Author Response

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

We thank the reviewers and editors for their constructive comments on the manuscript. We have extensively revised the manuscript based on these concerns and comments. The followings are the specific answers.

Public Reviews:

Reviewer #1 (Public Review):

In the manuscript "Long‐read single‐cell sequencing reveals expressions of hypermutation clusters of isoforms in human liver cancer cells", S. Liu et al present a protocol combining 10x Genomics single-cell assay with Element LoopSeq synthetic long-read sequencing to study single nucleotide variants (SNVs) and gene fusions in Hepatocellular carcinoma (HCC) at single‐cell level. The authors were the first to combine LoopSeq synthetic long‐read sequencing technology and 10x Genomics barcoding for single cell sequencing. For each cell and each somatic mutation, they obtain fractions of mutated transcripts per gene and per each transcript isoform. The manuscript states that these values (as well as gene fusion information) provide better features for tumor-normal classification than gene expression levels. The authors identified many SNVs in genes of the human major histocompatibility complex (HLA) with up to 25 SNVs in the same molecule of HLA‐DQB1 transcript. The analysis shows that most mutations occur in HLA genes and suggests evolution pathways that led to these hypermutation clusters. Yet, very little is said about novel isoforms and alternative splicing in HCC cells, differences in isoform ratio between cells carrying different mutations, or diversity of alternative isoforms across cells. While the manuscript by Liu et al. presents a promising combination of technologies, it lacks significant insights, a comprehensive introduction, and has significant problems with data description and presentation.

Answer: Thanks for the precious suggestion. Our long-read single-cell sequencing has discovered an average of 442 novel isoform transcripts per benign liver cell and 450 novel isoform transcripts per HCC cell per SCANTI v1.2 analysis. These are stated in the revised manuscript. The alternative splicing was detected by differential isoform expression as demonstrated in supplemental figures 6 and 7 and supplemental tables 8-11. The examples of differences in isoform ratio between cells carrying different mutations are now shown by DOCK8 and STEAP4 (figure 5 in the revised manuscript). A new section was added in the results to discuss the mutation expression of these two genes. The diversity of isoforms of the selected genes is shown in Supplemental Figure 10.

This study showed how mutations in the same allele evolved in liver cancer. In particular, HLA hypermutations were found to develop from some specific sites of the molecules into large clusters of mutations in the same molecules. A new paragraph of introduction was added about the role of mutations in human cancer development. We also revised the figures to present the information better. All the HLA genes expressed only one known isoform, as shown in Figure 4 and Supplemental Figure 3, regardless of mutations.

Major comments:

1. The introduction section is scarce. It lacks description of important previous works focused on clustered mutations in cancers (for example, PMID35140399), on deriving the process of cancer development through somatic evolution (PMID32025013, from single cell data PMID32807900). Moreover, some key concepts e.g. mutational gene expression and mutational isoform expression are not defined. The introduction and the abstract contain slang expressions e.g. "protein mutation', a combination of terms I teach my students not to use.

Answer: We appreciate the reviewer for the idea of more solid background introduction and term definition. We added a new paragraph in the introduction section to introduce the role of mutations and hypermutations in human cancers. Some important work has been cited. We added a new section in the "Methods" to define "mutation gene expression share" and "mutation isoform expression share". "Protein mutation" has been replaced by "genetic mutation".

1. In the results section, to select the mutations of interest, the authors apply UMAP dimensionality reduction to the mutation isoforms expression and cluster samples in UMAP space, then select the mutations that are present only in one cluster, then apply UMAP to the selected mutations only and cluster the samples again. The motivation for such a procedure seems unclear, could it be replaced with a more straightforward feature selection?

Answer: Thanks for raising up this important question. The goal of the analysis is an unbiased classification of the cell populations in the samples. We found that by removal of mutated isoform expressions that were at similar levels of all cells, the UMAP clustering generated clear segregation of three population cells. When the unique mutated isoform expressions from each group were applied, it generated highly distinct 8 groups of cells, with each group having a distinct mutation isoform expression pattern. If we force known knowledge into the mix of the analysis, it may generate unwanted bias. Specifically, the first UMAP was performed in an unbiased way to cluster cells, while the second step is a supervised approach by selecting the unique mutations in each cluster to identify the classifiers. The second UMAP matches the Benign/HCC labeling well.

1. As I understand, the first "mutated isoform"-based UMAP clustering was built from expression levels of 205 "mutational isoforms". What was the purpose and outcome of the second "mutated isoform"based UMAP clustering (Figure 2E)? In the manuscript the authors just describe the clusters and do not draw any conclusions or use the results of the clustering anywhere further.

Answer: Thanks for pointing this out. Figure 2E was generated from unique mutation isoform expressions in groups A, B, and C from Figure 2D. The purpose of Figure 2E is to investigate whether these unique mutation isoforms can further classify the cell populations free of prior biological knowledge. We added a sentence in the revision to clarify the purpose of the clustering. The conclusion from this analysis, including Figure 2F and Figure 3 (which is an extension of Figure 2E), is that HLA mutation isoform expressions dominated the classifications of cell populations.

1. The authors just cluster the data three times based on expression levels of different sets of "mutational isoforms" and describe the clusters. What do we need to gather from these clustering attempts besides the set of 113 mutations used for further analysis? What was the point of the reclusterings? Did the authors observe improvement of the classification at each step?

Answer: Thanks for asking this important question. The improvement of re-clustering to classify cell populations is the obvious segregation of 8 different groups of cells without any manual classification through prior knowledge. The distances among groups were far apart in comparison to the first clustering (figure 2B). Detailed subclassifications were achieved on cell populations that otherwise could not be segregated based on the first clustering.

1. The alignment of short reads generated from hypermutated transcriptomes is non-trivial. The proposed approach could address the issue without the need for whole genome sequencing and offer insights about the cancer development through somatic evolution. Why didn't the authors use modern phylogenetic approaches in the "Evolution of mutations in HLA molecules" section or at least utilize the already performed clustering to infer cell lineages?

Answer: We appreciate for the great question. For a single molecule mutation evolution, single gene clustering may not produce a desirable and robust effect. A simple evolution snowball chart in Figure 4B may be easier to be understood.

1. I am not sure I understood the definition of "mutated gene expression levels" and "mutated isoform expression levels" in the "Mutational gene expression and fusion transcript enhanced transcriptome clustering of benign hepatocytes and HCC" section. The authors mention that gene lists included all the isoforms within the same range of standard deviation. If I understand it correctly, they are equal if there is only one expressed transcript isoform. In that case, this overlap is not surprising at all.

Answer: We thank the reviewer for the great question. The definition of mutation gene expression level, mutation isoform expression level, and fusion gene expression level are now defined in the "Methods" section. In all HLA mutation transcripts, there were multiple transcripts with or without mutations for a single dominant isoform.

1. "To investigate the roles of gene expression alterations that were not accompanied with isoform expression changes, UMAP analyses were performed based on the non‐overlapped genes." Venn diagrams (Sup Figure 8) show that there are much less "non-overlapped genes" than "genes that showed both gene and isoform level changes" for each SD threshold (for example, for SD>=0.8 59 vs 275). Could that be the reason why clustering based on the former group is worse i.e the cancer and normal cells are separated less clearly?

Answer: The number of (attributes) genes could be a contributing factor in the segregation of cell populations. However, the number of attributes is not the underlying reason for worse performance for gene only classifier because much smaller isoforms/genes (22) overlap in SD>=1 outperformed a large number of genes (59) with SD>=0.8. It suggested that 59 gene expression classifier is less efficient in segregating the cell populations. To address this concern, we took SD>=0.8 as an example for demonstration if we subsampled the 275 overlapped genes/isoforms to 59 (equal to 59 non-overlapped genes in terms of number), we can still get better separation than the 59 DEG only. We repeated this subsampling process for three times. Similar results were found. The new data were inserted into supplemental Figure 8

Reviewer #2 (Public Review):

In the present study, Liu et al present an analysis of benign and HCC liver samples which were subjected to a new technology (LOOP-Seq) and paired WES. By integrating these data, the authors find isoforms, fusions and mutations which uniquely cluster within HCC samples, such as in the HLA locus, which serve as candidate leads for further investigation. The main appeal of the study is in the potential of LOOPSeq as a method to present isoform-resolved data without actually performing long-read sequencing. While this presents an exciting new method, the current study lacks systematic comparisons with other technologies/data to test the robustness, reproducibility and utility of LOOPSeq. Further, this study could be further improved by giving more physiologic context and examples from the analyses, thus providing a new resource to the HCC community. A few suggestions based on these are below:

Answer: We appreciate the reviewer to raise up all the important questions and the great suggestions. The LOOPseq technology was compared with Oxford nanopore and PacBio long-read sequencing in our previous study. We have cited analysis in the introduction section of the paper. HLA mutation clusters in the single molecules are our finding with major physiological significance since these mutations may help liver cancer cells evade immune surveillance. We have extensively discussed the potential impact of these mutations on cancer development in the discussion. In addition, we added a new section of DOCK8 and STEAP4 mutation expressions in the results (page 11, new Figure 5) that are highly relevant to the pathogenesis of HCC.

1. A primary consideration is that this seems to be the first implementation of LOOP-Seq, where the technology, while intriguing, has not been evaluated systematically. It seems like a standard 10x workflow is performed, where exons are selectively pulled down and amplified. Subsequent ultra-deep sequencing is assumed to give isoform-resolution of the sc-seq data. To demonstrate the utility of the approach it would benefit the study to compare the isoform-resolved results with studies where long-read sequencing was actually performed (ex: https://journals.lww.com/hep/Fulltext/2019/09000/Long_Read_RNA_Sequencing_Identifies_Alternativ e.19.aspx, https://www.jhep-reports.eu/article/S2589-5559(22)00021-0/fulltext, https://journals.plos.org/plosgenetics/article?id=10.1371/journal.pgen.1010342). Presumably, a fair amount of overlap should occur to justify the usage.

Answer: We have discussed the utility of the methodology in comparison with the previous studies by these three groups in the revision (results, page 12).

1. Related to this point, the sc-seq cell types and benign vs HCC genes should be compared with the wealth of data available for HCC sc-seq (https://www.nature.com/articles/s41467-022-322833, https://www.nature.com/articles/s41598-021-84693-w). These seem to be important to benchmark the technology in order to demonstrate that the probe-based selection and subsequent amplification does not bias cell type definition and clustering. In particular, https://www.nature.com/articles/s41586021-03974-6 seems quite relevant to compare mutational landscapes from the data.

Answer: This is a great point. The consistency probe-based analysis was demonstrated in our previous analyses and the analyses mentioned in the comments. We further discussed it in the results section of the paper (page 12).

1. From the initial UMAP clustering, it will be important to know what the identities are of the cells themselves. Presumably, there is quite a bit of immune cells and hepatocytes, but without giving identities, downstream mechanistic interpretation is difficult.

Answer: When mutation analyses were combined with cell marker analysis, i.e., immune marker positive but negative in HLA mutation, we found only one bona fide immune cell in the HCC sample. Thus, immune cells may not be significant in the current analysis.

1. In general, there are a fair amount of broad analyses, such as comparisons of hierarchical clustering of cell types, but very little physiologic interpretations of what these results mean. For example, among the cell clusters from Fig 6, knowing the pathways and cell annotations would help to contextualize these results. Without more biologically-meaningful aspects to highlight, most of the current appeal for the manuscript is dependent on the robustness of LOOP-seq and its implementation.

Answer: To address this comment, a new pathway analysis was performed on the cluster results of Figure 6. A new supplemental table was generated. The results are now discussed on page 13.

1. Many of the specific analyses are difficult and the methods are brief. Especially given that this technology is new and the dataset potentially useful, I would strongly recommend the authors set up a git repository, galaxy notebook or similar to maximize utility and reproducibility

Answer: The script file has been uploaded to GIT to facilitate the reproducibility of the analysis. We also added a new pipeline description script in the methods (pages 19-20).

