中文翻譯:騎士精神, meaning: "a code of conduct with medieval knighthood and the ideals of medieval knights." Synonym: Knightliness, Gallantry, Nobility.

So you can think sort of think of the clock speed like the throttle in a car, or perhaps engine speed like RPM.

https://kimberlyhirsh.com/uploads/2023/5976538a38.jpg

Putting a list of one's core values in the front of their notebook can be a useful reminder within their journaling or bullet journal practice.

In addition to using the three dots on the Analog system's cards to indicate how much one accomplished (modest value), Jeff Sheldon suggests using them to "link cards together", though he doesn't suggest how one could or should do this. Presumably he means to use the possible dot patterns as a code, but then one only has 2^3 or 8 ways of doing this, so the number of possible links is incredibly low. Some of this seems related to edge notched cards, though here, there's no suggestion of punching holes in these cards, so sorting or finding these cards isn't necessarily easy unless one otherwise indexes them, a functionality which falls outside of the minimalist scope of the product.

To expand on this method he also, more profitably, suggests adding titles to cards in the blank line at the top which is also frequently used for dating cards.

https://www.annahavron.com/handouts

In 2019, Carthusian monks went back to their values and decided to scale back their production of Chartreuse.

Code of Colorado Regulations 500 - Department of Human Services 2506 - Food Assistance Program (Volume 4B) Rule 10 CCR 2506-1 - RULE MANUAL VOLUME 4B, FOOD ASSISTANCE Section 10 CCR 2506-1-4.600 - ONGOING CASE MANAGEMENT Section 10 CCR 2506-1-4.606 - PUBLIC ASSISTANCE (PA) HOUSEHOLD CHANGES

B. When there is a change in a PA case and the local office has sufficient information to make the corresponding SNAP adjustment, the local office shall follow the guidelines listed below.

1. If the change in household circumstances requires a reduction or termination of both PA and SNAP, the following action will be required:

b. If a household requests a fair hearing any time prior to the effective date of the Notice of Adverse Action, and its certification period has not expired, the household's participation in the program shall be continued on the basis authorized immediately prior to the Notice of Adverse Action, unless the household specifically waives continuation of benefits. Continued benefits shall not be issued for a period beyond the end of the current certification period. c. If the household appeals only a PA adverse action and is granted interim relief, SNAP benefits authorized prior to the adverse action shall continue or be restored. However, the household must reapply if the SNAP certification period expires before the hearing process is completed.

REGARDING TANF: as stated in the state plan and no doubt somewhere in the CCR- "If the appellant is receiving financial assistance, medical assistance, social services, or basic cash assistance under the Colorado Works Program at any time a conference or hearing is requested, all benefits will continue, pending the outcome of the State level fair hearing "

This pretty much the basis for Kartik's criticism in Literate Programming: Knuth is doing it wrong.

Thought experiment: what if you elevated the documentation to first-class status rather than as low-stakes generated artifacts that can be blown away and regenerated? What if you modified your compiler to consume the documentation and produce the same binary as the one produced by what you now consider to be your source code?

same comment as given in the V2 doc

Author Response

Reviewer #1 (Public Review):

This manuscript is interesting because of the exploration of a novel model organisms utilizing next-generation sequencing approaches, such as single-cell-RNA-seq. Despite the authors' efforts the manuscript lacks a cohesive narrative and suffers from being extremely preliminary in nature. For example, most of the figures are cut and pasted directly from the computational programs with very little formatting or thought to creating new knowledge from the data generated. Essentially the manuscript consists of 2-3 experiments where the authors performed single-cell-RNA-seq on different anatomical locations in the pig and also on a couple of different pig types (The Chenghua and Large White). The authors used standard computational pipelines consisting of Seurat, Monocle, Cell Chat, and others to characterize differences in their data.

There is potential in this manuscript but the authors should improve upon the manuscript by mining the data better and generating a better understanding of anatomical positions of pig skin by evaluating the Hox genes.

(1) Thanks for the reviewer's positive evaluation for our article and providing valuable feedback to improve the quality of our manuscript. To provide a more cohesive narrative, we have edited throughout the manuscript.

(2) Meanwhile, we also modified and formatted some figures including Figures 2-6, Figure 4—figure supplement 1 and 2, Figure 5—figure supplement 1 and 2, and Figure 6—figure supplement 1.

(3) We have analyzed these data of regional- or species-based differences more extensively, and the added content are in Result Section of “Heterogeneity of skin FBs in different anatomic sites” and “Heterogeneity of skin cells in different pig populations”.

(4) However, in our study, we did not identify any Hox gene among these differentially expressed genes in skin fibroblasts from both different anatomical sites and different pig populations. The differences of Hox code expression patterns might come from the heterogeneity of different species.

Reviewer #2 (Public Review):

The authors aimed to analyze different dermal compositions of various skin regions, focusing on fibroblast, endothelium and smooth muscle cells. They collect skin samples from six different skin regions of adult pig skin including the head, ear, shoulder, back, abdomen, and leg skins. After dissociating the tissues into single cells, they perform single-cell RNA analyses. A total of 215 thousand cells were analyzed. The authors identified distinct cell clusters, enriched molecules within each cell cluster, and the dynamic of cell cluster transition and interactions. Based on their findings, they conclude that tenascin N, collagen 11A1, and inhibin A are candidate genes for facilitating extracellular matrix accumulation.

Strength:

The methodology they used to prepare scRNA data is appropriate. Bioinformatic analyses are solid. The authors emphasize the heterogeneous phenotypes and composition ratios of smooth muscle cells, endothelial cells and fibroblasts in each skin region. They identify potential cell communication pathways among cell clusters. Expression of selective molecules on tissue sections were done.

Weakness:

While tenascin, collagen and inhibin are highlighted as genes important for ECM accumulation, there is no functional evaluation data. The discussion section is a compilation of comparisons, and is somewhat fragmentary. More significance from this dataset could have been extracted.

(1) We appreciate the reviewer's suggestions for evaluating the functional significance further. In our next research, we will perform some experiments in vitro and in vivo to explore the functions of these identified key genes.

(2) The discussion section have been greatly modified and it shall be more logical and readable.

Reviewer #1 (Public Review):

The manuscript of Parab et al. reports a beautiful phenotype analysis of the vascular brain/meningeal anatomy in a variety of reporter lines and mutants for Wnt/β-catenin signaling and angiogenic cues (Vegfaa, Vegfab Vegfc, Vegfd) during zebrafish development.

The original finding is that a region-specific code of angiogenic cues controls fenestrated vessel formation. The authors show that fenestrated vessels form independently of Wnt/β-catenin signaling and BBB vascular development but require different combinations of Vegfa and Vegfc/d-dependent angiogenesis within and across brain regions. A previously unappreciated function of autocrine and paracrine Vegfc signaling is demonstrated in this brain region-specific regulation of fenestrated capillary development.

My only main concern is that no information is provided on the regional diversity of angiogenic receptor expression that may correlate with the regional angiogenic factor code. Without asking for a spatial transcriptomic study, the combination of Vegfr-reporter lines or in situ hybridization with a combination of receptor probes would allow for generating a comprehensive set of ligand/receptor data relative to the regional angiogenic signaling pattern involved in fenestrated vessel formation.

Reviewer #2 (Public Review):

The authors convert the AHBA dataset into a dense cortical map and conduct an impressively large number of analyses demonstrating the value of having such data.

I only have comments on the methodology. First, the authors create dense maps by simply using nearest neighbour interpolation followed by smoothing. Since one of the main points of the paper is the use of a dense map, I find it quite light in assessing the validity of this dense map. The reproducibility values they calculate by taking subsets of subjects are hugely under-powered, given that there are only 6 brains, and they don't inform on local, vertex-wise uncertainties). I wonder if the authors would consider using Gaussian process interpolation. It is really tailored to this kind of problem and can give local estimates of uncertainty in the interpolated values. For hyperparameter tuning, they could use leave-one-brain-out for that.

I know it is a lot to ask to change the base method, as that means re-doing all the analyses. But I think it would strengthen the paper if the authors put as much effort in the dense mapping as they did in their downstream analyses of the data.

It is nice that the authors share some code and a notebook, but I think it is rather light. It would be good if the code was better documented, and if the user could have access to the non-smoothed data, in case they was to produce their own dense maps. I was only wondering why the authors didn't share the code that reproduces the many analyses/results in the paper.

If possible we should add an error response sample code as well

this number needs to match with the one in the next code sample?

Possible reasons for the phenomenon of a small group of power users and a large group of lurkers include the fact that power users tend to have more knowledge and experience in a particular field, and they are more confident and motivated to make the vast majority of substantive contributions. For example, when developing a software, it is often a core group of dedicated developers who submit most of the code changes and improvements, while the larger user community benefits from their contributions without actively participating in the coding.

This is a very good example. For me, it highlights the ambiguity of "openness." In this case, there is a mostly public platform for open software development collecting private data for the benefit of one company. The useful product lures developers into the ecosystem, where they serve as a resource. It's a bit like farming, with animals you can't fully control.

Author Response

Reviewer #1 (Public Review):

The authors have compiled and analysed a unique dataset of patients with treatment-resistant aggressive behaviours who received deep brain stimulation (DBS) of the posterior hypothalamic region. They used established analysis pipelines to identify local predictors of clinical outcomes and performed normative structural and functional connectivity analyses to derive networks associated with treatment response. Finally, Gouveia et al. perform spatial transcriptomics to determine the molecular substrates subserving the identified circuits. The inclusion of data from multiple centres is a notable strength of this retrospective study, but there are current limitations in the methodology and interpretation of findings that need to be addressed.

1) The validation of findings is heterogeneous and inconsistent across analysis pipelines. While the authors performed non-parametric permutation testing during sweet-spot mapping, structural and functional connectivity were validated using a 'four-fold consistency analysis'. The latter consists of a visual representation of streamlines and peak intensities after randomly dividing data into four groups, the findings were not validated quantitatively. If possible, the authors should apply permutation analysis in alignment with sweet-spot mapping and demonstrate the predictive ability of their identified networks in a LOO or k-fold cross-validation paradigm as carried out by similar studies. Given that the data has been derived from multiple centers, the prediction of left-out cohorts based on models generated by the remaining cohorts could be another means of validation. If validation is not possible, the authors should clearly state the limitations of their approach.

We appreciate the comment. We have now improved the validation of our connectomics analyses and removed the four-fold consistency analysis. For the functional connectivity analysis, we performed a 1000 permutation test (p<0.05). Similar brain areas were detected in the corrected and uncorrected maps. For the structural connectivity analysis, we used False Discovery Rate (FDR) correction at a significant level of p<0.001, as it is not feasible to perform a 1000 permutation test with this data. The structural connectome is composed of 12 million fibres, and every single permutation takes approximately 4 hours to be completed using our most powerful computational system. To perform 1000 permutations, it would take at least 4000 hours (i.e. 167 days or 5.5 months) of uninterrupted analysis to complete the test. However, it is important to highlight that an FDR correction at the level of p<0.001 is an extremely stringent method. This means that of the 23,000 fibres detected as being touched by the VATs, only 23 would be incorrect, while the remaining 22,977 are correct. Here again, we observed many similarities between the uncorrected and corrected maps, with the main anatomical structures being detected in both. The Methods section and Figures 4 and 5 were revised to reflect these changes.

2) In addition to a 'four-fold consistency analysis', functional connectivity was evaluated using LOOCV in a priori identified ROIs. Their network analysis, however, revealed a far more extensive network encompassing cortical, subcortical, and cerebellar structures. To avoid selection bias the authors should incorporate identified structures into their analysis and apply appropriate means of validation.

We thank the reviewer for this valuable suggestion. We originally did not explore the various significant areas but performed a more focused analysis intended to demonstrate that regions of the known ‘aggression network’ are indeed implicated in our findings. We performed a new analysis exploring the correlation between symptom improvement and the functional connectivity of all the areas described in Figure 5 (i.e., functional connectivity map). To this aim, we extracted individual connectivity values from the peak within each significant region and performed the same additive linear model, incorporating the functional connectivity of each area as well as the age of the patients to estimate individual symptomatic improvement. In addition, we performed a complete exploratory analysis considering the connectivity of any 2 brain structures and age. The resulting matrix shows to what extent functional connectivity to any two areas can be used to estimate clinical outcomes. Interestingly, this new analysis revealed the Periaqueductal Grey matter (PAG) to be the most important functionally connected area when investigated alone or in combination with brain structures critically involved in the regulation of emotional responses, namely the amygdala, anterior cingulate cortex, bed nucleus of the stria terminalis, nucleus accumbens, orbitofrontal cortex and fusiform gyrus. Also, the significance of the PAG connectivity was retained during leave-one-out cross-validation (LOOCV). The Methods, Results, Discussion and Figure 6 were revised. In addition, we added a new Table 2 and Supplementary File 1 to describe the new analysis and results.

3) Functional connectivity mapping: how were R-maps generated? The authors mention that patient-specific R-maps were p-thresholded and corrected for multiple comparisons, but it is not clear how group-level maps were generated. How did the authors perform regression on these maps? Were voxels that did not survive thresholding excluded?

This is a multiple-step analysis. First, it is necessary to localize the electrodes in each patient’s brain and estimate the volume of activated tissue (VAT) observed when stimulation parameters associated with symptomatic improvement are used. The VATs are then used as seeds for the next steps, during which we investigate how much functional influence the VTAs have on the other areas of the brain (i.e., individual r-map). This is done by correlating the BOLD time course of the VAT’s seed with the BOLD time course of all other voxels in the brain. The individual r-maps are then corrected for multiple comparisons to exclude voxels with potentially spurious correlations, resulting in an individual r-map that only included voxels surviving Bonferroni correction at the level of p<0.05. Finally, to create group-level maps, a voxel-wise linear regression analysis was performed to investigate whether each voxel of the map exerts more or less influence (corrected individual r-map with the functional connectivity of the patient’s VAT) or is more or less related to the clinical outcome (i.e. individual improvement). The last step is a permutation correction resulting in a significant group-level functional connectivity map (ppermute<0.05). We modified the Methods section and added a new Figure 1-figure supplement 1 illustrating this analysis.

4) The authors determined that age was a significant prédictor of the outcome, but it is unclear whether certain age groups presented with distinct etiologies underlying their aggressiveness. For example, aggression in epilepsy may show a better response to DBS as opposed to schizophrenia. How does patient outcome change when stratifying according to etiology? How does model performance change when controlling for etiology? The authors should include the etiology of aggressiveness in Table 1.

This is an interesting point. We observed a similar distribution between the pediatric and adult populations in relation to the most common etiologies reported. Epilepsy was the most frequent diagnosis in both populations (pediatric: 50%, adult: 62%), followed by autism spectrum disorder (pediatric: 34%, adult: 24%). The remaining etiologies were largely composed of single cases. A similar proportion of intellectual disability was also observed in pediatric and adult populations. Severe cases were observed in 75% of pediatric and 85% of adult patients. Moderate disability was present in 25% of pediatric and 15% of adult patients. Since several diagnoses were unique to some patients, the addition of this information to Table 1 could result in the identification of the patient. Thus, to preserve anonymity, the diagnoses were added to the end of Table 1 from more to less frequent. We have also revised the Results and Discussion sections to address this concern.

5) Stimulation parameters. The authors report average pulse widths of 219 µs and 142µs respectively, which is up to 4-fold higher as compared to DBS settings used conventionally in movement disorders and will significantly alter the volume of activated tissue. Did the authors account for the drastic increases in pulse width during VAT modeling?

We thank the reviewer for raising this point about the volume of activated tissue (VAT) modelled and the unusual pulse width observed in some patients in this cohort. These patients presented stimulation-induced sympathetic side effects when DBS was set with higher frequencies (e.g. increased heart rate and blood pressure). The chosen final parameters were the ones associated with a clinical benefit without generating side effects. There are a multitude of ways to estimate the VATs, from advanced axon cable models – the gold standard, which simulate axon membrane dynamics and require patient-specific diffusion-weighted imaging and tremendous computing power 1 - to simple heuristics-based models that estimate the rough extent of a VAT based on stimulation parameters without constructing an actual spatial model 2–4. The model employed in our study (and in a number of previous publications by our group 5–10) was the FieldTripSimBio ‘E-field norm’ finite element method (FEM) model. This model, which was first described by Horn et al. 11 and is freely available in Lead-DBS (https://www.lead-dbs.org/), strikes a balance between the sophisticated axon cable models and the simpler heuristic models. In particular, it constructs an electric field (E-field, by applying an electric field strength threshold, or activation threshold) and calculates the VAT associated with specific voltage settings and contact configurations, taking into account the conductivity of surrounding brain tissue and electrode components. Notably, studies comparing VAT modelling techniques 12 showed that ‘E-field norm’ FEM models closely approximate (<0.1 mm difference) the gold standard axon cable models in terms of the size of VATs constructed for monopolar stimulation settings. However, it should be acknowledged that the FieldTripSimBio model in Lead-DBS does not allow the user to specifically enter values for pulse width. Instead, it employs a standard activation/electric field strength threshold (0.2 V/mm) that reflects a combination of commonly modelled axon diameters (roughly 3.5 μm) and pulse width values (i.e., 60-90 μs). This threshold is based on work by researchers such as Astrom et al. 13 and reflects a ‘middle ground’ value that takes into account the fact that any VAT model will necessarily be an imperfect approximation of how electrical stimulation interfaces with brain tissue, depending heavily on aspects such as the diameter of local axons. Nonetheless, it is certainly understood that increased pulse width does meaningfully increase the effective range of stimulation (thus translating to a larger VAT) by lowering the activation threshold of nearby axons 12.

Given that our patient cohort included a small number of patients who were stimulated with higher pulse widths than the values assumed by our model (90 μs), it is reasonable to wonder whether we underestimated the size of these patients’ VATs. To address this aspect, we modelled these patients’ VATs using a simpler heuristic model 2 that does allow specific pulse width values to be selected by the user. More specifically, we computed a range of VATs for these patients using varied pulse width values (ranging from 90 μs up to their actual values). Not surprisingly, this endeavour did yield larger VATs when higher pulse widths were used. On average, the absolute difference in VAT diameter between 90 μs and 450 μs (the largest pulse width observed in this cohort) versions of these patients’ VATs was 2 mm. To check whether or not this difference could have potentially impacted our results, we repeated our probabilistic mapping analysis using altered VATs (specifically, VATs that were enlarged by 2 mm in diameter) for the patients with higher pulse widths. This new repeat analysis yielded a very similar average map to the original analysis: the overall map pattern and location/values of the peak corresponding to the most efficacious area for maximal symptom alleviation remaining unaltered, and only a few voxels on the periphery of the map changing in value by a couple of percentage points. This new supplementary analysis indicates that our results were not meaningfully altered by the unusual pulse width observed in these patients. We modified the Methods section to address some of these aspects and added a new Figure 3-figure supplement 2 illustrating both voxel efficacy maps.

6) Imaging transcriptomics. The methods described lack detail: How did the authors account for differences in expression across donors, samples, and regions during preprocessing of the Allen Human Brain Atlas? How was expression data collapsed into regions of interest? Did the authors apply any normalization? Recent publications have introduced reproducible workflows for processing and preparing the AHBA expression data for analysis that is publicly available.

7) 'genes with similar patterns of spatial distribution to the TFCE map were compiled in an extensive list'. It is unclear why authors used TFCE maps for spatial transcriptomics as opposed to the functional connectivity map featured in Figure 5. How was similarity measured between the TFCE map and the AHBA? How were candidate genes identified? Please provide a more comprehensive description of the analysis pipeline.

We apologize for the short description of this analysis. We performed a gene set analysis using the abagen toolbox (https://abagen.readthedocs.io/en/stable/index.html) to investigate genes with a spatial pattern distribution similar to one of clinically relevant functional connectivity. For this analysis, we used the Allen Human Brain Atlas (https://alleninstitute.org/) microarray data describing the cortical, subcortical, brainstem and cerebellar localization of over 20,000 genes in the human brain (3702 anatomical locations from 6 neurotypical adult brains) 14–17, along with a cell-specific aggregate gene set 18. These data are provided preprocessed, with gene expression values normalized across all brains, and registered to standard MNI space, allowing for a direct comparison between the spatial pattern of gene expression and the functional connectivity map (https://human.brain-map.org/microarray/search) 15. The TFCE maps were used to create clusters of clinically relevant functional connectivity with a spatial extent that overlaps with the anatomical locations from which microarray data was obtained. We parcellated both datasets (results of functional connectivity analysis and Allen Gene Atlas) according to the Harvard-Oxford brain atlas and correlated the spatial distribution of gene expression with the spatial distribution of the results of the functional connectivity mapping. The resultant list of candidate genes was used as input in gene ontology tools to investigate the associated biological processes and cell types. It is important to highlight that this process involves 2 corrections for multiple comparisons using FDR at q<0.005; one correction occurs at the level of the gene list to include only the most significant genes in the gene ontology analysis; a second correction occurs at the level of the gene ontology analysis to consider only the most significant biological processes. We have included some of these details in the revised Methods section.

8) What do the bar plots in Figure 7 (left) represent? P-values? The authors should label the axes to make this clear to the reader.

9) Interprétation of imaging transcriptomics: The authors identify a therapeutic circuit associated with deep brain stimulation of the posterior hypothalamic area, however, it is unclear how to reconcile genes associated with hormones, inflammation, and plasticity in this context. The authors mention and discuss genes implicated in hormonal processing, specifically oxytocin. The results provided in Figure 7, however, do not support this finding and it is unclear how the authors identified genes linked to oxytocin. In addition, the authors identified reductions in the number of microglia and astrocytes, while oligodendrocytes were overexpressed relative to the expected distribution of genes per cell type. These findings were attributed to DBS effects, however, both connectomic and transcriptomic data are acquired from healthy subjects, which suggests a physiological deficit/enrichment in a therapeutic circuit. How do the authors interpret findings given that no electrode implantation and stimulation were performed?

The analysis of normative datasets (functional and structural connectomics and spatial transcriptomics) is based on the idea of better understanding mechanisms of treatment considering our current knowledge of the average human brain. Unlike patient-specific studies in which imaging is acquired from a single patient or genetic profiles are extracted from tissue samples, these normative analyses rely on high-quality “atlases” derived from healthy subjects. In the case of functional and structural connectivity, these atlases are calculated from very large cohorts of subjects (around 1000 brain scans). Thus, imaging connectomics investigates the pattern of brain activity and structural connectivity related to a specific area of the brain (in this case, the volume of tissue activated (VATs) with DBS) and correlate these data with clinical outcomes to shed light on potential mechanisms of action. Similarly, the spatial transcriptomic analysis identifies spatial correlations between patterns of gene expression and brain characteristics detected by MRI 19 (in this case, the spatial pattern of functional connectivity) to investigate possible genetic underlying mechanisms. It is important to highlight that previous studies have shown that normative analyses yield results that are similar to the ones observed using patient-specific data 20–22. In the specific case of imaging connectomics, It has been shown that normative datasets can be used to create probabilistic models of optimal connectivity associated with patients’ outcomes that are meaningful to predict outcomes in patient-specific connectivity data 21. Thus, these exploratory data-driven approaches strive to simulate the presumed fingerprint that a particular patient’s individualized DBS intervention might modulate. They also allow the investigation of possible mechanisms of action in a large, previously inaccessible cohort of patients whose individual data are available. We apologize for the inaccuracy in Figure 7. Along with improving the Discussion section of the manuscript, we included the label for the bar plots in the left panel to improve the clarity of the graph and added the missing result from the KEGG 2021 Human Library that shows the oxytocin signalling pathway.

10) Data availability. Code used for data processing should be made openly available or shared as source data along with the Figures that were generated using the code. Sweet-spot, structural, and functional connectivity maps should be shared openly.

All tools and codes necessary for localizing the electrodes, estimating the volume of activated tissues, and analyzing imaging connectomics are freely available in Lead-DBS (https://www.lead-dbs.org/), a toolbox designed for DBS electrode reconstructions and computer simulations based on postoperative imaging. All codes for spatial transcriptomics are freely available in abagen (https://abagen.readthedocs.io/en/stable/), a toolbox designed to analyze the Allen Brain Atlas genetics data. Along with the codes, the websites for these tools provide manuals describing the step-by-step procedure for successful analysis. The datasets were made freely available at Zenodo (doi: 10.5281/zenodo.7344268). We improved our Data Availability Statement to address this concern.

Reviewer #2 (Public Review):

Deep brain stimulation (DBS) is an important, relatively new approach for treating refractory psychiatric illnesses including depression, addiction, and obsessive-compulsive disorder. This study examines the structural and functional connections associated with symptom improvement following DBS in the posterior hypothalamus (pHyp-DBS) for severe and refractory aggressive behavior. Behavioral assessments, outcome data, electrode placements, and structural and functional (resting-state) imaging data were collected from 33 patients from 5 sites. The results show structural connections of the effective electrodes (91% of patients responded positively) were with sensorimotor regions, emotional regulation areas, and monoamine pathways. Functional connectivity between the target, periaqueductal gray, and amygdala was highly predictive of treatment outcome.

Strengths.

This dataset is interesting and potentially valuable.

Weaknesses.

