (e.g. on lifespan) without any confirmation that the different databases are comparable or not (and here I'm not just referring to the correlation between the databases, but also to a potential systematic bias (e.g. captivate-based sources likely consistently report longer lifespans). This is even more surprising, given that the authors raise the possibility of captivity effects in the discussion, and exploring this question would be extremely easy in their statistical models (a simple covariate in the MCMCglmms).

(3) The authors state that the measurement of one of the primary response variables (glycation) was measured without any replicability test or reference to the replicability of the measurement technique.

(4) The methods and results are very poorly presented. For instance, new model types and variables are popping up throughout the manuscript, already reporting results, before explaining what these are e.g. results are presented on "species average models" and "model with individuals", but it's not described what these are and why we need to see both. Variables, like "centered log body mass", or "mass-adjusted lifespan" are not explained. The results section is extremely long, describing general patterns that have little relevance to the questions raised in the introduction and would be much more efficiently communicated visually or in a table.

Reviewer #2 (Public review):

Summary

In this extensive comparative study, Moreno-Borrallo and colleagues examine the relationships between plasma glucose levels, albumin glycation levels, diet, and life-history traits across birds. Their results confirmed the expected positive relationship between plasma blood glucose level and albumin glycation rate but also provided findings that are somewhat surprising or contradicting findings of some previous studies (relationships with lifespan, clutch mass, or diet). This is the first extensive comparative analysis of glycation rates and their relationships to plasma glucose levels and life history traits in birds that are based on data collected in a single study and measured using unified analytical methods.

Strengths

This is an emerging topic gaining momentum in evolutionary physiology, which makes this study a timely, novel, and very important contribution. The study is based on a novel data set collected by the authors from 88 bird species (67 in captivity, 21 in the wild) of 22 orders, which itself greatly contributes to the pool of available data on avian glycemia, as previous comparative studies either extracted data from various studies or a database of veterinary records of zoo animals (therefore potentially containing much more noise due to different methodologies or other unstandardised factors), or only collected data from a single order, namely Passeriformes. The data further represents the first comparative avian data set on albumin glycation obtained using a unified methodology. The authors used LC-MS to determine glycation levels, which does not have problems with specificity and sensitivity that may occur with assays used in previous studies. The data analysis is thorough, and the conclusions are mostly well-supported (but see my comments below). Overall, this is a very important study representing a substantial contribution to the emerging field of evolutionary physiology focused on the ecology and evolution of blood/plasma glucose levels and resistance to glycation.

Weaknesses

My main concern is about the interpretation of the coefficient of the relationship between glycation rate and plasma glucose, which reads as follows: "Given that plasma glucose is logarithm transformed and the estimated slope of their relationship is lower than one, this implies that birds with higher glucose levels have relatively lower albumin glycation rates for their glucose, fact that we would be referring as higher glycation resistance" (lines 318-321) and "the logarithmic nature of the relationship, suggests that species with higher plasma glucose levels exhibit relatively greater resistance to glycation" (lines 386-388). First, only plasma glucose (predictor) but not glycation level (response) is logarithm transformed, and this semi-logarithmic relationship assumed by the model means that an increase in glycation always slows down when blood glucose goes up, irrespective of the coefficient. The coefficient thus does not carry information that could be interpreted as higher (when <1) or lower (when >1) resistance to glycation (this only can be done in a log-log model, see below) because the semi-log relationship means that glycation increases by a constant amount (expressed by the coefficient of plasma glucose) for every tenfold increase in plasma glucose (for example, with glucose values 10 and 100, the model would predict glycation values 2 and 4 if the coefficient is 2, or 0.5 and 1 if the coefficient is 0.5). Second, the semi-logarithmic relationship could indeed be interpreted such that glycation rates are relatively lower in species with high plasma glucose levels. However, the semi-log relationship is assumed here a priori and forced to the model by log-transforming only glucose level, while not being tested against alternative models, such as: (i) a model with a simple linear relationship (glycation ~ glucose); or (ii) a log-log model (log(glycation) ~ log(glucose)) assuming power function relationship (glycation = a * glucose^b). The latter model would allow for the interpretation of the coefficient (b) as higher (when <1) or lower (when >1) resistance in glycation in species with high glucose levels as suggested by the authors.

Besides, a clear explanation of why glucose is log-transformed when included as a predictor, but not when included as a response variable, is missing.

The models in the study do not control for the sampling time (i.e., time latency between capture and blood sampling), which may be an important source of noise because blood glucose increases because of stress following the capture. Although the authors claim that "this change in glucose levels with stress is mostly driven by an increase in variation instead of an increase in average values" (ESM6, line 46), their analysis of Tomasek et al.'s (2022) data set in ESM1 using Kruskal-Wallis rank sum test shows that, compared to baseline glucose levels, stress-induced glucose levels have higher median values, not only higher variation.

Although the authors calculated the variance inflation factor (VIF) for each model, it is not clear how these were interpreted and considered. In some models, GVIF^(1/(2*Df)) is higher than 1.6, which indicates potentially important collinearity; see for example https://www.bookdown.org/rwnahhas/RMPH/mlr-collinearity.html). This is often the case for body mass or clutch mass (e.g. models of glucose or glycation based on individual measurements).

It seems that the differences between diet groups other than omnivores (the reference category in the models) were not tested and only inferred using the credible intervals from the models. However, these credible intervals relate to the comparison of each group with the reference group (Omnivore) and cannot be used for pairwise comparisons between other groups. Statistics for these contrasts should be provided instead. Based on the plot in Figure 4B, it seems possible that terrestrial carnivores differed in glycation level not only from omnivores but also from herbivores and frugivores/nectarivores.

Given that blood glucose is related to maximum lifespan, it would be interesting to also see the results of the model from Table 2 while excluding blood glucose from the predictors. This would allow for assessing if the maximum lifespan is completely independent of glycation levels. Alternatively, there might be a positive correlation mediated by blood glucose levels (based on its positive correlations with both lifespan and glycation), which would be a very interesting finding suggesting that high glycation levels do not preclude the evolution of long lifespans.

Author response:

Reviewer #1:

(1) This concern is addressed in the ESM6, and partly in the ESM1. Indeed, many of the concerns raised by the reviewer later are already addressed on the multiple supplementary materials provided, so we kindly ask the reviewer to read them before moving forward into the discussion.

(2) This concern is reasonable, but its solution is not "extremely easy", as the reviewer states. The reviewer indicates the use of captive-based versus non-captive-based sources, remarking maximum lifespan, the main variable that is clearly expected to be systematically biased by the source of the data. Nevertheless, except for the ZIMS database, which includes only captive individuals, and some sources, as CNRS databases and EURING, which exclusively includes wild populations, the remaining databases, which are indeed where the vast majority of the data was collected from (i.e. Amniotes database, Birds of the World and AnAge) do not make any distinction. This means that they include just the maximum lifespan from the species as known by the authors of such databases' entries, regardless of provenance, which is also not usually made explicit by the database. Therefore, correcting for this would imply checking all the primary sources. Considering that these databases sometimes do not cite the primary source, but a secondary one, and that on several occasions such source is a specialized book that is not easily accessible, and still these referenced datasets may not indicate the source of the data, tracing all of this information becomes an arduous task, that would even render the usage of databases themselves useless. We will include some details about the concerns of database usage in the discussion to address this.

Furthermore, it remains relevant to indicate that what we discuss later about the possible effects of captivity is about our usage of animals that come from both sources, not about the provenance of the literature-extracted data used (i.e. captive or wild maximum lifespan, for example), which is an independent matter. We can test for the first for next submission, but very difficultly could we test for the second (as the reviewer seems to be pointing to). In any case, as we do not have in any case the same species from both a captive and a wild source, it would be difficult to determine if the effect tested comes from captivity or from species-specific differences.

(3) We will add data on the replicability of the glycation measurement in the next manuscript version. The CV for several individuals of different species measured repeated times is quite low (always below 2%).

(4) The reviewer remarks reported here are already addressed on the supplementary material (ESM6), given the lack of space in the main manuscript. We therefore kindly ask the reviewer to read the supplementary material added to the submission. If the editors agree, all or a considerable part of this could be transferred to the main text for clarity, but this would severely extend the length of a text that the reviewer already considered very long.

Reviewer #2:

Thanks for spotting this issue with the coefficient, as it is actually a redaction mistake. It is a remnant of a previous version of the manuscript in which a log-log relation was performed instead. Previous reviewers raised concerns about the usage of log transformation for glycation, this variable being (theoretically) a proportion variable (to which we argue that it does not behave as such), which they considered not to be transformed with a logarithm. After this, we still finally took the decision of not to transform this variable. In this line, the transformations of variables were decided generally by preliminary data exploration. In this particular case, both approaches lead to the same conclusion of higher glycation resistance in the species with higher glucose. Nevertheless, we will consider exploring the comparison of different versions for the resubmission.