1. The authors claim that clustering between benign and HCC samples was improved by including isoform & gene (Suppl fig 8). This seems like an important conclusion if true, especially to justify the use of longread implementation. Given that the combination of isoform + gene presents ~double the number of variables on which to cluster, it would be important to show that the improved separation on UMAP distance is actually due to the isoforms themselves and not just sampling more variables from either gene or isoform

Answer: The number of (attributes) genes could be a contributing factor in the segregation of cell populations. However, the number of attributes is not the underlying reason for worse performance for gene only classifier because much smaller isoforms/genes (22) overlap in SD>=1 outperformed a large number of genes (58) with SD>=0.8. It suggested that 58 gene expression classifier is less efficient in segregating the cell populations. To address this comment, we performed random subsampling to reduce the isoform/gene overlap iterates, similar results were obtained. A new supplemental figure was generated to reflect the new analyses.

1. SQANTI implementation to identify fusions relevant for the HCC/benign comparison. How do the fusions compare with those already identified for HCC? These analyses can be quite messy when performed on WES alone so it seems that having such deep RNA-seq would improve the capacity to see which fused genes are strongly expressed/suppressed. This doesn't seem as evident from current analysis. There are quite a bit of WES datasets which could be compared: https://www.nature.com/articles/ng.3252, https://www.nature.com/articles/s41467-01803276-y

Answer: Exome sequencing is not an ideal tool to identify fusion genes. Very few fusion genes have been discovered based on RNA sequencing so far. The fusion genes discovered in the study appeared mostly novel. No exome sequencing was involved in the identification of fusion genes.

1. Figure 4 is fairly unclear. The matrix graphs showing gene position mutations are tough to interpret and make out. Usually, gene track views with bars or lollipop graphs can make these results more readily interpretable. Also, how Figure 4 B infers causal directions from mutations is unclear.

Answer: We appreciate the reviewer for pointing this out. We have revised the diagram in Figure 4A to reflect the proper distance between the mutations in HLA-DQB1 NM_002123. Since these are the positions in the same alleles (protein), the gene track view or lollipop graph may not show that properly. The mutation clusters started from an isolated mutation, and mutation did not revert to wild type sequence after occurring. Based on these two principles, we showed several mutation accumulation pathways leading to hypermutation clusters.

Reviewer #3 (Public Review):

The Liu, et al. manuscript focuses on the interesting topic of evaluating in an almost genome-wide-scale, the number of transcriptional isoforms and fusion gene are present in single cells across the annotated protein coding genome. They also seek to determine the occurrences of single nucleotide variations/mutations (SNV) in the same isoform molecule emanating from the same gene expressed in normal and normal and hepatocellular carcinoma (HCC) cells. This study has been accomplished using modified LoopSeq long‐read technology (developed by several of the authors) and single cell isolation (10X) technologies. While this effort addresses a timely and important biological question, the reader encounters several issues in their report that are problematic.:

1. Much of the analysis of the evolution of mutations results and the biological effects of the fusion genes is conjecture and is not supported by empirical data. While their conclusions leave the reader with a sense that the results obtained from the LoopSeq has substantive biological implications. However, they are extended interpretations of the data. For example: The fusion protein likely functions as a decoy interference protein that negatively impacts the microtubule organization activity of EML4.(pg 9)... and other statements presented in a similar fashion.

Answer: We thank the reviewer for the helpful comment. The mutation results were experimentally validated by exome sequencing on the same samples. Furthermore, these mutations were filtered by requiring their presence in three different transcriptomes. The biological significance of these mutations is probably the subject of investigation in the next phase. Since a large number of HLA mutations did not occur overnight, the analysis of the accumulation pathways for these mutations was warranted, given the extensive evidence of such a process. The impact of mutations on HLA molecules appeared obvious and should be discussed. For ACTR2-EML4 fusion, we revised it as "The loss of microtubule binding domain may negatively impact the microtubule organization activity of EML4 domain of the fusion protein." We only discussed the obvious impact due to the loss of a large protein domain.

2, LoopSeq has the advantage of using short read sequencing analyses to characterize the exome capture results and thus benefits from low error rate compared to standard long-read sequencing techniques. However, there is no evidence obtained from standard long read sequencing that the isoforms observed with LoopSeq are obtained with parallel technologies such as long read technologies. It is not made clear how much discordance there is in comparing the LoopSeq results are with either PacBio or ONT long read technologies.

Answer: The comparative analyses among LOOPSeq, Oxford nanopore, and PacBio sequencing were performed in our previous study. We have cited the study in our introduction.

1. There is no proteome evidence (empirically derived or present in proteome databases) from the HCC and normal samples that confirms the presence or importance of the identified novel isoforms, nor is there support that indicate that changes in levels HLA genes translate to effects observed at the protein level. Since the stability and transport differences of isoforms from the same gene are often regulated at the post-transcriptional level, the biological importance of the isoform variations is unclear.

Answer: Given the transcriptome sequencing data, we can only focus on the isoform variation analysis but not directly link to the protein level variation because of the post-transcriptional level regulation. We discussed this in the revised manuscript (page 14).

4 It is unclear why certain thresholds were chosen for standard deviation (SD) <0.4 (page 5), SD >1.0 (pg 11).

Answer: The threshold is flexible and arbitrary. We showed different thresholds, and the same conclusion holds. We just choose the thresholds with better separation and a reasonable number of genes/isoforms for the downstream analysis. (Supplemental Figure 6-7 with different thresholds and supplemental tables 4-12).

1. HLA is known to accumulate considerable somatic variation. Of the many non-immunological genes determined to have multiple isoforms what are the isoform specific mutation rates in the same isoform molecule? Are the HLA genes unique in the number of mutations occurring in the same isoform?

Answer: We thank the reviewer for this important suggestion. We now show mutation expression patterns in isoforms of DOCK8 and STEAP4 in Figure 5. A new section is added to discuss the mutation expression of these two genes. As shown in supplemental figure 10, HLA-DQB1, HLA-DRB1, HLA-B, and HLA-C, have only one known isoform detected,

Editorial comments:

The present study pairs single-cell seq with LoopSeq synthetic long-read sequencing on samples of HCC and benign liver to identify mutations and fusion transcripts specific to cancer cells. The authors present a potentially important resource; however the overall support remains incomplete.

While the approach of evaluating isoform-specific changes at the cellular level to cancer seeks to address a timely and important topic, there is currently incomplete evidence in support of the major claims in the manuscript. In particular, major recommendations to provide stronger support for the combination of technologies and interpretation regarding cancer-associated genomic changes include: 1) systematic evaluation of UMAP-based clustering methods, to what subsets of data they are applied and subsequent interpretations, 2) direct comparisons of results with additional methods to quantify long-read sequencing data and those evaluating mutational consequences of HCC progression and 3) detailed expansion of the description of methods and rationale for selecting specific parameters and cell types for further analyses. Including these changes would significantly strengthen the support for utility of combining 10x single-cell with Loop-seq and provide compelling evidence for usage of this resource in dissecting HCC-associated molecular changes.

Answer: We appreciate the frank and constructive comments. The goal of UMAP is to obtain biological knowledge through unbiased data selection. Systematically, we select classifiers without any prior knowledge (blind to the samples). In our case, classifiers with high standard deviation across all the cells were chosen. We stressed this in the result section. The comparison among LOOPSeq, PacBio, and Oxford nanopore was made in our previous study. We cited that analysis in this paper. Analysis detail and pipelines were added in the revised manuscript to improve the reproducibility. The mutation expression analysis was quite clear-cut. The clustering classified the HCC and benign liver cells by itself and identified a few cancer cells in the benign liver sample. All these were accomplished without applying any knowledge.

Reviewer #1 (Recommendations For The Authors):

Overall, there are numerous problems with data presentation and insufficient description, which authors could fix.

1. Figure 4. A. It would be more clear if the figure showed the distribution of mutations in the molecule. Otherwise, it's hard to see if we see clusters of mutations or just 25 mutations spread uniformly across the transcript. B. It's unclear what the reader needs to take away from these columns of numbers.

Answer: The mutation positions are now presented as proportion to the location in a molecule. Column B is the distribution of mutation molecules from left panel in each cluster of cells (from Figure 3A) and their sample origin (HCC or benign liver). We clarify it a little more in the legend of Figure 4A.

1. As a reader, I did not understand how "mutated gene expression levels" and "mutated isoform expression levels" were calculated in terms of sequenced long reads

Answer: We defined the term and calculations in the methods section of the revised manuscript.

1. Page 6 "genes involving antigen presentation"

Answer: The full sentence of the subtitle is" Mutations of genes involving antigen presentation dominated the mutation expression landscape."

1. Page 6 "These unique mutational isoforms" - how are these isoforms unique?

Answer: We take away most of the "unique" adjectives to describe the non-redundant mutations.

1. Page 6. Unclear "All but one clusters contained cells co‐migrated with cells of their sources."

"Among 113 mutation isoforms, the major histocompatibility complex (HLA) was the most prominent with 68 iterations (60.2%) (Supplemental Table 3, Figure 3B)" There is nothing about HLA in Figure 3B.

Answer: We revised the sentence as "Cells in all but one clusters co-migrated with cells of their sources". The mutation isoform expressions were listed in supplemental Table 3. They are too small and become unreadable when put in the figure.

1. Page 10 "genes or isoforms that across all samples had with expression standard deviations less than" - probably "with" should not be there.

Answer: We correct the error and thank the reviewer for the comment.

1. Page 11 "UMAP analysis was performed using genes with standard deviations {greater than or equal to} 1.0 (182 wild‐type genes) and standard deviations >0.4 (282 mutated genes)". What do "wild-type" and "mutated" mean here?

Answer: We edited as "UMAP analysis was performed using gene expressions with standard deviations ≥ 1.0 (182 non-mutated genes) and gene mutation expression with standard deviations 0.4 (282 mutated genes)."

1. I could not find the description of Supplementary Tables.

Answer: The supplemental table legends are added in the revised manuscript.

1. In the Discussion section, the authors mention that mutations were mainly expressed in a specific isoform of a gene for a given cell. I suggest to emphasize this point in the Results section and illustrate it with a comparison of abundance of mutated and non-mutated isoforms

Answer: For HLA molecules, their expression appeared to be restricted to one known isoform, regardless of mutation status. This sentence is removed in the revision. A new section of DOCK8 and STEAP4 mutation expression is added to the result.

1. It is also mentioned that mutations may have an impact on the RNA splicing process. The authors should compare the observed isoform ratio to a prediction of the effect of variants on splicing by SpliceAI or similar tools

Answer: This sentence was removed from the discussion.

1. Figure 3c: triangles corresponding to HLA-positive cells are hard to distinguish

Answer: We provide a larger representation of the triangle and circle in figure 3c in the revision.

Reviewer #2 (Recommendations For The Authors):

Many of my comments could be addressed by spending time to provide the code/data and a walkthrough of analyses so that other users would be able to answer these questions on their own.

Answer: We have included a script section in the revision to ensure the reproducibility of the analysis. The raw data had been uploaded to GEO (see Methods).

Reducing the performance hit of copying data between processes:

Option #1: Just use threads

Processes have overhead, threads do not. And while it’s true that generic Python code won’t parallelize well when using multiple threads, that’s not necessarily true for your Python code. For example, NumPy releases the GIL for many of its operations, which means you can use multiple CPU cores even with threads.

``` # numpy_gil.py import numpy as np from time import time from multiprocessing.pool import ThreadPool

arr = np.ones((1024, 1024, 1024))

start = time() for i in range(10): arr.sum() print("Sequential:", time() - start)

expected = arr.sum()

start = time() with ThreadPool(4) as pool: result = pool.map(np.sum, [arr] * 10) assert result == [expected] * 10 print("4 threads:", time() - start) ```

When run, we see that NumPy uses multiple cores just fine when using threads, at least for this operation:

$ python numpy_gil.py Sequential: 4.253053188323975 4 threads: 1.3854241371154785

Pandas is built on NumPy, so many numeric operations will likely release the GIL as well. However, anything involving strings, or Python objects in general, will not. So another approach is to use a library like Polars which is designed from the ground-up for parallelism, to the point where you don’t have to think about it at all, it has an internal thread pool.

Option #2: Live with it

If you’re stuck with using processes, you might just decide to live with the overhead of pickling. In particular, if you minimize how much data gets passed and forth between processes, and the computation in each process is significant enough, the cost of copying and serializing data might not significantly impact your program’s runtime. Spending a few seconds on pickling doesn’t really matter if your subsequent computation takes 10 minutes.

Option #3: Write the data to disk

Instead of passing data directly, you can write the data to disk, and then pass the path to this file: * to the subprocess (as an argument) * to parent process (as the return value of the function running in the worker process).

The recipient process can then parse the file.