The figures seem to indicate that electrodes and symptom improvement is located lateral to the hypothalamus, perhaps in the subthalamic nucleus (STN). This is might explain why the streamlines from the tractography are strongest in motor regions. The inclusion of the monoaminergic based on the tractography is not warranted, as the resolution is not sufficient to demonstrate the distinction between the MFB (a relatively small bundle) and others flowing through this region to the brainstem.

This is an interesting point. The sweet spot identified in this work is indeed located in the posterior-inferior-lateral region of the posterior hypothalamic area, reaching the most superior part of the red nucleus, without including the STN. It is important to highlight that the voxel-efficacy mapping only shows voxels associated with a minimum of 50% symptomatic improvement following treatment. Thus, the areas not touching the red nucleus are also associated with excellent symptom alleviation. Although the structural connectivity mapping revealed tracts involved in motor and sensory information, it also showed tracts known to be involved in the regulation of emotions, such as the MFB, the Amygdalofugal Pathway and the ALIC. It is worth noting that this analysis is excellent for segregating the fibre tracts as relevant or not associated with a clinical improvement, but it is not capable of tearing apart the system to determine which of those are necessary for symptom alleviation. As a result, it is not possible to determine whether the motor projections are stronger or more relevant than others. However, the structural connectivity analysis presented here contributes to the body of knowledge on the network of aggressive behaviour and provides clinically relevant data that can be useful to improve future patient outcomes.

We agree with the reviewer that the engagement of the motor system is indeed highly relevant for the reduction of aggressive behaviours, as we have previously shown that aggressive behaviour is highly correlated with motor agitation 23,24. Additionally, in the context of ASD, self-injury behaviour is defined as a type of repetitive/stereotypic behaviour that results in physical injury to the patient’s own body. In relation to the involvement of the monoaminergic system, we would like to apologize for not being clear in the discussion of our findings. Although the functional and structural connectivity maps are related, they provide different means of exploring distinct aspects of the connectivity profile of each VAT. While the structural connectivity map may elucidate symptom improvement via direct fibre modulation (i.e. fibres that touch vs fibres that do not touch the VAT), the functional connectivity map investigates the functional dynamics of the network via BOLD signals (functional MRI). In this manuscript, we showed the functional connectivity (not fibre tracts) of the VATs with areas known to regulate monoamine production, such as the Raphe nuclei and the Substantia Nigra. Both serotonin and dopamine are critically involved in the control of aggressive behaviours, being the target of the main medications used to treat these symptoms in several patient populations. To address all the raised concerns, we incorporated a few sentences in the discussion, highlighting the relevance of the motor system and some limitations of our analysis. We also added a new Figure 3-figure supplement 1 and a discussion on the position of the sweet spot in relation to the red nucleus and subthalamic nucleus.

REFERENCES: 1. Gunalan, K., Howell, B. & McIntyre, C. C. Quantifying axonal responses in patient-specific models of subthalamic deep brain stimulation. Neuroimage 172, 263–277 (2018). 2. Dembek, T. A. et al. Probabilistic mapping of deep brain stimulation effects in essential tremor. Neuroimage Clin 13, 164–173 (2017). 3. Kuncel, A. M., Cooper, S. E. & Grill, W. M. A method to estimate the spatial extent of activation in thalamic deep brain stimulation. Clin. Neurophysiol. 119, 2148–2158 (2008). 4. Mädler, B. & Coenen, V. A. Explaining clinical effects of deep brain stimulation through simplified target-specific modeling of the volume of activated tissue. AJNR Am. J. Neuroradiol. 33, 1072–1080 (2012). 5. Elias, G. J. B. et al. Probabilistic Mapping of Deep Brain Stimulation: Insights from 15 Years of Therapy. Ann. Neurol. 89, 426–443 (2021). 6. Germann, J. et al. Brain structures and networks responsible for stimulation-induced memory flashbacks during forniceal deep brain stimulation for Alzheimer’s disease. Alzheimers. Dement. 17, 777–787 (2021). 7. Elias, G. J. B. et al. Mapping the network underpinnings of central poststroke pain and analgesic neuromodulation. Pain 161, 2805–2819 (2020). 8. Gouveia, F. V. et al. Case report: 5 Years follow-up on posterior hypothalamus deep brain stimulation for intractable aggressive behaviour associated with drug-resistant epilepsy. Brain Stimul. 14, 1201–1204 (2021). 9. Coblentz, A. et al. Mapping efficacious deep brain stimulation for pediatric dystonia. J. Neurosurg. Pediatr. 27, 346–356 (2021). 10. M Oliveira, L. et al. Probabilistic characterisation of deep brain stimulation in patients with tardive syndromes. J. Neurol. Neurosurg. Psychiatry (2021) doi:10.1136/jnnp-2020-324270. 11. Horn, A. et al. Connectivity Predicts deep brain stimulation outcome in Parkinson disease. Ann. Neurol. 82, 67–78 (2017). 12. Duffley, G., Anderson, D. N., Vorwerk, J., Dorval, A. D. & Butson, C. R. Evaluation of methodologies for computing the deep brain stimulation volume of tissue activated. J. Neural Eng. 16, 066024 (2019). 13. Astrom, M., Diczfalusy, E., Martens, H. & Wardell, K. Relationship between neural activation and electric field distribution during deep brain stimulation. IEEE Trans. Biomed. Eng. 62, 664–672 (2015). 14. Hawrylycz, M. J. et al. An anatomically comprehensive atlas of the adult human brain transcriptome. Nature 489, 391–399 (2012). 15. Arnatkeviciute, A., Fulcher, B. D. & Fornito, A. A practical guide to linking brain-wide gene expression and neuroimaging data. Neuroimage 189, 353–367 (2019). 16. Arnatkeviciute, A., Markello, R. D., Fulcher, B. D., Misic, B. & Fornito, A. Toward Best Practices for Imaging Transcriptomics of the Human Brain. Biol. Psychiatry 93, 391–404 (2023). 17. Markello, R. D. et al. Standardizing workflows in imaging transcriptomics with the abagen toolbox. Elife 10, (2021). 18. Seidlitz, J. et al. Transcriptomic and cellular decoding of regional brain vulnerability to neurogenetic disorders. Nat. Commun. 11, 3358 (2020). 19. Fornito, A., Arnatkevičiūtė, A. & Fulcher, B. D. Bridging the Gap between Connectome and Transcriptome. Trends Cogn. Sci. 23, 34–50 (2019). 20. Arnatkeviciute, A., Fulcher, B. D., Bellgrove, M. A. & Fornito, A. Imaging Transcriptomics of Brain Disorders. Biol Psychiatry Glob Open Sci 2, 319–331 (2022). 21. Wang, Q. et al. Normative vs. patient-specific brain connectivity in deep brain stimulation. Neuroimage 224, 117307 (2021). 22. Elias, G. J. B. et al. Normative connectomes and their use in DBS. in Connectomic Deep Brain Stimulation 245–274 (Elsevier, 2022). 23. Gouveia, F. V. et al. Bilateral Amygdala Radio-Frequency Ablation for Refractory Aggressive Behavior Alters Local Cortical Thickness to a Pattern Found in Non-refractory Patients. Front. Hum. Neurosci. 15, 653631 (2021). 24. Venetucci Gouveia, F. et al. Case report: 5 Years follow-up on posterior hypothalamus deep brain stimulation for intractable aggressive behaviour associated with drug-resistant epilepsy. Brain Stimul. (2021) doi:10.1016/j.brs.2021.07.062. 25. Gouveia, F. V. et al. Amygdala and Hypothalamus: Historical Overview With Focus on Aggression. Neurosurgery 85, 11–30 (2019). 26. Gouveia, F. V. et al. Longitudinal Changes After Amygdala Surgery for Intractable Aggressive Behavior: Clinical, Imaging Genetics, and Deformation-Based Morphometry Study-A Case Series. Neurosurgery 88, E158–E169 (2021). 27. Yan, H. et al. Deep brain stimulation for extreme behaviors associated with autism spectrum disorder converges on a common pathway: a systematic review and connectomic analysis. J. Neurosurg. 1–10 (2022). 28. Knotkova, H. et al. Neuromodulation for chronic pain. Lancet 397, 2111–2124 (2021). 29. Gray, A. M. et al. Deep brain stimulation as a treatment for neuropathic pain: a longitudinal study addressing neuropsychological outcomes. J. Pain 15, 283–292 (2014). 30. Miczek, K. A., Fish, E. W., De Bold, J. F. & De Almeida, R. M. M. Social and neural determinants of aggressive behavior: pharmacotherapeutic targets at serotonin, dopamine and gamma-aminobutyric acid systems. Psychopharmacology 163, 434–458 (2002). 31. Gouveia, F. V. et al. Reduction of aggressive behaviour following hypothalamic deep brain stimulation: involvement of 5-HT1Aand testosterone. bioRxiv (2023) doi:10.1101/2023.03.20.533520.

Reviewer #1 (Public Review):

The authors have compiled and analysed a unique dataset of patients with treatment-resistant aggressive behaviours who received deep brain stimulation (DBS) of the posterior hypothalamic region. They used established analysis pipelines to identify local predictors of clinical outcomes and performed normative structural and functional connectivity analyses to derive networks associated with treatment response. Finally, Gouveia et al. perform spatial transcriptomics to determine the molecular substrates subserving the identified circuits. The inclusion of data from multiple centres is a notable strength of this retrospective study, but there are current limitations in the methodology and interpretation of findings that need to be addressed.

1) The validation of findings is heterogeneous and inconsistent across analysis pipelines. While the authors performed non-parametric permutation testing during sweet-spot mapping, structural and functional connectivity were validated using a 'four-fold consistency analysis'. The latter consists of a visual representation of streamlines and peak intensities after randomly dividing data into four groups, the findings were not validated quantitatively. If possible, the authors should apply permutation analysis in alignment with sweet-spot mapping and demonstrate the predictive ability of their identified networks in a LOO or k-fold cross-validation paradigm as carried out by similar studies. Given that the data has been derived from multiple centers, the prediction of left-out cohorts based on models generated by the remaining cohorts could be another means of validation. If validation is not possible, the authors should clearly state the limitations of their approach.

2) In addition to a 'four-fold consistency analysis', functional connectivity was evaluated using LOOCV in a priori identified ROIs. Their network analysis, however, revealed a far more extensive network encompassing cortical, subcortical, and cerebellar structures. To avoid selection bias the authors should incorporate identified structures into their analysis and apply appropriate means of validation.

3) Functional connectivity mapping: how were R-maps generated? The authors mention that patient-specific R-maps were p-thresholded and corrected for multiple comparisons, but it is not clear how group-level maps were generated. How did the authors perform regression on these maps? Were voxels that did not survive thresholding excluded?

4) The authors determined that age was a significant prédictor of the outcome, but it is unclear whether certain age groups presented with distinct etiologies underlying their aggressiveness. For example, aggression in epilepsy may show a better response to DBS as opposed to schizophrenia. How does patient outcome change when stratifying according to etiology? How does model performance change when controlling for etiology? The authors should include the etiology of aggressiveness in Table 1.

5) Stimulation parameters. The authors report average pulse widths of 219 µs and 142µs respectively, which is up to 4-fold higher as compared to DBS settings used conventionally in movement disorders and will significantly alter the volume of activated tissue. Did the authors account for the drastic increases in pulse width during VAT modeling?

6) Imaging transcriptomics. The methods described lack detail: How did the authors account for differences in expression across donors, samples, and regions during preprocessing of the Allen Human Brain Atlas? How was expression data collapsed into regions of interest? Did the authors apply any normalization? Recent publications have introduced reproducible workflows for processing and preparing the AHBA expression data for analysis that is publicly available.

7) 'genes with similar patterns of spatial distribution to the TFCE map were compiled in an extensive list'. It is unclear why authors used TFCE maps for spatial transcriptomics as opposed to the functional connectivity map featured in Figure 5. How was similarity measured between the TFCE map and the AHBA? How were candidate genes identified? Please provide a more comprehensive description of the analysis pipeline.

8) What do the bar plots in Figure 7 (left) represent? P-values? The authors should label the axes to make this clear to the reader.

9) Interprétation of imaging transcriptomics: The authors identify a therapeutic circuit associated with deep brain stimulation of the posterior hypothalamic area, however, it is unclear how to reconcile genes associated with hormones, inflammation, and plasticity in this context. The authors mention and discuss genes implicated in hormonal processing, specifically oxytocin. The results provided in Figure 7, however, do not support this finding and it is unclear how the authors identified genes linked to oxytocin. In addition, the authors identified reductions in the number of microglia and astrocytes, while oligodendrocytes were overexpressed relative to the expected distribution of genes per cell type. These findings were attributed to DBS effects, however, both connectomic and transcriptomic data are acquired from healthy subjects, which suggests a physiological deficit/enrichment in a therapeutic circuit. How do the authors interpret findings given that no electrode implantation and stimulation were performed?

10) Data availability. Code used for data processing should be made openly available or shared as source data along with the Figures that were generated using the code. Sweet-spot, structural, and functional connectivity maps should be shared openly.

Author Response

Reviewer #1 (Public Review):

The strength of the manuscript is highlighted by the application of fractal formalism, which is commonly used in colloidal systems, in conjunction with MD simulation to study the phase separation of an IDP. The weakness lies in the fact that this study does not provide any discussion on how our understanding of the network structure and dynamical behavior of biomolecular condensates and their biological significance improves through this study. The experimental part remains weak, without any measurements of the dynamics of the condensates. Whether and how the formalism can distinguish between phase-separated condensates (WT) and classical protein aggregates (Y to A variant) remains unclear.

We thank the Reviewer for their careful reading of the manuscript and their appreciation of the link between IDP phase separation and colloid chemistry. Establishment of a quantitative framework behind this link, as given by the fractal formalism, and a multiscale model of the spatial organization of a biomolecular condensate, derived from MD simulations in combination with fractal scaling, are indeed two of our main contributions. In particular, to the best of our knowledge, ours is the first atomistically resolved model of the spatial organization of a biomolecular condensate at an arbitrary scale. The key features of the proposed model, as elaborated in the Discussion of the revised manuscript (p. 18, 20-21), are the coexistence of differently sized clusters inside a condensate, and a quantitative prediction of a particular scaling of mass with cluster size (Figure 5A), as further discussed below. Moreover, our results also point to the possible formation of pre-percolation clusters with sizes below the resolution limit of typical microscopy experiments, in agreement with recent observations (https://doi.org/10.1073/pnas.2202222119).

We agree that the full understanding of biomolecular condensates also requires a detailed treatment of the dynamical aspects. Following the Reviewer’s comments, we have provided significant new results in this regard and included an experimental characterization of fusion behavior (Videos 1, 2) and condensate dynamics by FRAP (Figure 1D, E and Figure 1—figure supplement 2) as well as a detailed analysis of diffusion and viscosity in the simulated systems (Figure 4C and Figure 4—figure supplement 1D-F). The newly performed FRAP experiments provide a direct measure of the condensate dynamics. Importantly, the measured recovery half-times for WT and R>K condensates resemble those of other well-characterized in vitro condensates. We have occasionally observed elongated, amorphous Y>A precipitates, albeit in low number and only at 50-fold higher concentration than the wild-type (45 mM and above, Figure 1C). While this may be consistent with the predictions of the fractal model and hint at the differences in mesoscopic organization between the WT and R>K condensates and the Y>A precipitates, the latter are rare and we are reluctant to draw major conclusions.

Furthermore, we could show that the WT diffusion coefficient is lower than for either mutant (Figure 4C, and Supplementary File 2). Clearly, this difference is not due to the effect of protein size or a higher solvent viscosity, but primarily indicates protein slow-down due to the more extensive interactions with partners (reflected also in higher average valency, Figure 2D, or probability of interactions Figure 2—figure supplement 1D). The fact that the WT diffusion coefficient drops by about 20% over the last 0.3 µs of the MD trajectory also correlates with the formation of a single percolating cluster in the system (Figure 2C). This is an expected effect on protein diffusion upon crossing the percolation threshold (https://doi.org/10.1038/ncomms11817, https://doi.org/10.1021/acs.jpcb.7b08785). Moreover, the difference in the recovery dynamics observed for WT and R>K mutant can be interpreted using the proposed model. Namely, accurate fitting of FRAP data was only possible if using at least two components (Figure 1—figure supplement 2). According to https://doi.org/10.1016/j.tcb.2004.12.001, these components indicate the contribution of particle diffusion and interaction (binding). Thus, recovery of the centrally bleached condensates is faster for WT than for the R>K mutant, which can be related to the higher compactness of the WT particles across scales as compared to R>K. On the other hand, the FRAP results for the condensates bleached in the peripheral area highlight the contribution of the binding component. Indeed, the recovery is about 3-fold faster for the R>K mutant, which could potentially be related to the lower valency of the interactions and the ease of the replacement of inactivated fluorescent species and/or exchange with proteins in the bulk. A further connection of the developed model and condensate dynamics concerns the multimodal description of diffusion in biomolecular condensates, together with multimodal fitting of FCS and FRAP data as used recently for interpreting single particle tracking results (https://doi.org/10.1016/j.bpj.2021.01.001). Namely, the polydisperse nature of the protein phase as suggested by the model translates to multimodal diffusion, reflecting the dynamics of protein clusters of different size. For instance, regularization fits used for DLS autocorrelation curves assume a multimodal character of the diffusion and are interpreted to reflect a multimodal distribution of cluster sizes in condensates (https://doi.org/10.1073/pnas.2202222119).

Finally, a way of testing the model prediction, which would merit a study in its own right, would involve static light scattering (SLS): a linearly decreasing scattering intensity as a function of the scattering vector in a log-log representation, as frequently seen for different colloidal systems, is expected by the fractal model. In fact, fractal dimension dF could directly be estimated from SLS experiments (https://doi.org/10.1038/339360a0) from the limiting value of scattering intensity for high values of the product of the scattering vector q and the average cluster size <Rg>. As a direct test of the predictions of the model, the experimental value of dF could then be compared with the predicted one. Moreover, techniques such as DLS and MALS could be used to measure independently masses and sizes of biomolecular condensates in vitro at different scales in order to test the validity of the particular scaling predicted by the fractal model. Such experiments are not trivial and are out of scope of the present study.

Reviewer #2 (Public Review):

A key aspect of the work is to use the simulations to explain differences between (i) dilute and dense phases and (ii) wild-type and mutant variants. Here, it would be important with a clearer analysis of convergence and errors to quantify which differences are significant.

Following the Reviewer’s suggestion, we now provide an analysis of convergence and statistical significance. Specifically, in Supplementary File 1 “Technical summary” we now report the average value, standard deviation and a block-average measure of convergence for all the key observables analyzed, including radius of gyration (Rg), valency (n), and compactness (), for all modeled systems. Furthermore, in the revised manuscript, we now also include the analysis of protein translational diffusion constants and solution viscosity for all modeled systems to assess the ability of the simulations to capture protein dynamics realistically (Figure 4C, Figure 4—figure supplement 1D-F, Supplementary File 2, see also above). Moreover, we include in the revised version a new figure depicting time evolution of average compactness in the 24-copy systems (Figure 4—figure supplement 1C). Thus, it can be seen that the two key model parameters derived from MD simulations of the 24-copy system – protein valency and compactness – reach a stable plateau over the last 0.3 µs (Figure 2D and Figure 4—figure supplement 1C), which were used for final analyses, with block-averaged deviations of less than 10% throughout (see below for details). All the differences in these parameters between single-copy and 24-copy simulations, as well as those between WT and mutation simulations, were found to be significant with p-values < 2.2 10-16 according to the Wilcoxon rank sum test with continuity correction (details in Supplementary File 1). Finally, considering the sampling limitations implicit in most MD studies, we clearly recognize the possibility that with longer simulation times or more protein copies per simulation box, the simulated systems may show a qualitatively different behavior. However, we emphasize that our derivation of the formalism that links the features of simulated ensembles on the scale of 10s of nanometers with their behavior on the scale of 100s of nanometers and beyond is independent of such limitations. Once longer, larger and more accurate simulations become available, one will be able to apply the formalism without alteration and obtain a model of the spatial organization of the condensate on an arbitrary scale, starting just from the local features of individual proteins. We now discuss these details on pp. 10, 11, 13 of the revised manuscript.

It would also be useful with a clearer description of how the analytical model is predictive, of which properties, and how they have been/can be validated. Which measurable quantities does the model predict?

As pointed above, the model predicts the existence and provides a quantitative description of pre-percolation finite-size clusters (https://doi.org/10.1016/j.molcel.2022.05.018, https://doi.org/10.1073/pnas.2202222119). More generally, the model provides the fractal dimension (dF) of protein clusters and enables evaluation of different scale-dependent properties of clusters of arbitrary size, including protein density as a function of cluster size (Figure 5—figure supplement 1C, Figure 5C). Importantly, the fractal dimension can be used in combination with local MD simulations and cluster–cluster aggregation algorithms to derive a detailed model of the 3D organization of fractal clusters of a chosen size at atomistic resolution (Figure 5A, B, and Videos 4, 5, and 6). Such detailed structural understanding of the interior organization of a condensate can, for example, be used to evaluate cavity sizes and interpret partitioning experiments. Since the differences in the morphology of WT and mutant protein clusters propagate across length scales, they can even be qualitatively characterized by the analysis of microscopic images (e.g. circularity, Figure 1—figure supplement 1C, see also discussion above). Finally, static light scattering (SLS) experiments give the possibility to test the model directly, which will be the subject of our future work. Namely, the fractal formalism predicts linear behavior in the log-log representation of the SLS intensity vs. scattering vector curves, while dF, which can directly be evaluated from such experiments, providing a quantitative point of comparison between theoretical predictions and experiment (see above).

In addition to these overall questions, a number of more specific suggestions follow below.

Major:

p. 7, line 120 (Fig. S1B) The proteins do not appear particularly pure based on the presented SDS PAGE analysis. How pure is the protein estimated to be, and is the presence of the other bands expected to affect e.g. the data presented in Fig. 1?

We have quantified the purity of the constructs by densitometry of the Coomassie stained gels and included it in Figure 1—figure supplement 1A: in the case of WT and R>K, we achieve purity higher than 91%. Importantly, the observed LLPS behavior of the constructs is consistent with the simulation and in agreement with other studies on R>K substitutions (https://doi.org/10.1073/pnas.2000223117; https://doi.org/10.1016/j.molcel.2020.01.025; https://doi.org/10.1073/pnas.2200559119; https://doi.org/10.1016/j.jmb.2019.08.008). In the case of Y>A, we have obtained the least pure protein (~65%), and must note that the precipitates observed in the experiments of Figure 1C are only present at the protein concentrations that are 50-fold higher as compared to WT (45 mM and above). Therefore, at such high total protein concentration, we cannot exclude the possibility that there might be some contamination affecting the behavior of this construct.

p. 7 & 8, lines 138-159: Has the method and energy function used to calculate the interact potential been validated by comparison to experiments, including studying the effect of varying the solvent? I see the computed error bars are very small, but am more interested in the average error when comparing to experiments. The numbers in water appear different from those e.g. reported by Krainer et al (https://doi.org/10.1038/s41467-021-21181-9), though the latter are also not immediately compared to experiments. Thus, it would be useful to know how much to trust these numbers.

We thank the Reviewer for raising this important point. To the best of our knowledge, the absolute binding free energies between Y-Y, Y-R or Y-K sidechain analogs or complete amino acids have never been determined experimentally, preventing a direct validation of the computed values and/or an evaluation of the average error when comparing to experiments. On the other hand, we did compare our data against the PMF curves presented by Krainer et al. (https://doi.org/10.1038/s41467-021-21181-9) for R-Y and Y-Y and the general trends are largely similar. In particular, in both analyses the R-Y interaction is stronger than the Y-Y interaction across different conditions, except at zero salt in Krainer et al. where the two are similar. When it comes to exact quantitative differences between the studies, it should first be pointed out that Krainer et al. studied capped amino-acids, while we used amino-acid side-chain analogs. The difference in the observed binding strengths is in part certainly related to the contribution of the capped backbone. Second, the values in Krainer et al. refer to the depth of the free energy minimum in the obtained PMFs and not to the resulting G values, as in our method. The latter includes integration over the PMF and an assumption of a standard-state concentration, which could also lead to significant differences. Finally, the differences could also be due to the intrinsic properties of the interaction potentials used. In particular, the prominent free-energy minima for the R-Y pair in the Krainer et al. study could only be obtained after refitting of the original AMBERff03ws charges on the Y bound to R via semi-empirical quantum-chemical calculations. On the other hand, the interaction potential used in our study was not adjusted to the system at hand, but rather comes from a published, widely used force field, the OPLS-AA (https://doi.org/10.1021/ja9621760), that was independently tested and validated experimentally in multiple studies. For example, OPLS-AA exhibits the low average error in absolute hydration free energy of ~0.5 kcal/mol, errors of only ~2% for heats of vaporization and densities (https://doi.org/10.1021/ja9621760), and a close agreement with osmotic coefficients (https://doi.org/10.1021/acs.jcim.9b00552) or a large range of organic compounds. This raises our confidence in the accuracy of the derived binding free energies, which directly or indirectly depend on these fundamental thermodynamic properties.