About the issue related to handling time, this variable is not available, for the reasons already exposed in the answer to the other reviewer. Moreover, Kruskal-Wallis test, by its nature, does not determine differences in medians between groups per se, as the reviewer claims, but just differences in ranks-sums. It can be equivalently used for that purpose when the groups' distributions are similar, but not when they differ, as we see here with a difference in variance. What a significant outcome in a Kruskal-Wallis test tells us, thus, is just that the groups differ (in their ranks-sums), which here is plausibly caused by the higher variance in the stressed individuals. Even if we conclude that the average is higher in those groups, mere comparisons of averages for groups with very different variances render different interpretations than when homoscedasticity is met, particularly more so when the distribution of groups overlaps. For example, in a case like this, where the data is left censored (glucose levels cannot be lower than 0), most of this higher variance is related to many values in the stressed groups lying above all the baseline values. This, of course, would increase the average, but such a parameter would not mean the same as if the distributions did not overlap.

Regarding the GVIFs, why the values are above 1.6 is not well known, but we do not consider this a major concern, as the values are never above 2.2, level usually considered more worrying. We will include a brief explanation of this in the results section. Also, we explicitly calculated life history variables adjusted for body mass, which should eliminate their otherwise strong correlation. There exist other biological and interpretational reasons justified in the ESM6 for using the residuals on the models, instead of the raw values, despite previously raised concerns.

Given the asseveration by the reviewer that credible intervals are not to be used for the post hoc comparisons, as this is what the whiskers shown in Figure 4B represent, the affirmation of this graph suggesting any difference between groups remains doubtful. New comparisons have now been made with the function HPDinterval() applied to the differences between each diet category calculated from the posterior values of each group, confirming no significant differences exist.

We do not understand the suggestion made in relation to the model shown in Table 2. Removing glucose from the model could have two results, as the reviewer indicates: 1. Maximum lifespan (ML) relates with glycation, potentially spuriously through the effect of glucose (in this case not included) on both; 2. ML does not relate to glycation, and therefore "high glycation levels do not preclude the evolution of long lifespans", which is what we are already showing with the current model, which also controls for glucose, in an attempt to determine if not just raw glycation values, but glycation resistance, relates to longevity. This is intended to asses if long-lived species may show mechanisms that avoid glycation, by showing levels lower than expected for a non-enzymatic reaction.

eLife Assessment

This study aims to investigate the RNA binding activities of a conserved heterochromatin protein (Swi6) and proposes an entirely new model for how heterochromatin formation is initiated in fission yeast. While the concept is interesting, the data provided are inadequate, both for support of the claims regarding the new RNA binding activities and for support of the new model. The paper requires extensive editing as well as the inclusion of numerous experiments with appropriately controlled conditions.

Reviewer #1 (Public review):

Summary:

This manuscript explores the RNA binding activities of the fission yeast Swi6 (HP1) protein and proposes a new role for Swi6 in RNAi-mediated heterochromatin establishment. The authors claim that Swi6 has a specific and high affinity for short interfering RNAs (siRNAs) and recruits the Clr4 (Suv39h) H3K9 methyltransferases to siRNA-DNA hybrids to initiate heterochromatin formation. These claims are not in any way supported by the incomplete and preliminary RNA binding or the in vivo experiments that the authors present. The proposed model also lacks any mechanistic basis as it remains unclear (and unexplored) how Swi6 might bind to specific small RNA sequences or RNA-DNA hybrids. Work by several other groups in the field has led to a model in which siRNAs produced by the RNAi pathway load onto the Ago1-containing RITS complex, which then binds to nascent transcripts at pericentromeric DNA repeats and recruits Clr4 to initiate heterochromatin formation. Swi6 facilitates this process by promoting the recruitment of the RNA-dependent RNA polymerase leading to siRNA amplification.

Weaknesses:

(1) The claims that Swi6 binds to specific small RNAs or to RNA-DNA hybrids are not supported by the evidence that the authors present. Their experiments do not rule out non-specific charged-based interactions. Claims about different affinities of Swi6 for RNAs of different sizes are based on a comparison of KD values derived by the authors for a handful of S. pombe siRNAs with previous studies from the Buhler lab on Swi6 RNA binding. The authors need to compare binding affinities under identical conditions in their assays. The regions of Swi6 that bind to siRNAs need to be identified and evidence must be provided that Swi6 binds to RNAs of a specific length, 20-22 mers, to support the claim that Swi6 binds to siRNAs. This is critical for all the subsequent experiments and claims in the study.

(2) The in vivo results do not validate Swi6 binding to specific RNAs, as stated by the authors. Swi6 pulldowns have been shown to be enriched for all heterochromatic proteins including the RITS complex. The sRNA binding observed by the authors is therefore likely to be mediated by Ago1/RITS.

Most of the binding in Figure S8C seems to be non-specific.

In Figure S8D, the authors' data shows that Swi6 deletion does not derepress the rev dh transcript while dcr1 delete cells do, which is consistent with previous reports but does not relate to the authors' conclusions.

Previous results have shown that swi6 delete cells have 20-fold fewer dg and dh siRNAs than swi6+ cells due to decreased RNA-dependent RNA polymerase complex recruitment and reduced siRNA amplification.

(3) The RIP-seq data are difficult to interpret as presented. The size distribution of bound small RNAs, and where they map along the genome should be shown as for example presented in previous Ago1 sRNA-seq experiments.

It is also unclear whether the defects in sRNA binding observed by the authors represent direct sRNA binding to Swi6 or co-precipitation of Ago1-bound sRNAs.

The authors should also sequence total sRNAs to test whether Swi6-3A affects sRNA synthesis, as is the case in swi6 delete cells.

(4) The authors examine the effects of Swi6-3A mutant by overexpression from the strong nmt1 promoter. Heterochromatin formation is sensitive to the dosage of Swi6. These experiments should be performed by introducing the 3A mutations at the endogenous Swi6 locus and effects on Swi6 protein levels should be tested.

(5) The authors' data indicate an impairment of silencing in Swi6-3A mutant cells but whether this is due to a general lower affinity for nucleosomes, DNA, RNA, or as claimed by the authors, siRNAs is unclear. These experiments are consistent with previous findings suggesting an important role for basic residues in the HP1 hinge region in gene silencing but do not reveal how the hinge region enhances silencing.

(6) RNase H1 overexpression may affect Swi6 localization and silencing indirectly as it would lead to a general reduction in R loops and RNA-DNA hybrids across the genome. RNaseH1 OE may also release chromatin-bound RNAs that act as scaffolds for siRNA-Ag1/RITS complexes that recruit Clr4 and ultimately Swi6.

(7) Examples of inaccurate presentation of the literature.<br /> a. The authors state that "RNA binding by the murine HP1 through its hinge domains is required for heterochromatin assembly (Muchardt et al, 2002). The cited reference provides no evidence that HP1 RNA binding is required for heterochromatin assembly. Only the hinge region of bacterially produced HP1 contributes to its localization to DAPI-stained heterochromatic regions in fixed NIH 3T3 cells.<br /> b. "... This scenario is consistent with the loss of heterochromatin recruitment of Swi6 as well as siRNA generation in rnai mutants (Volpe et al, 2002)." Volpe et al. did not examine changes in siRNA levels in swi6 mutant cells. In fact, no siRNA analysis of any kind was reported in Volpe et al., 2002.

Reviewer #2 (Public review):

The aim of this study is to investigate the role of Swi6 binding to RNA in heterochromatin assembly in fission yeast. Using in vitro protein-RNA binding assays (EMSA) they showed that Swi6/HP1 binds centromere-derived siRNA (identified by Reinhardt and Bartel in 2002) via the chromodomain and hinge domains. They demonstrate that this binding is regulated by a lysine triplet in the conserved region of the Swi6 hinge domain and that wild-type Swi6 favours binding to DNA-RNA hybrids and siRNA, which then facilitates, rather than competes with, binding to H3K9me2 and to a lesser extent H3K9me3.

However, the majority of the experiments are carried out in swi6 null cells overexpressing wild-type Swi6 or Swi63K-3A mutant from a very strong promoter (nmt1). Both swi6 null cells and overexpression of Swi6 are well known to exhibit phenotypes, some of which interfere with heterochromatin assembly. This is not made clear in the text. Whilst the RNA binding experiments show that Swi6 can indeed bind RNA and that binding is decreased by Swi63K-3A mutation in vitro (confusingly, they only much later in the text explained that these 3 bands represent differential binding and that II is likely an isotherm). The gels showing these data are of poor quality and it is unclear which bands are used to calculate the Kd. RNA-seq data shows that overall fewer siRNAs are produced from regions of heterochromatin in the Swi63K-3A mutant so it is unsurprising that analysis of siRNA-associated motifs also shows lower enrichment (or indeed that they share some similarities, given that they originate from repeat regions).