``` import pandas as pd import multiprocessing as mp from pathlib import Path from tempfile import mkdtemp from time import time

def noop(df: pd.DataFrame): # real code would process the dataframe here pass

def noop_from_path(path: Path): df = pd.read_parquet(path, engine="fastparquet") # real code would process the dataframe here pass

def main(): df = pd.DataFrame({"column": list(range(10_000_000))})

with mp.get_context("spawn").Pool(1) as pool:
    # Pass the DataFrame to the worker process
    # directly, via pickling:
    start = time()
    pool.apply(noop, (df,))
    print("Pickling-based:", time() - start)

    # Write the DataFrame to a file, pass the path to
    # the file to the worker process:
    start = time()
    path = Path(mkdtemp()) / "temp.parquet"
    df.to_parquet(
        path,
        engine="fastparquet",
        # Run faster by skipping compression:
        compression="uncompressed",
    )
    pool.apply(noop_from_path, (path,))
    print("Parquet-based:", time() - start)

if name == "main": main() `` **Option #4:multiprocessing.shared_memory`**

Because processes sometimes do want to share memory, operating systems typically provide facilities for explicitly creating shared memory between processes. Python wraps this facilities in the multiprocessing.shared_memory module.

However, unlike threads, where the same memory address space allows trivially sharing Python objects, in this case you’re mostly limited to sharing arrays. And as we’ve seen, NumPy releases the GIL for expensive operations, which means you can just use threads, which is much simpler. Still, in case you ever need it, it’s worth knowing this module exists.

Note: The module also includes ShareableList, which is a bit like a Python list but limited to int, float, bool, small str and bytes, and None. But this doesn’t help you cheaply share an arbitrary Python object.

A bad option for Linux: the "fork" context

You may have noticed we did multiprocessing.get_context("spawn").Pool() to create a process pool. This is because Python has multiple implementations of multiprocessing on some OSes. "spawn" is the only option on Windows, the only non-broken option on macOS, and available on Linux. When using "spawn", a completely new process is created, so you always have to copy data across.

On Linux, the default is "fork": the new child process has a complete copy of the memory of the parent process at the time of the child process’ creation. This means any objects in the parent (arrays, giant dicts, whatever) that were created before the child process was created, and were stored somewhere helpful like a module, are accessible to the child. Which means you don’t need to pickle/unpickle to access them.

Sounds useful, right? There’s only one problem: the "fork" context is super-broken, which is why it will stop being the default in Python 3.14.