Regarding the method to evaluate PMF profiles, we have used a classical all-atom Monte Carlo approach originally developed by Jorgensen and coworkers (see, e.g., https://doi.org/10.1021/ar00161a004 and https://doi.org/10.1021/ja00168a022), as implemented in the widely used BOSS program (v. 4.8) (https://doi.org/10.1002/jcc.20297). This approach has been extensively tested against experimental data on ΔΔG values of various compounds in environments of different polarity (e.g., 2). Moreover, we have previously successfully applied this methodology in studies of the free energy of association of amino acid residues (https://doi.org/10.1021/jp803640e) and other biologically important groups (https://doi.org/10.1021/acs.jcim.9b00193). The results obtained have been compared with the available experimental data and demonstrated a good agreement. As for the small error bars in the plots, the fairly good convergence achieved in our PMF calculations is a result of extensive sampling combined with small system size, although obviously this is not always the case – see, for example, PMFs in our recent work (https://doi.org/10.1021/acs.jcim.9b00193).

The above points have been discussed on pp 7-8 of the revised manuscript.

p. 8, lines 149-154: Following up on the above, the authors also write "Importantly, only in the latter case are the R-Y interactions slightly more favorable than the K-Y ones (Figure S1C). While this can potentially contribute to increasing of Csat for the R>K mutant as compared to WT, the estimated thermodynamic effect is not too strong, especially if one considers that these interactions take place in an environment with largely water-like polarity. Therefore, the effect of R>K substitution on LLPS should be further explored in the context of protein-protein interactions." In the absence of estimates of the accuracy of the predictions, these sentences are somewhat unclear. Also, it is unclear what the authors mean by that the effect of R>K should be studied; there are already several examples of this (https://doi.org/10.1016/j.cell.2018.06.006 [already cited], https://doi.org/10.1038/s41557-021-00840-w & https://doi.org/10.1073/pnas.2000223117 come to mind, but there are likely more).

As pointed above, the free-energy values were obtained using well-established computational techniques and are expected to reflect realistic trends. However, considering that there exist no equivalent experimental results to assess the accuracy of the predicted free energies, they indeed must clearly be understood as predictions. This is now stated on pp. 7-8 of the revised manuscript. Furthermore, it seems that the vague phrasing on our part in the above paragraph resulted in a misunderstanding. Namely, when we talk about “further exploration”, we only meant it in relation to our study, i.e. a connection with the MD part, and not in relation to a wider literature on the topic. In other words, we simply wanted to refer to the fact that our binding free energies for individual residues do not provide sufficient information about interactions between Lge11-80 protein chains. Following the Reviewer’s comment, we have rephrased this part and included additional references on the known role of R and K residues on phase separation.

p. 8, lines 161-162: The authors perform MD simulations of Lge1 and variants using 24 copies and a box that gives them protein concentrations "in the mM concentration range". I realize that there's a concern about what is computationally feasible, but it would be important with an argument for this choice. Why is 24 expected to be enough to represent a condensate (I expect that there could be substantial finite-size effects)? What is the exact protein concentration in the simulations of the 24 chains [and of the 1-chain simulations]? How does this protein concentration compare to that in the condensates? The authors performed simulations in the NPT ensemble; how stable were the box dimensions?

The effective protein concentration for different 24-copy systems is 6-7 mM, depending on the system (Figure 2—figure supplement 1A). This concentration range was selected in order to get a reasonable system size for microsecond all-atom MD simulations, while still being approximately one order of magnitude lower than the semi-dilute regime of the protein at hand. As a testament to the internal consistency of our framework, the fractal model predicts the concentration inside WT condensates of the size observed in the experiment to indeed be in the mM range. Moreover, as seen in many other systems, the concentration inside the observed droplets is expected to be significantly higher than Csat (https://doi.org/10.1101/2020.10.25.352823). Here, we should again emphasize that we did not aim to model the process of phase separation in our all-atom MD. We rather use multicopy simulations for the analyses of the organization of the protein crowded phase and specifically, the mode of intermolecular interactions, and then use the fractal scaling to derive a model of the internal organization of condensates at arbitrary scales.

Regarding the experimental determination of the protein concentration in the condensates, we have used different approaches to estimate Csat and CD values: spin-down analyses (https://doi.org/10.1126/science.aaw8653), volumetry analysis (https://doi.org/10.1038/nchem.2803), estimation of concentration by fluorescent intensity of the condensates (https://doi.org/10.1016/j.molcel.2018.12.007; https://doi.org/10.1016/j.cell.2019.08.008), FCS (https://doi.org/10.1038/nchem.2803; https://doi.org/10.1016/j.cell.2019.10.011; https://doi.org/10.1126/science.aaw8653). However, different approaches yield values that vary by several orders of magnitude. That is the reason why we did not report definitive numbers. In general, there are uncertainties in the field about how to reliably measure protein concentrations in a condensate, necessitating the development of new approaches (https://doi.org/10.1101/2020.10.25.352823).

With regard to the convergence and potential finite-size effects, we agree that this is an important issue and have addressed it in the revised version. In general, the convergence of our observables such as valency or compactness (Figure 2C, D and Figure 4—figure supplement 1C) gives confidence that the simulations are at least in a local equilibrium, especially when it comes to short-range properties such as contact preferences as further elaborated in our reply to the Reviewer’s specific comment about convergence below (please, see also above for our response to Editor’s comment #5). Importantly, in all 24-copy systems, the average separation between protein images lies in the 12-15 nm range, and no instances of self-interaction between images due to PBC were observed (Supplementary File 1). Finally, analysis of fluctuations in box dimensions shows that they are all in the range of picometers and largely negligible when it comes to the analysis at hand (Supplementary File 1).

In order to highlight the realistic behavior of the simulated systems in the revised version, we now also report a detailed analysis of protein translational diffusion in MD simulations (Figure 4C and Figure 4—figure supplement 1D-F and Supplementary File 2). According to this analysis, single-molecule translational diffusion coefficients of Lge11-80 variants obtained from fitting of MSD curves with applied finite-size PBC correction and rescaling by the solvent viscosity (see Methods for details) are typically in the range of 100 µm2/s (Figure 4C and Supplementary File 2), which corresponds to experimentally measured values for different proteins of similar size. Importantly, the requisite finite-size corrections applied in the case of 24-copy systems are relatively small and amount to about 35-60%, while this is almost an order of magnitude higher (450-530%) for the single-copy simulations (Supplementary File 2). Please, see also the reply above to the Editor’s statements above for more details.

Also, did the authors include the Strep- and His-tags in the simulations? If not, why not?

We did not simulate the constant part of the constructs in order to: 1. expedite computation and 2. more directly expose the effect of different mutations. Since our comparison between simulation and experiment concerned largely qualitative observables, we have primarily focused on the relative differences between the three Lge11-80 variants. Importantly, the effect of mutations on the full-length protein and its different variants was analyzed in vivo in a previous publication (https://doi.org/10.1038/s41586-020-2097-z).

Throughout: One of my major concerns about this work is the general lack of analysis of convergence of the simulations. The authors must present some solid analysis of which results are robust given the relatively short simulations and potential for bias from the chosen starting structures.

First, we would like to emphasize that we did not attempt to capture the process of phase separation or characterize two coexisting phases, for which much larger ensembles and/or simulation times would be needed. Rather, our aim was to study the conformational behavior of individual protein chains in the context of a crowded protein mixture, taken as a model for the dense phase, and then use fractal scaling to provide a model of spatial organization of a condensate at an arbitrary length scale. Having said this, it is absolutely important to address how converged the key observables are, given the finite size of the all-atom simulation setup and the limited sampling used. In the revised manuscript, we have included an additional analysis of convergence of our simulations and could show that both key MD-derived parameters required by the fractal model, protein compactness and valency, display convergent behavior over the last third of 0.3 µs MD in the 24-copy systems (Figure 4—figure supplement 1C) and all analyses were performed over this region. In particular, the block averages of compactness and valency exhibit a standard deviation of only 2-4% and 4-8%, respectively, over the last 0.3 µs of MD simulations. Moreover, since we are interested in single-chain features in the context of a crowded mixture, our sampling corresponds effectively to 24 x 0.3 µs = 7.2 µs. Finally, a detailed analysis of convergence in conformational sampling was performed for single-copy simulations using calculations of configurational entropy as evaluated by the MIST formalism (Figure 4—figure supplement 1B). For instance, in the case of the weakly self-interacting Y>A, we do observe a close convergence in terms of the configurational entropy between two independent replicas on 1 µs MD trajectory (Figure 4—figure supplement 1B). However, we still recognize the possibility that with longer simulation times and/or more protein copies per simulation, the simulated systems may show a qualitatively different behavior, as discussed on pp. 10, 11, and 13 of the revised manuscript. Finally, we would like to reiterate the point that our derivation of the formalism that links the features of simulated ensembles on the scale of 10s of nanometers with their behavior on the scale of 100 s of nanometers and beyond is independent of such limitations. Once longer, larger and more accurate simulations become available, one will be able to apply the formalism without alteration and obtain a model of the spatial organization of the condensate on an arbitrary scale, starting just from the local features of individual proteins. We now discuss these details on pp. 10, 11, and 13 of the revised manuscript.

As an example, on p. 8 the authors discuss a potential asymmetry between the interactions found in the dilute (single-copy) and dense (24-mer) phases. These observations are somewhat in contrast to other observations in the field, namely that it is the same interactions that drive compaction of monomers as those that drive condensate formation.

Obviously, both the results in the literature and those presented here could be true. But in order to substantiate the statements made here, the authors should show some substantial statistical analyses to make it clear which differences are robust. The above holds for all parts of the computational/simulation work (e.g. other aspects of Fig. 2)

Note: this comment by the Reviewer echoes in several respects the comment 7 by the Editor. Because of this, our reply in some parts is identical to that given above to the Editor. We have decided to include it here for the ease of reading and completeness.

An expectation of the symmetry between intra- and intermolecular modes of interaction emerged from the background of polymer theory, which was primarily aimed to describe the behavior of homopolymers. In the case of heteropolymers such as proteins, the asymmetry in the aforementioned modes is rather intuitive. For instance, if there is only a single Y in a protein, then Y-Y contacts will not be possible in the intramolecular context, but could occur in multichain interactions. However, we agree with the Reviewer that this is an important issue and have deepened the analysis of this phenomenon in the revised manuscript.

First, our analysis shows that the observed asymmetry between intra- and intermolecular contexts is statistically significant and is likely not a consequence of limited sampling (pp. 10-11, Figure 3—figure supplement 1B-C). Moreover, the observed symmetry breaking is in line with the recent studies by Bremer et al. (https://doi.org/10.1038/s41557-021-00840-w) and Martin et al. (https://doi.org/10.1126/science.aaw8653), which have delineated the key requirements for the symmetry between single-chain and collective phase behavior to hold. Specifically, we have compared in detail the sequence composition of Lge11-80 with that of A1-LCD variants studied by Bremer et al. When it comes to aromatic composition, Lge1 is most similar to the -12F+12Y mutant of A1-LCD, and by this token, i.e. the high frequency of stickers tyrosines, should exhibit a strong coupling between single-chain and phase behavior. However, the net charge per residue (NCPR) in Lge11-80 of 0.075 is greater than that of A1-LCD (0.059) and this could contribute to the extent of decoupling, as suggested by Bremer et al. Moreover, Lge1 is extremely abundant in Arg (13.5 % as compared to 7.4 % in A1-LCD), and is in this sense most similar to the +7R A1-LCD mutant, which showed the greatest degree of decoupling between single-chain and phase behavior in Bremer et al., in agreement with what we see here. While these authors have demonstrated that NCPR is the primary determinant of decoupling in the case of A1-LCD mutants, their analysis showed that the nature of positive and negative residues involved also makes a significant difference. In particular, the significant excess of Arg residues, as context-dependent auxiliary stickers, could create the asymmetry between interactions that determine single-chain dimensions vs. collective phase behavior.

Furthermore, Martin et al. (https://doi.org/10.1126/science.aaw8653) have shown that an approximately uniform distribution of stickers along the sequence is required for the correspondence between the driving forces behind coil-to-globule transitions and phase separation to hold. We have analyzed the patterning of Tyr residues along the Lge11-80 sequence using Waro parameter used by Martin et al. (note that Tyr is the only aromatic in the Lge11-80 sequence). Interestingly, Lge11-80 exhibits a highly non-uniform patterning of Tyr residues, with Waro of the native Lge1 sequence (0.47) falling in the middle of the distribution for its shuffled variants (p=0.57). This is in contrast to the highly patterned sequences such as that of A1-LCD with p>0.99. Taken together, in addition to the relatively high NCPR, symmetry breaking in the case of Lge11-80 could be a consequence of its complex sequence composition, including both the non-uniform patterning of tyrosines and a high abundance of arginines. Provided that our simulations are long enough to provide an equilibrium picture and are on the length-scale of a single protein not strongly influenced by finite-size effects (these potential artifacts cannot be discounted), they actually can be seen as a demonstration of such symmetry breaking in a heteropolymer.

Furthermore, analysis of pairwise contacts suggests that intra- and intermolecular interactions rely on a similar pool of contacts by amino-acid type, but differ significantly if one analyzes specific sequence location of the interacting residues involved (Figure 2—figure supplement 1B and C). For example, one observes a high correlation between the frequencies of different contacts by amino-acid type when comparing intramolecular contacts in single-copy simulations and intermolecular contacts in 24-copy simulations (Figure 3—figure supplement 1B). This correlation is completely lost (Figure 3—figure supplement 1C) if one analyzes position-resolved statistics (2D pairwise contacts maps) or statistically defined interaction modes (Figure 3A, and Figure 3—figure supplement 1A). For example, although Tyr-Tyr interactions dominate in both cases, in single-copy simulations of WT Lge11-80 the C-terminal Tyr80 barely participates in any intramolecular interactions with other residues (Figure 3—figure supplement 1A), while in 24-copy simulations it is one of the most intermolecularly interactive residues (Figure 3). In other words, while the symmetry between intra- and intermolecular interactions can be observed at the level of pairwise contact types (similar type contact used for both), the distribution of these contacts along the peptide sequence is clearly different in the two cases. Finally, it should be mentioned that the parallel between single-copy and phase behavior in both homopolymers and heteropolymers is observed primarily at the level of thermodynamic variables such as LLPS critical temperature (Tc), coil-to-globule transition temperature (Tq) or the Boyle temperature (TB). It is possible that the noted correspondence extends primarily to such and similar thermodynamic variables, while and more structural, topological features of the globule in the single-molecule case and the network in the collective phase case remain uncoupled.

Interestingly, the core of intramolecular interactions observed for a single molecule at infinite dilution and in the crowded context remain approximately the same as reflected in the high correlation between intramolecular modes obtained in single and multichain simulations. Namely, proteins keep core self-contacts and establish new ones with neighbors, but do not donate everything to the intermolecular network losing “self-identity”, as in homopolymer melts. Similar effects have also been observed elsewhere: https://doi.org/10.1073/pnas.2000223117, https://doi.org/10.1073/pnas.1804177115.

Similarly, how were the errors of the radius of gyration for WT, R>K and Y>A mutants calculated? Is the Rg for WT significantly smaller than the values for the two mutants? And are the differences in Rg between single-copy and multi-copy simulations statistically significant? I am asking since converging the Rg of IDPs of this length in all-atom MD is not easy.

The errors for Rg values correspond to the standard deviations of the underlying distributions and are reported in Figure 4A and B, together with the corresponding means and an assessment of statistical significance of the difference. In particular, the character of the distributions (especially, for 24-copy systems) also suggests significant differences. In order to deepen this part in the revised version, we have added a new supplementary table (Supplementary File 1 “Technical summary) where we have included the average values of Rg together with the standard deviations for all modeled systems. Due to distributions being non-Gaussian, we have estimated the significance of the differences in Rgs between single-copy and multicopy simulations, as well as WT and mutants, using Wilcoxon rank sum test with continuity correction, with the resulting p-values < 2.2 10-16 for all cases.

p. 12, line 251: Has the MIST formalism been validated for IDPs; if so please provide a reference.

In the present work, we have evaluated the configurational entropy using a mutual information expansion approach with maximum-spanning-tree (MIST) approximation in internal-coordinate (bond-angle-torsion) representation. The latter is particularly well-suited for the analysis of IDPs as it allows one to avoid a number of artifacts (e.g., due to fitting of disordered ensembles to the average structure) associated with the more widely-used Cartesian-coordinate-based quasi-harmonic approaches. In particular, the MIST approach was used previously for the analysis of disordered protein ensembles (https://doi.org/10.1021/acs.jctc.8b00100). Here it should also be noted that, since intramolecular couplings are in general lower in IDPs, this makes them even better suited for MIST as compared to globular proteins. We have highlighted these points on p. 13 of the revision.

p. 5, line 105, p. 16 line 334 and p. 18 line 283: It is not completely clear what the predictions are and what/which experiments they are compared to. On p. 16, exactly what does the analytical model predict? As far as I understand, the results from the MD simulations are input to the model, but I am probably missing something. Which concrete and testable predictions does the model enable?

A key contribution of the present work is the development of a quantitative model that treats the spatial organization of a biomolecular condensate across scales using two key properties of individual polymer chains in the condensate - their average valency and compactness. The main predictions of the model concern the presence of a particular scaling of condensate mass with its radius, M(R), as captured by the fractal dimension, and the consequences this has on condensate morphology across scales. In the present manuscript, we have taken the first steps in testing these predictions in four different contexts. First, we could show that the MD simulations indeed match the predictions of fractal scaling for the three smallest clusters, which relates to the discussion on p. 16 that the Reviewer refers to. Here, it is important to understand that MD simulations in the first instance just give the average valency and compactness of individual chains in the dense phase. These values are then input into the fractal scaling formalism, which is conceptually fully independent from MD simulations, to obtain the dependence of condensate mass on its radius, M(R), at any desired length scale. The analysis presented in Figure 5—figure supplement 1B and discussed on p. 16 shows that the predictions of fractal scaling for the first three smallest clusters indeed correspond to what is seen in MD. This is a non-trivial correspondence and can be taken as direct evidence that fractal organization is present even at the shortest scale, i.e. at the level of MD simulation boxes.

Second, the model was used to reconstruct the spatial organization of clusters of arbitrary size at the atomistic level (Figure 5A and B, Videos 4, 5, and 6), enabling a structural understanding of the organization of condensate interior. One direct practical application of such understanding concerns the nature of cavity sizes and interpretation of dextran partitioning experiments (p. 20). Moreover, as pointed above, differences in morphology of protein clusters propagate across scales, and can be qualitatively characterized by the analysis of microscopic images (see also discussion above). In particular, the model correctly predicts the difference in the behavior of WT and R>K as opposed to Y>A variants, solely based on the predicted fractal dimension they exhibit. Ultimately, however, static light scattering experiments would give the best possibility to test the model directly and will be the topic of our future work. In particular, the fractal formalism predicts significant regions of linear behavior in such curves in log-log representation, while the fractal dimension df, provides a quantitative point of comparison between theoretical predictions and experimental measurements (Figure 5C). These points have been further discussed on p. 21 of the revised manuscript.

p. 19, lines 408-411: The authors find that when building clusters of Y>A from the simulations they find filamentous structures that they suggest explain the aggregation of the Y>A variant at high concentrations. While that sounds like an intriguing suggestion, it would be useful with a bit more detail about the robustness of this observation. For example, the simulations of Y>A appear similar to that of R>K; are the differences in topology really significantly different?

Fractal dimension, dF, is the key parameter that defines self-similar organization of differently sized protein clusters according to the fractal model. Consequently, the difference in morphology between R>K and Y>A mutants is reflected in different values of dF for the two. In particular, with a dF of 1.63, the Y>A mutant is predicted to form low-dimensional clusters, straddling the range between a linear (1-dimensional) and a planar (2-dimensional object), unlike WT and R>K variants, which both exhibit dF values greater than 2. The qualitative behavior of the three variants, whereby WT and R>K result in spherical condensates and Y>A does not, is consistent with this. Notably, we have observed sporadic precipitates at high protein concentration in the Y>A mutant, which may be consistent with the predictions of the fractal model. However, the material properties and possible influence of sample impurities in the Y>A case at high concentrations remain unclear. Moreover, the sporadic nature of Y>A precipitates prevents an adequate statistical analysis. Hence, in the revised manuscript we refrain from commenting on these infrequently observed precipitates.

Regarding MD simulations, the morphological differences between Y>A and R>K proteins can already be seen at the level of individual proteins in multicopy simulations, highlighted by the significantly different distribution of Rg (Figure 4B). This distribution in the case of Y>A has a prominently long tail, which indicates the possibility of adopting significantly more elongated configurations. Due to the self-similarity principle, such differences in morphology may propagate across length scales. Importantly, a recent publication included the experimental study of the possibility of IDRs to form low dimensional fractal systems upon disruption of the LLPS tendency by polyalanine insertion in synthetic elastin-like polypeptides (Roberts et al., Nature Materials, 2018).

Finally, I would suggest that the authors make their code and data available in electronic format.

All sharable data has been made available as part of the article package. Due to the heterogeneous character of our analysis, we do not have a single master code to be shared, but rather a collection of different scripts in combination with different software packages as indicated in the Methods section of the manuscript (GROMACS, MATLAB, R, FracVAL).

Reviewer #2 (Public Review):

A key aspect of the work is to use the simulations to explain differences between (i) dilute and dense phases and (ii) wild-type and mutant variants. Here, it would be important with a clearer analysis of convergence and errors to quantify which differences are significant.

It would also be useful with a clearer description of how the analytical model is predictive, of which properties, and how they have been/can be validated. Which measurable quantities does the model predict?

In addition to these overall questions, a number of more specific suggestions follow below.

Major:

p. 7, line 120 (Fig. S1B)<br /> The proteins do not appear particularly pure based on the presented SDS PAGE analysis. How pure is the protein estimated to be, and is the presence of the other bands expected to affect e.g. the data presented in Fig. 1?

p. 7 & 8, lines 138-159:<br /> Has the method and energy function used to calculate the interact potential been validated by comparison to experiments, including studying the effect of varying the solvent? I see the computed error bars are very small, but am more interested in the average error when comparing to experiments. The numbers in water appear different from those e.g. reported by Krainer et al (https://doi.org/10.1038/s41467-021-21181-9), though the latter are also not immediately compared to experiments. Thus, it would be useful to know how much to trust these numbers.

p. 8, lines 149-154:<br /> Following up on the above, the authors also write "Importantly, only in the latter case are the R-Y interactions slightly more favorable than the K-Y ones (Figure S1C). While this can potentially contribute to increasing of Csat for the R>K mutant as compared to WT, the estimated thermodynamic effect is not too strong, especially if one considers that these interactions take place in an environment with largely water-like polarity. Therefore, the effect of R>K substitution on LLPS should be further explored in the context of protein-protein interactions."<br /> In the absence of estimates of the accuracy of the predictions, these sentences are somewhat unclear. Also, it is unclear what the authors mean by that the effect of R>K should be studied; there are already several examples of this (https://doi.org/10.1016/j.cell.2018.06.006 [already cited], https://doi.org/10.1038/s41557-021-00840-w & https://doi.org/10.1073/pnas.2000223117 come to mind, but there are likely more).

p. 8, lines 161-162:<br /> The authors perform MD simulations of Lge1 and variants using 24 copies and a box that gives them protein concentrations "in the mM concentration range". I realize that there's a concern about what is computationally feasible, but it would be important with an argument for this choice. Why is 24 expected to be enough to represent a condensate (I expect that there could be substantial finite-size effects)? What is the exact protein concentration in the simulations of the 24 chains [and of the 1-chain simulations]? How does this protein concentration compare to that in the condensates? The authors performed simulations in the NPT ensemble; how stable were the box dimensions?

Also, did the authors include the Strep- and His-tags in the simulations? If not, why not?

Throughout:<br /> One of my major concerns about this work is the general lack of analysis of convergence of the simulations. The authors must present some solid analysis of which results are robust given the relatively short simulations and potential for bias from the chosen starting structures.

As an example, on p. 8 the authors discuss a potential asymmetry between the interactions found in the dilute (single-copy) and dense (24-mer) phases. These observations are somewhat in contrast to other observations in the field, namely that it is the same interactions that drive compaction of monomers as those that drive condensate formation.

Obviously, both the results in the literature and those presented here could be true. But in order to substantiate the statements made here, the authors should show some substantial statistical analyses to make it clear which differences are robust.

The above holds for all parts of the computational/simulation work (e.g. other aspects of Fig. 2)

Similarly, how were the errors of the radius of gyration for WT, R>K and Y>A mutants calculated? Is the Rg for WT significantly smaller than the values for the two mutants? And are the differences in Rg between single-copy and multi-copy simulations statistically significant? I am asking since converging the Rg of IDPs of this length in all-atom MD is not easy.

p. 12, line 251:<br /> Has the MIST formalism been validated for IDPs; if so please provide a reference.

p. 5, line 105, p. 16 line 334 and p. 18 line 283:<br /> It is not completely clear what the predictions are and what/which experiments they are compared to. On p. 16, exactly what does the analytical model predict? As far as I understand, the results from the MD simulations are input to the model, but I am probably missing something.<br /> Which concrete and testable predictions does the model enable?

p. 19, lines 408-411:<br /> The authors find that when building clusters of Y>A from the simulations they find filamentous structures that they suggest explain the aggregation of the Y>A variant at high concentrations. While that sounds like an intriguing suggestion, it would be useful with a bit more detail about the robustness of this observation. For example, the simulations of Y>A appear similar to that of R>K; are the differences in topology really significantly different?