The experiments are seemingly linked yet fail to substantiate their overall conclusions. For instance, the authors show that the Swi63K-3A mutant displays reduced siRNA binding in vitro (Figure 1D) and that H3K9me2 levels at heterochromatin loci are reduced in vivo (Figure 3C-D). They conclude that Swi6 siRNA binding is important for Swi6 heterochromatin localization, whilst it remains entirely possible that heterochromatin integrity is impaired by the Swi63K-3A mutation and hence fewer siRNAs are produced and available to bind. Their interpretation of the data is really confusing.

The authors go on to show that Swi63K-3A cells have impaired silencing at all regions tested and the mutant protein itself has less association with regions of heterochromatin. They perform DNA-RNA hybrid IPs and show that Swi63K-3A cells which also overexpress RNAseH/rnh1 have reduced levels of dh DNA-RNA hybrids than wild-type Swi6 cells. They interpret this to mean that Swi6 binds and protects DNA-RNA hybrids, presumably to facilitate binding to H3K9me2. The final piece of data is an EMSA assay showing that "high-affinity binding of Swi6 to a dg-dh specific RNA/DNA hybrid facilitates the binding to Me2-K9-H3 rather than competing against it." This EMSA gel shown is of very poor quality, and this casts doubt on their overall conclusion.

Unfortunately, the manuscript is generally poorly written and difficult to comprehend. The experimental setups and interpretations of the data are not fully explained, or, are explained in the wrong order leading to a lack of clarity. An example of this is the reasoning behind the use of the cid14 mutant which is not explained until the discussion of Figure 5C, but it is utilised at the outset in Figure 5A.

Another example of this lack of clarity/confusion is that the abstract states "Here we provide evidence in support of RNAi-independent recruitment of Swi6". Yet it then states "We show that...Swi6/HP1 displays a hierarchy of increasing binding affinity through its chromodomain to the siRNAs corresponding to specific dg-dh repeats, and even stronger binding to the cognate siRNA-DNA hybrids than to the siRNA precursors or general RNAs." RNAi is required to produce siRNAs, so their message is very unclear. Moreover, an entire section is titled "Heterochromatin recruitment of Swi6-HP1 depends on siRNA generation" so what is the author's message?

The data presented, whilst sound in some parts is generally overinterpreted and does not fully support the author's confusing conclusions. The authors essentially characterise an overexpressed Swi6 mutant protein with a few other experiments on the side, that do not entirely support their conclusions. They make the point several times that the KD for their binding experiments is far higher than that previously reported (Keller et al Mol Cell 2012) but unfortunately the data provided here are of an inferior quality and thus their conclusions are neither fully supported nor convincing.

Author response:

In this manuscript, we have addressed one of the possible modes of recruitment of Swi6 to the putative heterochromatin loci.

Our investigation was guided by earlier work showing ability of HP1 a to bind to a class of RNAs and the role of this binding in recruitment of HP1a to heterochromatin loci in mouse cells (Muchardt et al). While there has been no clarity about the mechanism of Swi6 recruitment given the multiple pathways being involved, the issue is compounded by the overall lack of understanding as to how Swi6 recruitment occurs only at the repeat regions. At the same time, various observations suggested a causal role of RNAi in Swi6 recruitment.

Thus, guided by the work of Muchardt et al we developed a heuristic approach to explore a possibly direct link between Swi6 and heterochromatin through RNAi pathway. Interestingly, we found that the lysine triplet found in the hinge domain in HP1, which influences its recruitment to heterochromatin in mouse cells, is also present in the hinge domain of Swi6, although we were cautious, keeping in mind the findings of Keller et al showing another role of Swi6 in binding to RNAs and channeling them to the exosome pathway. 

Accordingly, we envisaged that a mode of recruitment of Swi6 through binding to siRNAs to cognate sites in the dg-dh repeats shared among mating type, centromere and telomere loci could explain specific recruitment as well as inheritance following DNA replication. In accordance we framed the main questions as follows: i) Whether Swi6 binds specifically and with high affinity to the siRNAs and the cognate siRNA-DNA hybrids and whether the Swi63K-3A mutant is defective in this binding, ii) whether this lack of binding of Swi63K-3A affects its localization to heterochromatin, iii) whether the this specificity is validated by binding of Swi6 but not Swi63K-3A  to siRNAs and siRNA-DNA hybrids in vivo and iv) whether the binding mode was qualitatively and quantitatively different from that of Cen100 RNA or random RNAs, like GFP RNA.

We think that our data provides answers to these lines of inquiry to support a model wherein the Swi6-siRNA mediated recruitment can explain a cis-controlled nucleation of heterochromatin at the cognate sites in the genome. We have also partially addressed the points raised by the study by Keller et al by invoking a dynamic balance between different modes of binding of Swi6 to different classes of RNA to exercise heterochromatin formation by Swi6 under normal conditions and RNA degradation under other conditions.

While we aver about our hypothesis, we do acknowledge the need for more detailed investigation both to buttress our hypothesis and address the dynamics of siRNA binding and recruitment of Swi6  and how Swi6 functions fit in the context of other components of heterochromatin assembly, like the HDACs and Clr4 on one hand and exosome pathway on the other. Our future studies will attempt to address these issues.

Public Reviews:

Reviewer #1 (Public review):

Summary:

This manuscript explores the RNA binding activities of the fission yeast Swi6 (HP1) protein and proposes a new role for Swi6 in RNAi-mediated heterochromatin establishment. The authors claim that Swi6 has a specific and high affinity for short interfering RNAs (siRNAs) and recruits the Clr4 (Suv39h) H3K9 methyltransferases to siRNA-DNA hybrids to initiate heterochromatin formation. These claims are not in any way supported by the incomplete and preliminary RNA binding or the in vivo experiments that the authors present. The proposed model also lacks any mechanistic basis as it remains unclear (and unexplored) how Swi6 might bind to specific small RNA sequences or RNA-DNA hybrids. Work by several other groups in the field has led to a model in which siRNAs produced by the RNAi pathway load onto the Ago1-containing RITS complex, which then binds to nascent transcripts at pericentromeric DNA repeats and recruits Clr4 to initiate heterochromatin formation. Swi6 facilitates this process by promoting the recruitment of the RNA-dependent RNA polymerase leading to siRNA amplification.

Weaknesses:

(1) a) The claims that Swi6 binds to specific small RNAs or to RNA-DNA hybrids are not supported by the evidence that the authors present. Their experiments do not rule out non-specific charged-based interactions.

We disagree. We have used synthetic siRNAs of 20-22 nt length to do EMSA assay, as mentioned in the manuscript. Further, we have sequenced the small RNAs obtained after RIP experiments to validate the enrichment of siRNA in Swi6 bound fraction as compared to the mutant Swi6-bound fraction. These results are internally consistent regardless of the mode of binding. In any case the binding occurs primarily through the chromodomain although it is influenced by the hinge domain (see below).

Furthermore, we have carried out EMSA experiments using Swi6 mutants carrying all three possible double mutations of the K residues in the KKK triplet and found that there was no difference in the binding pattern as compared to the wt Swi6: only the triple mutant “3K-3A” showed the effect. These results suggest that that the bdining is not completely dependent on the basic residues. These results will be included in the revised version.

We also have some preliminary data from SAXS study showing that the CD of wt Swi6 shows a change in its structure upon binding to the siRNA, while the “3K-3A” mutant of Swi6 has a compact, folded structure that occludes the binding site of Swi6 in the chromodomain.” We propose to mention this preliminary finding in the revised version as unpublished data.

b) Claims about different affinities of Swi6 for RNAs of different sizes are based on a comparison of KD values derived by the authors for a handful of S. pombe siRNAs with previous studies from the Buhler lab on Swi6 RNA binding. The authors need to compare binding affinities under identical conditions in their assays.

Thus, the EMSA data do suggest sequence specificity in binding of Swi6 to specific siRNA sequences (Figure S5) and implies specific residues in Swi6 being responsible for that. Thus, Identification of the residues in Swi6 involved in siRNA binding in the CD would definitely be interesting, as also the experimental confirmation of the consensus siRNA sequence. It may however be noted that as against the binding of Swi6 to siRNAs occurs through CD, that of Cen100 or GFP RNA was shown be through the hinge domain by Keller et al.

The estimation of Kd by the Buhler group was based on NMR study, which we are not in a position to perform in the near future. Nonetheless, we did carry out EMSA study using the ‘Cen100’ RNA, same as the one used by the Keller et al study. Surprisingly, in contrast with the result of EMSA in agarose gel showing binding of Swi6 to “Cen100” RNA as reported by Keller et al, we fail to observe any binding in EMSA done in acrylamide gel. (The same is true of the RevCen 100). While this raises issues of why the Keller et al chose to do EMSA in agarose gel instead of the conventional approach of using acrylamide gel, it does lend support to our claim of stronger binding of Swi6 to siRNAs. Another relevant observation of binding of Swi6 to the “RevCen” RNA precursor RNAs but a detectable binding to siRNAs denoted as VI-IX (as measured by competition experiments, that are derived from RevCen RNA; Figure S4 and S7), which are derived by Dcr1 cleavage of the ‘’RevCen’’ RNA.