Consider the following program:

``` import threading import sys from multiprocessing import Process

def thread1(): for i in range(1000): print("hello", file=sys.stderr)

threading.Thread(target=thread1).start()

def foo(): pass

Process(target=foo).start() ```

On my computer, this program consistently deadlocks: it freezes and never exits. Any time you have threads in the parent process, the "fork" context can cause in potential deadlocks, or even corrupted memory, in the child process.

You might think that you’re fine because you don’t start any threads. But many Python libraries start a thread pool on import, for example NumPy. If you’re using NumPy, Pandas, or any other library that depends on NumPy, you are running a threaded program, and therefore at risk of deadlocks, segfaults, or data corruption when using the "fork" multiprocessing context. For more details see this article on why multiprocessing’s default is broken on Linux.

You’re just shooting yourself in the foot if you take this approach.

Multiple threads let you run code in parallel, potentially on multiple CPUs. On Python, however, the global interpreter lock makes this parallelism harder to achieve.

Multiple processes also let you run code in parallel—so what’s the difference between threads and processes?

All the threads inside a single process share the same memory address space. If thread 1 in a process stores some memory at address 0x7f0cd1a88810, thread 2 can access the same memory at the same address. That means passing objects between threads is cheap: you just need to get the pointer to the memory address from one thread to the other. A memory address is 8 bytes: this is not a lot of data to move around.

In contrast, processes do not share the same memory space. There are some shared memory facilities provided by the operating system, typically, and we’ll get to that later. But by default, no memory is shared. That means you can’t just share the address of your data across processes: you have to copy the data.

they are defining a normative aspect of perception, which need not be intrinsic to perception itself: they are defining what it takes to perceive well and accurately, the accuracy conditions of perception, which is this tacit knowledge of the sensorimotor contingencies. but this knowledge is a skill because it requires continued and varied perceptual experiences. So - in order to perceive well - one first needs to perceive, and so knowing the conditions of perceiving well will not necessarily impart knowledge of perception as such. 'familiarity' and knowledge of 'contingencies', an imposition of order on 'what would otherwise be' 'patterns of changing stimulation'.

I.E., the skill invokes precisely the kind of cognitive processing they are seeking to explain? Are they describing an aspect of cognition by describing sensory motions that involve cognitive processing? Is this circular?

Reviewer #1 (Public Review):

Summary:

This study aims to provide imaging methods for users of the field of human layer-fMRI. This is an emerging field with 240 papers published so far. Different than implied in the manuscript, 3T is well represented among those papers. E.g. see the papers below that are not cited in the manuscript. Thus, the claim on the impact of developing 3T methodology for wider dissemination is not justified. Specifically, because some of the previous papers perform whole brain layer-fMRI (also at 3T) in more efficient, and more established procedures.

The authors implemented a sequence with lots of nice features. Including their own SMS EPI, diffusion bipolar pulses, eye-saturation bands, and they built their own reconstruction around it. This is not trivial. Only a few labs around the world have this level of engineering expertise. I applaud this technical achievement. However, I doubt that any of this is the right tool for layer-fMRI, nor does it represent an advancement for the field. In the thermal noise dominated regime of sub-millimeter fMRI (especially at 3T), it is established to use 3D readouts over 2D (SMS) readouts. While it is not trivial to implement SMS, the vendor implementations (as well as the CMRR and MGH implementations) are most widely applied across the majority of current fMRI studies already. The author's work on this does not serve any previous shortcomings in the field.

The mechanism to use bi-polar gradients to increase the localization specificity is doubtful to me. In my understanding, killing the intra-vascular BOLD should make it less specific. Also, the empirical data do not suggest a higher localization specificity to me.

Embedding this work in the literature of previous methods is incomplete. Recent trends of vessel signal manipulation with ABC or VAPER are not mentioned. Comparisons with VASO are outdated and incorrect.

The reproducibility of the methods and the result is doubtful (see below).

I don't think that this manuscript is in the top 50% of the 240 layer-fmri papers out there.

3T layer-fMRI papers that are not cited:<br /> Taso, M., Munsch, F., Zhao, L., Alsop, D.C., 2021. Regional and depth-dependence of cortical blood-flow assessed with high-resolution Arterial Spin Labeling (ASL). Journal of Cerebral Blood Flow and Metabolism. https://doi.org/10.1177/0271678X20982382

Wu, P.Y., Chu, Y.H., Lin, J.F.L., Kuo, W.J., Lin, F.H., 2018. Feature-dependent intrinsic functional connectivity across cortical depths in the human auditory cortex. Scientific Reports 8, 1-14. https://doi.org/10.1038/s41598-018-31292-x

Lifshits, S., Tomer, O., Shamir, I., Barazany, D., Tsarfaty, G., Rosset, S., Assaf, Y., 2018. Resolution considerations in imaging of the cortical layers. NeuroImage 164, 112-120. https://doi.org/10.1016/j.neuroimage.2017.02.086

Puckett, A.M., Aquino, K.M., Robinson, P.A., Breakspear, M., Schira, M.M., 2016. The spatiotemporal hemodynamic response function for depth-dependent functional imaging of human cortex. NeuroImage 139, 240-248. https://doi.org/10.1016/j.neuroimage.2016.06.019

Olman, C.A., Inati, S., Heeger, D.J., 2007. The effect of large veins on spatial localization with GE BOLD at 3 T: Displacement, not blurring. NeuroImage 34, 1126-1135. https://doi.org/10.1016/j.neuroimage.2006.08.045

Ress, D., Glover, G.H., Liu, J., Wandell, B., 2007. Laminar profiles of functional activity in the human brain. NeuroImage 34, 74-84. https://doi.org/10.1016/j.neuroimage.2006.08.020

Huber, L., Kronbichler, L., Stirnberg, R., Ehses, P., Stocker, T., Fernández-Cabello, S., Poser, B.A., Kronbichler, M., 2023. Evaluating the capabilities and challenges of layer-fMRI VASO at 3T. Aperture Neuro 3. https://doi.org/10.52294/001c.85117

Scheeringa, R., Bonnefond, M., van Mourik, T., Jensen, O., Norris, D.G., Koopmans, P.J., 2022. Relating neural oscillations to laminar fMRI connectivity in visual cortex. Cerebral Cortex. https://doi.org/10.1093/cercor/bhac154

Strengths:

See above. The authors developed their own SMS sequence with many features. This is important to the field. And does not leave sequence development work to view isolated monopoly labs. This work democratises SMS.<br /> The questions addressed here are of high relevance to the field: getting tools with good sensitivity, user-friendly applicability, and locally specific brain activity mapping is an important topic in the field of layer-fMRI.

Weaknesses:

1. I feel the authors need to justify why flow-crushing helps localization specificity. There is an entire family of recent papers that aim to achieve higher localization specificity by doing the exact opposite. Namely, MT or ABC fRMRI aims to increase the localization specificity by highlighting the intravascular BOLD by means of suppressing non-flowing tissue. To name a few:

Priovoulos, N., de Oliveira, I.A.F., Poser, B.A., Norris, D.G., van der Zwaag, W., 2023. Combining arterial blood contrast with BOLD increases fMRI intracortical contrast. Human Brain Mapping hbm.26227. https://doi.org/10.1002/hbm.26227.

Pfaffenrot, V., Koopmans, P.J., 2022. Magnetization Transfer weighted laminar fMRI with multi-echo FLASH. NeuroImage 119725. https://doi.org/10.1016/j.neuroimage.2022.119725

Schulz, J., Fazal, Z., Metere, R., Marques, J.P., Norris, D.G., 2020. Arterial blood contrast ( ABC ) enabled by magnetization transfer ( MT ): a novel MRI technique for enhancing the measurement of brain activation changes. bioRxiv. https://doi.org/10.1101/2020.05.20.106666

Based on this literature, it seems that the proposed method will make the vein problem worse, not better. The authors could make it clearer how they reason that making GE-BOLD signals more extra-vascular weighted should help to reduce large vein effects.

The empirical evidence for the claim that flow crushing helps with the localization specificity should be made clearer. The response magnitude with and without flow crushing looks pretty much identical to me (see Fig, 6d).<br /> It's unclear to me what to look for in Fig. 5. I cannot discern any layer patterns in these maps. It's too noisy. The two maps of TE=43ms look like identical copies from each other. Maybe an editorial error?

The authors discuss bipolar crushing with respect to SE-BOLD where it has been previously applied. For SE-BOLD at UHF, a substantial portion of the vein signal comes from the intravascular compartment. So I agree that for SE-BOLD, it makes sense to crush the intravascular signal. For GE-BOLD however, this reasoning does not hold. For GE-BOLD (even at 3T), most of the vein signal comes from extravascular dephasing around large unspecific veins, and the bipolar crushing is not expected to help with this.

2. The bipolar crushing is limited to one single direction of flow. This introduces a lot of artificial variance across the cortical folding pattern. This is not mentioned in the manuscript. There is an entire family of papers that perform layer-fmri with black-blood imaging that solves this with a 3D contrast preparation (VAPER) that is applied across a longer time period, thus killing the blood signal while it flows across all directions of the vascular tree. Here, the signal cruising is happening with a 2D readout as a "snap-shot" crushing. This does not allow the blood to flow in multiple directions.<br /> VAPER also accounts for BOLD contaminations of larger draining veins by means of a tag-control sampling. The proposed approach here does not account for this contamination.

Chai, Y., Li, L., Huber, L., Poser, B.A., Bandettini, P.A., 2020. Integrated VASO and perfusion contrast: A new tool for laminar functional MRI. NeuroImage 207, 116358. https://doi.org/10.1016/j.neuroimage.2019.116358

Chai, Y., Liu, T.T., Marrett, S., Li, L., Khojandi, A., Handwerker, D.A., Alink, A., Muckli, L., Bandettini, P.A., 2021. Topographical and laminar distribution of audiovisual processing within human planum temporale. Progress in Neurobiology 102121. https://doi.org/10.1016/j.pneurobio.2021.102121

If I would recommend anyone to perform layer-fMRI with blood crushing, it seems that VAPER is the superior approach. The authors could make it clearer why users might want to use the unidirectional crushing instead.

3. The comparison with VASO is misleading.<br /> The authors claim that previous VASO approaches were limited by TRs of 8.2s. The authors might be advised to check the latest literature of the last years.<br /> Koiso et al. performed whole brain layer-fMRI VASO at 0.8mm at 3.9 seconds (with reliable activation), 2.7 seconds (with unconvincing activation pattern, though), and 2.3 (without activation).<br /> Also, whole brain layer-fMRI BOLD at 0.5mm and 0.7mm has been previously performed by the Juelich group at TRs of 3.5s (their TR definition is 'fishy' though).

Koiso, K., Müller, A.K., Akamatsu, K., Dresbach, S., Gulban, O.F., Goebel, R., Miyawaki, Y., Poser, B.A., Huber, L., 2023. Acquisition and processing methods of whole-brain layer-fMRI VASO and BOLD: The Kenshu dataset. Aperture Neuro 34. https://doi.org/10.1101/2022.08.19.504502

Yun, S.D., Pais‐Roldán, P., Palomero‐Gallagher, N., Shah, N.J., 2022. Mapping of whole‐cerebrum resting‐state networks using ultra‐high resolution acquisition protocols. Human Brain Mapping. https://doi.org/10.1002/hbm.25855

Pais-Roldan, P., Yun, S.D., Palomero-Gallagher, N., Shah, N.J., 2023. Cortical depth-dependent human fMRI of resting-state networks using EPIK. Front. Neurosci. 17, 1151544. https://doi.org/10.3389/fnins.2023.1151544

The authors are correct that VASO is not advised as a turn-key method for lower brain areas, incl. Hippocampus and subcortex. However, the authors use this word of caution that is intended for inexperienced "users" as a statement that this cannot be performed. This statement is taken out of context. This statement is not from the academic literature. It's advice for the 40+ user base that wants to perform layer-fMRI as a plug-and-play routine tool in neuroscience usage. In fact, sub-millimeter VASO is routinely being performed by MRI-physicists across all brain areas (including deep brain structures, hippocampus etc). E.g. see Koiso et al. and an overview lecture from a layer-fMRI workshop that I had recently attended: https://youtu.be/kzh-nWXd54s?si=hoIJjLLIxFUJ4g20&t=2401

Thus, the authors could embed this phrasing into the context of their own method that they are proposing in the manuscript. E.g. the authors could state whether they think that their sequence has the potential to be disseminated across sites, considering that it requires slow offline reconstruction in Matlab?<br /> Do the authors think that the results shown in Fig. 6c are suggesting turn-key acquisition of a routine mapping tool? In my humble opinion, it looks like random noise, with most of the activation outside the ROI (in white matter).

4. The repeatability of the results is questionable.<br /> The authors perform experiments about the robustness of the method (line 620). The corresponding results are not suggesting any robustness to me. In fact, the layer profiles in Fig. 4c vs. Fig 4d are completely opposite. The location of peaks turns into locations of dips and vice versa.<br /> The methods are not described in enough detail to reproduce these results.<br /> The authors mention that their image reconstruction is done "using in-house MATLAB code" (line 634). They do not post a link to github, nor do they say if they share this code.

It is not trivial to get good phase data for fMRI. The authors do not mention how they perform the respective coil-combination.<br /> No data are shared for reproduction of the analysis.

5. The application of NODRIC is not validated.<br /> Previous applications of NORDIC at 3T layer-fMRI have resulted in mixed success. When not adjusted for the right SNR regime it can result in artifactual reductions of beta scores, depending on the SNR across layers. The authors could validate their application of NORDIC and confirm that the average layer-profiles are unaffected by the application of NORDIC. Also, the NORDIC version should be explicitly mentioned in the manuscript.

Akbari, A., Gati, J.S., Zeman, P., Liem, B., Menon, R.S., 2023. Layer Dependence of Monocular and Binocular Responses in Human Ocular Dominance Columns at 7T using VASO and BOLD (preprint). Neuroscience. https://doi.org/10.1101/2023.04.06.535924

Knudsen, L., Guo, F., Huang, J., Blicher, J.U., Lund, T.E., Zhou, Y., Zhang, P., Yang, Y., 2023. The laminar pattern of proprioceptive activation in human primary motor cortex. bioRxiv. https://doi.org/10.1101/2023.10.29.564658

Reviewer #2 (Public Review):

Summary<br /> Song et al investigate the role of the frontal eye field (FEF) and the intraparietal sulcus (IPS) in mediating the shift in ocular dominance (OD) observed after a period of dichoptic stimulation during which attention is selectively directed to one eye. This manipulation has been previously found to transiently shift OD in favor of the unattended eye, similar to the effect of short-term monocular deprivation. To this aim, the authors combine psychophysics, fMRI, and transcranial magnetic stimulation (TMS). In the first experiment, the authors determine the regions of interest (ROIs) based on the responses recorded by fMRI during either dichoptic or binocular stimulation, showing selective recruitment of the right FEF and IPS during the dichoptic condition, in line with the involvement of eye-based attention. In a second experiment, the authors investigate the causal role of these two ROIs in mediating the OD shift observed after a period of dichoptic stimulation by selectively inhibiting with TMS (using continuous theta burst stimulation, cTBS), before the adaptation period (50 min exposure to dichoptic stimulation). They show that, when cTBS is delivered on the FEF, but not the IPS or the vertex, the shift in OD induced by dichoptic stimulation is reduced, indicating a causal involvement of the FEF in mediating this form of short-term plasticity. A third control experiment rules out the possibility that TMS interferes with the OD task (binocular rivalry), rather than with the plasticity mechanisms. From this evidence, the authors conclude that the FEF is one of the areas mediating the OD shift induced by eye-selective attention.

Strengths<br /> 1. The experimental paradigm is sound and the authors have thoroughly investigated the neural correlates of an interesting form of short-term visual plasticity combining different techniques in an intelligent way.

2. The results are solid and the appropriate controls have been performed to exclude potential confounds.

3. The results are very interesting, providing new evidence both about the neural correlates of eye-based attention and the involvement of extra-striate areas in mediating short-term OD plasticity in humans, with potential relevance for clinical applications (especially in the field of amblyopia).

Weaknesses<br /> 1. Ethics: more details about the ethics need to be included in the manuscript. It is only mentioned for experiment 1 that participants "provided informed consent in accordance with the Declaration of Helsinki. This study was approved by the Institutional Review Board of the Institute of Psychology, Chinese Academy of Sciences". (Which version of the Declaration of Helsinki? The latest version requires the pre-registration of the study. The code of the approved protocol together with the code and date of the approval should be provided.) There is no mention of informed consent procedures or ethics approval for the TMS experiments. This is a huge concern, especially for brain stimulation experiments!

2. Statistics: the methods section should include a sub-section describing in detail all the statistical analyses performed for the study. Moreover, in the results section, statistical details should be added to support the fMRI results. In the current version of the manuscript, the claims are not supported by statistical evidence.

3. Interpretation of the results: the TMS results are very interesting and convincing regarding the involvement of the FEF in the build-up of the OD shift induced by dichoptic stimulation, however, I am not sure that the authors can claim that this effect is related to eye-based attention, as cTBS has no effect on the blob detection task during dichoptic stimulation. If the FEF were causally involved in eye-based attention, one would expect a change in performance in this task during dichoptic stimulation, perhaps a similar performance for the unattended and attended eye. The authors speculate that the sound could have an additional role in driving eye-based attention, which might explain the lack of effect for the blob discrimination task, however, this hypothesis has not been tested.

4. Writing: in general, the manuscript is well written, but clarity should be improved in certain sections.

a. fMRI results: the first sentence is difficult to understand at first read, but it is crucial to understand the results, please reformulate and clarify.

b. Experiment 3: the rationale for experiment one should be straightforward, without a long premise explaining why it would not be necessary.

c. Discussion: the language is a bit familiar here and there, a more straightforward style should be preferred (one example: p.19 second paragraph).

5. Minor: the authors might consider using the term "participant" or "observer" instead of "subject" when referring to the volunteers who participated in the study.

it's literally just iB, is it not? Why do I need a reference for that, and why isn't it just directly stated?

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 #3

Evidence, reproducibility and clarity

Summary:

Haj Abdullah Alieh at al., describe re-analysis of an existing short read RNA-Seq dataset consisting of 3 replicates of 3 FAC sorted cell populations of the E14.5 Btg2::RFP/Tubb3::GFP mouse cortex: neural stem cells (NSC; RFP-/GFP-), neural precursors (NP; RFP+/GFP-) and neurons (N; GFP+), for the purpose of investigating alternative splicing isoform switching during neuronal cell-type specification. They generate a one replicate PacBio dataset of these same sorted cells, with the aim of identifying full-length transcript isoforms, which are difficult to discern with short-read data alone. The key conclusions are the discovery of ~50,000 novel transcript isoforms containing ~2,500 novel splice junctions; the discovery of isoform switches between NSC -> neuron that contain a high proportion of microexon inclusion events and the finding that many of these switches are predicted by Alphafold2 to have a structural impact.

The data is interesting and the bioinformatics approach of investigating potential impacts of splice variants on protein structure using Alphafold2 is also interesting, however at present the paper would be better presented as a resource, unless effort is undertaken to experimentally validate some potential biological findings. However, for the paper to be useful as a resource, links to newly generated data and analysis code need to be provided. The capacity for exploration of these newly identified splice isoforms, or further analysis using the new GTF, could then be one of the attractions of this work.

Major comments:

• Are the key conclusions convincing?
• Should the authors qualify some of their claims as preliminary or speculative, or remove them altogether?
• Would additional experiments be essential to support the claims of the paper? Request additional experiments only where necessary for the paper as it is, and do not ask authors to open new lines of experimentation. Figure 1 The discovery of ~50,000 novel transcript isoforms containing ~2,500 novel splice junctions As far as I can see the description of novelty is based on them being not present in either Ensembl (GRCm38.p6), NCBI_RefSeq, or Gencode (vM10) - note here the numbers are genome assembly versions and do not refer to the GTF annotation versions compared against - these should be provided as they are frequently updated. The claim is that they are not present in these references because the unique cell samples have not been analysed before. For transcript isoforms to be included in these references they must have a good level of support. I have a couple of concerns about the support for these isoforms: The numbers in figure 1A do not add up. For long read sequencing two pipelines are used resulting in 76,077 and 80,782 isoforms - in the venn diagram 1A the overlapping circle has two numbers of isoforms in it: 70,658 and 71,760 so it is unclear, are 70,658 isoforms found by both pipelines or 71,760? Then we are told the union of these transcripts is taken forward to the next venn diagram. However this diagram is labelled with 82,046 transcript isoforms. Pipeline 1 has labelled 5419 unique isoforms, pipeline 2 has 9,022 unique isoforms so 5419 + 9022 + 70658(71760) = 85,099(86201) not 82,046 - perhaps some extra filtering has occurred that should be labelled/described? Again the final number of transcripts at the end of everything is off - if the 82,046 transcripts from long read are combined with the 16,070 unique to the short read this equals 98,116, not 97,240. The authors decide to use long read sequencing to assemble the isoforms as short-read sequencing is unreliable for assembling full length isoforms - however for their final list they merge isoforms assembled by StringTie from short read data with the isoforms assembled from the PacBio long read data, it seems likely that the isoforms detected only by short-read Stringtie assembly would be unreliable and shouldn't be included in the final total. The authors perform only one biological replicate of PacBio long read sequencing of three different samples, so it is not possible to easily determine the reproducibility of the findings. I appreciate PacBio is expensive, the authors could consider other ways to evaluate the reproducibility - perhaps by looking at the detection of transcripts expected to be uniformly expressed between the different conditions? The authors provide no quality information for their PacBio sequencing run - eg. length distribution of reads, how many reads are left after quality filtering, quality across the length of reads, ie. I do not know if most isoforms reported are supported by 5 full length isoform reads, or if it is rare in the dataset to get full length isoform reads .etc is the quality comparable across the three PacBio samples? How many of the novel isoforms are supported by both short read and long read data? How many of the novel isoforms are supported only by short reads? How many isoforms are found in all three PacBio samples? Does gene expression measured with the PacBio data match the previous results of measuring gene expression in the short read data? Adding these kinds of analyses would give more confidence in the results. This section of methods is confusing, I don't really understand what has been done or what part of the manuscript this refers to: "​​Events were assigned to an inclusion isoform if their coordinates overlapped, at least partially, with an exon or to an exclusion isoform if they were located within an intron. AS events without a corresponding inclusion or exclusion isoform were assigned to an Ensembl or NCBI_RefSeq isoform using the criteria above. Only AS events assigned to at least one inclusion and one exclusion isoform were considered for further analysis." VastDB is a splicing database created by Manuel Irimia/Ben Blencowe containing a lot of neural samples across development - how many of the 'novel' splice sites are present in VastDB? Similarly, how many of the 'novel' splice isoforms were previously detected by Zhang et al., 2016, Cell.

Figure 2: over neuronal maturation the major splicing change is for cassette exons to become more included, 50% of those measured being microexons Overall this section is strongest, the conclusions are well supported. Figure 2D - there are no genome coordinates given to allow the reader to check the highlighted events out for themselves. Figure 2F is very confusing, consider an alternative way to present this. Figure 2G, the premise of this analysis is interesting! But confused on the numbers - in 2F its shown that 226 exons become more included between both NSC->NP->N, so why are 441 exons plotted in 2G? Whilst I appreciate genes must be expressed in both NSCs and neurons to be able to calculate differential splicing, one thing not addressed is whether expression of a lot of these genes also goes up in neurons, i.e. could it be that when these genes are lowly expressed in NSCs their splicing is not particularly well regulated but it doesn't really matter because they are not really required in NSCs? This becomes relevant later where you start to address the functionality of isoform switches - if the gene is expressed to the same degree in NSC vs. N this would suggest that both isoforms are functional, if a gene is very lowly expressed in NSC but highly expressed in N, then maybe only the N isoform needs to be functional. Gene ontology methodology is not described in the methods. What were the spliced genes compared against? Given these are neural samples, lots of expressed genes will have neural functions, so is this really informing us about the alternatively spliced genes? The manuscript would benefit by integration of its data with other published datasets - especially with the microexons - how do these behave in other datasets of neuronal maturation (such as those from vastdb or zhang 2016)? The authors could consider looking at motifs around regulated microexons to try and establish if any specific RBPs might be involved in this regulation, although this would benefit from follow up experiments.

Figure 3: exon inclusion in neuronal specific transcripts confers different structures to translated proteins, suggesting these events are important functionally Here, Alphafold2 is used to predict the structures of switching isoforms, whilst an interesting approach to inform further experiments, presented alone, it remains hypothetical. Hook2 is highlighted as one example, where inclusion of a microexon introducing two amino acids to the translated protein is predicted to cause a structural change that will impact its binding to microtubules. It's hard to determine if this really will have a functional impact without doing experiments in the lab. For this manuscript to serve as a research (rather than resource) article, it would benefit from an example experiment expressing neuronal vs. NSC Hook2 isoform in a cell line and measuring co-localisation with microtubules via IF microscopy, or something similar to address the proposed function. In the second half of this figure, more subtle local structural changes are investigated and the example of an alpha-helix to beta-strand switch predicted in Kctd13 is presented. The figure would benefit from showing the splicing change at the RNA level and relating that to the change seen at the protein sequence level as it is a bit confusing - the region of deletion is labelled as 'AS REGION' however, two amino acids preceding this box are different between the two isoforms (KVEF vs. KVRG) - so presumably the splicing change starts earlier than denoted? In the discussion the authors state: "While these regions are long known to exist, their structural switch was assumed to be dependent on substantial changes in their structural and sequence contexts (Gendoo and Harrison, 2011; W. Li et al., 2015) as opposed to, as observed in our study, being triggered by small perturbations within nearly identical sequence contexts." It's not clear whether these small local predictions are accurate and would require some additional structural data to validate. - Are the suggested experiments realistic in terms of time and resources? It would help if you could add an estimated cost and time investment for substantial experiments. Suggestions of additional computational analysis are very realistic and shouldn't take longer than a month or two. The addition of experimental data to support Figure 3 would take considerable time and resources, potentially collaboration with other labs. Perhaps focusing on making this dataset an accessible resource would be a better route to publication. - Are the data and the methods presented in such a way that they can be reproduced? No, no source code, software versions or supplementary data/materials is provided. - Are the experiments adequately replicated and statistical analysis adequate? Having one replicate of the PacBio experiment is a bit concerning, but I am aware that it is expensive. Given they have three samples of different conditions with PacBio data perhaps showing the quality control of the libraries, reproducibility of transcripts that don't change in the three conditions, etc. would give more confidence in the data.

Minor comments:

• Specific experimental issues that are easily addressable. Made above.
• Are prior studies referenced appropriately? Yes. Except for this section of introduction: "While great effort is being made to overcome these limitations, capturing cell type-specific AS dynamics that is both quantitative and comprehensive of full-length transcript information currently requires combination of both SRS and LRS performed in parallel on the same cell pool. This was seldom attempted (Gupta et al., 2018; Joglekar et al., 2021) and, to the best of our knowledge, never for specific cell types of the developing mammalian brain. Even more limiting, systematic assessment of the consequences of AS on protein structure and putative function in cell fate commitment is entirely lacking. "

LRS has allowed for whole transcriptome determination and quantification in a number of cases, especially in non-model organisms, below I mention some examples from human and mouse: Nanopore use in GTEX + short reads: Glinos et al., 2022 Nature https://www.nature.com/articles/s41586-022-05035-y PacBio SMRT-Seq + short reads human and mouse cortex: Leung et al. Cell Reports 2021 https://www.cell.com/cell-reports/pdf/S2211-1247(21)01504-7.pdf PacBio IsoSeq + short reads in human and mouse sperm: Sun et al., 2021 Nature Communications https://www.nature.com/articles/s41467-021-21524-6 Single cell long read RNA-Seq has also been described in several scenarios and is worth referencing in the introduction: Samples from various mouse and human sources: Tian et al., 2021 Genome Biology https://link.springer.com/article/10.1186/s13059-021-02525-6 differential isoform usage in myeloma cell lines: Phillpott et al., 2021, Nature Biotech https://www.nature.com/articles/s41587-021-00965-w Single cell long read isoform analysis in human immune cells: Volden and Vollmers, 2022, Genome Biology https://genomebiology.biomedcentral.com/articles/10.1186/s13059-022-02615-z - Are the text and figures clear and accurate? Mostly, I've highlighted where numbers in figures don't make sense to me. Generally the text could use some going over and tightening up (eg. sentence on page 12 needs revising for clarity and typo "The fact that within this helical packing resides the protein domain essential for Hook2 function to bind microtubules, implies that such a negligible AS switch by two ammino acids may result in a completely altered function. ") - Do you have suggestions that would help the authors improve the presentation of their data and conclusions? I have made suggestions above about figures that are unclear to me.

Referees cross-commenting

After reading the reviews of other reviewers, it seems we are much in agreement over the main concerns relating to this manuscript. Namely: concerns over the PacBio being single replicate, concerns over indiscriminately merging PacBio and SRS transcripts, concerns about lack of validation of the structural changes predicted by AlphaFold2. On the question of novelty and significance we also seem to be aligned.

Significance

• Describe the nature and significance of the advance (e.g. conceptual, technical, clinical) for the field.

The main general findings of the work have been described elsewhere: that microexon inclusion increases in many transcripts during neuronal cell fate commitment has previously been described, the suggestions of important isoform structural changes in Hook2 and Kctd13 are not backed up by any experimental data and so are not reliable. The description of a huge number of novel isoforms is not particularly useful because it's not clear if these have been found by other similar studies, because the data is not compared, furthermore we have no information about these isoforms to be able to pursue further research about them. The main output of the work would be the data and transcript annotations for other people to follow up on, but this is not provided in any accessible way. The paper might be better reframed as a resource, if it is not possible to follow up on the biological conclusions. - Place the work in the context of the existing literature (provide references, where appropriate).

Previously, alternative splicing has been studied in purified cell types of the developing mouse cortex using short read sequencing eg. in Zhang 2016, Cell. In this previous study, VZ NPCs (EGFP−) and non-VZ cells (EGFP+) were isolated from E14.5 Tbr2-EGFP mouse cerebral cortex. The double reporter mouse model used in the present study allows for better cell sorting into NSC, NPC and neurons, and the long read sequencing allows for whole transcript identification, however the present study has made no effort to compare the data, so it's not clear how much new biology this leads to. In Zhang 2016, the authors also predict disruption to protein domains caused by AS, but go further to perform experiments to validate the impact of some of these predictions. - State what audience might be interested in and influenced by the reported findings.

Researchers of this cell fate transition might want to look at their favourite genes to see if there are novel isoforms reported (however this is currently not possible because this information is not provided). Researchers of Hook2 or Kctd13 may want to further explore the described predicted structural changes. Researchers generally studying alternative splicing may want to include the novel isoforms in their analyses (again currently not possible because they are not provided). Generally this paper would probably be best seen as a resource. - Define your field of expertise with a few keywords to help the authors contextualize your point of view.

Indicate if there are any parts of the paper that you do not have sufficient expertise to evaluate. Bioinformatics, Splicing, RBP biology

Interesting: using ternary operator in except clause

a thread pool and imap_unordered():