Finally, I would suggest that the authors make their code and data available in electronic format.

should this not be in code format too?

P208: Variables that exist at the beginning and executable code are loaded into the heap.

acknowledging cultural differences

It would be great if you could make the polished assemblies or assembled contigs analyzed in this study available since it takes quite a bit of work to get to that point

Using Silktide's Screen Reader Simulator for Chrome, I tested the webpage. The featured listing are especially confusing to understand using the screen reader because although there are only 10 featured listings, the screen reader repeats the featured listings three times (reads all 10 listings, then reads them again, then again). Screen readers are used by people who are blind or visually impaired, so the webpage has to be made in a way that makes it easy for a screen reader to navigate through, which is not the case with this webpage. This violates the "Robust" principal of the Web Content Accessibility Guidelines.

Solution: Look through the code to see if there is any repeated code that is causing the screen reader to read the featured listings multiple times. If there is no repeated code, then it may be a problem with how the featured listings section has been coded, so the format of the featured listings section may need to change. After making the appropriate changes, test the webpage using a screen reader before publishing to make sure that the webpage is easily navigable by a screen reader.

w

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

Learn more at Review Commons

Reply to the reviewers

Reviewer #1 (Evidence, reproducibility and clarity (Required)):

The manuscript presents a detailed numerical model of blood flow in a region of the zebrafish vasculature.

The results section is quite intense and detailed. it is difficult to understand what the authors are after. I think a rewrite would beneficial. The authors present simulations for a wild type and a couple of phenotypes. For each of these they speculate on the possible adaptation mechanism leading to the discussed phenotype, as preservation of constant wall shear stress. However, the comparison between experiments and numerical simulations is really elusive as the conclusions on those mechanisms. Overall we suggest a rewrite with clearer organisation in a way that the reader is not overflown with useless details.

We thank the reviewer for the advice on the general writing standard and data organization. We have reanalyzed experiment data and interpreted the findings more conservatively for application into the simulation models. As a result, some conclusions to the results sections have changed. Accordingly, we have done a major revision of the entire Results, Discussion and Models and Methods sections in the paper to articulate these reinterpretations while removing superfluous details that may obfuscate the data.

It is not always clear what info of the experiments are used in the simulations on top of the anatomy. Our understanding is that the pressure boundary conditions are set to match the red blood cel velocity observed in experiments. Is this always the case for the three phenotypes and which vessels ?

We thank the reviewer for the question. Only WT and Marcksl1 KO have been matched for peak velocities in the CA, CV and ISVs between experiments and simulations. WT results were compared to both the experimental reference of 27 embryos in Table 3 and also to the current experiment pool of WT (5 embryos) in Table 6. Marcksl1 KO simulation models 1, 2 and 3 were compared against the average level seen in the low and moderate perfusion Marcksl1 KO phenotypes (8 embryos) from the experiment (Table 5 and Table 6). Additionally, we also have represented the similar level of RBC hematocrit in the CA for WT model to WT experiment data from the reference cited in Table 3.

In addition to the velocity comparisons, we now use the experimentally observed trend of decreased flow rate in the CA of Marcksl1 KO experiment data to assess the model boundary conditions amongst Marcksl1 KO models 1, 2 and 3 that best reflect the experimental observations:

Page 11, lines 1 to 20

The Marcksl1 OE cannot be matched because we do not have the experiment data for that, the same goes for PlxnD1 where we have no experiment flow data. These two networks represent more conceptual discussions, particularly in PlxnD1 case where we have explicitly stated in the new discussion section:

Page 15, lines 24 to 34

There are about 7 inlets and outlets where to impose pressure boundary conditions. Can the author comment on the uniqueness of this problem?

Can different combination of pressure boundary condition leading to the same result ? In how many points/vessels is the measured velocity matched ?

We thank the reviewer on this insightful concern. Indeed, the uniqueness of flow and pressure field can be a problem without careful consideration. We have tried to address this to some extent, because CA, CV are connected by the ISV and DLAV network, to match flow velocity in all regions, the pressure distribution ought to be unique to the particular setting we employed.

As shown in table 3, average systolic peak flow velocities in the entire CA and CV encompassing the 5 ISV segment domain is matched between the simulation and the population-averaged experimental data from the experimental reference (27 fish sampled in the cited reference) for the same regions in WT network. Average systolic peak flow velocities for the 10 ISVs in the simulation were matched against WT experiment population-averaged systolic peak flow velocities in arterial and venous ISVs in the same caudal region.

Additionally, we also compared the flow velocities to the experiment conducted within this study (5 WT, and embryos). This comparison data is shown in Table 6. Admittedly the discrepancy was large for CV and ISVs regions likely due to a smaller data set sampled in this study and biological variations that happen from one experiment to another. We have acknowledged this deficiency in the revised discussion section:

Page 15, lines 3 to 9

The argument that similar beating frequency in the WT and GATA1 MO suggest pressure does not change is not clear. If the heart was a volumetric pump it would impose the same flow rate, not the same pressure. It would be more useful to measure the cardiac output in terms of flow rate in the Dorsal Aorta. Previous measurements by Vermot suggested the latter would not change much in gata1 MO. It could be that the cardiac output is the same but the vasculature network is different in a way that the shear stress remain the same. It does not look like this was checked by the authors.

We thank the reviewer for this insight. In accordance with the reviewer’s suspicion, we have estimated the flow rates in the CA of gata1 MO injected embryos and found the level to be similar to WT. This supports the reviewer’s opinion that the heart rate similarity indicates cardiac output similarity and not arterial pressure similarity as we previously put forward. Furthermore, we have checked that the gata1 morphants do in fact present reduced ISV diameters. In light of this reinterpretation, we performed an additional zero hematocrit model (model 3 in section 2.1). We have consequently rewritten the entire section on how RBC hematocrit modulates hemodynamics in a microvascular network:

Page 6, line 18 to page 8 line 10.

Additionaly, it would be useful to provide an effective viscosity for the different vessels, and an effective hydraulic impedance relating DP and Q to interpret the results.

We have followed the reviewer’s advice and have analyzed for vessel hydraulic impedance and effective viscosity in all the network models presented. This is included in the main figures and discussion. The vessel impedances are discussed for the various models in these following parts of the manuscript:

Page 9, lines 20 to 29

Page 11, lines 28 to 30

Page 12, line 1 to page 13 line 10

Is the hydraulic impedance of the vessels kept constant in the smooth-geometry model? This needs clarification

The SGM diameters have been determined based on geometric averages and not impedance equivalency. The reason why we did this is because the impedance will not be known until the CFD is performed for the WT network. This is because without a pressure distribution (which cannot be determined experimentally) we cannot calculate vessel impedance since only flow can be measured and both flow and pressure are requirements to impedance calculation. Our intention with the SGM is to highlight how geometric averaging of morphological characteristics lead to incorrect flow and stress predictions. However, we understand the reviewer’s sensibility and have revised the entire section of the SGM results. We have now discussed three SGM models with varying degrees of geometry simplification. The SGM1 in the revised manuscript matches WT network impedance in the ISVs by including both axial variation in lumen diameter of the WT network and the elliptical fit representation of cross-sectional skewness seen in WT ISV lumens. SGM 2 has representation of axial variation but not luminal skewness and SGM3 has only geometric average similarity to WT ISVs. The new findings and discussion can be found in the revised manuscript here:

Page 8, line 19 to page 9 line 36.

As mentioned by the authors they propose a very complex and time expensive simulation. However the results they report are kind of intuitive. Given the availability of the experimental results, would it be useful to use a simpler red blood cell model in the future, to make their simulation more practical? Or clarify when such demanding simulations can add something new?

We agree that the intuition feedback depends on the expertise of the investigator. The boundary condition selection is intuitive from the experimental findings and key data like pressures in the network cannot be measured. Furthermore, population-averaged flow data does not always match the flow-to-geometry situations that vary from sample to sample, thus demonstrated by the high margin of prediction discrepancy for flow velocities in table 6. We have discussed these challenges and our recommendations for improvement in the Discussion section:

Page 15, lines 3 to 9

Page 15, lines 35 to 40

Page 16, lines 12 to 15

On the topic of RBC model simplification, we agree with the reviewer that our work suggests the methodology would benefit from a further coarse-graining approach to the RBC phase. Accordingly, we discussed the possibility of using a low-dimensional RBC model already published in literature:

Page 14, lines 13 to 17

The authors should check their references as this is not the first time work has been done on the topic. Would be good to have a check in the work of Freund JB and colleagues, as well as Dickinson and colleagues and Franco and colleagues to discuss how the work compares. There may be interesting work in modelling cardiac flow forces in the embryo too.

Thank you for referring us to other publications that are related to our study. To our knowledge and after performing publication search on these authors, we find that although Dickinson and colleagues performed experiments to examine the effects of perturbed blood flow on vessel remodelling (Udan et al., 2013), they did not perform any numerical modelling to calculate hemodynamic forces such as WSS and luminal pressure. Instead, changes in vessel morphogenetic process were only correlated with blood flow velocity. In our study, we attempt to quantitatively correlate WSS and pressure distributions within a vascular network. Franco and colleagues (Bernabeu et al., 2014) developed PoINet to model haemodynamic forces in mouse retina model of angiogenesis. From what we understand, PoINet is different from our 3D CFD model by 1) not having red blood cells incorporated in their model and as such, the blood viscosity prediction is modelled using shear-rate dependent formulation and not through red blood cell hematocrit, 2) cross sections of blood vessels are assumed to be circular and therefore have no irregularity and 3) live imaging of blood flow is difficult in mouse retina therefore preventing accurate boundary conditions for the model.

We have included the reference to work of Franco and colleagues:

Page 14, line 28 to line 31

Page 9, lines 12 to 14

Freund JB indeed has had extensive work on RBC and cellular flow in microvessels. We have included a reference of his work in:

Page 14, lines 22 to 25.

Reviewer #1 (Significance (Required)):

The authors discuss the applicability of a detailed numerical model of blood flow in a region of the zebrafish vasculature.

We are not expert in the lattice boltzmann method used here, but the results are what it would be expected from a physical stand point, and together with the information from the method section, we do not have major concerns about the numerics.

Reviewer #2 (Evidence, reproducibility and clarity (Required)):

Summary: The authors report corroborating numerical-experimental studies on the relationship between morphological alterations (e.g. vessel lumen dilation/constriction, network mispatterning) and hemodynamical changes (e.g. variation in flow rate, pressure, wall shear stress) in the vascular network of zebrafish trunk circulation. Various physiological or pathological adaptation scenarios were proposed and tested, with a range of simulation and experiment models. Where I found it a solid piece of work supported by abundant data, certain aspects need to be clarified/enhanced to improve the scientific rigor and potential impact of the manuscript. Below are my detailed comments in the hope of helping the authors improve the manuscript's quality.

Major comments:

1. Cellular blood flow in vascular networks has been extensively studied in recent years by existing computational models (some of which were published open-source) with similar methods and features to the one proposed by the present work. Can the authors be more explicit about the original contributions of the current model, and provide evidence accordingly (e.g. Github repository or code resources)

The RBC model is essentially the model developed by Fedosov and colleagues (Fedosov, et al., 2010). Likewise, the LBM solver for fluid flow calculation is not. Following the reviewer’s advice, we have removed the details of these non-novel aspects of the methodology and placed them in sections E and F of supplementary material instead. The new Models and methods now show condensed descriptions of the three numerical solvers used and the addition of a grid independence matrix discussion section:

Page 17, line 8 to page 20, line 33.

Crucial details for the simulation setup and model configuration are missing. What were the exact boundary conditions (e.g. inlet and outlet pressures) and initial conditions (e.g. feeding hematocrit of RBCs), and how the numerical-experimental validation process of "to match the velocities of various segments of the network by iteratively altering the pressure inputs ..." as stated on page 13 (lines 1-2) was performed for simulations in this work?

We apologize for the vagueness of our description on how numerical to experimental validations were performed. As replied to reviewer 1 for a similar clarification, we have indicated in Table 3 how average systolic peak flow velocities in the entire CA and CV encompassing the 5 ISV segment domain were matched between the simulation and the population-averaged experimental data for the same regions in WT network. Average systolic peak flow velocities for the 10 ISVs in the simulation were matched against WT experiment population-averaged systolic peak flow velocities in arterial and venous ISVs in the same caudal region.

With regards to what iterative alterations of pressure inputs mean, we monitored the average systolic peak velocities and hematocrit levels in CA, CV and ISVs in intervals of 5 cardiac cycle intervals before manually correcting the pressure input levels to better match average systolic peak velocities in these vessels from the experiment averages. Since we are using population averaged flow data, we do not expect their levels to match the levels in a particular fish-specific geometry, the degree of discrepancy between experiment averages and the model predictions of systolic velocities can be large (Table 6). Admittedly, this is one of the weaknesses of our approach and this limitation is stated in the Discussion section:.

Page 15, lines 3 to 9

As RBC flow typically requires roughly 5 cardiac cycles of flow to reach flow development this process of iterative correction typically takes place over 10 to 20 cardiac cycles. We understand that validation may be a subject of keen interest to readers, hence we have now briefly described the solution initialization and flow development protocol in our modeling approach here:

Page 6, lines 5 to 8

What lattice resolution was used for the flow solver and was the RBC membrane mesh chosen accordingly? Were there any sensitivity analysis (regarding pressure input) or grid-independence study (regarding lattice resolution)

We originally decided on the grid (∆X) and time (∆T) discretization resolutions (0.5 µm and 0.5 µs) based on the acceptable computing turnaround time for each model within our scale of resources. We have now included a section on the grid independence matrix in Models and Methods:

Page 19, line 20 to page 20, line 33

Details of the statistical tests (type of tests used, assessment of data normality, sample size etc.) should be given in the figure caption where applicable (e.g. Fig. 3C, Figs. 7-9).

We apologize for the lack of clarity. All statistical tests used have now been mentioned at least once in each section of results and also in Figure captions wherever significance bars are displayed.

The regression models should also be used with caution, e.g. in Fig. 4B, why should data from two different fish types, namely Gata1 MO and WT, be grouped to fit a linear regression model?

We understand the reviewer’s concern that two population data sets should not be carelessly pooled together for regression analysis without adequate justification. In this case we are utilizing gata1 morpholino injection as a means to alter hematocrit level. There is no reported side-effect as to the best of our knowledge, only hematocrit and possibly hemodynamics and morphological response related to hematocrit level should be affected. Moreover, we have mislabelled the companion set to the gata1 morpholino as WT, the data is in fact data from control morphants and not WT. This change has been applied to Fig. 3 graphs and Table 4 and results section:

Page 7, lines 3 to 16

Finally, as we want to generate a continuum range of varying hematocrit for embryos of the same developmental age. In this regard, we think that within the scope of our intentions and well-accepted usage of gata1 morpholino as a hematocrit reduction protocol it is reasonable to pool the two data sets together for regression analysis.

4.I found the data presented in Fig. 7 insufficient to confidently exclude the numerical models 2, 3 but favor model 1 as the adaptation scenario for the Marcksl1KO case. The first question is, how are the threshold RBC perfusion levels determined to categorize the experimented Marcksl1KO fishes into four groups, i.e. "high", "moderate", "low", "zero"? The authors also need to justify why the "high", "moderate", "low" groups can be mapped to the three modelling scenarios (namely models 1, 2, 3) is it just because "a qualitative match with the experimental trend of ascending CA blood velocity" (Fig. 7F)?

We thank the reviewer for his interpretation of our results. Firstly, we apologize for generating the confusion but we are not trying to map simulation models 1, 2 and 3 to high moderate and low groups respectively in Fig. 7. The high, moderate and low categorizations of experimental Marcksl1 KO phenotypes are based on RBC flux levels observed experimentally. We are trying to ascertain which Marcksl KO phenotype the models 1, 2 and 3 fit, if they do fit the experiment trend at all.

Second, in Fig. 7C, it is shown that no significant difference exists between the "high" group and the WT in their average ISV diameter, then what is defining that group as Marcksl1KO type ?

We apologize for the confusion generated. High flow phenotype is similar to WT flow, the diameter is also similar to WT. In Marcksl1 KO mutants we don’t always see clear phenotyping and often a range is presented from mutant to mutant. Hence the high group is essentially morphometrically and hemodynamically similar to WT, the only reason we know it is a mutant because we have genotyped the zebrafish (marcksl1a-/-;marcksl1b1-/-).

Third, a central assumption here is using heart rate as a measure of the pressure drop in different fish individuals (Fig. 7D). Can't two fishes with similar heart rate have distinct pressure drops in the trunk due to difference in network architecture and topology, vice versa?

We agree with the reviewer’s opinion and now feel that our initial proposition was naïve. After addressing the interpretation of heart rate similarity in the gata1 morphants with more convincing CA flow rate estimations, we now believe that heart rates might not be useful indicators of flow or pressure levels in the network. Instead, cardiac output in the form of CA flow rate as reviewer 1 has suggested might be a better indicator. As the reanalysis has dismantled the earlier interpretation, and found that based on the flow rate estimation for the CA, Marcksl1 KO networks have reduced blood flow rates in the CA.

Page 11, lines 9 to 20

This finding has been incorporated into the consideration of flow adaptation scenarios predicted by the simulation models accordingly in the revised manuscript:

Page 12, line 1 to page 13, line 10

Fourth, the authors should explain why a power-law fit (note that it is not "exponential" as stated on page 10, line 3) should be adopted for the regression analysis in Figs. 7E-v,vi (a useful reference may be Joseph et al. eLife 2019: 10.7554/eLife.45077).

We thank the reviewer for the useful reference and the careless mislabeling of regression curve used. This figure has been redone and a linear regression is instead used that does not attempt to imply any physical law for a power or exponential fitting.

Change made: Fig. 7C

Minor comments:

1. The state of art of cell-resolved blood flow models employed to simulate microcirculatory hemodynamics is not accurately described in the introduction (page 4). More recent works should be cited and critically reviewed to present a fair view on the novelty of the computational model developed herein.

We apologize that the models were mentioned in a passing manner. However ,the need for brevity in introduction somewhat limits their expansion. We have instead gave more direct discussion on similar studies and their relevance to our present work in the Discussion section:

Page 14, lines 13 to 31

It is unclear what "realistic representation of local topologies in the network" (page 7, lines 28-31) means as a claim of novelty. If it means vessel "diameter variation", this geometric feature has been modeled by the works the author referenced (namely Roustaei et al. 2022, Zhou et al. 2021). If it means something else, for example, unsmooth or non-circular vessel surface (or "irregularity of the local endothelium surface" as mentioned on page 5, line 2), then strangely the effects of such features are actually not described in the manuscript.

We apologize for not meeting the expectation of novelty as claimed. We see value in the SGM study matrix have now generated data on three SGM scenarios. The SGM1 in the revised manuscript matches WT network impedance in the ISVs by including both axial variation in lumen diameter of the WT network and the elliptical fit representation of cross-sectional skewness seen in WT ISV lumens. SGM 2 has representation of axial variation but not luminal skewness and SGM3 has only geometric average similarity to WT ISVs. Essentially the comparison between SGM1 and SGM2 highlights the role of luminal cross-sectional shape skewness while SGM2 to SGM3 highlights the role of axial variation in luminal diameter. With this new SGM data set, we think we can better qualify the aspiration of demonstrating how vessel shape “irregularities” can alter network hemodynamics. The new findings and discussion can be found in the revised manuscript here:

Page 8, line 19 to page 9 line 36.

Why should Fig. 8 contain data from Marcksl1KO model 2? The scenario underlying model 2 was rejected earlier in the manuscript (see point 6 above), and the Marcksl1KO model 2 data are not mentioned in the text when describing the results of Fig. 8, either.

We have reanalyzed the experiment trend and rewritten the outcome of this results section. In summary, both models 1 and model 2 meet the trend of flow rate reduction (with respect to WT levels) in the CA observed in the experiment. Hence, model 2 inclusion is relevant to the WSS analysis. The changes pertaining to this can be found here:

Page 11, line 9 to page 13 line 10.

It is a dense article with loads of data, which is an advantage but only if appropriately streamlined. More subheadings should be considered, especially for section 2.3 (for which the current subsections appear mistaken, 2.3.1 followed by 2.4.2) The manuscript could also benefit from restructuring through optimal combination of simulation visualizations and quantitative analyses. For example, in Fig. 6, not all simulation snapshots are needed here (it is difficult to visually compare the changes between different cases), whereas some quantification in the form of histograms or boxplots will be handy for the readers to note the variation of WSS magnitudes and ranges.

Thank you for the advice, we removed the unnecessary graphical plots and refer to simulation videos in supplementary data instead for such cases. The bad indexing of results subsections has been fixed, while new subsections have been made for better directional narrative to the paper. These changes are colored in red throughout the revised results section:

Page 4, line 37 to page 13 line 39

Related to point 8, the authors could also consider integrating or synthesizing the analyses for individual aISVs and vISVs presented in various figures. Current descriptions for the ISV data appear scattered with frequent exceptions to the summarized trends or relationships. Some minor formatting issues should also be addressed, e.g. the confusing color codes in Figs. 9D-i, E-i.

Thank you for the advice, we have now pooled aISVs together into one group and vISVs into another, instead of discussing data trends on each of the 10 ISVs.

The mispattening case presented in the end of the results section (section "2.4.2") is interesting but appears loosely connected to the preceding contents. Also, it seems not even mentioned in the discussion section.

We agree that the mispatterning case has been only tangentially relevant to the rest of the manuscript. We have linked the topic thematically by network alterations transforming network flows. It is also now included in the discussion section here:

Page 15, lines 30 to 34

Finally, apart from the effect of topological features on local blood flow, the authors should consider the global flow redistribution arising from the network structure (useful refs. Include Chang et al. PLOS Computational Biology 2017: 10.1371/journal.pcbi.1005892; Meigel et al. Physical Review Letters 2019: 10.1103/PhysRevLett.123.228103; Schmid et al. eLife 2021: 10.7554/eLife.60208).

Thank you for the additional references. These are solid pieces of work that have been added to the discussion here:

Page 16, lines 3 to 10

**Referees cross-commenting**

This review report resonates with mine from an experimental perspective and I agree with all points made regarding issues of the current manuscript that the authors need to address with a revised version.

Reviewer #2 (Significance (Required)):

Significance: The particular merit of the work lies in its comprehensiveness of design and abundance of data, which will be of great interest to both the computational and experimental communities in this research field. However, some crucial details (especially with respect to the modelling aspects) are missing, thus hampering the scientific rigor and potential impact of the work. Furthermore, certain justifying statements appear speculative and inconclusive to explain the obtained data, especially regarding the effect of boundary conditions and systemic parameters. The citation of references (some not cited, some cited already but not properly discussed) also needs to be enhanced with engaging discussions to better bridge the findings of the current work (e.g. RBC partitioning in vascular network, effect of WSS on vasculature morphogenesis) with recent works on this research topic.

References

Fedosov DA, Caswell B, Karniadakis GE. 2010. A Multiscale Red Blood Cell Model with Accurate Mechanics, Rheology, and Dynamics. Biophys J 98:2215–2225. doi:10.1016/j.bpj.2010.02.002

Freund JB, Goetz JG, Hill KL, Vermot J. 2012. Fluid flows and forces in development: functions, features and biophysical principles. Dev Camb Engl 139:1229–45. doi:10.1242/dev.073593

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

Learn more at Review Commons

Referee #2

Evidence, reproducibility and clarity

Summary: The authors report corroborating numerical-experimental studies on the relationship between morphological alterations (e.g. vessel lumen dilation/constriction, network mispatterning) and hemodynamical changes (e.g. variation in flow rate, pressure, wall shear stress) in the vascular network of zebrafish trunk circulation. Various physiological or pathological adaptation scenarios were proposed and tested, with a range of simulation and experiment models. Where I found it a solid piece of work supported by abundant data, certain aspects need to be clarified/enhanced to improve the scientific rigor and potential impact of the manuscript. Below are my detailed comments in the hope of helping the authors improve the manuscript's quality.