We also disagree that we carried out EMSA with a small bunch of siRNAs. As indicated in Figure 1 and S1, we synthesized nearly 12 siRNAs representing the dg-dh repeats at Cen, mat and tel loci and measured their specificity of binding to Swi6 using EMSA assay by labeling the ones labelled “D”, “E” and “V” directly and those of the remaining ones by the latter’s ability to compete against the binding (Figure 1, S4). These results point to presence of a consensus sequence in siRNAs that shows highly specific and strong binding to Swi6 in the low micromolar range.

Further, our claim of binding of Swi6 and not Swi63K>3A to siRNA in vivo is validated by RIP experiments, as shown in Fig 2 and S9.

c) The regions of Swi6 that bind to siRNAs need to be identified and evidence must be provided that Swi6 binds to RNAs of a specific length, 20-22 mers, to support the claim that Swi6 binds to siRNAs. This is critical for all the subsequent experiments and claims in the study.

We have provided both in vitro data, which is va;idiated in vivo by RIP experiments, as mentioned above. However, we agree that it wpuld be very interesting to identify the residues in Swi6 chromdomain responsible for binding to siRNA. However, such an investigation is beyond the scope of the present study.

(2) a) The in vivo results do not validate Swi6 binding to specific RNAs, as stated by the authors. Swi6 pulldowns have been shown to be enriched for all heterochromatic proteins including the RITS complex. The sRNA binding observed by the authors is therefore likely to be mediated by Ago1/RITS.

We disagree with the first comment. Our RIP experiments do validate the in vitro results (Fig 1, 2, S4 and S9), as argued above. The observation alluded to by the reviewer “Swi6 pulldowns have been shown to be enriched for all heterochromatic proteins including the RITS complex” is not inconsistent with our observation; it is possible that the siRNA may be released from the RITS complex and transferred to Swi6, possibly due to its higher affinity.

Thus, we would like to suggest that the role of Swi6 is likely to be coincidental or subsequent to that of Ago1/RITS (see below). We think that the binding by Swi6 to the siRNA and siRNA-DNA hybrid and could be also carried out in cis at the level of siRNA-DNA hybrids.

This point needs to be addressed in future studies.

b) Most of the binding in Figure S8C seems to be non-specific.

We would like to point out that the result in Figure S8C needs to be examined together with the Figure S8B, which shows RNA bound by Swi6 but not Swi63K-3A to hybridize with dg, dh and dh-k probes.

c) In Figure S8D, the authors' data shows that Swi6 deletion does not derepress the rev dh transcript while dcr1 delete cells do, which is consistent with previous reports but does not relate to the authors' conclusions.

The purpose of results shown in Figure S8D is just to compare the results of Swi6 with that of Swi63K-3A.

d) Previous results have shown that swi6 delete cells have 20-fold fewer dg and dh siRNAs than swi6+ cells due to decreased RNA-dependent RNA polymerase complex recruitment and reduced siRNA amplification.

This result is consistent with our results invoking a role of Swi6 in binding to, protecting and recruiting siRNAs to homologous sites.

To find if the overall production of siRNA is compromised in swi6 3K->3A mutant, we i) calculated the RIP-Seq read counts for swi6 3K->3A , swi6+ and vector control in 200 bp genomic bins , ii) divided the Swi6 3K->3A and swi6+ signals by that of control, iii) removed the background using the criteria of signal value < 25% of max signal, and iv) counted the total reads (in excess to control) in all peak regions in both samples.  This revealed a total count of 10878 and 8994 respectively for Swi6 3K->3A  and swi6+ samples, possibly implying that the overall siRNA production is not compromised in the Swi6 3K->3A mutant.

(3) a) The RIP-seq data are difficult to interpret as presented. The size distribution of bound small RNAs, and where they map along the genome should be shown as for example presented in previous Ago1 sRNA-seq experiments.

Please see the response to 2(d).

b) It is also unclear whether the defects in sRNA binding observed by the authors represent direct sRNA binding to Swi6 or co-precipitation of Ago1-bound sRNAs.

The correspondence between our in vivo and in vitro results suggests that the binding to Swi6 would be direct. We do not observe a complete correspondence between the Swi6- and Ago-bound siRNAs. We think Swi6 binding may be coincident with or following RITS complex formation.

This point will be discussed in the Revision.

The authors should also sequence total sRNAs to test whether Swi6-3A affects sRNA synthesis, as is the case in swi6 delete cells.

Please see response to 2(d) above.

(4) The authors examine the effects of Swi6-3A mutant by overexpression from the strong nmt1 promoter. Heterochromatin formation is sensitive to the dosage of Swi6. These experiments should be performed by introducing the 3A mutations at the endogenous Swi6 locus and effects on Swi6 protein levels should be tested.

Although we agree, we think that the heterochromatin formation is occurring in presence of nmt1-driven Swi6 but not Swi63K>3A, as indicated by the phenotype and Swi6 enrichment at otr1R::ade6, imr1::ura4 and his3-telo (Figure 3) and mating type (Fig. S10). Furthermore, the both GFP-Swi6 and GFPSwi63K>3A are expressed at similar level (Fig. S8A).

(5) The authors' data indicate an impairment of silencing in Swi6-3A mutant cells but whether this is due to a general lower affinity for nucleosomes, DNA, RNA, or as claimed by the authors, siRNAs is unclear. These experiments are consistent with previous findings suggesting an important role for basic residues in the HP1 hinge region in gene silencing but do not reveal how the hinge region enhances silencing.

Our study aims to correlate the binding of Swi6 but not Swi63K-3A to siRNA with its localization to heterochromatin. A similar difference in binding of Swi6 but not Swi63K-3A to siRNA-DNA hybrid, together with sensitivity of silencing and Swi6 localization to heterochromatin to RNaseH support the above correlations as being causally connected.

In terms of mechanism of binding, we need to clarify that the primary mode of binding is through the CD and not the hinge domain, although the hinge domain does influence this binding. This result is different from those of Keller et al.

We have some structural data based on preliminary SAXS experiment supporting binding of siRNA to the CD and influence of the hinge domain on this binding. However, this line of investigation need to be extended and will be subject of future investigations.

(6) RNase H1 overexpression may affect Swi6 localization and silencing indirectly as it would lead to a general reduction in R loops and RNA-DNA hybrids across the genome. RNaseH1 OE may also release chromatin-bound RNAs that act as scaffolds for siRNA-Ag1/RITS complexes that recruit Clr4 and ultimately Swi6.

These are formal possibilities. However, the correlation between swi6 binding to siRNA-DNA hybrid and delocalization upon RNase H1 treatment argues for a more direct link.

(7) Examples of inaccurate presentation of the literature.

a) The authors state that "RNA binding by the murine HP1 through its hinge domains is required for heterochromatin assembly (Muchardt et al, 2002). The cited reference provides no evidence that HP1 RNA binding is required for heterochromatin assembly. Only the hinge region of bacterially produced HP1 contributes to its localization to DAPI-stained heterochromatic regions in fixed NIH 3T3 cells.

Noted. Statement will be corrected.

b) "... This scenario is consistent with the loss of heterochromatin recruitment of Swi6 as well as siRNA generation in rnai mutants (Volpe et al, 2002)." Volpe et al. did not examine changes in siRNA levels in swi6 mutant cells. In fact, no siRNA analysis of any kind was reported in Volpe et al., 2002.

Correct.  We only say that Swi6 recruitment is reduced in rnai mutants and correlate it with ability of SWi6 to bind to siRNA generated by RNAi and subsequently to siRNA-DNA hybrid.

Reviewer #2 (Public review):

The aim of this study is to investigate the role of Swi6 binding to RNA in heterochromatin assembly in fission yeast. Using in vitro protein-RNA binding assays (EMSA) they showed that Swi6/HP1 binds centromere-derived siRNA (identified by Reinhardt and Bartel in 2002) via the chromodomain and hinge domains. They demonstrate that this binding is regulated by a lysine triplet in the conserved region of the Swi6 hinge domain and that wild-type Swi6 favours binding to DNA-RNA hybrids and siRNA, which then facilitates, rather than competes with, binding to H3K9me2 and to a lesser extent H3K9me3.

However, the majority of the experiments are carried out in swi6 null cells overexpressing wild-type Swi6 or Swi63K-3A mutant from a very strong promoter (nmt1). Both swi6 null cells and overexpression of Swi6 are well known to exhibit phenotypes, some of which interfere with heterochromatin assembly. This is not made clear in the text.