``` pool = multiprocessing.dummy.Pool(2)

for result in pool.imap_unordered(do_stuff, things_to_do): print(result) ```

Here, concurrency is limited by the fixed number of threads.

Async only makes sense if you're doing IO.

There's ZERO benefit in using async to stuff like this that is CPU-bound:

``` import asyncio

async def sum_two_numbers_async(n1: int, n2: int) -> int: return n1 + n2

async def main(): await sum_two_numbers_async(2, 2) await sum_two_numbers_async(4, 4)

asyncio.run(main()) ```

Your code might even get slower by doing that due to the Event Loop.

That's because Python async only optimizes IDLE time!

Reviewer #1 (Public Review):

Summary:<br /> Huang and Luo investigated whether regularities between stimulus features can be exploited to facilitate the encoding of each set of stimuli in visual working memory, improving performance. They recorded both behavioural and neural (EEG) data from human participants during a sequential delayed response task involving three items with two properties: location and colour. In the key condition ('aligned trajectory'), the distance between locations of successively presented stimuli was identical to their 'distance' in colour space, permitting a compression strategy of encoding only the location and colour of the first stimulus and the relative distance of the second and third stimulus (as opposed to remembering 3 locations and 3 colours, this would only require remembering 1 location, 1 colour, and 2 distances). Participants recalled the location and colour of each item after a delay.

Consistent with the compression account, participants' location and colour recall errors were correlated and were overall lower compared to a non-compressible condition ('misaligned trajectory'). Multivariate analysis of the neural data permitted decoding of the locations and colours during encoding. Crucially, the relative distance could also be decoded - a necessary ingredient for the compression strategy.

Strengths:<br /> The main strength of this study is a novel experimental design that elegantly demonstrates how we exploit stimulus structure to overcome working memory capacity limits. The behavioural results are robust and support the main hypothesis of compressed encoding across a number of analyses. The simple and well-controlled design is suited to neuroimaging studies and paves the way for investigating the neural basis of how environmental structure is detected and represented in memory. Prior studies on this topic have primarily studied behaviour only (e.g., Brady & Tenenbaum, 2013).

Weaknesses:<br /> The main weakness of the study is that the EEG results do not make a clear case for compression or demonstrate its neural basis. If the main aim of this strategy is to improve memory maintenance, it seems that it should be employed during the encoding phase. From then on, the neural representation in memory should be in the compressed format. The only positive evidence for this occurs in the late encoding phase (the re-activation of decoding of the distance between items 1 and 2, Fig. 5A), but the link to behaviour seems fairly weak (p=0.068). Stronger evidence would be showing decoding of the compressed code during memory maintenance or recall, but this is not presented. On the contrary, during location recall (after the majority of memory maintenance is already over), colour decoding re-emerges, but in the un-compressed item-by-item code (Fig. 4B). The authors suggest that compression is consolidated at this point, but its utility at this late stage is not obvious.

Impact:<br /> This important study elegantly demonstrates that the use of shared structure can improve capacity-limited visual working memory. The paradigm and approach explicitly link this field to recent findings on the role of replay in structure learning and will therefore be of interest to neuroscientists studying both topics.

I appreciate the attention to detail of returning nil if that's what should be returned (rather than accidentally just returning the return value from the last line of code, errors.add.

superficially and carelessly?

"Core performance" would probably describe this better, since this talks about the performance in the core competence dimension of a software developer. A generic name like "Performance" can be misleading since performance of a software engineer is much more than the quality of code.

In this case, the activity could be an input metric, like the # of hours a developer is working, or an output metric, like the # of lines of code, or the # of PRs and so on, both of which are tricky to be used as sole indicators of productivity.

Adding to this, developer activity is often seen through a narrow lens of code. While that's true for junior engineers, software engineering is much more than writing code, especially in senior positions, and involves planning, documenting, communicating/collaborating, execution, monitoring and maintenance.

OK, in that case I would wait until next chapter to introduce the R code. There isn't much value in showing someone how to get an answer in R that differs from your hand worked example

There are markdown documentation frameworks, but I still want to try sphinx.

ensure when you extracted it just has one release folder, so three files in the release folder

A Tale of Two Codes我會翻《雙典記》 code 法典 encode 編成法典

姑且不談code該怎麼翻，如原po律師指出，作者明顯泛稱的laywer一詞，譯者顯然帶著一種死腦筋，硬相信自己幾十年的淺度學習記憶，lawyer一定就是「律師」不能有他義，於是翻出令人好笑的意思。例如，犯了原文 lawyers ACT AS judges...翻成「律師的舉止有如法官」，連act as意思是「擔任、充當」如此清晰，都會變成「舉止有如」，這句意思是「法律人擔任法官（時）...」。

「最終形成公共意見」，「形成」應作「形塑」（影響、左右），原文是shape，不是form、make up。

「法律傾向...政治過程」那個病句真的令人納悶，編輯根本睡著了，原文有一個字漏翻。