Major comments:

1. Cellular blood flow in vascular networks has been extensively studied in recent years by existing computational models (some of which were published open-source) with similar methods and features to the one proposed by the present work. Can the authors be more explicit about the original contributions of the current model, and provide evidence accordingly (e.g. Github repository or code resources)?
2. Crucial details for the simulation setup and model configuration are missing. What were the exact boundary conditions (e.g. inlet and outlet pressures) and initial conditions (e.g. feeding hematocrit of RBCs), and how the numerical-experimental validation process of "to match the velocities of various segments of the network by iteratively altering the pressure inputs ..." as stated on page 13 (lines 1-2) was performed for simulations in this work? What lattice resolution was used for the flow solver and was the RBC membrane mesh chosen accordingly? Were there any sensitivity analysis (regarding pressure input) or grid-independence study (regarding lattice resolution)?
3. Details of the statistical tests (type of tests used, assessment of data normality, sample size etc.) should be given in the figure caption where applicable (e.g. Fig. 3C, Figs. 7-9). The regression models should also be used with caution, e.g. in Fig. 4B, why should data from two different fish types, namely Gata1 MO and WT, be grouped to fit a linear regression model? 4.I found the data presented in Fig. 7 insufficient to confidently exclude the numerical models 2, 3 but favor model 1 as the adaptation scenario for the Marcksl1KO case. The first question is, how are the threshold RBC perfusion levels determined to categorize the experimented Marcksl1KO fishes into four groups, i.e. "high", "moderate", "low", "zero"? The authors also need to justify why the "high", "moderate", "low" groups can be mapped to the three modelling scenarios (namely models 1, 2, 3); is it just because "a qualitative match with the experimental trend of ascending CA blood velocity" (Fig. 7F)? Second, in Fig. 7C, it is shown that no significant difference exists between the "high" group and the WT in their average ISV diameter, then what is defining that group as Marcksl1KO type? Third, a central assumption here is using heart rate as a measure of the pressure drop in different fish individuals (Fig. 7D). Can't two fishes with similar heart rate have distinct pressure drops in the trunk due to difference in network architecture and topology, vice versa? Fourth, the authors should explain why a power-law fit (note that it is not "exponential" as stated on page 10, line 3) should be adopted for the regression analysis in Figs. 7E-v,vi (a useful reference may be Joseph et al. eLife 2019: 10.7554/eLife.45077).

Minor comments:

1. The state of art of cell-resolved blood flow models employed to simulate microcirculatory hemodynamics is not accurately described in the introduction (page 4). More recent works should be cited and critically reviewed to present a fair view on the novelty of the computational model developed herein.
2. It is unclear what "realistic representation of local topologies in the network" (page 7, lines 28-31) means as a claim of novelty. If it means vessel "diameter variation", this geometric feature has been modeled by the works the author referenced (namely Roustaei et al. 2022, Zhou et al. 2021). If it means something else, for example, unsmooth or non-circular vessel surface (or "irregularity of the local endothelium surface" as mentioned on page 5, line 2), then strangely the effects of such features are actually not described in the manuscript.
3. Why should Fig. 8 contain data from Marcksl1KO model 2? The scenario underlying model 2 was rejected earlier in the manuscript (see point 6 above), and the Marcksl1KO model 2 data are not mentioned in the text when describing the results of Fig. 8, either.
4. It is a dense article with loads of data, which is an advantage but only if appropriately streamlined. More subheadings should be considered, especially for section 2.3 (for which the current subsections appear mistaken, 2.3.1 followed by 2.4.2). The manuscript could also benefit from restructuring through optimal combination of simulation visualizations and quantitative analyses. For example, in Fig. 6, not all simulation snapshots are needed here (it is difficult to visually compare the changes between different cases), whereas some quantification in the form of histograms or boxplots will be handy for the readers to note the variation of WSS magnitudes and ranges.
5. Related to point 8, the authors could also consider integrating or synthesizing the analyses for individual aISVs and vISVs presented in various figures. Current descriptions for the ISV data appear scattered with frequent exceptions to the summarized trends or relationships. Some minor formatting issues should also be addressed, e.g. the confusing color codes in Figs. 9D-i, E-i.
6. The mispattening case presented in the end of the results section (section "2.4.2") is interesting but appears loosely connected to the preceding contents. Also, it seems not even mentioned in the discussion section.
7. Finally, apart from the effect of topological features on local blood flow, the authors should consider the global flow redistribution arising from the network structure (useful refs. Include Chang et al. PLOS Computational Biology 2017: 10.1371/journal.pcbi.1005892; Meigel et al. Physical Review Letters 2019: 10.1103/PhysRevLett.123.228103; Schmid et al. eLife 2021: 10.7554/eLife.60208).

Referees cross-commenting

This review report resonates with mine from an experimental perspective and I agree with all points made regarding issues of the current manuscript that the authors need to address with a revised version.

Significance

The particular merit of the work lies in its comprehensiveness of design and abundance of data, which will be of great interest to both the computational and experimental communities in this research field. However, some crucial details (especially with respect to the modelling aspects) are missing, thus hampering the scientific rigor and potential impact of the work. Furthermore, certain justifying statements appear speculative and inconclusive to explain the obtained data, especially regarding the effect of boundary conditions and systemic parameters. The citation of references (some not cited, some cited already but not properly discussed) also needs to be enhanced with engaging discussions to better bridge the findings of the current work (e.g. RBC partitioning in vascular network, effect of WSS on vasculature morphogenesis) with recent works on this research topic.

the problem is summed up in the domain name :stackify

https://hyp.is/6LMTSvCUEe2p6nOS-NOIwQ/stackify.com/dependency-injection/

存在一个可计算函数，用于计算一个可判定近似器的不精确性（witness）

ALL PHOTOS 1. It is not clear if there is Alt text for any of the photos on this page. (I tried looking in the HTML code and using a screen reader, but it was still unclear to me.) 1. Accessibility features are usually hidden in the code and only people with assistive technology will notice them. If websites could be 'certified' accessible people would not waste time searching websites that do not take their needs under consideration.

Reviewer #1 (Public Review):

A typical path from preprocessed data to findings in systems neuroscience often includes a set of analyses that often share common components. For example, an investigator might want to generate plots that relate one time series (e.g., a set of spike times) to another (measurements of a behavioral parameter such as pupil diameter or running speed). In most cases, each individual scientist writes their own code to carry out these analyses, and thus the same basic analysis is coded repeatedly. This is problematic for several reasons, including the waste of time, the potential for errors, and the greater difficulty inherent in sharing highly customized code.

This paper presents Pynapple, a python package that aims to address those problems.

Strengths:

The authors have identified a key need in the community - well-written analysis routines that carry out a core set of functions and can import data from multiple formats. In addition, they recognized that there are some common elements of many analyses, particularly those involving timeseries, and their object-oriented architecture takes advantage of those commonalities to simplify the overall analysis process.

The package is separated into a core set of applications and another with more advanced applications, with the goal of both providing a streamlined base for analyses and allowing for implementations/inclusion of more experimental approaches.

Weaknesses:

There are two main weaknesses of the paper in its present form.

First, the claims relating to the value of the library in everyday use are not demonstrated clearly. There are no comparisons of, for example, the number of lines of code required to carry out a specific analysis with and without Pynapple or Pynacollada. Similarly, the paper does not give the reader a good sense of how analyses are carried out and how the object-oriented architecture provides a simplified user interaction experience. This contrasts with their GitHub page and associated notebooks which do a better job of showing the package in action.

Second, the paper makes several claims about the values of object-oriented programming and the overall design strategy that are not entirely accurate. For example, object-oriented programming does not inherently reduce coding errors, although it can be part of good software engineering. Similarly, there is a claim that the design strategy "ensures stability" when it would be much more accurate to say that these strategies make it easier to maintain the stability of the code. And the authors state that the package has no dependencies, which is not true in the codebase. These and other claims are made without a clear definition of the properties that good scientific analysis software should have (e.g., stability, extensibility, testing infrastructure, etc.).

There is also a minor issue - these packages address an important need for high-level analysis tools but do not provide associated tools for preprocessing (e.g., spike sorting) or for creating reproducible pipelines for these analyses. This is entirely reasonable, in that no one package can be expected to do everything, but a bit deeper account of the process that takes raw data and produces scientific results would be helpful. In addition, some discussion of how this package could be combined with other tools (e.g., DataJoint, Code Ocean) would help provide context for where Pynapple and Pynacollada could fit into a robust and reliable data analysis ecosystem.

Author Response:

We would like to thank the reviewers and editor for their insightful comments and suggestions. We will update the manuscript accordingly. We are particularly glad to read that our software package constitutes a set of “well-written analysis routines” which have “the potential to become very valuable and foundational tools for the analysis of neurophysiological data”. Both reviewers have identified a number of weaknesses in the manuscript, and we would like to take this opportunity to provide a response to some of the remarks and clarify the objectives of our work. We would like to stress that this kind of toolkit is in continual development, and the manuscript offered a snapshot of the package at one point during this process. Since the initial submission several months ago, several improvements have been implemented and further improvements are in development by our group and a growing community of contributors. The manuscript will be updated to reflect these more recent changes, some which will directly address the reviewers’ remarks.

It was first suggested that the manuscript should better showcase the value of the analysis pipeline. As noted by the first reviewer, the online repository (i.e. GitHub page) conveys a better sense of how the toolbox can be used than the present manuscript. Our original intention was to illustrate some examples of data analysis in Figure 4 by adding the corresponding Pynapple command above each processing step. Each step takes a single line of code, meaning that, for example, one only needs to write three lines of code to decode a feature from population activity using a Bayesian decoder (Fig. 4a), or to compute a cross-correlograms of two neurons during specific stimulus presentation (Fig. 4b), or to compute the average firing rate of two neurons around a specific time of the experimental task (Fig. 4c). In our revision, we will include code snippets which will clearly show the required steps for each of these analyses. In addition, we will more clearly point the reader to the online tools (e.g. Jupyter notebooks), which offer an easier and clearer way to demonstrate the use of the toolbox.

Another remark concerns our claim that the package does not have dependencies. We agree that this claim was not well-worded. Our intention was to say that the package exclude dependencies such as scikit-learn, tensorflow or pytorch, which are often used in signal processing and which can be tedious to install. Pynapple still depends on a few packages including the most common ones: Numpy, Scipy, and Pandas. We will rephrase this statement in the manuscript and emphasize the importance of minimal dependencies for long-term backwards-compatibility in scientific computing.

We will complete the bibliography to make sure we properly reference all the packages designed for similar purpose. To note, some are not citable per se (i.e. no associated paper) but will be discussed.

It was suggested that the manuscript should better describe the integration of Pynapple into a full experimental data pipeline. This is an interesting point, which was briefly mentioned in the third paragraph of the discussion. Pynapple was not originally designed to pre-process data. However, it can load any type of data stream after the necessary pre-processing steps. Overall, this modularity is a key aspect of the Pynapple framework, and this is also the case for the integration with data pre-processing pipelines, for example spike sorting in electrophysiology and detection of region of interest in calcium imaging. We do not think there should be an integrated solution to the problem but, instead, to make it possible that any piece of code can be used for data irrespective of how the dataset was acquired. This is why we focused on making data loading straightforward and easy to adapt to any situation. This feature enables any user with any data modality and any long-established (often in-house) pre-processing scripts/software to utilize Pynapple in the analysis phase of their pipeline. Overall, not imposing a certain format compatibility from data acquisition phase is a strength for any analysis package.  

Finally, the reviews raised the issue of data and intermediate result storage. We agree that this is a critical issue. In the long term, we do not believe that the current implementation of NWB is the right answer for data involved in active analysis, as it is not possible to overwrite a NWB file. This would require the creation of a new NWB file each time an intermediate result is saved, which will be computationally intensive and time consuming, further increasing the odds of writing error. Theoretically, users who need to store intermediate results in a flexible way could use any methods they prefer, writing their own data files and wrappers to reload these data into Pynapple object. However, it is desirable for the Pynapple ecosystem to have a standardized format for storing data. We are currently improving this feature by developing save and loads methods for each Pynapple core object. We aim to provide an output format that is very simple to read in future Pynapple releases. This feature will be available in the coming weeks and will be described in the revised manuscript.

Je ne partage pas entièrement votre impression. Ainsi, typiquement, Déprise n'est pas une oeuvre revendiquée comme "codée", et il en est ainsi de beaucoup d'autres relevant de la littérature hypermédiatique (période Flash et Director notamment). Sans même parler des livres numériques enrichis !

Cet recommandation semble être en contradiction avec une autre recommandadion

Author Response

Reviewer #1 (Public Review):

Estimating the effects of mutations on the thermal stability of proteins is fundamentally important and also has practical importance, e.g, for engineering of stable proteins. Changes can be measured using calorimetric methods and values are reported as differences in free energy (dG) of the mutant compared to wt proteins, i.e., ddG. Values typically range between -1 kcal/mol through +7 kcal/mol. However, measurements are highly demanding. The manuscript introduces a novel deep learning approach to this end, which is similar in accuracy to ROSETTA-based estimates, but much faster, enabling proteomewide studies. To demonstrate this the authors apply it to over 1000 human proteins.

The main strength here is the novelty of the approach and the high speed of the computation. The main weakness is that the results are not compared to existing machine learning alternatives.

We thank Prof. Ben-Tal for taking the time to assess our work, and for his comments and suggestions below.

Reviewer 2 (Public Review):

Summary:

This work presents a new machine-learning method, RaSP, to predict changes in protein stability due to point mutations, measured by the change in folding free energy ΔΔG.<br /> The model consists of two coupled neural networks, a 3D selfsupervised convolutional neural network that produces a reduceddimensionality representation of the structural environment of a given residue, and a downstream supervised fully-connected neural network that, using the former network's structural representation as input, predicts the ΔΔG of any given amino-acid mutation. The first network is trained on a large dataset of protein structures, and the second network is trained using a dataset of the ΔΔG values of all mutants of 35 proteins, predicted by the biophysics-based method Rosetta.

The paper shows that RaSP gives good approximations of Rosetta ΔΔG predictions while being several orders of magnitude faster. As compared to experimental data, judging by a comparison made for a few proteins, RaSP and Rosetta predictions perform similarly. In addition, it is shown that both RaSP and Rosetta are robust to variations of input structure, so good predictions are obtained using either structures predicted by homology or structures predicted using AlphaFold2.<br /> Finally, the usefulness of a rapid approach such as RaSP is clearly demonstrated by applying it to calculate ΔΔG values for all mutations of a large dataset of human proteins, for which this method is shown to reproduce previous findings of the overall ΔΔG distribution and the relationship between ΔΔG and the pathological consequences of mutations. The RaSP tool and the dataset of mutations of human proteins are shared.

Strengths:

The single main strength of this work is that the model developed, RaSP, is much faster than Rosetta (5 to 6 dex), and still produces ΔΔG predictions of comparable accuracy (as compared with Rosetta, and with the experiment). The usefulness of such a rapid approach is convincingly demonstrated by its application to predicting the ΔΔG of all single-point mutations of a large dataset of human proteins, for which using this new method they reproduce previous findings on the relationship between stability and disease. Such a large-scale calculation would be prohibitive with Rosetta. Importantly, other researchers will be able to take advantage of the method because the code and data are shared, and a google colab site where RaSP can be easily run has been set up. An additional bonus is that the dataset of human proteins and their RaSP ΔΔG predictions, annotated as beneficial/pathological (according to the ClinVar database) and/or by their allele frequency (from the gnomAD database) are also made available, which may be very useful for further studies.

Weaknesses:

The paper presents a solid case in support of the speed, accuracy, and usefulness of RaSP. However, it does suffer from a few weaknesses.

The main weakness is, in my opinion, that it is not clear where RaSP is positioned in the accuracy-vs-speed landscape of current ΔΔGprediction methods. The paper does show that RaSP is much faster than Rosetta, and provides evidence that supports that its accuracy is comparable with that of Rosetta, but RaSP is not compared to any other method. For instance, FoldX has been used in large-scale studies of similar size to the one used here to exemplify RaSP. How does RaSP compare with FoldX? Is it more accurate? Is it faster? Also, as the paper mentions in the introduction, several ML methods have been developed recently; how does RaSP compare with them regarding accuracy and CPU time? How RaSP fares in comparison with other fast approaches such as FoldX and/or ML methods will strongly affect the potential usefulness and impact of the present work.

Second, this work being about presenting a new model, a notable weakness is that the model is not sufficiently described. I had to read a previous paper of 2017 on which this work builds to understand the self-supervised CNN used to model the structure, and even so, I still don't know which of 3 different 3D grids used in that original paper is used in the present work.

A third weakness is, I think, that a stronger case needs to be made for fitting RaSP to Rosetta ΔΔG predictions rather than experimental ΔΔGs. The justification put forward by the authors is that the dataset of Rosetta predictions is large and unbiased while the dataset of experimental data is smaller and biased, which may result in overfitting. While I understand that this may be a problem and that, in general, it is better to have a large unbiased dataset in place of a small biassed one, it is not so obvious to me from reading the paper how much of a problem this is, and whether trying to fix it by fitting the model to the predictions of another model rather than to empirical data does not introduce other issues.

Finally, the method is claimed to be "accurate", but it is not clear to me what this means. Accuracy is quantified by the correlation coefficient between Rosetta and RaSP predictions, R = 0.82, and by the Mean Absolute Error, MAE = 0.73 kcal/mol. Also, both RaSP and Rosetta have R ~ 0.7 with experiment for the few cases where they were tested on experimental data. This seems to be a rather modest accuracy; I wouldn't claim that a method that produces this sort of fit is "accurate". I suppose the case is that this may be as accurate as one can hope it to be, given the limitations of current experimental data, Rosetta, RaSP, and other current methods, but if this is the case, it is not clearly discussed in the paper.

We thank the reviewer for their detailed comments and suggestions.

As discussed in our general comments above and also below, we have now added additional benchmarking, making it easier to compare the accuracy of RaSP with other methods. Regarding the model description, we have now added a more detailed description of also the 3D CNN.

Regarding whether to fit the model to experiments or computational data, we agree that it is not clear cut that the former would also not work. Indeed, a main problem is that in both cases it is hard to answer which approach is better because of the scarcity of experimental data. One major problem with the larger sets of experimental data is, as we mention, the bias and variability; another is the provenance. While some databases exist, they are rarely exactly raw data, and for example may contain ∆∆G values estimated from ∆Tm values. In the revised manuscript we now explain better why we chose to target Rosetta, but also acknowledge that one might also have used experiments.

As to the question of accuracy, we agree completely that the methods could be better. One problem, however, is that it is very difficult to answer how much better because of problems with experiments. As mentioned also by reviewer 1, variation across different experiments suggest that even a “perfect” predictor would only achieve Pearson correlation coefficients in the range 0.7–0.8 (https://doi.org/10.1093/bioinformatics/bty880). Clearly, this is an issue with imperfect data curation (it is possible to measure ∆∆G quite accurately), but in the absence of larger and better curated experiments, one will not expect much better accuracy than what we report here. This is now discussed in the revised manuscript.

Reviewer 3 (Public Review):

The authors present a machine learning method for predicting the effects of mutations on the free energy of protein stability. The method performs similarly to existing methods, but has the advantage that it is faster to run. Overall this is reasonable and a faster method will likely have some potential uses. However, not improving performance beyond the reasonable but not great performance of existing methods of course makes this a less useful advance. The authors provide predictions for a set of human proteins, but the impact of their method would be much greater if they provided predictions for all substitutions in all human proteins, for example. In places the text somewhat overstates the performance of computational methods for predicting free energy changes and is potentially misleading about when ddGs are predicted vs. experimentally measured. In addition, the comparison to existing methods is rather slim and there isn't a formal evaluation of how well RASP discriminates pathological from benign variants.

We thank the reviewer for taking time to read our work and for their various suggestions.

theme

I'm curious about what specifically they're referring to and where that quote comes from. If it's from the book, it has no real meaning. If it's a reference to a quote from a person in a position of "the elite" then that's at least potentially a different story but the meaning would still need to be evidenced. Otherwise, it take a giant leap of assumption for the reader to go from one contextless quote to a conclusion.

1. 一个包含 400 种资源和工具的精选目录，供初创公司使用
2. “来自筹集了超过 4 亿美元的真实初创公司的真实套牌
3. 查找与域相关的所有电子邮件地址
4. 营销资源和工具的精选目录
5. 为您的活动免费搜索影响者和成本估算。
6. 轻松地为您的内容添加动画。
7. 创业前需要阅读的文章。
8. 简单、响应迅速的单页网站，几乎适用于任何内容。
9. 将人类插图与设计库混合搭配。
10. 一个手绘插画库。
11. 世界上最好的创业建议，精选且可搜索。
12. 接受支持的一种免费、快速、美观的方式 ☕ 💰
13. 在 10 分钟内创建非常好的社交视频
14. 一键发现运行网站的内容 🖥️ 🔍
15. 制造商交付出色产品的工具包 ⛵
16. 在没有大预算的情况下制作像大品牌一样的视频广告
17. 为您的初创公司设计的简单登陆页面
18. 免费的超现实主义插图 💡
19. 适合所有人的插图
20. 完全免费的高质量股票视频。
21. 基于AI技术的网站审计监控平台
22. 100% 自动删除视频和 GIF 背景
23. 构建人们想要的产品的多合一平台
24. DemandCurve.com 和 Julian Shapiro 的增长指南
25. 在几秒钟内自动生成可视站点地图。
26. 在线图形设计应用程序主要专注于产品模型。
27. 瞬间将屏幕截图变成精美的图像
28. 干净、简单、无忧的标志设计
29. 精选的增长黑客工具集合
30. Neil Patel 的免费关键字研究工具
31. 适用于您的视频、游戏和 UI 的免费库存音乐
32. 最大的免费营销资源集合
33. 启动有利可图的副业项目的终极资源
34. 呈现、协作和管理 - 集于一身。
35. 2018 年使用的顶级增长黑客（2019 年更新）
36. 25 多个增长技巧（2019 年更新）
37. 内容营销资源的精选目录
38. 创建营销视频的最简单方法，来自 Slidely
39. 在 60 秒内创建病毒式等待列表
40. 易于构建，免费自定义登录页面
41. 适用于网站的 Turbo 原型制作工具 🚀
42. 人工智能驱动的标志制造商
43. 一个应用程序来取代它们
44. 获取您使用和喜爱的产品的创业股份
45. 从 3000 多种元素创建免费的个性化插图
46. 2019 年 30 多个增长黑客案例研究
47. 现实世界营销示例的画廊
48. 免费插图以获得更好的用户体验
49. 潜在客户生成登陆页面作为聊天机器人
50. 网站分析器和即时曝光检测系统
51. 免费弹出窗口生成器，将您的流量转化为销售和潜在客户
52. 🔎预览并准备您的 Product Hunt 提交
53. 免费粗略插图的集合
54. 无需编码即可为 iOS、Android 和 Web 构建应用程序💥
55. 帮助您的新创业公司的工具
56. 900 多个最佳 SaaS 登录页面的屏幕截图
57. 适用于初创公司的免费和高级登陆页面模板
58. 直接在浏览器中创建精彩视频
59. 创建漂亮的推销平台以完成交易
60. 打造属于自己的美丽品牌
61. 发起杀手级活动的免费资源
62. 基于网站的 Chatbot 构建器 => 获得更多潜在客户
63. 创建交互式图像的最快和最简单的方法
64. 发展业务的权威手册
65. 将 100 个单词变成电影介绍
66. 计划、跟踪和分析您的 Instagram 营销
67. 演示文稿您可以👀和👂
68. 一键让你的幻灯片闪闪发光✨
69. 2017 年的 LinkedIn 黑客攻击（现已收录在 THE BAMF BIBLE 中）
70. 只需单击一下，即可在 100 个目录中推广您的业务
71. 为每个人无限定制的图形
72. 一款可立即为您的目标网页评分的免费工具
73. 查看 100 多个创业公司宣传资料或提交您自己的资料。
74. 在 Facebook Messenger 上发送时事通讯
75. 通过发布前分析提升您的营销活动绩效
76. 数十亿美元的公司如何获得第一批用户
77. 初创公司的病毒式传播变得容易💪
78. 无需电子表格或代码即可自动执行手动流程
79. 你有一个 MVP。 怎么办？
80. 使用这个有用的工具构建 Dropbox 风格的推荐 ✌️
81. 在 Web 上创建交互式故事
82. 为您的公司制作专业的新闻资料袋
83. 来自自动完成建议的内容创意
84. 只需点击几下即可免费创建产品屏幕截图
85. 您的内容和社交团队的指挥中心🚩
86. 为初创公司提供的 750 多种资源的精选列表
87. 适用于 Slack 的 Google Analytics 机器人
88. 来自 30 位应用营销专家的 55 堂应用增长课程
89. 在几秒钟内创建令人惊叹的产品屏幕截图！
90. 一种可扩展的与客户交谈的方式
91. 用于增加应用安装和自然流量的 ASO 工具 📈
92. 免费生成全长网页截图
93. 创建、编辑和管理很棒的电子邮件通讯😻
94. 使用动态个性化图像将您的电子邮件点击率提高 2 倍。
95. 团队成长协作软件📈
96. 了解开发人员如何编写自己的薪水
97. 关键字量、每次点击费用和竞争的免费浏览器插件
98. 更好的用户入职。 免费代码。
99. 有利可图的在线业务：经过审查并交给您
100. 免费的优质链接构建工具，可帮助您获得更多流量
101. 一个简单易用的无代码 API 市场
102. 发送和接收有针对性的付费微任务
103. 使您的网站转换更好的 41 点清单
104. 免费书籍，其中包含来自 top P 的实用技巧

低/无代码工具

文章介绍了一位 Twitter 用户 @marckohlbrugge 向 GPT-4 提出请求，让它制作 Nomad List，结果成功了。其他用户也在评论中提出了一些问题和讨论，包括 Unsplash API 密钥是否真实以及 Starlink 卫星互联网网络等话题。 https://twitter.com/levelsio/status/1635994524286881792?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1635994524286881792%7Ctwgr%5E21ad34182282b4fe87da85a9af0c7e5345cf7715%7Ctwcon%5Es1_c10&ref_url=https%3A%2F%2Firis-brush-b04.notion.site%2FGPT-4-Master-Course-6d5eec61db4a4a51bed80ca5e274f1c2

文章介绍了一位用户使用 GPT-4 对其聊天机器人代码进行了漏洞和错误扫描，并分享了扫描结果和观察到的现象。用户发现 GPT-4 响应速度较慢，上下文窗口较大，比以前的模型更准确，输出变化较小。用户的代码库只有一个 “前端” 文件。文章还附带了一个 GitHub 仓库链接，用于验证漏洞和解决方案。 https://twitter.com/mayowaoshin/status/1635757442859671553?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1635757442859671553%7Ctwgr%5E21ad34182282b4fe87da85a9af0c7e5345cf7715%7Ctwcon%5Es1_c10&ref_url=https%3A%2F%2Firis-brush-b04.notion.site%2FGPT-4-Master-Course-6d5eec61db4a4a51bed80ca5e274f1c2

一位开发者在 Twitter 上分享了他使用 GPT-4 写代码的经历，GPT-4 仅用了 3 小时就完成了一个开发者需要 2 周才能完成的任务，而且只花费了 0.11 美元。GPT-4 不仅编写了脚本，还提供了详细的设置和运行说明，每个脚本都一次性通过测试。此外，GPT-4 还可以添加注释和修复错误。文章列举了两个例子，展示了 GPT-4 的编程能力。 https://twitter.com/joeprkns/status/1635933638725451779?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1635933638725451779%7Ctwgr%5E21ad34182282b4fe87da85a9af0c7e5345cf7715%7Ctwcon%5Es1_c10&ref_url=https%3A%2F%2Firis-brush-b04.notion.site%2FGPT-4-Master-Course-6d5eec61db4a4a51bed80ca5e274f1c2

一位 Twitter 用户展示了如何使用 Chat GPT-4 和 @Replit，在不到 20 分钟内零基础编写了一个贪吃蛇游戏。他只需向 GPT-4 提出请求，获取游戏所需的 HTML、CSS 和 JavaScript 代码，然后将其粘贴到 @Replit 中运行即可。虽然游戏存在一些问题，但他只需向 GPT-4 提出修改请求，GPT-4 就能够快速响应并解决问题。作者认为，GPT-4 是一个不错的编程老师。 https://twitter.com/ammaar/status/1635754631228952576?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1635754631228952576%7Ctwgr%5Ef5c6d2f248973e51fc61a7995b55ae169800d1dc%7Ctwcon%5Es1_c10&ref_url=https%3A%2F%2Fbefitting-share-28e.notion.site%2FThe-Ultimate-GPT-4-Cheatsheet-f87942eb2c95458881d593dc2203aeb5

该视频是关于使用 React JS 和 OpenAI 编写更好版本的 ChatGPT。 该视频首先讨论了 ChatGPT 的局限性，以及构建您自己的 AI 版本如何让您个性化聊天机器人并使用更精致的提示。 创建者还为那些想要构建自己的聊天机器人的人提供了 OpenAI 入门工具包和 ChatGPT 入门工具包。 然后视频继续展示使用 Visual Studio Code 构建网站前端并创建侧边菜单和主要部分的过程。 最后，该视频讨论了通过删除标题并为其指定类名称“聊天框”同时将文本颜色更改为白色来使聊天框充分利用整个空间。

Note to self: getting into WP Gutenberg is a shift deeper into JS and less PHP. My usually entry into creating something for myself is to base it on *AMP (MAMP now) so I can re-use what I have in PHP and MySQL as a homecook.