We think that the argument is not valid as we show that swi6 but not Swi63K-3A could restore silencing at imr1::ura4, otr1::ade6 and his3-telo (Fig 3) and mating type (Fig. S10), when transformed into a swi6D strain.

Whilst the RNA binding experiments show that Swi6 can indeed bind RNA and that binding is decreased by Swi63K-3A mutation in vitro (confusingly, they only much later in the text explained that these 3 bands represent differential binding and that II is likely an isotherm). The gels showing these data are of poor quality and it is unclear which bands are used to calculate the Kd.

We disagree with the comment about the quality of EMSA data. We think it is of similar quality or better than that of Keller et al, except in some cases, like Fig 1D, a shorter exposure shown to distinguish the slowest shifted band has caused the remaining bands to look fainter.

RNA-seq data shows that overall fewer siRNAs are produced from regions of heterochromatin in the Swi63K-3A mutant so it is unsurprising that analysis of siRNA-associated motifs also shows lower enrichment (or indeed that they share some similarities, given that they originate from repeat regions).

Please see response to comment 2(d) of the first reviewer above.

It is not clear which bands are being alluded to. However, we‘ll rectify any gaps in information in the revision.

The experiments are seemingly linked yet fail to substantiate their overall conclusions. For instance, the authors show that the Swi63K-3A mutant displays reduced siRNA binding in vitro (Figure 1D) and that H3K9me2 levels at heterochromatin loci are reduced in vivo (Figure 3C-D). They conclude that Swi6 siRNA binding is important for Swi6 heterochromatin localization, whilst it remains entirely possible that heterochromatin integrity is impaired by the Swi63K-3A mutation and hence fewer siRNAs are produced and available to bind. Their interpretation of the data is really confusing.

Our argument is that the lack of binding by Swi63K>3A to siRNA can explain the loss of recruitment to heterochromatin loci and thus affect the integrity of heterochroamtin; the recruitment of Swi6 can occur possibly by binding initially to siRNA and thereafter as siRNA-DNA hybrid. However, the overall level of siRNAs is not affected, as in 2(D) above. This interpretation is supported by results of ChIP assay and confocal experiments, as also by the effect of RNaseH1 in the recruitment of Swi6.

The authors go on to show that Swi63K-3A cells have impaired silencing at all regions tested and the mutant protein itself has less association with regions of heterochromatin. They perform DNA-RNA hybrid IPs and show that Swi63K-3A cells which also overexpress RNAseH/rnh1 have reduced levels of dh DNA-RNA hybrids than wild-type Swi6 cells. They interpret this to mean that Swi6 binds and protects DNA-RNA hybrids, presumably to facilitate binding to H3K9me2. The final piece of data is an EMSA assay showing that "high-affinity binding of Swi6 to a dg-dh specific RNA/DNA hybrid facilitates the binding to Me2-K9-H3 rather than competing against it." This EMSA gel shown is of very poor quality, and this casts doubt on their overall conclusion.

We do agree with the reviewer about the quality of EMSA (Fig. 5B). However, as may be noticed in the EMSA for siRNA-DNA hybrid binding  (Fig 4A), the bands of Swi6-bound siRNA-DNA hybrid are extremely retarded. Hence the EMSA for subsequent binding by H3-K9-Me peptides required a longer electrophoretic run, which led to reduction in the sharpness of the bands. Nevertheless, the data does indicate binding efficiency in the order H3K9-Me2> H3-K9-Me3 > H3-K9-Me0. Having said that, we plan to repeat the EMSA or address the question by other methods, like SPR.

Unfortunately, the manuscript is generally poorly written and difficult to comprehend. The experimental setups and interpretations of the data are not fully explained, or, are explained in the wrong order leading to a lack of clarity. An example of this is the reasoning behind the use of the cid14 mutant which is not explained until the discussion of Figure 5C, but it is utilised at the outset in Figure 5A.

We tend to agree somewhat and will attempt to submit a revised version with greater clarity, as also the explanation of experiment with cid14D strain.

Another example of this lack of clarity/confusion is that the abstract states "Here we provide evidence in support of RNAi-independent recruitment of Swi6". Yet it then states "We show that...Swi6/HP1 displays a hierarchy of increasing binding affinity through its chromodomain to the siRNAs corresponding to specific dg-dh repeats, and even stronger binding to the cognate siRNA-DNA hybrids than to the siRNA precursors or general RNAs." RNAi is required to produce siRNAs, so their message is very unclear. Moreover, an entire section is titled "Heterochromatin recruitment of Swi6-HP1 depends on siRNA generation" so what is the author's message?

The reviewer has correctly pointed out the error. Indeed, our results actually indicate an RNAi-dependent rather than independent mode of recruitment. Rather, we would like to suggest an H3-K9-Me2-indpendnet recruitment of Swi6. We will rectify this error in our revised manuscript.

The data presented, whilst sound in some parts is generally overinterpreted and does not fully support the author's confusing conclusions. The authors essentially characterise an overexpressed Swi6 mutant protein with a few other experiments on the side, that do not entirely support their conclusions. They make the point several times that the KD for their binding experiments is far higher than that previously reported (Keller et al Mol Cell 2012) but unfortunately the data provided here are of an inferior quality and thus their conclusions are neither fully supported nor convincing.

We have used the method of Heffler et al (2012) to compute the Kd from EMSA data.

https://typewriterdatabase.com/196x-remington-superriter.23240.typewriter

Annotated photo of Remington's Fold-A-Matic screw locations.

The Remington De Luxe KMC by [[Richard Polt]]

Picking up the changes in colors in different scene adds an extra layer of understanding to just how deep the message of love goes. Well noticed.

I find it interesting that you never mention the end of the video or the result of their final interaction with this cube that is chasing them---as that is technically not part of the point. The message is clear: the reader is meant to focus on the love and support that carries them through the frightening journey regardless of the result.

In all the images they (IU and V) are seen wearing wedding clothing (wedding dress and suit, respectively). Is it safe to assume that they are supposed to be (either about to get or already are) married? This might be something worth mentioning.

I find it interesting that you showed both IU trying to protect V and the other way around. This drives home your idea that love goes both ways and people can support one another.

Including pictures from the music video is a great idea that helps the reader to do less work in regards to envisioning the scenes that contribute to IU's message AS they are reading.

User participation may also require education about platform operations for informed decision-making. A model like this could instill more trust and be in line with user values but risks inefficiency or conflicts among diverse user interests.

Remington 24: Darth Vader's other typewriter by [[Richard Polt]]

This is entirely true such as in Pakistan, my hometown which is considered to be a poorer country comparatively. Over there I gained coding experience in Visual Basic for two years but that did not help me at all in States because here they prefer Python or Java. But I feel like there should not be this discrimination as the ones gaining experience in poor countries would surely want to work in a bigger international market such as Silicon Valley.

The Italian Super-Riter by [[Richard Polt]]

also group stability

percolation

https://www.youtube.com/watch?v=puMqkq6jG0o

Points forts de la vidéo "Parlez-vous français : pour une relation entre citoyens et services publics sans jargon" avec timestamps

00:00:00 - Introduction

Présentation de l'atelier sur la simplification du langage administratif et son importance pour la relation entre citoyens et services publics. Les enjeux de la simplification du langage administratif : accessibilité, compréhension, confiance des usagers. Coûts de l'inintelligibilité du langage administratif pour les usagers et les services publics.

03:30 - Plan gouvernemental "Parlez-nous français"

Lancement d'un plan gouvernemental pour lutter contre le jargon administratif. Capitalisation sur les actions déjà menées et les initiatives des services publics. Centré sur les écrits administratifs (courriers, formulaires, sites internet, démarches en ligne). Articulé avec la suppression et la numérisation des formulaires administratifs (SERFA).

07:30 - Difficultés de la lutte contre le jargon administratif

Le langage administratif est souvent technique et centré sur l'administration elle-même. Nécessité d'un renversement de perspective pour se mettre à la place de l'usager. La norme "Langage clair" vise à communiquer des informations claires et utiles aux usagers.

10:30 - Exemple de France travail

Réécriture des courriers avec des ergonomes et des usagers. Confrontation des courriers aux usagers pour tester leur compréhension. Utilisation de la notion de "parcours" pour simplifier les démarches administratives.

15:30 - Rôle de l'IA dans la simplification du langage administratif

Potentiel de l'IA pour générer des textes clairs et accessibles. Nécessité de prendre en compte les biais et les limites de l'IA. Importance de l'évaluation et de la validation humaine des textes générés par l'IA.

20:00 - Conclusion

Importance de la simplification du langage administratif pour la relation entre citoyens et services publics.

Engagement du gouvernement à travers le plan "Parlez-nous français".

Rôle de la DITIP pour accompagner les services publics dans cette démarche.

Appel à continuer les efforts de simplification et de communication claire.