``` import re import time

import httpx import trio

urls = [ "https://www.bitecode.dev/p/relieving-your-python-packaging-pain", "https://www.bitecode.dev/p/hype-cycles", "https://www.bitecode.dev/p/why-not-tell-people-to-simply-use", "https://www.bitecode.dev/p/nobody-ever-paid-me-for-code", "https://www.bitecode.dev/p/python-cocktail-mix-a-context-manager", "https://www.bitecode.dev/p/the-costly-mistake-so-many-makes", "https://www.bitecode.dev/p/the-weirdest-python-keyword", ]

title_pattern = re.compile(r"<title[^>]>(.?)</title>", re.IGNORECASE)

user_agent = ( "Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:109.0) Gecko/20100101 Firefox/116.0" )

async def fetch_url(url): start_time = time.time()

async with httpx.AsyncClient() as client:
    headers = {"User-Agent": user_agent}
    response = await client.get(url, headers=headers)
    match = title_pattern.search(response.text)
    title = match.group(1) if match else "Unknown"
    print(f"URL: {url}\nTitle: {title}")

end_time = time.time()
elapsed_time = end_time - start_time
print(f"Time taken for {url}: {elapsed_time:.4f} seconds\n")

async def main(): global_start_time = time.time()

# That's the biggest API difference
async with trio.open_nursery() as nursery:
    for url in urls:
        nursery.start_soon(fetch_url, url)

global_end_time = time.time()
global_elapsed_time = global_end_time - global_start_time
print(f"Total time taken for all URLs: {global_elapsed_time:.4f} seconds")

if name == "main": trio.run(main) ```

Because it doesn't create nor schedule coroutines immediately (notice the nursery.start_soon(fetch_url, url) is not nursery.start_soon(fetch_url(url))), it will also consume less memory. But the most important part is the nursery:

# That's the biggest API difference async with trio.open_nursery() as nursery: for url in urls: nursery.start_soon(fetch_url, url)

The with block scopes all the tasks, meaning everything that is started inside that context manager is guaranteed to be finished (or terminated) when it exits. First, the API is better than expecting the user to wait manually like with asyncio.gather: you cannot start concurrent coroutines without a clear scope in trio, it doesn't rely on the coder's discipline. But under the hood, the design is also different. The whole bunch of coroutines you group and start can be canceled easily, because trio always knows where things begin and end.