It is different from the source code

A lot of the magic of this merge seems to hide in this versionDiff function, worth outlining the rough process at least. From the code it seems like it produces the least common ancestor version and is considered empty iff the versions are the same

Ive always found dress code to be controversial a women should be able to dress comfortable in a way they feel the most safe it becomes an issue when boundaries are cross such as unwanted interactions or comments.

Again, this is a tradeoff - do you want the tool to be purpose-built for your specific goal (and to build lots of purpose-built tools), or do you want a chat interface that can do anything? It's okay to do some of the work while the technology is in its infancy - right now it's more important to explore than anything else.

The programmatic version (noGUI) version of this tool could be made a little easier to use by accepting inputs from the command line, such as with argparse.

It should be somewhat straightforward to modify the code to make it accept command line arguments. A tutorial can be found here: https://realpython.com/command-line-interfaces-python-argparse/

Are there any cover coverage tools that tell you what proportion of your code is run how many times by your test suite and give you a coverage histogram?

L'article est passionnant et traite d'un sujet classique dans les sciences du langage, philosophie de la connaissance et intelligence artificielle. Sujet classique mais toujours pertinent et au coeur des interrogations contemporaines.

L'auteur verra que je ne suis pas d'accord du tout avec son approche malgré tout l'intérêt que j'ai pris à le lire. En effet, sont mobilisées dans l'article des hypothèses sur la langue, les langues, qui ne sont pas suffisamment explicitées pour permettre de mieux situer la proposition d'IEML.

Le principal point de discussion est la nature de langue philologique revendiquée pour IEML, au sens où une langue philologique permet de tout dire. IEML est un système de codage fondé sur des primitives combinées et décorées par des liens supplémentaires. Que le potentiel de description soit infini est clair, qu'il soit celui d'une langue naturelle l'est moins. Une langue naturelle peut tout dire car elle n'a pas de primitives, qu'elle est sensible au contexte, qu'elle est son propre méta-langage (cf. Rastier "sémantique interprétative" PUF, et "Sémantique et recherches cognitives" PUF).

IEML n'est pas ainsi une alternative aux méthodes formelles souvent critiquées dans l'article mais une variante. Mais c'est ce qui fait son intérêt : sa capacité expressive permet de capter nombre de domaines de spécialité, dont la sémantique est normalisée donc formalisables et modélisables. La question qui se poserait alors serait le ratio expressivité / complexité car IEML est difficile à utiliser. Comme nombre de modésliation, elle rejoint la catégorie des langues dont le spécialiste est le seul locuteur.

Il serait intéressant également que l'auteur considère les autres tentatives de formalisation de la langue naturelle, par exemple la sémantique de Montague. C'est le lambda calcul typé qui est dans ce cas l'outil de référence, et il apporte précisément ce que revendique IEML : grammaticalité, primitive, combinatoire.

Enfin, une discussion devrait être ouverte sur la question de ce qu'est un concept, qui flotte dans l'article entre un mot du langage IEML, une notion, un signifié. L'usage en sémantique linguistique est de considéré que le concept est un signifié normé. Ce point de vue n'est pas discuté ou abordé dans l'article : on a l'impression que le mot de la langue est brutalement rabattu sur le mot en IEML pour en faire un concept. Brutal car la correspondance s'effectue sur un arbitraire de codage, choix des mots de IEML qu'on fait correspond à des mots de langue. Choix que l'on comprend, mais qu'on ne saurait refaire à l'identique, le codage reflétant l'arbitraire du codeur et non la norme sémantique adoptée.

C'est ce qui rend IEML difficile à aborder, le fait que le code produit semble résulter d'un arbitraire propre au codeur. On pourrait certes tout coder, mais toujours de façon différente. Ce qui irait à l'encontre de l'objectif d'interopérabilité revendiqué (sans doute à tort si notre sentiment est exact) et mais à raison (besoin pratique des bases de connaissances).

Au final, je soutiens la publication de l'article qui présente un point de vue qui, pour n'être pas le mien, est intéressant et apporte une contribution aux débats en cours. Il faut cependant des révisions pour situer les arguments linguistiques utilisés qui restent trop implicites ou trop détachés des traditions linguistiques et sémantiques (très peu de linguistes sont cités contrairement aux logiciens par exemple).

On obtient une belle systématicité, qui fonctionne bien sur le langage ainsi créé, mais l'affectation à des signifiés reste arbitraire : il s'agit plus d'un code algébrique, codant donc un matériau sémantique, sans pour autant justifier que ce codage reflète la structure primitive du sens, ou une représentation complète sur une base de primitives données.

Software product lines and platforms are software products

That's because you didn't clean the data in this example. In the following code, astype() didn't "convert objects to floats." You called dropna(), which eliminated the "objects" because those rows also happened to have blank columns.

Behavior factors seems like an artificial distinction, at least based on these examples. These would be better classified as Knowledge factors. Drawing a pattern that you've memorized is conceptually no different than typing a code. Or should I point out that typing a code is also a behavior? You have to press your fingers in a certain location on your keyboard and in a certain order.

It is important that it work on Windows, but three-quarters of all PCs do not run Windows. Three-quarters of all traditional laptops and desktops? Sure, but most personal computers these days are mobile phones.

Reviewer #1 (Public Review):

The paper by Dr. Ter-Ovanesyan et. all discussing a very important topic in the field of extracellular vesicles: how to enrich EVs compare to more abundant other circulating particles like lipoproteins, especially VLDL and LDL, which overlap in size and density with EVs and make the purification process challenging. The authors discussed several approaches, including size exclusion chromatography, density-gradient centrifugation, and methods combining charge and size separation. They also proposed the Tri-Mode Chromatography (TMC) method as a good alternative to conventional SEC separation. However, the results provided for the TMC method do not fully support the claim. TEM images provided show the presence of lipoprotein particles at a higher rate than EVs. In addition, proteomics data suggest that lipoproteins and free proteins are still overrating ones associated with EVs.

The importance of this paper is the code available for an automated device for simultaneous fraction collection, which can be very useful for researchers with limited resources since commercial devices are quite expensive.

@17:03

The idea of portability is not that you take your C code and recompile it and hope it compiles and hope the compilers have the same bugs in them.

Reviewer #1 (Public Review):

This study investigates the context-specificity of facial expressions in three species of macaques to test predictions for the 'social complexity hypothesis for communicative complexity'. This hypothesis has garnered much attention in recent years. A proper test of this hypothesis requires clear definitions of 'communicative complexity' and 'social complexity'. Importantly, these two facets of a society must not be derived from the same data because otherwise, any link between the two would be trivial. For instance, if social complexity is derived from the types of interactions individuals have, and different types of signals accompany these interactions, we would not learn anything from a correlation between social and communicative complexity, as both stem from the same data.

The authors of the present paper make a big step forward in operationalising communicative complexity. They used the Facial Action Coding System to code a large number of facial expressions in macaques. This system allows decomposing facial expressions into different action units, such as 'upper lid raiser', 'upper lip raiser' etc.; these units are closely linked to activating specific muscles or muscle groups. Based on these data, the authors calculated three measures derived from information theory: entropy, specificity and prediction error. These parts of the analysis will be useful for future studies.

The three species of macaque varied in these three dimensions. In terms of entropy, there were differences with regard to context (and if there are these context-specific differences, then why pool the data?). Barbary and Tonkean macaques showed lower specificity than rhesus macaques. Regarding predicting context from the facial signals, a random forest classifier yielded the highest prediction values for rhesus monkeys. These results align with an earlier study by Preuschoft and van Schaik (2000), who found that less despotic species have greater variability in facial expressions and usage.

Crucially, the three species under study are also known to vary in terms of their social tolerance. According to the highly influential framework proposed by Bernard Thierry, the members of the genus Macaca fall along a graded continuum from despotic (grade 1) to highly tolerant (grade 4). The three species chosen for the present study represent grade 1 (rhesus monkeys), grade 3 (Barbary macaques), and grade 4 (Tonkean macaques).

The authors of the present paper define social complexity as equivalent to social tolerance - but how is social tolerance defined? Thierry used aggression and conflict resolution patterns to classify the different macaque species, with the steepness of the rank hierarchy and the degree of nepotism (kin bias) being essential. However, aggression and conflict resolution are accompanied by facial gestures. Thus, the authors are looking at two sides of the same coin when investigating the link between social complexity (as defined by the authors) and communicative complexity. Therefore, I am not convinced that this study makes a significant advance in testing the social complexity for communicative complexity hypothesis. A further weakness is that - despite the careful analysis - only three species were considered; thus, the effective sample size is very small.

Reviewer #2 (Public Review):

This is a well-written manuscript about a strong comparative study of diversity of facial movements in three macaque species to test arguments about social complexity influencing communicative complexity. My major criticism has to do with the lack of any reporting of inter-observer reliability statistics - see comment below. Reporting high levels of inter-observer reliability is crucial for making clear the authors have minimized chances of possible observer biases in a study like this, where it is not possible to code the data blind with regard to comparison group. My other comments and questions follow by line number:

38-40. Whereas I am an advocate of this hypothesis and have tested it myself, the authors should probably comment here, or later in the discussion, about the reverse argument - greater communicative complexity (driven by other selection pressures) could make more complicated social structures possible. This latter view was the one advocated by McComb & Semple in their foundational 2005 Biology Letters comparative study of relationships between vocal repertoire size and typical group size in non-human primate species.

72-84 and 95-96. In the paragraph here, the authors outline an argument about increasing uncertainty / entropy mapping on to increasing complexity in a system (social or communicative). In lines 95-96, though, they fall back on the standard argument about complex systems having intermediate levels of uncertainty (complete uncertainty roughly = random and complete certainty roughly = simple). Various authors have put forward what I think are useful ways of thinking about complexity in groups - from the perspective of an insider (i.e., a group member, where greater randomness is, in fact, greater complexity) vs from the perspective of an outside (i.e., a researcher trying to quantify the complexity of the system where is it relatively easy to explain a completely predictable or completely random system but harder to do so for an intermediately ordered or random system). This sort of argument (Andrew Whiten had an early paper that made this argument) might be worth raising here or later in the discussion? (I'm also curious where the authors sentiments lie for this question - they seem to touch on it in lines 285-287, but I think it's worth unpacking a little more here!)

115-129. See also:<br /> Maestripieri, D. (2005). "Gestural communication in three species of macaques (Macaca mulatta, M. nemestrina, M. arctoides): use of signals in relation to dominance and social context." Gesture 5: 57-73.<br /> Maestripieri, D. and K. Wallen (1997). "Affiliative and submissive communication in rhesus macaques." Primates 38(2): 127-138.<br /> On that note, it is probably worth discussing in this paragraph and probably later in the discussion exactly how this study differs from these earlier studies of Maestripieri. I think the fact that machine learning approaches had the most difficulty assigning crested data to context is an important methodological advance for addressing these sorts of questions - there are probably other important differences between the authors' study here and these older publications that are worth bringing up.

220-222. What is known about visual perception in these species? Recent arguments suggest that more socially complex species should have more sensitive perceptual processing abilities for other individuals' signals and cues (see Freeberg et al. 2019 Animal Behaviour). Are there any published empirical data to this effect, ideally from the visual domain but perhaps from any domain?

274-277. I am not sure I follow this - could not different social and non-social contexts produce variation in different affective states such that "emotion"-based signals could be as flexible / uncertain as seemingly volitional / information-based / referential-like signals? This issue is probably too far away from the main points of this paper, but I suspect the authors' argument in this sentence is too simplified or overstated with regard to more affect-based signals.

288 on. Given there are only three species in this study, the chances of one of the species being the 'most complex' in any measure is 0.33. Although I do not believe this argument I am making here, can the authors rule out the possibility that their findings related to crested macaques are all related to chance, statistically speaking?

329-330. The fact that only one male rhesus macaque was assessed here seems problematic, given the balance of sexes in the other two species. Can the authors comment more on this - are the gestures they are studying here identical across the sexes?

354-371. Inter-observer reliability statistics are required here - one of the authors who did not code the original data set, or a trained observer who is not an author, could easily code a subset of the video files to obtain inter-observer reliability data. This is important for ruling out potential unconscious observer biases in coding the data.

Prompt chaining is when you iterate through multiple steps from an input to a final result, where the output of intermediate steps is input for the next. This is what AutoGPT does too. Appleton's employer Ought is working in this area too. https://www.zylstra.org/blog/2023/05/playing-with-autogpt/

You can customise the challan generated for bank transfers and add custom fields and disclaimers. You can do this by passing specific parameters to the Razorpay Checkout code.

x

Il faut décrire les résultats. Dans le caption du graphique il faut donner le contenu du graphique, idem pour les tableaux.

Mon effort d'interprétation va jusqu'à aller lire le code pour comprendre le rapport.

Reviewer 2: Armin Scheben

The authors present the web app DivBrowse for visualizing genomic variant data. Their code is publicly available, and their web app is well-documented and provides several demonstration implementations for human, mouse and barley. The manuscript is well-written and concisely covers the key features of DivBrowse and summarizes the implementation of the software.

I was able to test the demonstration website and was impressed with how smoothly everything ran and was set up. Due to time constraints, I was not able to test the installation and set up of DivBrowse but the documentation looks sufficient to allow easy set up by experts. Overall, I think this is a useful contribution to the community. One key issue I believe the authors should address, however, is that the manuscripts presents DivBrowse in a vaccum, not providing much mention of or comparison with existing software with overlapping functionality. Below I provide some further details illustrate my point and how it might be addressed, as well as listing several other minor comments.

Main comment

The authors rightly indicate in their introduction that the growing amounts of genomic data generated require robust solutions for visualization and exploration that does not require use of the command-line. But the authors fail to mention that there exists a considerable ecosystem of software that already does this. Moreover, some of the software available offers substantially expanded features compared to DivBrowse.

To help readers better decide when DivBrowse might be the right choice for their needs compared to other options, the authors could cite existing software and provide some comparison. My knowledge of all available software is not exhaustive, but Wang et al. 2020 (https://doi.org/10.1093/gigascience/giaa060) in their publication of SnpHub provide a comparison table including SnpHub itself and Jbrowse. I would consider both of these tools for exploration and visualization of SNPs and additional data, similar to DivBrowse. Jbrowse is relatively widely used and considerably more feature-rich. The standalone offline tool TASSEL (https://academic.oup.com/bioinformatics/article/23/19/2633/185151) also offers many options for visualisation and exploration and analysis of VCF data offline. There may also be other tools I am not aware of, and readers would likely benefit from some brief overview of the landscape and the pros and cons of each piece of software and what differentiates DivBrowse.

Minor comments

The authors can consider the minor comments below as 'take it or leave it' comments. I do not think it is essential to address these, but in my view they may enhance the manuscript.

1) In the discussion, the authors point out the efficiency and low latency of DivBrowse, however this is not quantified in the manuscript. If it were technically feasible without substantial effort, it might be useful to quantify in some way just how efficient DivBrowse can be, especially if this could be one of the stand-out features of DivBrowse.

2) The authors use divergence Bezier curves to increase the amount of variant calls that can be visualized. This is helpful and a useful default. However, invariant sites can also be of considerable evolutionary and breeding/medicinal interest. When collapsing invariant sites, they become indistinguishable from unmapped regions. This is a fundamental issue and many VCF files may not encode information on invariant sites, so it may not be possible to develop robust functionality that allows users to also show invariant sites optionally. Still, this point may be worth briefly mentioning in the discussion, if the authors agree it is noteworthy.

3) One advantage of visualization of relatively raw data like SNPs is that it can reveal patterns that are less obvious in other types of data exploration. To fully take advantage of this tools like Jbrowse allow export of the browser window in SVG format, allowing users to incorporate images into high-resolution figures. I don't expect the authors to necessarily implement this feature for this review, but it may be worth adding it to the list of potential enhancements that could be implemented based on user demand.

It is almost funny in the situation that Elon Musk mentions. At the end of the day, recommendations are a sequence of code so "hate-posting" or replying would only get more content of that kind pushed to you, unless you explicitly report or ask the social media site to stop recommending the content relevant.

On the open source recommendation algorithms, people have spotted that it's Elon Musk who made the post. The post will have received a higher score than others. Moreover, it will execute an entirely different chuck of code.

Cause for concern. The laws surrounding LLC, 501c3s, and 501c4s exist for a reason. The lack of transparency itself is an argument for closing loopholes in the tax code.

It's a good remainder of not taking Free Libre Open Source Software (FLOSS) as granted, as a commons we don't need to fight for in an increasing world of extractivism, expropriation and platform surveillance capitalism against the commons. So even with all the indirection and friction behind software building, delivery and modification, having FLOSS should not given for granted.

On another note, there is already an intermediate place between hypercard and FLOSS, with pretty agile development/prototyping cycles in things like Pharo/GT. It's for the desktop, not yet into iPhone, but with betas in progress to the more more open Android ecosystem and with possibilities to run on on with PharoJS

重构目的：抽象与复用

il faut décrire les résultats. Dans le caption du graphique il faut donner le contenu du graphique idem pour les tableaux. Mon effort d'interpretation va jusque aller lire le code pour comprendre le rapport. Merci de faire un tableau avec les investissements puis expliquer la méthode d'annualisation. Lpour les overlay, merci de choisir les digits et le vocabulaire.

In the schools I have worked in, this has not been an issue for most students because they receive a strong password from the school, and in elementary they just receive a QR code to scan for their password. I think the more important thing to teach about digital citizenship is the sharing of passwords, as it is much more likely to be an issue among students.

A compelling insight.

Ever the challenge of relying on dependencies. A necessary tradeoff, but a tradeoff non-the-less. This loops back in to something that I've been thinking recently which is that our means of sharing code is critical to the overall stability of software. No code sharing solution obviously isn't ideal but modern package management leaves something to be desired. I think languages that handle this as a first class feature (https://www.unison-lang.org/, https://scrapscript.org/) have a leg up in this regard.

ok

Better code today and everyday!

Does this mean this will be available only via API and not dashboard and webhooks?

Below are the sample status details responses that appears as part of the status details object.

Download the table to view exhaustive list of status details, their corresponding payout statuses and also possible steps you can take to proceed further.

duality between shards/crystal should be straightened. But: shards is not a task runner.

shards supports any git repository, not just ones hosted on GitHub or GitLab. And also mercurial and fossil repositories.

Mentions the protocol also mention in the Washington Post article and what would happen after her death.

this makes me think back to the novel The Hate U Give, which largely centers around the main character’s code-switching, between her mostly white private school and her hometown. It is an opposite situation, which is interesting given an historical context.

Capitalize

code format

do you want to add the Code of Ethics here too?

This highlighted section states one of the tips the author mentions which would aid becoming a software engineer. It mentions to "start writing code every day." I think this is true because to become a good software engineer, it takes a lot of time and practice. If someone takes time out of their daily lives to write some code, their knowledge of the subject will stay sharp and there will be a less likely chance that the author forgets some of the rules of coding. This interests me because if I am teaching this to a class, it would be great for the students to understand this fact so they know how much time and practice it takes to become a software engineer.

Many of the republicans are refusing to elaborate on the position of gun laws after the Nashville shooting.

Pleading the 5th, maybe if I just don't say anything, they'll leave me alone.

Definitely biases, but I think they're staying quiet because its obvious what the correct thing to do is. If they are caught and have to answer questions, they'll have to explain themselves and their thinking, and it only makes them sound stupid.

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

Learn more at Review Commons

Reply to the reviewers

General Statements

Thank you for providing an initial assessment of our manuscript. We went through all the raised comments and suggestions aiming to improve our manuscript. Our manuscript will benefit from addressing them.

Our main impression is that the concerns regarding the novelty of our work by Reviewers #1 and #3 come from the fact that we apply a known flexible statistical framework (group factor analysis) to novel applications in single-cell data analysis, namely the estimation of multicellular programs and sample-level unsupervised analysis. The core methodology of our work is indeed based on the popular tool Multi-omics factor analysis (MOFA). We see the novelty of our study in the formulation of these relatively new applications within this framework, and the demonstration of the added value that this formulation provides building on MOFA’s strengths, in particular by expanding the possibilities of downstream analysis of single-cell data including the meta-analysis of distinct single-cell patient cohorts and its integration to complementary bulk and spatial data modalities.

The simultaneous estimation of multicellular programs together with sample-level unsupervised analysis is only possible with a single available tool, scITD, which is limited by its modeling strategy, based on tensor decomposition: with tensor decomposition, multicellular programs can not be estimated from distinct feature sets across cell-types, making this method less flexible and sensitive to technical effects, such as background expression. We compared our proposed methodology with scITD and showed the benefits of using group factor analysis as implemented in MOFA for this task. Moreover, as of now, no other methodology is able to estimate multicellular programs and perform sample-level unsupervised analysis, simultaneously in multiple independent single-cell atlases. We also showed how multicellular programs are traceable in bulk transcriptomics data and show that they are better fit to classify heart failure patients compared to classic cell-type deconvolution approaches.

Altogether, we believe that our current manuscript complements existing literature and puts forward an approach with distinct features to analyze single-cell atlases. We will edit the text to make more explicit the novelty and advantages of our proposed methodology, and we will emphasize that our work does not mean to propose a new method, but rather demonstrate how group factor analysis can be used for novel sample-level analysis of single-cell data. We plan to incorporate the suggestions by Reviewer #1 regarding the inclusion of additional datasets, model validations, and novel applications involving a direct modeling of cell-compositions and spatial organization of cells. Moreover, we plan to discuss perspectives on how cell communications can be incorporated in the analysis of multicellular programs as suggested by Reviewer #2. Additionally, we will correct all the figure and text typos identified by the reviewers. Finally, we will provide an R package (https://github.com/saezlab/MOFAcellulaR) and python implementations (https://liana-py.readthedocs.io/en/latest/notebooks/mofacellular.html) that facilitate the use of our approach.