Résumé de la vidéo "Parlez-vous français : pour une relation entre citoyens et services publics sans jargon" après 00:20:00 avec timestamps

Voici un résumé de la vidéo "Parlez-vous français : pour une relation entre citoyens et services publics sans jargon" après 00:20:00 avec des timestamps :

00:20:00 Introduction de l'atelier et présentation des intervenants.

00:25:22 Gisèle Doriano, chef du service expérience usager à la DITIP, explique les raisons de l'atelier :

• Le langage administratif est un problème pour les usagers, notamment les plus vulnérables.
• Il y a un coût pour les usagers et les services publics.
• Le gouvernement a lancé un plan appelé "Parlez-nous français" pour simplifier le langage administratif.

00:31:42 Discussion sur les difficultés de la simplification du langage administratif :

• Il est souvent plus simple pour les agents d'utiliser un langage technique.
• L'administration a tendance à se centrer sur elle-même plutôt que sur l'usager.

00:35:22 Présentation des actions menées par France Travail pour simplifier ses courriers :

• Réécriture des courriers avec des ergonomes.
• Confrontation des courriers aux usagers.
• 00:39:42 Discussion sur l'utilisation de l'IA pour simplifier le langage administratif :
• L'IA peut être un outil utile, mais il faut être vigilant sur les biais et l'uniformisation.
• L'IA doit être utilisée comme un appui pour l'intervention humaine.

00:44:22 Cécile Barouat, Défenseur des droits, souligne l'importance de la simplification du langage administratif :

• Il faut identifier les objets les plus compliqués pour les usagers.
• Le plan gouvernemental "Parlez-nous français" est une bonne initiative.
• La DITIP a un rôle important à jouer pour accompagner les services publics.
• 00:48:22 Conclusion de l'atelier par Gisèle Doriano :

Il y a une dynamique de simplification en cours.

Le plan "Parlez-nous français" est un engagement fort du gouvernement.

La DITIP est là pour aider les services publics à simplifier leur langage.

00:51:22 Fin de l'atelier.

Résumé de la vidéo [00:00:00][^1^][1] - [00:23:53][^2^][2] : Ce webinaire, animé par Alice Pierre-François, se concentre sur l'animation d'un collectif SISM (Semaines d'Information sur la Santé Mentale) en France. Il aborde les stratégies pour engager les membres sur le long terme, les partenariats possibles, et les méthodes d'animation pour susciter la motivation. Des intervenants partagent leurs expériences en matière de coordination d'événements SISM et d'animation de collectifs locaux.

Points saillants : + [00:00:00][^3^][3] Introduction et objectifs du webinaire * Présentation par Alice Pierre-François * Discussion sur l'engagement des membres et l'animation des collectifs * Conseils pour la gestion des collectifs SISM + [00:01:04][^4^][4] Intervenants et leurs expériences * Partage d'expériences par divers intervenants * Exemples de coordination et d'animation de collectifs * Importance de l'engagement et de la communication + [00:03:26][^5^][5] Règles d'échange et modération du webinaire * Modération par Léa Sonet, responsable communication du Psycom * Rappel des règles pour le bon déroulement du webinaire * Encouragement à l'interaction via le chat + [00:07:35][^6^][6] Historique et importance des SISM * Explication des SISM, un rendez-vous annuel sur la santé mentale * Objectifs et organisation des SISM * Rôle du collectif national et des collectifs locaux + [00:11:21][^7^][7] Présentation de Widad l Wafi sur les SISM à Vichy * Organisation des SISM par le collectif de Vichy communauté * Diversité des acteurs et événements organisés * Exemples d'actions menées lors des SISM 2023 + [00:22:15][^8^][8] Présentation de Mélissa sur les SISM dans le département de l'Ain * Contexte géographique et démographique de l'Ain * Adaptation des événements SISM aux spécificités du département * Importance de l'accès aux soins et de la communication

Résumé de la vidéo [00:23:55][^1^][1] - [00:48:17][^2^][2]:

Cette vidéo présente un webinaire sur l'animation d'un collectif SISM (Semaines d'Information sur la Santé Mentale) en juin 2024. Elle aborde l'évolution des SISM dans le département de l'Indre depuis leur création en 2013, leur intégration dans le projet territorial de santé mentale en 2020, et la coordination par le service de santé mentale de l'Indre depuis 2021. La vidéo met en lumière l'importance de la mutualisation des moyens, la participation des membres du collectif, et l'évaluation de la satisfaction des participants.

Points forts: + [00:23:55][^3^][3] Historique et évolution des SISM * Création en 2013 par un petit groupe * Évolution et intégration dans le projet territorial de santé mentale en 2020 * Coordination par le service de santé mentale de l'Indre depuis 2021 + [00:26:01][^4^][4] Participation et organisation * Environ 48 partenaires en 2023 * Réalisation de 26 événements en 2023 * Types d'événements variés : ateliers, conférences, débats, etc. + [00:29:28][^5^][5] Le collectif EO et ses objectifs * Existence depuis 2016 * Objectifs de décloisonnement et de renforcement des liens entre acteurs * Organisation de manifestations variées en 2023 + [00:39:10][^6^][6] Rôles et partenariats au sein des collectifs * Importance de la clarté des rôles et des missions * Mutualisation des moyens et participation active des membres * Évaluation de la satisfaction et amélioration continue

Résumé de la vidéo [00:48:20][^1^][1] - [01:11:41][^2^][2]:

Cette vidéo présente un webinaire sur l'animation d'un collectif SISM (Semaines d'Information sur la Santé Mentale) en juin 2024. Les intervenants discutent des méthodes d'organisation, de la diversité des acteurs impliqués, et de l'importance de l'interconnaissance et du soutien mutuel pour le succès des initiatives.

Points forts: + [00:48:20][^3^][3] Organisation et partenariats * Importance de l'offre et de la demande de ressources * Exemple d'un débat universitaire facilité par la disponibilité d'une salle * Émergence de beaux partenariats + [00:49:16][^4^][4] Rôle et diversité au sein du collectif * Composition variée du collectif inscrite dans la charte * Représentation des structures hospitalières, associations d'usagers, et autres * Deux sous-groupes : coordination et communication + [00:51:57][^5^][5] Interconnaissance et engagement * Interconnaissance préalable entre certains membres * Cultivation de liens à travers différents projets * Partage d'expériences et soutien dans les actions + [00:56:21][^6^][6] Importance de la présence politique * Impact de la présence politique sur la valorisation des actions * Objectif futur de renforcer le lien avec les élus + [00:59:32][^7^][7] Méthodes d'animation d'un collectif * Présentation d'outils d'animation pour faciliter l'engagement * Exemple d'un appel à participation pour élargir le collectif + [01:07:59][^8^][8] Animation et réunions plénières du collectif * Cinq réunions plénières annuelles pour l'organisation * Présentiel privilégié pour l'accueil et la convivialité * Partage d'expériences et création de partenariats lors des réunions

Résumé de la vidéo [01:11:45][^1^][1] - [01:23:14][^2^][2]:

Cette partie du webinaire se concentre sur l'animation d'un collectif SISM en juin 2024, mettant en lumière les stratégies de communication, les outils de coordination et les pratiques d'engagement des membres.

Points forts: + [01:11:45][^3^][3] Communication et visibilité * Distribution de flyers et programmes communs * Utilisation de QR codes et cartes pour localiser les actions * Soutien logistique par les coordinateurs + [01:14:55][^4^][4] Facilitation et soutien aux membres * Simplification de la participation au collectif * Prise en charge interne de la production de matériel promotionnel * Financement de la convivialité et des réunions par la communauté + [01:17:01][^5^][5] Planification et organisation des réunions * Utilisation d'outils participatifs comme Doodle pour planifier * Rotation des lieux de réunion pour une meilleure connaissance mutuelle * Création d'un padlet pour partager les coordonnées et informations + [01:21:00][^6^][6] Conseils et recommandations pour l'animation * Importance de l'horizontalité, convivialité et partage d'expérience * Bienveillance, suppression des rapports de force et rappel des enjeux * Créativité dans l'animation du collectif pour renforcer l'identité

Video summary [00:00:00][^1^][1] - [00:54:09][^2^][2]:

Cette vidéo présente une discussion approfondie sur la zététique, l'esprit critique, et les croyances, avec Samuel Buisseret.

Il aborde son parcours personnel, ses critiques du milieu sceptique, et son livre "Arrêter de croire n'importe quoi".