As soon as things get complicated, code with curio-like design become radically simpler than ones with asyncio-like design.

``` import re import time

import gevent from gevent import monkey

monkey.patch_all() # THIS MUST BE DONE BEFORE IMPORTING URLLIB

from urllib.request import Request, urlopen

urls = [ "https://www.bitecode.dev/p/relieving-your-python-packaging-pain", "https://www.bitecode.dev/p/hype-cycles", "https://www.bitecode.dev/p/why-not-tell-people-to-simply-use", "https://www.bitecode.dev/p/nobody-ever-paid-me-for-code", "https://www.bitecode.dev/p/python-cocktail-mix-a-context-manager", "https://www.bitecode.dev/p/the-costly-mistake-so-many-makes", "https://www.bitecode.dev/p/the-weirdest-python-keyword", ]

title_pattern = re.compile(r"<title[^>]>(.?)</title>", re.IGNORECASE)

user_agent = ( "Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:109.0) Gecko/20100101 Firefox/116.0" )

We move the fetching into a function so we can isolate it into a green thread

def fetch_url(url): start_time = time.time()

headers = {"User-Agent": user_agent}

with urlopen(Request(url, headers=headers)) as response:
    html_content = response.read().decode("utf-8")
    match = title_pattern.search(html_content)
    title = match.group(1) if match else "Unknown"

    print(f"URL: {url}\nTitle: {title}")

end_time = time.time()
elapsed_time = end_time - start_time

print(f"Time taken: {elapsed_time:.4f} seconds\n")

def main(): global_start_time = time.time()

# Here is where we convert synchronous calls into async ones
greenlets = [gevent.spawn(fetch_url, url) for url in urls]
gevent.joinall(greenlets)

global_end_time = time.time()
global_elapsed_time = global_end_time - global_start_time

print(f"Total time taken: {global_elapsed_time:.4f} seconds")

main() ```

No async, no await. No special lib except for gevent. In fact it would work with the requests lib just as well. Very few modifications are needed, for a net perf gain.

The only danger is if you call gevent.monkey.patch_all() too late. You get a cryptic error that crashes your program.

asyncio

asyncio is the modern module for asynchronous network programming provided with the python stdlib since 3.4. In other words, it's the default stuff at your disposal if you want to code something without waiting on the network.

asyncio replaces the old deprecated asyncore module. It is quite low level, so while you can manually code most network-related things with it, you are still at the level of TCP or UDP. If you want higher-level protocols, like FTP, HTTP or SSH, you have to either code it yourself, or install a third party library or module.

Because asyncio is the default solution, it has a the biggest ecosystem of 3rd party libs, and pretty much everything async strives to be compatible with it directly, or through compatibility layers like anyio.

Twisted

20 years ago, there was no asyncio, there was no async/await, nodejs didn't exist and Python 3 was half a decade away. Yet, it was the .com bubble, everything needed to be connected now. And so was born twisted, the grandfather of all the asynchronous frameworks we have today. Twisted ecosystem grew to include everything, from mail to ssh.

To this day, twisted is still a robust and versatile tool. But you do pay the price of its age. It doesn't follow PEP8 very well, and the design lean on the heavy size.

Tornado

Tornado was developed after Twisted, by FriendFeed, at this weird 2005-2015 web dev period where everything needed to be social web scale. It was like Twisted, but tooted to be faster, and was higher level. Out of the box, the HTTP story is way nicer.

Today, you are unlikely to use Tornado unless you work at Facebook or contribute to jupyter. After all, if you want to make async web things, the default tool is FastAPI in 2023.

gevent

Gevent came about in 2009, the same year as Tornado, but with a fundamentally different design. Instead of attempting to provide an asychronous API, it decided to do black magic. When you use gevent, you call from gevent import monkey; monkey.patch_all() and it changes the underlying mechanism of Python networking, making everything non-blocking.

Author Response

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

Public Reviews:

Reviewer #1 (Public Review):

This is an interesting, timely and informative article. The authors used publicly available data (made available by a funding agency) to examine some of the academic characteristics of the individuals recipients of the National Institutes of Health (NIH) k99/R00 award program during the entire history of this funding mechanism (17 years, total ~ 4 billion US dollars (annual investment of ~230 million USD)). The analysis focuses on the pedigree and the NIH funding portfolio of the institutions hosting the k99 awardees as postdoctoral researchers and the institutions hiring these individuals. The authors also analyze the data by gender, by whether the R00 portion of the awards eventually gets activated and based on whether the awardees stayed/were hired as faculty at their k99 (postdoctoral) host institution or moved elsewhere. The authors further sought to examine the rates of funding for those in systematically marginalized groups by analyzing the patterns of receiving k99 awards and hiring k99 awardees at historically black colleges and universities.

The goals and analysis are reasonable and the limitations of the data are described adequately. It is worth noting that some of the observed funding and hiring traits are in line with the Matthew effect in science (https://www.science.org/doi/10.1126/science.159.3810.56) and in science funding (https://www.pnas.org/doi/10.1073/pnas.1719557115). Overall, the article is a valuable addition to the research culture literature examining the academic funding and hiring traits in the United States. The findings can provide further insights for the leadership at funding and hiring institutions and science policy makers for individual and large-scale improvements that can benefit the scientific community.

Thank you for these comments. We have incorporated the articles referenced on the Matthew effect into the first paragraph of the Discussion our revised preprint.

Reviewer #2 (Public Review):

Early career funding success has an immense impact on later funding success and faculty persistence, as evidenced by well-documented "rich-get-richer" or "Matthew effect" phenomena in science (e.g., Bol et al. 2018, PNAS). Woitowich et al. examined publicly available data on the distribution of the National Institutes of Health's K99/R00 awards - an early career postdoc-to-faculty transition funding mechanism - and showed that although 85% of K99 awardees successfully transitioned into faculty, disparities in subsequent R01 grant obtainment emerged along three characteristics: researcher mobility, gender, and institution. Men who moved to a top-25 NIH funded institution in their postdoc-to-faculty transition experienced the shortest median time to receiving a R01 award, 4.6 years, in contrast to the median 7.4 years for women working at less well-funded schools who remained at their postdoc institutions. This result is consistent with prior evidence of funding disparities by gender and institution type. The finding that researcher mobility has the largest effect on subsequent funding success is key and novel, and enhances previous work showing the relationship between mobility and ones' access to resources, collaborators, or research objects (e.g., Sugimoto and Larivière, 2023, Equity for Women in Science (Harvard University Press)).

These results empirically demonstrate that even after receiving a prestigious early career grant, researchers with less mobility belonging to disadvantaged groups at less-resourced institutions continue to experience barriers that delay them from receiving their next major grant. This result has important policy implications aimed at reducing funding disparities - mainly that interventions that focus solely on early career or early stage investigator funding alone will not achieve the desired outcome of improving faculty diversity.

The authors also highlight two incredible facts: No postdoc at a historically Black college or university (HBCU) has been awarded a K99 since the program's launch. And out of all 2,847 R00 awards given thus far, only two have been made to faculty at HBCUs. Given the track record of HBCUs for improving diversity in STEM contexts, this distribution of awards is a massive oversight that demands attention.

At no fault of the authors, the analysis is limited to only examining K99 awardees and not those who applied but did not receive the award. This limitation is solely due to the lack of data made publicly available by the NIH. If this data were available, this study would have been able to compare the trajectory of winners versus losers and therefore could potentially quantify the impact of the award itself on later funding success, much like the landmark Bol et al. (2018) paper that followed the careers of winners of an early career grant scheme in the Netherlands. Such an analysis would also provide new insights that would inform policy.

Although data on applications versus awards for the K99/R00 mechanism are limited, there exists data for applicant race and ethnicity for the 2007-2017 period, which were made available by a Freedom of Information Act request through the now defunct Rescuing Biomedical Research Initiative: https://web.archive.org/web/20180723171128/http://rescuingbiomedicalresearch.org/blog/examining-distribution-k99r00-awards-race/. These results are not presently discussed in the paper, but are highly relevant given the discussion of K99 award impacts on the sociodemographic composition of U.S. biomedical faculty. From 2007 to 2017, the K99 award rate for white applicants was 31.0% compared to 26.7% for Asian applicants and 16.2% for Black applicants. In terms of award totals, these funding rates amount to 1,384 awards to white applicants, 610 to Asian applicants, and 25 to Black applicants for the entire 2007-2017 period. And in terms of R00 awards, or successful faculty transitions: whereas 77.0% of white K99 awardees received an R00 award, the conversion rate for Asian and Black K99 awardees was lower, at 76.1% and 60.0%, respectively. Regarding this K99-to-R00 transition rate, Woitowich et al. found no difference by gender (Table 2). These results are consistent with a growing body of literature that shows that while there have been improvements to equity in funding outcomes by gender, similar improvements for achieving racial equity are lagging.

The conclusions are well-supported by the data, and limitations of the data and the name-gender matching algorithm are described satisfactorily.

One aspect that the authors should expand or comment on is the change in the rate of K99 to R00 conversions. Since 2016, while the absolute number of K99 and R00 awards has been increasing, the percentage of R00 conversions appears to be decreasing, especially in 2020 and 2021. This observation is not clearly stated or shown in Figure 1 but is an important point - if the effectiveness of the K99/R00 mechanism for postdoc-to-faculty transitions has been decreasing lately, then something is undermining the purpose of this mechanism. This result bears emphasis and potentially discussion for possible reasons for why this is happening.

Thank you for these insightful comments. We now calculate a rolling conversion rate for K99 to R00 awards which shows there is not as much of a decline in conversion from K99 to R00 (Fig 1B). We still see a slight decline in 2021 and 2022. 468 K99 awards are from 2020 or later so they may still convert to the R00 phase. Thus it is difficult to draw conclusions about 2021/2022 yet. As more time passes, we may better be able to determine whether or not significant alteration from normal occurred in these years, presumably due to pressures from the Covid-19 pandemic. We also thank you for providing the details of the FOIA request. We have included a discussion of these data in the discussion.

Reviewer #3 (Public Review):

The researchers aim add to the literature on faculty career pathways with particular attention to how gender disparities persist in the career and funding opportunities of researchers. The researchers also examine aspects of institutional prestige that can further amplify funding and career disparities. While some factors about individuals' pathways to faculty lines are known, including the prospects of certain K award recipients, the current study provides the only known examination of the K99/R00 awardees and their pathways.

Strengths:

The authors establish a clear overview of the institutional locations of K99 and R00 awardees and the pathways for K99-to-R00 researchers and the gendered and institutional patterns of such pathways. For example, there's a clear institutional hierarchy of hiring for K99/R00 researchers that echo previous research on the rigid faculty hiring networks across fields, and a pivotal difference in the time between awards that can impact faculty careers. Moreover, there's regional clusters of hiring in certain parts of the US where multiple research universities are located. Moreover, documenting the pathways of HBCU faculty is an important extension of the Wapman et al. study (among others from that research group), and provides a more nuanced look at the pathways of faculty beyond the oft-discussed high status institutions. (However, there is a need for more refinement in this segment of the analyses as discussed further below.). Also, the authors provide important caveats throughout the manuscript about the study's findings that show careful attention to the complexity of these patterns and attempting to limit misinterpretations of readers.

Weaknesses:

The authors reference institutional prestige in relation to some of the findings, but there's no specific measure of institutional prestige included in the analyses. If being identified as a top 25 NIH-funded institution is the proximate measure for prestige in the study, then more justification of how that relates to previous studies' measures of institutional prestige and status are needed to further clarify the interpretations offered in the manuscript.

The identification of institutional funding disparities impacting HBCUs is an important finding and highlights another aspect of how faculty at these institutions are under resourced and arguably undervalued in their research contributions. However, a lingering question exists: why compare HBCUs with Harvard? What are the theoretical and/or methodological justifications for such comparisons? This comparison lends itself to reifying the status hierarchy of institutions that perpetuate funding and career inequalities at the heart of the current manuscript. If aggregating all HBCU faculty together, then a comparable grouping for comparison is needed, not just one institution. Perhaps looking at the top 25 NIH funded institutions could be one way of providing a clearer comparison. Related to this point is the confusing inclusion of Gallaudet in Figure 6 as it is not an officially identified HBCU. Was this institution also included in the HBCU-related calculations?

Thank you for this comment. We agree this comparison perpetuates the perception of the prestige hierarchy and is problematic. We now compare all institutions in the top 25 NIH funding category to all HBCUs. Thank you also for identifying our error in mis-coding Gallaudet as an HBCU. We have corrected this in the current version.

There is a clear connection that is missed in the current iteration of the manuscript derived from the work of Robert Merton and others about cumulative advantages in science and the "Matthew effect." While aspects of this connection are noted in the manuscript such as well-resourced institutions (those with the most NIH funding in this circumstance) hire each others' K99/R00 awardees, elaborating on these connections are important for readers to understand the central processes of how a rigid hierarchy of funding and career opportunities exist around these pathways. The work the authors build on from Daniel Larremore, Aaron Clauset, and their colleagues have also incorporated these important theoretical connections from the sociology of knowledge and science, and it would provide a more interdisciplinary lens and further depth to understanding the faculty career inequalities documented in the current study.

Reviewer #1 (Recommendations For The Authors):

Comments to authors:

1. For the benefit of general reader, it would be informative to mention the amount of annual NIH investment in the k99 funding mechanism in the text (230 awards representing a ~ 230 million US dollars investment).

Thank you for this suggestion. We have added that this is ~$25 million investment annually.

1. It is worth noting that some of the observed funding and hiring traits resemble the Matthew effect, discussed in: The Matthew effect in science: https://www.science.org/doi/10.1126/science.159.3810.56

The Matthew effect in science funding: https://www.pnas.org/doi/10.1073/pnas.1719557115

It would be of value to cite these for further context for the readers.

Thank you for this suggestion. We have included these references and briefly discussed the Matthew effect in the first paragraph of the Discussion.

1. Figs 3, 6 and Fig S1 are hard to read without zooming in due to their format and don't work great within a letter size page but can work if they are also linked to a zoomable web version. It would make sense to have an online navigable/searchable/selectable version. But when the reader zooms out, there are patterns that reflect what points the authors are making (though those could be illustrated differently). These figures are really made for online webapp visualization (such as Shiny in R).

We agree with this comment and have used the “googleVis()” package in R to put together interactive Sankey diagrams. These can be found at: https://dantyrr.github.io/K99-R00-analysis/ and they are referenced in the manuscript.

1. The abstract states 85% of awardees get R00 awards. That appears to come from 198/234 (page 6) though it's not explicitly stated, and other ratios give different answers (e.g., 1-304/3475 = 91%) but the 85% seems to be the right one. That first paragraph of the results could be clearer. Also, in the middle of page three the number given is 90% so something is inconsistent. For Figure 1A, given the methodology it should be possible to calculate a rolling conversion rate as "R00(t) / K99(t-1)" (and a similarly-calculated cumulative rate).

Thank you for catching these errors. These were introduced because there are R00 awardees that did not have extramural K99 awards. These are intramural NIH K99 awardees but there is no public data on these awardees. The correct number is 78% of K99 awardees that transitioned to the R00 phase. We have also calculated the rolling conversion rate which is 89% if you exclude the first 2 years of the program (when the first awardees were within the 2-yr K99 period) and final 2 years (when most recent K99 awardees were still within their first 2 years of the K99 period).

1. Assuming that 85% is the correct number, is there any information/insight into why ~1/6 of awardees do not continue to R00, which seems high given that only two years passes - that's a lot of awardees not getting R00 positions.

We are unsure of why these don’t convert. In the revised version of the manuscript, we speculate on this in the 4th paragraph of the discussion:

The factors that prevented the other 302 K99 awardees from 2019 and earlier unable to convert their K99-R00 grants is cause for concern within our greater academic community. Possible explanations include leaving the biomedical workforce, accepting tenure-track positions or other positions abroad, or by simply not successfully securing a tenable tenure-track offer.

1. It looks like perhaps a non-zero number of K99s are just one year and not two (e.g., see 2006 in Fig 1A, which should not appear if all 2006 awards were 2 years). What is the typical percentage of K99s not activated for a second year, and is this a sizable % of the 15% not converting to R00?

This is an interesting question. We didn’t originally look into this and the dataset that we originally downloaded from NIH reporter included a significant number of duplicates for the grants because year 1 of the K99 was listed on its own line and year 2 was listed on a different line. The first step in curating the data was to delete the duplicate values so we only had one entry per person. Unfortunately based on sorting of the data tables, sometimes the year 1 appeared above year 2 and at other times year 2 appeared before year 1. Because none of the data we were interested in are benchmarked to K99 start date, we removed the duplicate values non-specifically. With the dataset we currently have, we would not be able to tell which individuals dropped out (didn’t convert to R00) during the first or second year of the K99. In order to do this we would have to download the raw data from NIH reporter again and curate it again. We may do this in the future but for the purpose of publishing the current manuscript we prefer to focus our efforts on other aspects of the revision.

1. Further down page 3, the authors state that "men typically experience 2-3% greater funding success rates" is ambiguous, as rates are themselves a percentage. So, is it 2-3% greater as in 23% vs 20%, or is it 2-3% greater as in 20.6% vs 20%? Please clarify the language.

Thank you for asking for this clarification. We have updated the text here to reflect that we mean “23% vs 20%”.

1. Metrics such as time to first R01 are compared internally within the study set, which yields interesting insights, but more could be done to benchmark these metrics to non-K99 scientists.

We agree with the reviewer that this would be ideal; however, we feel that it is out of the scope of this manuscript. We may examine this in the future.

1. In the text, several times percentages are being referred to when the figures cited do not show percentages. For example (page 6) 'proportion of awardees that stayed at the same institution declined to about 20% where it has remained consistent (Fig 1B)' - Figure 1B does not show percentages, instead the reader would need to work out from the raw numbers what the pattern of percentages might look like. It's fine (great even) to provide the raw numbers, but would be great to show the percentages as well. This happened for multiple graphs.

Thank you for this comment. We agree that showing the percentage would be beneficial so we have included the percentages in Figure 1 for the conversion rate. We also added a standalone figure panel for the rolling conversion rate for Figure 1. For Figure 4, we have also included a right Y-axis to better indicate the % women.

1. Figure 4 - putting the %women on a 0-250 scale makes it difficult to see the changes in that curve. Please replot it as a separate graph with an appropriate scale (30-50%? 30-70%?)

Thank you for this comment. We have made this edit.

1. Figure 5 - The table appears inconsistent - the Moved/Stayed HR is 1.411 suggesting that moving is better for reducing time to R01, but then Woman/Man is 1.208, so one of these pairs needs to be written in the opposite order to have the table make sense (intended to be listed as 'better/worse'?)

Thank you for noticing this. In the revised manuscript we have re-run the cox proportional hazard model using the R package “survival” and the function “coxph()”. There were minor differences in the hazard ratios using this package instead of Graphpad prism; however, the R package is much more widely used compared to prism for these types of analysis. We present the new data in the table in Figure 5B in the revised manuscript. We now present the “detrimental” cox hazard value for each variable (i.e. 0.7095 for the mobility [moved/stayed]). We also underlined the variable which was detrimental to receiving an R01 award earlier.

1. Figure 5's graph appears strange. All the lines have an appearance of stochasticity but are actually multiples of each other, rising exactly in sync. Are these actually modeled lines? If so, why not instead actually draw the lines based on the real data from the real groups depicted, and give the n for each group?

Thank you for picking this up. The software we originally used to plot the graphs did plot modeled lines instead of the actual data. We have re-run the cox proportional hazard model using the R “survival” package v3.5-5 and the coxph() and survfit() functions. The updated data are in Figure 5 of the revised manuscript.

1. Table 1 should note that each column sums to 100%.

This is a good suggestion. In the revised manuscript, we have added a row to the table to indicate the column total N and %.

1. The authors discuss how k99/R00 grant reviewing process may have to change but the k99 awards also impact the faculty hiring ecosystem as well. There are faculty hiring job ads explicitly requesting or indicating preference towards k99 holders and the results described in this article show that k99 awarding is biased towards particular demographics at select wealthy institutions. Of course, collective/central action is almost always more effective/impactful (especially in shorter time line) than individual elective action. In other words, NIH changing granting patterns would likely work better than encouraging faculty searches to change the weight they give to K99s, because there are many searches and just one NIH. But these are not mutually exclusive and individual action can still help when central action isn't done (if the NIH does not change the k99/R00 grant review process for more inclusive funding and does not increase the number of annual k99 awards hence the annual budget for this award mechanism) and it would be good to have this discussed in the manuscript.

Thank you for this comment and thoughtful insights. We have included additional discussion on this in the final paragraph of the discussion.

Reviewer #2 (Recommendations For The Authors):

Thank you for conducting this important work. On top of some thoughts I have described in the public review (in particular, Chris Pickett's FOIA data on K99/R00 outcomes by applicant race and ethnicity), I only have a few comments for potential improvements to this paper:

1. The comparison of K99-R00 transition rates by gender was interesting. However, I missed the analysis on the K99-R00 transition rates by institution (by type or by top-25 NIH funded institution versus not). I think this analysis may be buried somewhere in the more nuanced descriptions about faculty flows from one institution type to another, but I was not able to locate it. I wonder if the authors could consider dedicating a subsection to specifically describing the transition rate by institution type, creating a table equivalent to Table 2. This section would probably fit best somewhere before the authors dive into the nuances of self-hires and faculty flows.

Said another way: As I was reading, I felt I was missing an answer to a simple question - are there differences in conversion rates by institution type (however you define institution type, as an MSI or non MSI, or top-25 NIH funded versus not)?

Thank you for this suggestion. We have created the table (Table 3 and Table 4) in the revised manuscript. We also made a new figure (now figure 5 in the revised manuscript). This was an interesting way to look at the data and it is very clear that the number of K99 and R00 awards is heavily concentrated within the institutions that have the highest NIH funding. We have added a paragraph in the results in a new section entitled “K99 and R00 awards are concentrated within the highest funded institutions”.

1. Regarding the comparison of HBCUs and Harvard: this analysis was elucidating, but I am not sure if the framing of this analysis as pertaining to "systematically marginalized groups" - see second sentence in the section, "Faculty doctorates differ between Harvard and HBCUs" is appropriate. While it is true that proportionally more faculty at HBCUs are from marginalized groups, there are also many faculty at HBCUs who are from privileged or advantaged backgrounds (e.g., white, men, educated at elite institutions). It would be more accurate to rephrase the second sentence to say something along the lines of, "We sought to examine the rates of funding for those at historically under-funded institutions." I recommend that the authors comb the paper for any other potential places in the text that conflate systemic marginalization with institution type, and rephrase as needed for accuracy.

Thank you for pointing this out. This is an extremely important point and we have removed any instances we could find where we conflate systemically marginalized groups with institution type.

1. I strongly recommend Sugimoto and Larivière (2023)'s new book, Equity for Women in Science, which has an entire section dedicated to previous work investigating how researcher mobility impacts access to resources, collaborations, et cetera (Chapter 5 on Mobility; other chapters on Funding are also relevant but I hone in on Mobility since this is such a key result of this work). I think this chapter would provide significant food-for-thought and background that could strengthen the Discussion section of the paper.

Thank you for this suggestion. We have added some discussion of mobility in the first paragraph of the Discussion.

1. I appreciated the subsection headings that described key results (e.g., "Institutions with the most NIH funding tend to hire K99/R00 awardees from other institutions with the most funding"; "K99/R00 awardee self-hires are more common at institutions with the top NIH funding.") This paper structure made it easier for me to ensure that I was getting the intended takeaway from a figure or section. But partway through the paper, the subheadings changed to being less declarative and therefore less informative (e.g., "Gender of K99/R00 awardees"; "Factors influencing K99/R00 awardee future funding success"). It would be great to rephrase these boilerplate subsection headers to be more declarative, like earlier subsection headings. For example, maybe say "Men receive the majority of K99 awards" or "No gender difference in the rate of conversion from K99 to R00" or something to that effect, depending on what result the authors wish to emphasize.

Thank you for this comment. This is a very good point. We have re-worded the more generic headings in the revised version.

1. Lastly, I would like to share a question that came to my mind that involves an additional analysis, but is work that is (probably) out-of-the-scope of this paper, but could instead be a separate paper or product. Circling back to Chris Pickett's FOIA-ed data on K99/R00 funding outcomes by applicant race and ethnicity (https://web.archive.org/web/20180723171128/http://rescuingbiomedicalresearch.org/blog/examining-distribution-k99r00-awards-race/): Given that Pickett's numbers provide incontrovertible information on the number of awards to various racial and ethnic groups, I wonder if it is possible to use this information as an "answer key" to (1) check the accuracy of an algorithm that assigns race based on name for applications in your analysis but for 2007-2017 period, and, (2) if the results are reasonable, then examine the dataset with race and ethnicity information. Some recent papers performing large-scale bibliometric analyses have applied such algorithms (e.g., see Kozlowski et al. 2022 PNAS Intersectional inequalities in science) and I wonder if they could be useful, or at least tested, here. Again, Pickett's data would serve as the benchmark to see if the algorithm produces numbers that are consistent with the actual funding outcomes; if they're not wildly off, or perhaps accurate for some groups but not others, there might be something here.

This is a really insightful comment. We have discussed whether we could assign ethnicity based on an algorithm and check based on Chris Pickett’s data. We agree that it is beyond the scope of this article, but has potential for future research.

Reviewer #3 (Recommendations For The Authors):

-In the methods section, it would be helpful to provide an overview of the number of universities, departments, and faculty represented in the data analyzed in the study.

Thank you for this comment. We agree with the reviewer. We have added a section to the results discussing the distribution of different types of institutions. We also added Table 3 and Table 4 and a new Figure 5 describing these. Regarding the faculty, we have discussed the demographics of the K99 and R00 awardees as best as we could. We do not have data on which faculty laboratories the K99 awardees were in when they received their awards. This information is not available through NIH reporter.

-I would consider incorporating, or at least citing, Jeff Lockhart and colleagues' recent paper Nature Human Behavior article "Name-based demographic inference and the unequal distribution of misrecognition" about to provide readers with an additional resource and more information about the likelihood of misattribution and general cautionary notes about using gender and race/ethnicity ascription/imputation approaches and tools for research.

Thank you for bringing this reference to our attention. We have incorporated this into the methods section describing our name-based gender determination.

-In the next to last sentence under the final paragraph of the methods section, there looks to be a typo as it should read "K99 or R00," not "K00" as currently written.

Thank you for catching this. We have now corrected it.

-Clarifying some of the data and measures used are necessary to limit confusion and misinterpretations of the study's findings.

Thank you. We have significantly updated the revised manuscript and hope that it is more clear.

-Elaborating more on the gender inequality notable in the Cox proportional hazard model would strengthen the authors' point about persistent gender inequalities within the K99/R00 funding mechanism and pathways. In its current iteration, the findings are somewhat buried by the discussion of institutional differences, but when we look at the findings and the plot associated with the model, we notice that men have more advantages than women in funding and institutional location.

Thank you for highlighting this. This is true and we have elaborated on the gender inequality in the revised version of the manuscript.

-Also for the Cox proportional hazard model, I would consider exploring the inclusion of data that can further clarify the biomedical research infrastructure of institutions. For example, in the conversation about the differences between Princeton and other universities including other Ivies, it's important to note that Princeton does not have a medical school. Moreover, other institutions do not operate or are affiliated with a hospital. Adding more data to the model that can better contextualize the research infrastructure around researchers with NIH awards beyond the size of the NIH portfolio can shed light on possibly other important institutional differences that undergird these inequalities.

Thank you for this comment. We have added additional details about the institutional type; however, to examine whether institutions are attached to a hospital (or are themselves as hospital like MGH etc.) or whether institutions include a medical school may be difficult. We would have to manually code these and then determine whether or not the award recipient was affiliated with a department within that entity or not. We believe that this is a fascinating question but that it is out of the scope of the present manuscript. This is something that we will look into for potential future publications.

-Throughout the manuscript there's usage of "elite" and "prestigious" that are somewhat ambiguous regarding what exactly they are referring to about institutional characteristics. This is a common issue in the literature, but trying to clarify what these terms specifically mean for the current study and checking for consistent usage with limited interchangeability that can add confusion for readers about what is being referred to would give added strength to the conversation provided by the authors.

Thank you for this suggestion. Based on these comments and those by the other reviewers, in the revised version of the manuscript, we have limited the use of “elite” and “prestigious” to describe institutions in order not to perpetuate biases toward certain institutions.

-In relation to the discussion at the end of the manuscript of the longer time to award noted for researchers who stay at the same institutions, another possibility for the disparity could be their reliance for service work (e.g., hiring committees, departmental committees, supporting graduate students through mentoring and/or dissertation committee work, etc.) in their institutions given their knowledge of and experience within it.

Thank you for this suggestion. We have added 2 sentences to the discussion reflecting this possibility.

-Engaging with how STEM professional cultures can perpetuate these funding disparities and related hiring and career outcomes could enhance the contributions of the study. In relation to STEM professional cultures, engaging with the work of Mary Blair-Loy and Erin Cech in their recent book, Misconceiving Merit, could help provide additional insights for readers.

Thank you for these comments. We have incorporated edits to the revised manuscript reflecting the work of Erin Cech and Mary Blair-Loy.

Different section