Please find below the point-by-point response to the reviewers in blue, numbered for convenience.

__Reviewer #1 (Evidence, reproducibility and clarity (Required)): __

Remark to authors

Flores et al. present a pipeline in which they leverage MOFA framework, a matrix factorization algorithm to infer multi-cellular programs (MCPs). Learning and using MCP has already been proposed by others. Yet, authors pursue a similar goals by using MOFA, providing a cell*sample matrix for different cell types as different views (instead of multiple modalities/views) as the input. They later apply MOFA using this data format on a series of applications to analyze acute and chronic human heart failure single-cell datasets using MCPs. Authors further try to expand their analysis by incorporating other modalities.

Major points:

1.1 As briefly outlined in the remarks, the current manuscript needs novel findings and methodology to grant a research article which I can' see here. The underlying matrix factorization is the original MOFA (literally imported in the code) with no modification to further optimize the method toward the task. While I appreciate and acknowledge the author's efforts resulting in a detailed analysis of heart samples, I think all of these could have been part of MOFA's existing tutorials.

Response 1.1 As the reviewer correctly states, we used the framework and code of MOFA. The novelty lies in its application for the unsupervised analysis of samples from cross-condition single-cell data and the inference of MCPs. MOFA is a statistical framework implementing a generalization of group factor analysis with fast inference and its current version fits the task of MCP inference and unsupervised analysis of samples across cell-types that provides a more flexible modeling alternative than current available methods (as presented in Table 1 of the manuscript). Current work on MCP inference is based on the premise of multi-view factorization with distinct statistical modeling alternatives. As mentioned in the discussion of our manuscript, three main points distinguish our discussed methodology from present alternatives and provide evidence about its relevance and uniqueness over available tools:

Simultaneous unsupervised analysis of samples across cell-types and inference of MCPs, together with comprehensive interpretable descriptions of the reconstruction of the original multi-view dataset. This is only currently possible with scITD (Mitchel et al, 2022) and is compared in the manuscript. DIALOGUE (Jerby-Arnon & Regev, 2022) is limited to the generation of MCPs and Tensor-cell2cell (Armingol et al, 2021) is only focused in cell-communications with limited interpretability.

Flexible non-overlapping feature set that handles better technical effects such as background expression, as discussed in section “__2.2 Multicellular factor analysis for an unsupervised analysis of samples in single-cell cohorts”. __Moreover, as mentioned by the reviewer in a later point (Reviewer comment 1.2), this enables joint modeling of distinct aspects of the tissue, such as cell compositions, cell communications (preliminary work: https://liana-py.readthedocs.io/en/latest/notebooks/mofatalk.htm) and spatial organization.

Joint-modeling of independent atlases that enables meta-analysis at the sample level of cross-condition single-cell data. No currently available methodology is capable of performing similar modeling. For these reasons, we believe that our work is worth being discussed and presented to the community as a research article. We will modify the discussion to put more emphasis on the added value of group factor analysis as implemented in MOFA.

Moreover, we now provide an R package (https://github.com/saezlab/MOFAcellulaR) and python implementations within our analysis framework LIANA (https://liana-py.readthedocs.io/en/latest/notebooks/mofacellular.html) that facilitates the usage of our proposed methodology. The R and python implementations are compatible with current Bioconductor and scverse pipelines, respectively.

Application of our methodology to heart failure datasets also revealed novel knowledge about heart disease processes:

In myocardial infarction, we found that our MCPs associated with cardiac remodeling capture cell-state-independent gene expression changes. This provides a novel understanding on the effect of disease contexts in the expression profiles of specialized cells. This finding was not reported in the original atlas publication.

In chronic heart failure, we identified a conserved MCP of cardiac remodeling across patient cohorts and etiologies, suggesting a common chronic phase between distinct initial causes of heart failure.

Moreover, we showed that deconvoluted chronic heart failure MCPs from bulk transcriptomics better classify patients in comparison to classic cell-type composition deconvolution of bulk data. To our knowledge, this finding was not presented in any of the manuscripts of other methodologies focused on MCPs.

Altogether, our current work shows a novel application of group factor analysis for the simultaneous estimation of MCPs and the sample-level unsupervised analysis of cross-condition single cell data. We showed the unique features compared to current available tools. Distinct post-hoc analysis in combination with other data modalities shows the biological relevance of our proposed methodology to complement the tissue-centric knowledge of disease.

1.2 How can you explain that the results in donor-level analyses are not due to technical artifacts (batch variation)? Can this be used to infer a new patient similarity map? For example, I would test this by leaving out a few patients from training, projecting them, and seeing where they would end up in the manifold or classifying disease conditions for new patients and explaining the classification by MCPs responsible for that condition.

Response 1.2 When knowledge of the technical batches is available it is possible to test for association between these labels and the factors encoding MCPs as shown in Figure 2.

In our current applications, we additionally showed the biological relevance of our estimated MCPs by mapping them to spatial and bulk data sets, which is a direct way of testing how generalizable were our findings:

In the application of MOFA to human myocardial infarction data, we mapped the gene loadings conforming the MCP associated with cardiac remodeling to paired spatial transcriptomics datasets. We showed that in general, the cell-type specific expression of the MCP of cardiac remodeling encompassed larger areas in ischemic and fibrotic samples compared to myogenic (control) samples.

In the application of MOFA to chronic human end-stage heart failure data, we mapped the gene loadings conforming the MCP associated with cardiac remodeling to 16 independent bulk transcriptomics datasets of heart failure. There we showed that the cell-type specific expression of the MCP of cardiac remodeling separates heart failure patients from control individuals. Regarding the generation of new patient similarity maps, it is possible to estimate the positions of new samples in the manifold formed by the factors representing the MCPs. As suggested by the reviewer we will show this by classifying heart failure single-cell samples using MCPs of two independent patient cohorts (presented in section 2.7).

1.3 The bulk and spatial analysis are used posthoc after running MOFA, I think since MOFA can use non-overlapping features set, it would be interesting to see if deconvoluted bulk or ST data can be encoded as another view (one view from scRNAseq data for each cell-type and another view from bulk RNA-seq or ST, you can get normalized expression per spot (for ST) or per sample (for bulk) and use them as input.

Response 1.3 Thanks for the suggestion. We agree that the possibility of using non-overlapping features opens options of complex models that include the cell-type compositional and organizational aspects of tissues. However these features must be quantified in the same sample, thus it is limited to samples profiled simultaneously at different scales.

We will present the results of a sample-level joint model of multicellular programs together with cell-proportions and spatial dependencies using the myocardial infarction dataset presented in section 2.2. For this dataset based on our previous work we have the compositions of major cell-types and their spatial relationships based on spatially contextualized models (Kuppe et al, 2022). We will run a MOFA model and show how it can be used to find factors associated with structural and molecular features of tissues.

__Minor: __

1.4 Some figure references are not correct (e.g., "the single-cell data into a multi-view data representation by estimating pseudo bulk gene expression profiles for each cell-type across samples (Figure 1b)." should be figure 2b)

Response 1.4 Thanks for pointing this out. We apologize for these mistakes and we will adjust all labels correctly.

1.5 The paper is well written, but there could be some more clarifications about what authors consider as cell-type and cell-state, condition, MCPs which I think is critical to current analysis (see here https://linkinghub.elsevier.com/retrieve/pii/S0092867423001599) for the reader not familiar with those concepts.

Response 1.5 We agree with the reviewer that it is important to introduce these concepts in more detail to avoid confusion. We will adapt the current manuscript to incorporate these definitions in the introduction.

__Reviewer #1 (Significance (Required)): __

1.6 While I find the concept of MCPs interesting, the current work seems like a series of vignettes and tutorials by simply applying MOFA on different datasets (The authors rightfully state this). However, It needs to be clarified what the novelty is since there is no algorithmic improvement to current MCP methods (because there is no new method) nor novel biological findings. Additionally, even in the current form, the applications are limited to the heart, and the generalization of this proposed analysis pipeline to other tissues and datasets is not explored. Overall, the paper lacks focus and novelty, which is required to grant a publication at this level.

Response 1.6 As mentioned in response 1.1, we show that group factor analysis as implemented in MOFA has advantages given its flexibility of the feature space, the joint-modeling of independent datasets, and the interpretability of the model. We will make these advantages clearer in the discussion, and we will explicitly mention the disadvantages and lack of functionalities of available methods.

The applications were mainly done in heart data for consistency although they represent four distinct single-cell datasets, one spatial transcriptomics dataset, and 16 independent bulk transcriptomics datasets. For completeness, as suggested by the reviewer, we will show the application of our methodology to peripheral blood mononuclear cell data of lupus samples (preliminary results: https://liana-py.readthedocs.io/en/latest/notebooks/mofacellular.html)

__expertise: Computational biology, single-cell genomics, machine learning __

__Reviewer #2 (Evidence, reproducibility and clarity (Required)): __

Summary:

The authors use MOFA, an unsupervised method to analyze multi-omics data, to create multicellular programs of cross-condition multi-sample studies. First, for each cell-type, a pseudobulk expression matrix per sample is created. The cell-type now functions as the separate view, typically reserved for the different omics layers in MOFA. This then results in a latent space with a certain number of factors across samples. The factors, representing coordinated gene expression changes across cell-types, can then be checked for associations with covariates of interest across the samples.

MOFA is well-suited for this task, as it can handle missing data and it is a linear model facilitating the interpretation of the factors. Users should be aware that MOFA can estimate the number of factors, but the pseudobulk profiles require a rigorous selection of cell-type specific marker genes. The result will be most suited for downstream analysis if there is a clear association with one factor and a clinical covariate of interest. In a final step, a positive or negative gene signature can be created by setting a cut-off on the gene weights for that specific factor.

The method is applied on 3 separate data sets of heart disease, each time demonstrating that at least one of the factors is associated with a disease covariate of interest. The authors also compare the method to a competitor tool, scITD, and explore to what extent a factor mainly captures variance associated with (i) a general condition covariate or rather (ii) specific cell states.

The multicellular programs are also mapped to spatial data with spot resolution. Though this analysis does not bring any novel biological insight in the use case, it does support the claim that the programs are associated with the covariate of interest.

The most interesting applications of MOFA are in my opinion the potential for meta-analysis of single-cell studies and validation of cell-type specific gene signatures with publicly available bulkRNAseq data sets.

The authors provide various data sets and data types to support their claims and the paper is well written. The relevant code and data has been made available.

We thank the reviewer for the positive comments to our work.

__Major comments __

2.1 What is the added value of the gene signatures obtained from MOFA compared to e.g. a naive univariate approach? In theory, a similar collection of genes or gene signature could be obtained by running a differential gene expression analysis across the samples for each cell-type (e.g. myogenic vs ischemic ) and applying a set of relevant cut-offs or filters on the results. In other words, does MOFA detect genes that would otherwise be missed?

Response 2.1 Thank you for the relevant comment. The original motivation of our work is the unsupervised analysis of samples based on a manifold formed by a collection of multicellular molecular programs. We envisioned that this unsupervised analysis would be relevant in situations where a clear histological or clinical classification of samples is not possible with reliability. As mentioned by Reviewer #1 in comment 1.2, one advantage of these approaches is that they create patient similarity maps, which have been shown useful to stratify patients in a recent analogous work in multiple sclerosis (Macnair et al, 2022). The cell-type signatures obtained from relevant factors explaining the patient stratification avoid the likelihood of performing “double dipping” by avoiding the need of a direct differential expression analysis between newly formed groups.

In our applications, the generation of cell-type signatures (here called multicellular programs) associated to a specific clinical covariate (eg. control vs perturbation) are post-hoc analyses of the generated manifold. And as the reviewer correctly points out, these signatures should be similar to performing direct differential expression analysis between those patient conditions. In the related work of scITD (Mitchel et al, 2022) the authors showed high concordance between the cell-type signatures and the results of differential expression analysis. For completion, we will similarly quantify the degree of overlap between genes of our generated signatures with the ones coming from differential expression analysis.

It is relevant to mention that in complex experimental designs with multiple conditions, our approach facilitates patient ordering, which allows the understanding of one condition in the context of all the others, avoiding the need of multiple testing and the definition of multiple contrasts, as mentioned in the text.

We will incorporate these points in the discussion section of the manuscript.

2.2 Could scITD also be used for meta-analysis or could the obtained gene signatures of that method also be mapped to bulkRNAseq data? If so, it would be interesting to show the relative performance with MOFA. If not, this specific advantage should be highlighted.

Response 2.2 Thank you for pointing this out. scITD does not provide a group-based model to perform meta-analysis, and this feature is one of the main advantages of group factor analysis as currently implemented in MOFA. We will highlight this feature in Table 1 and in the discussion.

Although scITD signatures of a single study could be mapped to bulk transcriptomics data, the stringent tensor representation leads to the generation of signatures that may be influenced by technical effects as shown in the manuscript section 2.2. Thus we believe that the flexibility of the feature space in MOFA is an advantage for this task. We will add this observation to the discussion.

2.3 Users need to specify gene set signatures based on the weights for a factor of interest. This might suggest a limitation to categorical covariates of interest. If the authors see potential for a continuous covariate of interest, this should at least be highlighted in the text and if possible demonstrated on a use case.

Response 2.3 In our applications we limited ourselves to categorical variables, however, it is possible to associate factors to continuous variables. An implementation of the association with continuous variables is already available in our newly created R package “MOFAcellulaR”: https://github.com/saezlab/MOFAcellulaR/blob/main/R/get_associations.R.

The datasets we analyzed have no continuous clinical covariates to showcase this functionality, but as suggested by the reviewer we will highlight this feature in the text.

__Minor comments __

2.4 In Figure 2c the association between factor 2 and the technical factor shows a very strong outlier. Please verify that the association is still significant after applying a more robust statistical test (e.g. non-parametric test as Wilcoxon).

Response 2.4 Thanks for the observation, we will test these differences with a non-parametric test.

2.5 For mapping the cell-type specific factor signatures to bulk transcriptomics, the exact performed comparison or model is unclear. There are seven cell-type signatures for each sample in every study. Was there a t-test run for each cell-type or was a summary measure taken across the cell-types? he thresholding is also rather lenient (adj. p-val 0.1).

Response 2.5 We are sorry for not being clear about our procedure. After identifying the multicellular program associated with heart failure estimated from the two single cell studies meta-analyzed, we calculated the weighted mean expression of the seven cell-type signatures independently to every sample of the 16 bulk studies. In other words each sample within each bulk study will be represented by a vector of 7 values representing the relative expression of a cell-type specific signature (Figure 6D-left). For each bulk transcriptomics study, first, we centered the gene expression data before calculating the weighted mean.

In supplementary figure 4-e we show the results of performing a t-test of the cell-type scores between heart failure and control samples within each study. Given the relative low sample size of most of the studies (affecting the power of the test), we chose a not so stringent adjusted p-value. For completion, we will show the results of a more classical threshold (adj. p-value

2.6 typo in abstract: In sum, our framework serves as an exploratory tool for unsupervised analysis of cross-condition single-cell ***atlas*** and allows for the integration of the measurements of patient cohorts across distinct data modalities

Response 2.6 Thanks for pointing out this typo. We will modify the text.

2.7 In Figure 4a it is not clear to me why on the one hand we see marker enrichment vs loading enrichment with healthy and disease.

Response 2.7 We apologize, this is a typo after editing the labels. Both should contain the marker enrichment label. We will fix this.

2.8 IN Figure 4b it would help if the same color scheme would be maintained throughout the paper (here now black and white) and if for the cell states the boxplots would be connected per condition, emphasizing the (absence) of change across cell states within a condition.

Response 2.8 We thank the reviewer for the suggestion. We will reorganize the panels showing the gene expression per condition and fix the color scheme.

__Reviewer #2 (Significance (Required)): __

__General assessment: __

2.9 MOFA is well-suited for detecting multicellular programs because it can handle missing data and allows for easy interpretation of the factors as a linear method. It might have particular potential for meta-analysis across multiple studies and reevaluating bulkRNAseq data sets, but in the current manuscript it is unclear to what extent this is a specific advantage of MOFA or could also be done with competitors. The authors show how the obtained results and associations with clinical covariates can be validated across multiple data types. How the resulting multicellular programs can provide additional biological insight or form the starting point for additional downstream analysis (e.g. cell communication) is not covered in the paper.

Response 2.9 We thank the reviewer for highlighting the methodological advantages of group factor analysis for the estimation of multicellular programs and the unsupervised analysis of samples from cross-condition single-cell atlas. As mentioned in response 1.1 and 2.2, the added value of our methodology is the flexibility of feature views (that goes beyond gene expression) and simultaneous modeling of independent single-cell datasets, a feature not present in any of the currently available methods that facilitates the meta-analysis of datasets across modalities.

While we interpret the presented multicellular programs in the context of cellular functions and the division of labor of cell states, it is true that we did not attempt to provide mechanistic hypotheses, for example, via cell-cell communication, on how this coordination across cell-types emerges.

Previous work of the related tool Tensor-cell2cell (Armingol et al, 2021) has presented the idea of the estimation of multicellular programs from cell-cell communications and group factor analysis can also be used for this task (preliminary work: https://liana-py.readthedocs.io/en/latest/notebooks/mofatalk.html). We will discuss in the text perspectives on how the estimation of multicellular programs can be linked to the inference of cell communications from single-cell data together with analysis alternatives previously proposed by scITD and Tensor-cell2cell. However, we believe that this question requires further work and it is out of scope of our current manuscript.

__Audience: This paper will be mainly of interest to a specialized public interested in unsupervised methods for large scale multi-sample and multi-condition studies. __

__Reviewer: main background in the analysis of scRNAseq data. __

__Reviewer #3 (Evidence, reproducibility and clarity (Required)): __

This manuscript by Saez-Rodriguez and colleagues proposes to repurpose Multi-Omics Factor Analysis for the use of single cell data. The initial open problem stated by the paper is the need for a framework to map multicellular programs (such as derived from factor analysis) to other modalities such as spatial or bulk data. The authors propose to repurpose MOFA for use in single cell data. Case studies involve human heart failure datasets (and focuses on spatial and bulk comparisons).

There are particular issues with clarity regarding the key methodological contribution (and assessment of it), discussed under significance.

__Reviewer #3 (Significance (Required)): __

3.1 I am very puzzled by the repeated claims the manuscript makes that their central methodological contribution and innovation is to use MOFA for single cell data. One of their citations for MOFA is to MOFA+, which is precisely that (in a relatively popular manuscript published by the original authors of MOFA and not overlapping with the present authors). I am left to wonder what I missed.

Response 3.1 We apologize for the misunderstanding, as mentioned in the response to review 1.1 and explained by reviewer 2’s summary, the main objective of our work is to use the statistical framework of group factor analysis for the inference of multicellular programs and the sample-level unsupervised analysis of cross-condition single-cell data, which is a distinct task to multimodal integration (Argelaguet et al, 2021).

While it is true that MOFA+ introduced expansions to the model for the modeling of single-cell data, namely fast inference and group-based modeling, the main focus in their applications is the multimodal integration of data, where each cell is represented by a collection of distinct collection of features (e.g. chromatin accessibility and gene expression). Unlike multimodal integration, here we propose a different approach to analyze single-cell data at the sample level instead of the cell level, without modifying the underlying statistical model (see section 2.1 of the manuscript).

In detail, what we assume is that samples of single-cell transcriptomics data (e.g. tissue from a patient) can be represented by a collection of independent vectors collecting the gene expression information of cell types composing the tissue analyzed. Decomposition of these multiple views with group factor analysis produces a manifold that captures multicellular programs (coordinated expression processes across cell-types), or shared variability across cell-types simultaneously. Altogether, this represents a novel usage of group factor analysis in an application for the inference of multicellular programs, where the main focus is not at the cell-level but at the patient level.

As a side note, Britta Velten, one of main developers of MOFA and coauthor of both the MOFA and MOFA+ papers, is a contributor and coauthor of this manuscript, and Ricard Argelaguet, who also led both versions of MOFA, gave us helpful feedback and is acknowledged as such on this work.

3.2 Multimodal integration methods are fairly numerous and even if they're not all exactly factor analyses, it's strange to argue that MOFA fills some unique conceptual gap. I agree it fills something of an interesting gap (except for MOFA+ already filling it), but it's not like the quite popular spatial to single-cell integration approaches aren't doing similar things. If this is a methods paper (as it is presented) then there would have to be very substantially more comparative evaluation to these other approaches.

Response 3.2 As presented in the previous response (3.1) our current work is not focused on multimodal integration, but rather the inference of multicellular programs and the sample-level unsupervised analysis of single-cell data. Given this, in the current manuscript we compared our proposed methodology with the only three other available methods that address at least partially the inference of multicellular programs (see Table 1 in our manuscript). In response 1.1 and 3.2 we discussed the advantages of our proposed methodology compared to available methods. In the manuscript section 2.2 we compared group factor analysis with tensor decomposition and showed that the former better deals with technical artifacts and better identifies known patient groups.

We will distinguish our work from multimodal integration explicitly in the introduction and the manuscript section 2.1 to avoid confusions.

3.3 The biological use cases are comparatively interesting and dominate the manuscript (but are still presented principally as use cases rather than a compelling biological narrative of their own).

Response 3.3 The focus of our manuscript was the reintroduction of group factor analysis for the novel applications of the inference of multicellular programs and the sample-level unsupervised analysis from single-cell data. Given the distinct possibilities of post-hoc analyses, we mainly used acute and chronic heart failure data to showcase the utility of MOFA to connect spatial and bulk modalities with single-cell data.

That said, as discussed in response 1.1, our analyses allowed to generate novel hypotheses of these datasets:

In myocardial infarction, we found that our estimated multicellular programs associated with cardiac remodeling capture cell-state-independent gene expression changes. This provides a novel understanding of the effect of disease contexts in the expression profiles of specialized cells. In other words, we found that cell-states, regardless of their specialized function, share a common response in the tissue context.

In chronic heart failure, we identified a conserved multicellular program of cardiac remodeling across patient cohorts and etiologies, suggesting a common chronic phase between distinct initial causes of heart failure, which again may be linked to the dominating response to the tissue context that is shared across etiologies.

These two results support the observation that deconvoluted chronic heart failure multicellular programs from bulk transcriptomics better classify patients in comparison to classic cell-type composition deconvolution of bulk data. To our knowledge, this finding was not presented in any of the manuscripts of other methodologies focused on MCPs. We summarize these results in the third paragraph of the discussion in the manuscript:

“In an application to a collection of public single-cell atlases of acute and chronic heart failure, we found evidence of dominant cell-state independent transcriptional deregulation of cell-types upon myocardial infarction. This may suggest that while certain functional states within a cell-type are more favored in a disease context, most of the cells of a specific type have a shared transcriptional profile in disease tissues. If part of this shared transcriptional profile is interpreted as a signature of the tissue microenvironment that drives cells in tissues towards specific functions, this result may also indicate that a major source of variability across tissues, besides cellular composition, is the degree in which the homeostatic transcriptional balance of the tissue is disturbed. By combining the results of multicellular factor analysis with spatial transcriptomics datasets, we explored this hypothesis and identified larger areas of cell-type-specific transcriptional alterations in diseased tissues. Given these observations on global alterations upon myocardial infarction, we meta-analyzed single-cell samples from two additional studies of healthy and heart failure patients with multiple cardiomyopathies. Here, we found a conserved transcriptional response across cell-types in failing hearts, despite technical and clinical variability between patients. Further, we could find traces of these cell-type alterations in independent bulk data sets. These observations suggest that our approach can estimate cell-type-specific transcriptional changes from bulk data that, together with changes in cell-type compositions, describe tissue pathophysiology. Altogether, these results highlight how MOFA can be used to integrate the measurements of independent single-cell, spatial, and bulk datasets to measure cell-type alterations in disease.”

To fully assess the relevance of these observations, they should be investigated in more datasets and analyses, where shared functional cell-states across distinct heart failure etiologies are identified and then compared at their compositional and molecular level. This, in our opinion, represents an independent study on its own.

3.4 Altogether, I found the framing of this manuscript very puzzling. It is possible the result would be more clearly presented if the use case was the major focus rather than the more conceptual point about factor analysis.

Response 3.4 Thanks for the suggestion. The major aim of this manuscript is to highlight the versatility of the generalization of group factor analysis as implemented in MOFA for novel applications in single-cell data analysis, beyond multimodal integration of single cells. The definition of multicellular programs from single-cell data and its sample-level unsupervised analysis are relatively new analyses in the field, and thus we believe that it is timely to show how a known statistical framework can be used for these applications.

We believe that a detailed analysis of single-cell datasets of heart failure deserves its own focus and it is out of scope of our current objective with this manuscript. We apologize for the apparent misunderstanding of the objective of our methodology. We will add these distinctions in the introduction of the manuscript.