Highlights: + [00:00:00][^3^][3] Introduction et présentation * Samuel Buisseret se présente * Discussion sur la zététique et l'esprit critique * Annonce de l'arrêt de sa chaîne YouTube + [00:02:26][^4^][4] Critique du milieu sceptique * Distinction entre outil et application * Sensibilité particulière de Samuel en tant qu'ancien complotiste * Importance de l'autocritique dans la zététique + [00:04:02][^5^][5] Genèse du livre de Samuel * Commande des éditions de bouc supérieures * Synthèse de huit années de pratique zététique * Révision et contextualisation de ses opinions + [00:23:32][^6^][6] Création de la chaîne YouTube * Motivation personnelle et événement déclencheur * Première vidéo et découverte de l'esprit critique * Importance de la prudence épistémique + [00:52:14][^7^][7] Résultats en parapsychologie * Expériences et résultats significatifs * Importance de la rigueur méthodologique * Contribution de Renaud Evrard et Jean-Michel Abrassart

outweigh = l'emporter sur

I wonder if social media hasn't reached this point because it's a newer industry. I'm curious to see what happens over the coming years, especially with the potential for some competition to be removed with the possible banning of tik tok.

Points forts de la vidéo "La fabrique de la connaissance territoriale en région" avec timestamps:

00:00 - 05:00 Introduction et présentation des intervenants * Alix Roche, vice-président de l'Unadel, présente le contexte et les objectifs de la rencontre. * Jean-Baptiste Chabert, directeur de la connaissance de la région Sud Provence-Alpes-Côte d'Azur, présente son service et son approche de la connaissance territoriale. * Joseph Compera, chef du service prospective en région Bourgogne-Franche-Comté, présente son service et son approche de la prospective territoriale. * Stéphane Ambert, chef du service prospective en région Hauts-de-France, présente son service et son approche de la prospective territoriale.

05:00 - 25:00 Le rôle de la connaissance territoriale dans les politiques publiques * Les intervenants discutent de l'importance de la connaissance territoriale pour éclairer les politiques publiques locales et régionales. * Ils soulignent que la connaissance territoriale doit être accessible et partagée par tous les acteurs du territoire. * Ils évoquent les défis de la production et de la diffusion de la connaissance territoriale.

25:00 - 40:00 Les outils et méthodes pour la fabrique de la connaissance territoriale * Les intervenants présentent les différents outils et méthodes utilisés pour produire de la connaissance territoriale. * Ils évoquent notamment les enquêtes, les études documentaires, la modélisation et la simulation. * Ils soulignent l'importance de la collaboration entre les chercheurs et les acteurs du territoire.

40:00 - 50:00 Les enjeux de la prospective territoriale * Les intervenants discutent des enjeux de la prospective territoriale, notamment face aux transitions en cours (climatique, numérique, etc.). * Ils soulignent l'importance de la prospective pour anticiper les changements et préparer l'avenir des territoires. * Ils évoquent les défis de la construction de scénarios prospectifs.

50:00 - 53:00 Conclusion et perspectives * Les intervenants appellent à un renforcement de la coopération entre les acteurs de la connaissance territoriale. * Ils soulignent le rôle essentiel de la connaissance territoriale pour construire des politiques publiques durables et justes.

Remarques

• La vidéo est riche en informations et en exemples concrets.
• Les intervenants sont des experts reconnus dans le domaine de la connaissance territoriale.
• La vidéo est bien structurée et facile à suivre.

J'espère que ces points forts vous seront utiles. N'hésitez pas à me contacter si vous avez d'autres questions.

Notez que

• Les timestamps sont approximatifs et peuvent varier légèrement en fonction de la façon dont vous regardez la vidéo.
• Je n'ai pu inclure que les points forts les plus importants dans ce résumé. Si vous souhaitez des informations plus détaillées sur un sujet particulier, n'hésitez pas à me le faire savoir.

Résumé de la vidéo [00:00:00][^1^][1] - [01:54:14][^2^][2]:

Cette vidéo présente une réunion de l'Institut Bertrand Schwartz, axée sur la participation des jeunes et l'implication des élus locaux dans les missions locales.

Les intervenants discutent des bénéfices et des risques de la proximité avec les administrés, ainsi que des changements nécessaires dans les pratiques des élus pour favoriser la participation citoyenne.

Points forts : + [00:00:00][^3^][3] Introduction et déroulé de la réunion * Présentation des intervenants * Objectifs de la réunion * Importance de la participation des jeunes + [00:02:00][^4^][4] Rappel de la démarche et des principes * Implication des élus locaux * Contribution des jeunes aux politiques * Importance de la décentralisation + [00:05:00][^5^][5] Recherche-action pour la participation des jeunes * Trois types d'acteurs : jeunes, professionnels, élus * Objectifs des webinaires * Changement de posture des élus + [00:10:00][^6^][6] Outil de mesure de la participation * Types de participation : consultants, collaborateurs, pilotes * Importance de la non-participation assumée * Risques de fausse participation + [00:31:20][^7^][7] Discussion sur l'implication des élus * Importance des échanges directs avec les jeunes * Changement de contexte avec la garantie jeune * Rôle des missions locales comme médiateurs

vidéo "Décrypter la recherche - ep#2 "Les formes de soutien à la vie associative : du national au local""

0h00min - 0h05min : Introduction et contexte de la recherche

• Présentation de l'épisode et de l'invitée Mathilde Rininassi, chargée de recherche.
• Contexte de la recherche : questionner l'action publique de soutien à la vie associative en France.
• Objectifs de l'étude : comprendre les formes de soutien à la vie associative au niveau national et les interactions entre les acteurs publics et associatifs.

0h05min - 0h15min : Illisibilité et diversité des dispositifs de soutien

• Difficulté pour les associations de naviguer dans la multitude de dispositifs de soutien.
• Existence de dispositifs portés par différents ministères et organismes d'État.
• Manque de coordination et de lisibilité des politiques publiques de soutien à la vie associative.

0h15min - 0h25min : Méthodologie de la recherche

• Étude qualitative basée sur 42 entretiens semi-directifs.
• Rencontre avec des agents de 15 ministères ou organismes d'État.
• Interrogation de 19 têtes de réseau ou associations.

0h25min - 0h35min : Formes de soutien à la vie associative

• Typologie des formes de soutien :
  • Soutien financier: subventions, appels à projets, etc.
  • Soutien organisationnel: accompagnement à la structuration, formation, etc.
  • Soutien informationnel: mise à disposition de ressources, communication, etc.
• Variation des formes de soutien selon les ministères et les types d'associations.

0h35min - 0h45min : Rôle des têtes de réseau et des fédérations

• Implication des têtes de réseau dans la mise en œuvre des politiques publiques.
• Coordination et structuration du secteur associatif par les fédérations.
• Dialogue entre les acteurs publics et associatifs.

0h45min - 0h55min : Constats et perspectives

• Absence d'une politique publique unifiée de soutien à la vie associative.
• Nécessité d'une meilleure coordination entre les acteurs publics.
• Simplification des procédures administratives et des financements.
• Renforcement du dialogue entre les acteurs publics et associatifs.

0h55min - 0h59min : Conclusion et remerciements

• Remerciements aux participants et aux intervenants.
• Invitation à l'épisode suivant.

Points forts de la vidéo [Webinaire] Renforcement du monde associatif - Ep.1 "Face aux attaques des libertés associatives"

Introduction

Le contexte actuel est marqué par des attaques croissantes contre les libertés associatives.

Le RNMA et le Collectif des associations citoyennes organisent un cycle de webinaires pour explorer les risques et construire des réponses collectives.

Ce premier épisode se concentre sur la question de la réponse collective face aux menaces qui pèsent sur les libertés associatives.

Les libertés associatives en France en 2024

Définition des libertés associatives

Etat des lieux des menaces et des attaques contre les libertés associatives

Acteurs et enjeux de la défense des libertés associatives Illustration des menaces par deux organisations

Le Planning familial et son expérience face aux attaques anti-avortement.

La Fédération d'associations de protection de la nature et les difficultés de financement.

Stratégies de réponse collective

Mobilisation et solidarité entre associations.

Travail interassociatif et création de coalitions.

Outillage juridique des associations. Développement de contentieux stratégiques.

Construction d'une jurisprudence favorable aux associations.

Conclusion

Importance de la réflexion et de l'action collective pour défendre les libertés associatives.

Le RNMA propose un cycle de webinaires pour approfondir la réflexion et construire des stratégies communes.

Ressources

Lien vers la production de Vox Public sur "Anticiper et oui des raps entre guillemets". Livre de Jean-Baptiste Jobart sur les libertés associatives.

Fiche réflex sur les stratégies de défense des libertés associatives. Adresse mail du RNMA pour recevoir les ressources et les dates des prochains webinaires.

Moments clés

• 00:00 - Introduction et présentation des intervenants
• 05:30 - Définition des libertés associatives et contexte actuel
• 10:30 - Exposé de Jean-Baptiste Jobart sur les menaces aux libertés associatives
• 25:00 - Témoignage de Karine sur l'expérience du Planning familial
• 32:30 - Témoignage de Stéphane Giroud sur les difficultés des associations de protection de la nature
• 42:00 - Discussion sur les stratégies de réponse collective
• 57:00 - Conclusion et présentation des prochains webinaires

If social media couldn't profit from personal data, then platforms would move toward subscription models or ethical partnerships, which would likely improve privacy but limit access for those unwilling or unable to pay.

Résumé vidéo [00:00:05][^1^][1] - [00:21:55][^2^][2]:

Cette vidéo présente une recherche sur les prises de position des élèves sur le bien-être animal en élevage, une question socialement vive.

Elle montre comment les élèves articulent des connaissances, des émotions et des valeurs pour construire leurs arguments, et comment ils utilisent parfois des stratégies de défense pour justifier leurs pratiques.

Elle propose des leviers pédagogiques pour accompagner les élèves à avoir un regard réflexif et à réduire les dissonances.

Points forts: + [00:00:05][^3^][3] La présentation de la chercheuse et de son domaine de recherche * Enseignante chercheuse à l'ENSFEA * Spécialiste de la didactique des Questions Socialement Vives * Auteure d'un chapitre d'ouvrage sur le sujet + [00:01:27][^4^][4] La définition et les caractéristiques des Questions Socialement Vives * Questions à enjeux de société controversées dans les champs de référence, la société et la classe * Exemple du changement climatique * Articulation entre domaine cognitif, émotionnel et axiologique + [00:03:30][^5^][5] Le cas du bien-être animal en élevage et le scénario pédagogique mis en œuvre * Problème des interventions douloureuses sur les animaux * Séances de cours, de TP et de retours d'expérience * Recueil des discours des élèves lors du dilemme éthique et professionnel + [00:06:12][^6^][6] Deux exemples d'élèves illustrant des situations contrastées * Aymeric, fils d'éleveur, qui change de position selon les contextes * Rudy, élève peu impliqué, qui propose des pratiques douloureuses * Analyse des connaissances, des émotions et des valeurs mobilisées + [00:18:43][^7^][7] Les leviers pour mieux comprendre les prises de position des élèves * Nécessité des connaissances mais pas suffisantes * Accompagnement du regard réflexif, de la verbalisation des émotions et des valeurs * Diversification du vécu et confrontation d'alternatives * Réduction des dissonances et ouverture de nouveaux possibles + [00:21:24][^8^][8] La référence du chapitre d'ouvrage et l'invitation à échanger en mars * Ouvrage Educagri : "L'éthique dans l'enseignement agricole" * Chapitre : "Points de vue des élèves sur leur bien-être en classe" * Auteure : Amélie Lipp

DOI: 10.7554/eLife.70506

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.7554/eLife.50294

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.4414/pc-d.2015.01147

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.4414/pc-d.2013.00543

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.3390/ph16101460

Resource: None

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_023446

---

What is this?

DOI: 10.26689/pbes.v5i3.3960

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.21769/bioprotoc.4532

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.17487/rfc6355

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.17487/rfc6355

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

Traceback (most recent call last): File "/home/ubuntu/dashboard/py/create_release_tables.py", line 54, in format_anno_for_release parsedanno = HypothesisAnnotation(anno) File "/home/ubuntu/dashboard/py/hypothesis.py", line 231, in init self.links = row['document']['link'] TypeError: string indices must be integers

DOI: 10.14309/01.ajg.0001045324.24382.8f

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.14309/01.ajg.0001044272.51888.91

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.14309/01.ajg.0001044272.51888.91

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.14309/01.ajg.0001042584.98405.b8

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.14309/01.ajg.0001042584.98405.b8

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.14309/01.ajg.0001042584.98405.b8

Resource: (IMSR Cat# JAX_007181,RRID:IMSR_JAX:007181)

Curator: @AlecAsdourian

SciCrunch record: RRID:IMSR_JAX:007181

---

What is this?

DOI: 10.14309/01.ajg.0001042584.98405.b8

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.14309/01.ajg.0001042584.98405.b8

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.14309/01.ajg.0001042584.98405.b8

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.14309/01.ajg.0001042584.98405.b8

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.1108/s1534-085620150000017002

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.1101/2024.11.28.625840

Resource: RRID:IMSR_JAX:007909

Curator: @AlecAsdourian

SciCrunch record: RRID:IMSR_JAX:007909

---

What is this?

DOI: 10.1101/2024.11.28.625840

Resource: RRID:IMSR_JAX:006660

Curator: @AlecAsdourian

SciCrunch record: RRID:IMSR_JAX:006660

---

What is this?

DOI: 10.1101/2024.11.26.625526

Resource: (DSMZ Cat# ACC-305, RRID:CVCL_0045)

Curator: @AlecAsdourian

SciCrunch record: RRID:CVCL_0045

---

What is this?

DOI: 10.1101/2024.09.18.613698

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.1101/2024.04.12.589206

Resource: Ottawa Hospital Research Institute StemCore Laboratories Core Facility (RRID:SCR_012601)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_012601

---

What is this?

Traceback (most recent call last): File "/home/ubuntu/dashboard/py/create_release_tables.py", line 54, in format_anno_for_release parsedanno = HypothesisAnnotation(anno) File "/home/ubuntu/dashboard/py/hypothesis.py", line 231, in init self.links = row['document']['link'] TypeError: string indices must be integers

Traceback (most recent call last): File "/home/ubuntu/dashboard/py/create_release_tables.py", line 54, in format_anno_for_release parsedanno = HypothesisAnnotation(anno) File "/home/ubuntu/dashboard/py/hypothesis.py", line 182, in init if 'link' in row['document']: TypeError: argument of type 'NoneType' is not iterable

Traceback (most recent call last): File "/home/ubuntu/dashboard/py/create_release_tables.py", line 54, in format_anno_for_release parsedanno = HypothesisAnnotation(anno) File "/home/ubuntu/dashboard/py/hypothesis.py", line 182, in init if 'link' in row['document']: TypeError: argument of type 'NoneType' is not iterable

Traceback (most recent call last): File "/home/ubuntu/dashboard/py/create_release_tables.py", line 54, in format_anno_for_release parsedanno = HypothesisAnnotation(anno) File "/home/ubuntu/dashboard/py/hypothesis.py", line 182, in init if 'link' in row['document']: TypeError: argument of type 'NoneType' is not iterable

DOI: 10.1039/D2CC07004A

Resource: None

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_017362

---

What is this?

DOI: 10.1039/D0CP04274A

Resource: None

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_017362

---

What is this?

DOI: 10.1038/s41598-023-38858-4

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.1038/s41596-024-01072-1

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

Traceback (most recent call last): File "/home/ubuntu/dashboard/py/create_release_tables.py", line 54, in format_anno_for_release parsedanno = HypothesisAnnotation(anno) File "/home/ubuntu/dashboard/py/hypothesis.py", line 231, in init self.links = row['document']['link'] TypeError: string indices must be integers

DOI: 10.1038/s41467-024-54920-9

Resource: (RRID:CVCL_0320)

Curator: @AlecAsdourian

SciCrunch record: RRID:CVCL_0320

---

What is this?

DOI: 10.1038/s41467-024-54920-9

Resource: (DSMZ Cat# ACC-357, RRID:CVCL_0292)

Curator: @AlecAsdourian

SciCrunch record: RRID:CVCL_0292

---

What is this?

DOI: 10.1038/s41467-024-54920-9

Resource: (KCB Cat# KCB 200848YJ, RRID:CVCL_0546)

Curator: @AlecAsdourian

SciCrunch record: RRID:CVCL_0546

---

What is this?

DOI: 10.1038/s41467-024-54920-9

Resource: (IMSR Cat# JAX_005557,RRID:IMSR_JAX:005557)

Curator: @AlecAsdourian

SciCrunch record: RRID:IMSR_JAX:005557

---

What is this?

DOI: 10.1038/s44318-023-00010-3

Resource: University of California at San Francisco Advanced Microscopy Core Facility (RRID:SCR_025781)

Curator: @dhovakimyan1

SciCrunch record: RRID:SCR_025781

---

What is this?

DOI: 10.1038/s41419-024-07243-y

Resource: (BCRC Cat# 60420, RRID:CVCL_0041)

Curator: @AlecAsdourian

SciCrunch record: RRID:CVCL_0041

---

What is this?

Traceback (most recent call last): File "/home/ubuntu/dashboard/py/create_release_tables.py", line 54, in format_anno_for_release parsedanno = HypothesisAnnotation(anno) File "/home/ubuntu/dashboard/py/hypothesis.py", line 231, in init self.links = row['document']['link'] TypeError: string indices must be integers

Traceback (most recent call last): File "/home/ubuntu/dashboard/py/create_release_tables.py", line 54, in format_anno_for_release parsedanno = HypothesisAnnotation(anno) File "/home/ubuntu/dashboard/py/hypothesis.py", line 231, in init self.links = row['document']['link'] TypeError: string indices must be integers