References

Argelaguet R, Cuomo ASE, Stegle O & Marioni JC (2021) Computational principles and challenges in single-cell data integration. Nat Biotechnol 39: 1202–1215

Armingol E, Baghdassarian H, Martino C, Perez-Lopez A, Knight R & Lewis NE (2021) Context-aware deconvolution of cell-cell communication with Tensor-cell2cell. BioRxiv

Jerby-Arnon L & Regev A (2022) DIALOGUE maps multicellular programs in tissue from single-cell or spatial transcriptomics data. Nat Biotechnol 40: 1467–1477

Kuppe C, Ramirez Flores RO, Li Z, Hayat S, Levinson RT, Liao X, Hannani MT, Tanevski J, Wünnemann F, Nagai JS, et al (2022) Spatial multi-omic map of human myocardial infarction. Nature 608: 766–777

Macnair W, Calini D, Agirre E, Bryois J, Jaekel S, Kukanja P, Stokar-Regenscheit N, Ott V, Foo LC, Collin L, et al (2022) Single nuclei RNAseq stratifies multiple sclerosis patients into three distinct white matter glia responses. BioRxiv

Mitchel J, Gordon MG, Perez RK, Biederstedt E, Bueno R, Ye CJ & Kharchenko P (2022) Tensor decomposition reveals coordinated multicellular patterns of transcriptional variation that distinguish and stratify disease individuals. BioRxiv

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

Learn more at Review Commons

Referee #2

Evidence, reproducibility and clarity

Summary:

The authors use MOFA, an unsupervised method to analyze multi-omics data, to create multicellular programs of cross-condition multi-sample studies. First, for each cell-type, a pseudobulk expression matrix per sample is created. The cell-type now functions as the separate view, typically reserved for the different omics layers in MOFA. This then results in a latent space with a certain number of factors across samples. The factors, representing coordinated gene expression changes across cell-types, can then be checked for associations with covariates of interest across the samples. MOFA is well-suited for this task, as it can handle missing data and it is a linear model facilitating the interpretation of the factors. Users should be aware that MOFA can estimate the number of factors, but the pseudobulk profiles require a rigorous selection of cell-type specific marker genes. The result will be most suited for downstream analysis if there is a clear association with one factor and a clinical covariate of interest. In a final step, a positive or negative gene signature can be created by setting a cut-off on the gene weights for that specific factor. The method is applied on 3 separate data sets of heart disease, each time demonstrating that at least one of the factors is associated with a disease covariate of interest. The authors also compare the method to a competitor tool, scITD, and explore to what extent a factor mainly captures variance associated with (i) a general condition covariate or rather (ii) specific cell states. The multicellular programs are also mapped to spatial data with spot resolution. Though this analysis does not bring any novel biological insight in the use case, it does support the claim that the programs are associated with the covariate of interest. The most interesting applications of MOFA are in my opinion the potential for meta-analysis of single-cell studies and validation of cell-type specific gene signatures with publicly available bulkRNAseq data sets. The authors provide various data sets and data types to support their claims and the paper is well written. The relevant code and data has been made available.

Major comments

• What is the added value of the gene signatures obtained from MOFA compared to e.g. a naive univariate approach? In theory, a similar collection of genes or gene signature could be obtained by running a differential gene expression analysis across the samples for each cell-type (e.g. myogenic vs ischemic ) and applying a set of relevant cut-offs or filters on the results. In other words, does MOFA detect genes that would otherwise be missed?
• Could scITD also be used for meta-analysis or could the obtained gene signatures of that method also be mapped to bulkRNAseq data? If so, it would be interesting to show the relative performance with MOFA. If not, this specific advantage should be highlighted.
• Users need to specify gene set signatures based on the weights for a factor of interest. This might suggest a limitation to categorical covariates of interest. If the authors see potential for a continuous covariate of interest, this should at least be highlighted in the text and if possible demonstrated on a use case.

Minor comments

• In Figure 2c the association between factor 2 and the technical factor shows a very strong outlier. Please verify that the association is still significant after applying a more robust statistical test (e.g. non-parametric test as Wilcoxon).
• For mapping the cell-type specific factor signatures to bulk transcriptomics, the exact performed comparison or model is unclear. There are seven cell-type signatures for each sample in every study. Was there a t-test run for each cell-type or was a summary measure taken across the cell-types? he thresholding is also rather lenient (adj. p-val 0.1).
• typo in abstract: In sum, our framework serves as an exploratory tool for unsupervised analysis of cross-condition single-cell atlas and allows for the integration of the measurements of patient cohorts across distinct data modalities
• In Figure 4a it is not clear to me why on the one hand we see marker enrichment vs loading enrichment with healthy and disease.
• IN Figure 4b it would help if the same color scheme would be maintained throughout the paper (here now black and white) and if for the cell states the boxplots would be connected per condition, emphasizing the (absence) of change across cell states within a condition.

Significance

General assessment:

MOFA is well-suited for detecting multicellular programs because it can handle missing data and allows for easy interpretation of the factors as a linear method. It might have particular potential for meta-analysis across multiple studies and reevaluating bulkRNAseq data sets, but in the current manuscript it is unclear to what extent this is a specific advantage of MOFA or could also be done with competitors. The authors show how the obtained results and associations with clinical covariates can be validated across multiple data types. How the resulting multicellular programs can provide additional biological insight or form the starting point for additional downstream analysis (e.g. cell communication) is not covered in the paper.

Audience: This paper will be mainly of interest to a specialized public interested in unsupervised methods for large scale multi-sample and multi-condition studies.

Reviewer: main background in the analysis of scRNAseq data.

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

Learn more at Review Commons

Referee #1

Evidence, reproducibility and clarity

Remark to authors

Flores et al. present a pipeline in which they leverage MOFA framework, a matrix factorization algorithm to infer multi-cellular programs (MCPs). Learning and using MCP has already been proposed by others. Yet, authors pursue a similar goals by using MOFA, providing a cell*sample matrix for different cell types as different views (instead of multiple modalities/views) as the input. They later apply MOFA using this data format on a series of applications to analyze acute and chronic human heart failure single-cell datasets using MCPs. Authors further try to expand their analysis by incorporating other modalities.

Major points:

As briefly outlined in the remarks, the current manuscript needs novel findings and methodology to grant a research article which I can' see here. The underlying matrix factorization is the original MOFA (literally imported in the code) with no modification to further optimize the method toward the task. While I appreciate and acknowledge the author's efforts resulting in a detailed analysis of heart samples, I think all of these could have been part of MOFA's existing tutorials.

How can you explain that the results in donor-level analyses are not due to technical artifacts (batch variation)? Can this be used to infer a new patient similarity map? For example, I would test this by leaving out a few patients from training, projecting them, and seeing where they would end up in the manifold or classifying disease conditions for new patients and explaining the classification by MCPs responsible for that condition.

The bulk and spatial analysis are used posthoc after running MOFA, I think since MOFA can use non-overlapping features set, it would be interesting to see if deconvoluted bulk or ST data can be encoded as another view (one view from scRNAseq data for each cell-type and another view from bulk RNA-seq or ST, you can get normalized expression per spot (for ST) or per sample (for bulk) and use them as input.

Minor:

Some figure references are not correct (e.g., "the single-cell data into a multi-view data representation by estimating pseudo bulk gene expression profiles for each cell-type across samples (Figure 1b)." should be figure 2b)

The paper is well written, but there could be some more clarifications about what authors consider as cell-type and cell-state, condition, MCPs which I think is critical to current analysis (see here https://linkinghub.elsevier.com/retrieve/pii/S0092867423001599) for the reader not familiar with those concepts.

Significance

While I find the concept of MCPs interesting, the current work seems like a series of vignettes and tutorials by simply applying MOFA on different datasets (The authors rightfully state this). However, It needs to be clarified what the novelty is since there is no algorithmic improvement to current MCP methods (because there is no new method) nor novel biological findings. Additionally, even in the current form, the applications are limited to the heart, and the generalization of this proposed analysis pipeline to other tissues and datasets is not explored. Overall, the paper lacks focus and novelty, which is required to grant a publication at this level.

expertise: Computational biology, single-cell genomics, machine learning

This doesn't sound to me like they have actually been doing pair programming as I have always understood it. Neither participant needs to "engage" in those (admittedly distracting) things—least of all the person at the keyboard.

In pair programming as I have had it laid out—and not as a consequence of hearing "pair programming" and extrapolating or assuming what it involves—one person is writing the code just like when they're alone, except they're not actually controlling the computer. That's the other person's job. The first person is controlling the person who is controlling the computer. Part of the job of the second person involves shutting the fuck up and just following what the other person is saying to do. This pattern only ever breaks when the pair decides to switch places or the person dictating runs into an issue, at which point the person at the keyboard (who has been thinking all the while as an observer of what the two have been producing and is expected to know what the problem is, having already recognized the problem the first time around) should speak up. When switching roles or after reaching milestones, the two can confer about high-level concerns,immediate and distant plans to deal with things overlooked or set aside in the last round, etc.

I am aware that "two people working at a single computer" is how most people understand pair programming (and that there seems to be academic work covering the topic which lays it out in a way that contradicts what I've described here), but I regard that as wrong—for all the obvious reasons, including and especially the ones described by the commenter here...

可以对部分列进行操作，不需要对所有列都进行操作（填补缺失值）

I like this aspect espeically within coding because there are so many different ways to create and buld these functions. This gives engineers a lot of space to be creative. The resuse of code can be confusing and may even cause errors in the long run

I think this is yet to be seen. Although it is true that the computer always gives the same output given the same input code, a biased network with oppressive ideologies could simply transform, instead of change, our current human judiciary enforcement of the law.

harmful

appdef needs to be changed here, paraprobe-toolbox==0.4 currently reports ROI without a clear group (which should be changed in the paraprobe-nanochem source code) but also the fields signed_distance and isotope are not scalars but arrays of the same length [k] although that length may different for each instance of a ROI

It would be possible to add a default plot for every single ROI in this way the classical proxigram could be displayed directly using H5Web, the key overhead which comes with this is the need for very many small data entries which likely blow up the HDF5 file unnecessarily, for this reason currently such default plot is not exported

this can be considered a bug in both the appdef and the implementation of paraprobe-parmsetup-nanochem: paraprobe-toolbox=v0.4 expects here a field named isotope_whitelist(NX_INT) of length (n_filter, 32) the docstring/meaning of decorating_iontypes_filter group is correct and how v0.4 interprets the isotope_whitelist e.g. setting a isotope_whitelist of 5, 0 ... (with 32-1 zeros following) instructs the code to take all molecular ions with boron and assume that these ions decorate the defect, a subset of the point cloud including these ions is computed and the interface model fitted

Author Response

Reviewer #1 (Public Review):

The introduction does not clearly set up the background for the key questions that the manuscript addresses. One of the key parts of the manuscript is to attempt to determine whether locomotory behaviour evolves because of direct or indirect selection of the traits. However, the authors don't provide an argument for why a salty environment would select for locomotory traits. Indeed, in the discussion, the authors point out that it is likely an unmeasured trait (body size) correlated with locomotory traits that are under selection. They present arguments for why this might be the case and point to un-included data that show body size significantly genetically covaries with all of the traits studied. Since the authors appear to have these data, and one of their key questions is comparing direct vs. indirect responses to selection, it would be more powerful to include the body size data and estimate selection on all traits together.

We now include body size in all of our phenotypic and genetic analyses. We also include estimates of selection gradients from the ancestral selection differentials and the Gmatrix. We detail in the Introduction the biological significance of locomotion traits and their potential relationship with body size, in low and high salt environments. The experimental results show that divergence in locomotion traits (Figure 6) correlates with adaptation (Figure 5), because of direct and indirect selection (Figure 9).

Phenotypic plasticity was estimated from a series of univariate models, with estimates arranged in a vector. As the authors point out in the manuscript, traits that are not included in a model but covary with traits that are can largely bias estimates of the traits that are included. For this reason, it would make sense to estimate phenotypic plasticity using a multivariate model, as has been done for G matrices.

We analyze the ancestral phenotypic plasticity and the phenotypic divergence during evolution using a multivariate approach (MANOVA). This approach simplifies the text as from the eigen decomposition of the SSCP matrices we can estimate canonical traits of ancestral phenotypic plasticity (pmax; see Table 1 with notation definitions) and phenotypic divergence in the new target high salt environment (dmax). We continue to do the univariate analysis as it allows us to estimate BLUPs for each inbred line (used for visual representation), as well as the significance of phenotypic divergence at each replicate population relative to the ancestral population (delta_q). Both multivariate and univariate approaches led to similar results (shown as supplementary figures).

The estimation and interpretation of G matrices are a critical part of the manuscript. The authors state that broad sense estimates of G are a good proxy for additive genetic variation in this system, but in the Discussion they also state that overdominance was likely important during evolution to the salt environment, leading to some lack of clarity on whether dominance is important or not.

We are sorry for the lack of clarity. We have eliminated the discussion on overdominance as it was peripheral to our results. Broad-sense genetic variances should be a good proxy for additive genetic variances when there is no inbreeding depression and no directional dominance or dominance epistasis; cf. Lynch and Walsh 1998. We previously showed that there is no inbreeding depression for the trait we use as surrogate for relative fitness (self-fertility) and also that there is no directional dominance for locomotion behavior traits. We now explain our use of broad-sense genetic (co)variances as a proxy for additive genetic (co)variances in the Introduction and Methods.

It is also unclear how uncertainty in estimated G matrices was assessed. Showing that G differs from noise is critical to the majority of the results presented. The authors cite Morrissey and Bonnet (2019) as providing the method for generating the null distribution of G, however, this paper does not appear to propose or describe a method to do this.

Thanks for this comment. Morrissey and Bonnet (J Heredity, 2019) was incorrectly cited and the explanation for finding the expected noise distributions was misleading. In brief, we produced a set of 1000 G-matrices each computed after shuffling the line ID and the block ID from the phenotypic dataset. This was done to produce random expectations of the genetic variances as the MCMC estimates are positive-definite. We computed the posterior mode for each of these 1000 G-matrices to obtain a null distribution (shown in orange). To infer significance, we compared the posterior mode of the empirical estimate with the 95% CI of the posterior mode distribution obtained from the randomized G-matrices. When determining which eigenvectors explain standing genetic variation we also used the distribution of posterior modes of the randomized G-matrices. However, as pointed out by Sztepanacz and Blows (Genetics, 2017), the eigenvalues of the eigenvectors do not follow a uniform distribution, as would be expected by chance. Because of this we asked the question of whether the amount of variance in the eigenvectors of the empirical G-matrix (gmax, g2, etc.) was expected, by projecting the random G-matrices onto these eigenvectors. This is a null that is conditional on the observed data. We show these results in Figure 2 - supplement figure 3. Both approaches are similar, particularly for the first 2 eigenvectors. There is now a paragraph in the Discussion about finding potential consequences for adaptation of traits with little genetic variance.

Although the figure captions state that they are showing estimates of genetic variances, it appears to be heritability (bounded between 0 and 1). Whether the authors are studying heritability or genetic variance is an important difference, particularly in the context of a changing environment and phenotypic plasticity, where environmental variation is important and expected to change. For example, the result that G is smaller in evolved populations could simply be due to their being larger environmental variance in the salt environment (as you would expect). This is unrelated to an evolutionary response.

There might have been some confusion because transition rates are positive and not normally distributed. To achieve normality they were log transformed. We have not reported estimates of heritability, all estimates presented are of genetic variances, unscaled. The only exception is body size where the raw data was multiplied by 50 in order to have a similar phenotypic scale as the transition rates when estimating genetic (co)variances, not heritability. We agree that the evolution of environmental stochastic variance is interesting but not immediately relevant to the questions we address.

It seems that comparisons to the ancestral population were done for A160, not the founding population for each evolved line at G0. It is not clear whether the founder effects of each replicate are important and if this is the most appropriate comparison (the Discussion suggests that founder effects are important).

We have better detailed in the Methods, and also with an introductory section in the Results section, the derivation of the experimental populations. The population acronyms might have been misleading. The A6140 is a population that was domesticated to the lab conditions for 140 generations (replicate #6 of the domestication process). We report the evolution of 3 GA populations, which were all derived from A6140 with minimal sampling problems for the estimated effective population sizes (sampled 10^4 individuals from A6140 for each GA, for Ne of 1000 during domestication - Chelo and Teotónio Evolution 2013 -). Therefore, GA populations after 50 generations of evolution are appropriately compared with their (unique) ancestor population. We no longer discuss potential founder effects.

Overall, there is much interesting data collected and analysed in this manuscript, addressing a valuable question. However, it is not obvious whether the estimates of G matrices are different from noise, and heritability may not be the most appropriate scale to ask questions about phenotypic plasticity and evolution in a novel stressful environment that may affect levels of environmental variation.

Please see previous replies. Our ancestral G-matrix estimates indicate that at least 3 eigentraits are different from random expectation in both environments (Figure 2, supplement figure 3), and in high salt evolved populations continue to have more than expected genetic variance at 3-5 eigentraits (Figure 7, supplement 2). We are conservative in these estimates as depending on the null we could consider more eigentraits. In the previous version of the manuscript we concluded that only 2 ancestral eigentraits were orthogonal due to an error in the code (we did not divide by 2 the null expectations). But even presuming that only one eigentrait (gmax) has genetic variance in the ancestral population, we previously reported that mutational variance is not in the same trait (see Mallard et al., G3, 2023; and mmax in Table 3), and further that the trait under selection is neither gmax or mmax (compared in Table 3 the selection gradients with gmax or mmax). At a minimum there are 3 genetically or environmentally independent traits. As noted in previous replies, we estimate and present genetic variances throughout. We do not present estimates of environmental variances and feel that doing so would make the manuscript overly complicated.

Reviewer #2 (Public Review):

Response to selection: It was not clear to me that it was appropriate to interpret locomotor behavior as having evolved in response to the salinity environment. Specifically, where is the evidence that any change in trait means is a (direct or indirect) response to selection imposed by increased salinity rather than the neutral drift of a trait due to the reduction in population size caused by the salinity? Strong evidence of adaptive evolution would be provided by all 3 replicates significantly diverging from the ancestor in the same direction. Model 2 seems to aim to test the null hypothesis that the three replicates diverged from one another via a random effects model - but with only three replicates, there is very low power, and variance is likely to be estimated as zero. I'm not sure what is shown in Tables 3 & 4, or how these results relate to models 2 & 3, so my interpretation of the information may be incorrect. Nonetheless, and noting that the errors around estimates are not presented, there seems to be considerable heterogeneity in size and direction of divergence between replicates for most of the traits. Is this study really dissecting responses to directional selection, or is it dissecting drift?

We have modified the statistical modelling of the phenotypic data. Model 2 is no longer presented. We provide a MANOVA multivariate analysis equivalent to model 2 (with replicate populations as fixed effects) but now including both environments, together with the univariate models. MANOVA results show that all traits are significantly different across populations (i.e., at least two populations differ from one another). The fitted estimates from the MANOVA are not reported with errors in R but it is obvious that not all traits evolved in each replicate GA population (Figure 6). We therefore tested the difference between each of the evolved populations and the ancestral population using a univariate approach (Figure 6, supplemental source data table 2). In this univariate analysis, block was modeled as having random effects (which we could not model with MANOVA). In the high salt environment, the replicates GA 1,2,4 differed significantly for respectively 4, 6 and 4 transitions rates (out of 6). The traits are all evolving in the same direction, and this even when the trait difference between evolved and ancestral populations is not significant. We provide compelling evidence of parallel evolution and thus selection (see review about how to infer selection in evolution experiments in Teotónio et al. Genetics 2017). We tried to be exhaustive in our statistical reporting but would happily provide additional details if requested.

What are the traits, and what is the confidence in G? My outsider's interpretation of these results is that defining 6 transition states is a way of getting at a single behavioral trait, and I was not convinced that these data were suitable for addressing questions about multivariate evolution. Genetic parameters were estimated using MCMCglmm, which imposes boundaries on estimates. The authors state that they followed Morrissey and Bonnet 2019, but I was unable to infer what this means with respect to accounting for the contribution of sampling error to covariances (or how they accounted for the positive variance constraint). Because I was unsure how sampling error was being assessed for G, I was not confident about the interpretation of statistical support for individual parameters, or for eigenvalues of G. Following this forward, if the measured characteristics constitute a single trait, with an entirely shared genetic basis, then the results of strong alignment of everything with gmax makes complete sense - there is a single trait, that is heritable and plastic, and for which the mean evolved.

Our initial draft was misleading and we now provide more detailed description (see also replies #5 and #12 above). We computed 1000 randomized G matrices to account for the constraints imposed by the MCMCglmm algorithm. This should account for the bias inherent with variance estimation and the eigen decomposition we did given our sample sizes. You will find that all 6 transition rates show genetic variance (Figure 2, supplement figure 2) and that up to three eigentraits have more genetic variance than the randomized G-matrices (Figure 2, supplement figure 3).

The 6 transition rates are the mathematical description of changing movement states in 1-dimensional space (under memoryless assumptions). A priori we do not know how many relevant traits there are, if they are genetically or environmentally independent. To help the reader, we provide a Table 3 with the trait loadings for the several canonical traits of phenotypic plasticity, divergence and selection. The first canonical trait of standing genetic variation, gmax, is indeed aligned with phenotypic divergence (dmax; Figure 8, panels A and B) and with the axis of genetic variance reduction during evolution (emax; Figure 8, panels C and D), but not with ancestral plasticity (pmax; Figure 3) or mutational variance (mmax, from Mallard et al. G3 2023). pmax, for example, is aligned with g3, the third eigenvector of the ancestral G matrix. Note, however, that we do not have any power to detect the influence of g2 or g3 on phenotypic divergence or genetic divergence (Figure 8), though they together explain about 15% of the genetic variance. This is because performing such a test would require an alignment of the deviations in divergence not explained by gmax with g2 or g3. We now mention this issue in the Discussion. Overall, however, there are clearly several behavioral traits.

AB: Should we color-code this table to make it more clear that these are associated to either the forebrain, midbrain and hindbrain. I.e., first 3 rows one color for forebrain, next 3 rows another color for midbrain, and last 3 rows a different color for hindbrain.

Reviewer #2 (Public Review):

In this work Ushio et al. combine environmental DNA metabarcoding with novel statistical approaches to demonstrate how fish communities respond to changing sea temperatures over a seasonal cycle. These findings are important due to the need for new techniques that can better measure community stability under climate change. The eDNA metabarcoding dataset of 550 water samples over two years is, I feel, of sufficient scale to provide power to detect fine-scale ecological interactions, the experiments are well controlled, and the statistical analysis is thorough.

The major strengths of the manuscript are: (1) the magnitude of the dataset, which provides densely replicated sampling that can overcome some of the noise associated with eDNA metabarcoding data and scale up the number of data points to make unique inferences; (2) the novel method of transforming the metabarcode reads using endogenous qPCR "spike-in" data from a common reference species to obtain estimates of DNA concentration across other species; and (3) the statistical analysis of time-series and network data and translating it into interaction strengths between species provides a cross-disciplinary dimension to the work.

I feel like this kind of study showcases the power of eDNA metabarcoding to answer some really interesting questions that were previously unobtainable due to the complexities and cost of such an exercise. Notwithstanding the problems associated with PCR primer bias and PCR stochasticity, the qPCR "spike-in" method is easy to implement and will likely become a standardised technique in the field. Further studies will examine and improve on it.

Overall I found the manuscript to be clear and easy to follow for the most part. I did not identify any serious weaknesses or concerns with the study, although I am not able to comment on the more complex statistical procedures such as the "unified information-theoretic causality" method devised by the authors. The section on limitations of the study is important and acknowledges some issues with interpretation that need to be explained. The methods, while brief in parts, are clear. The code used to generate the results has been made available via a GitHub repository. The figures are clear and attractive.

The bank code.

Comments make the code intuitive.

Yes, this is very concerning. This is a reason to be worried, and a reason to be concerned that AI alignment efforts are unlikely to succeed.

this guy was arrested in 1964 with a law meant to target a specific movement in 1865???

https://hyp.is/sfJJkNz7Ee2wLk808A8T7g/opensource.googleblog.com/2023/03/introducing-service-weaver-framework-for-writing-distributed-applications.html

What he is saying is that IT is a total disaster. He does not say, that it is deliberate, but 60, 50 years ago there were all the germs if the ideas that we needed. In 85 I went back to 20 years earlier. doing some computing archeology to find them. A better future had been invented back then. Those ideas were already buried under detritus and the worse is better. I am sorry to say, but they were right, There is a good way of going about things and there is the mess we are in. I admire the subtle way he is conveying the message, that our present is a total insanity

while computer's capacity grows exponentially we are engineering to waste human brain power units at exponentially growing ways. We are building APIs which is insane

Programming today is the opposite of diamond mining. In diamond mining you dig up a lot of dirt to find a small bit of value. With programming you start with the value, the real intention, and bury it in a bunch of dirt. - Charles Simonyi

bret victor

Colorado Department of Health Care Policy and Financing. (2020, June 4). COVID-19 Guidance for Pharmacies 6/4/2020. https://hcpf.colorado.gov/sites/hcpf/files/COVID%20Guidance%20for%20Pharmacies%20060420.pdf

landmark vector cells

This ending part overviews his perspective as a software engineer. He mentions how "my job is to alleviate frustration of real people and help them lead more enjoyable and productive lives." He also mentions a goal and what he wants to accomplish after he's all done with his career.

This seems counter-education, as foreign language is a requirement for many high schools anyway. There should be a way to better accommodate people who were raised with a different language rather than simply erasing this part of their identity.

• generating output as a side effect of "recognizing a shape"-

• rule actions to be defined in RDF itself