```python from concurrent.futures import ThreadPoolExecutor, as_completed

with ThreadPoolExecutor(max_workers=5) as executor: executor.submit(do_something_blockint) ```

The true distinction that we should be teaching students is the difference between properties of languages that can be determined statically—that is, by just staring at the code without running it—and properties that can only be known dynamically, during runtime.

Notice that I said “properties” and not “languages”. Every programming language chooses its own set of properties that can be determined either statically or dynamically, and taken together, this makes a language more “dynamic” or more “static”. Static versus dynamic is a spectrum, and yes, Python falls on the more dynamic end of the spectrum. A language like Java has far more static features than Python, but even Java includes things like reflection, which is inarguably a dynamic feature.

For instance, JavaScript is commonly lumped into the “interpreted language” category. But for a while, JavaScript running in Google Chrome would never be interpreted—instead, JavaScript was compiled directly to machine code! As a result, JavaScript can keep pace with C++.

$ python3 -m compileall .

This will place the compiled bytecode of all Python files in the current directory in pycache/ and show you any compiler errors.

The CPython interpreter really is an interpreter. But it also is a compiler. Python must go through a few stages before ever running the first line of code:

1. scanning
2. parsing

Older versions of Python added an additional stage:

1. scanning
2. parsing
3. checking for valid assignment targets

Let’s compare this to the stages of compiling a C program:

1. ~~preprocessing~~
2. lexical analysis (another term for “scanning”)
3. syntactic analysis (another term for “parsing”)
4. ~~semantic analysis~~
5. ~~linking~~

Here is the buggy program:

python 1 / 0 print() = None if False ñ = "hello

Each line of code generates a different error message:

• 1 / 0 will generate ZeroDivisionError: division by zero.
• print() = None will generate SyntaxError: cannot assign to function call.
• if False will generate SyntaxError: expected ':'.
• ñ = "hello will generate SyntaxError: EOL while scanning string literal.

The question is… which will be reported first?

Spoilers: the specific version of Python matters (more than I thought it would) so keep that in mind if you see different results.

The first error message detected is on the last line of source code. What this tells us is that Python must read the entire source code file before running the first line of code. If you have a definition in your head of an “interpreted language” that includes “interpreted languages run the code one line at a time”, then I want you to cross that out!

1. preprocessing
2. lexical analysis
3. syntactic analysis
4. semantic analysis
5. linking

In this article you started implementing your own version of Python. To do so, you needed to create four main components:

A tokenizer: * accepts strings as input (supposedly, source code); * chunks the input into atomic pieces called tokens; * produces tokens regardless of their sequence making sense or not.

A parser: * accepts tokens as input; * consumes the tokens one at a time, while making sense they come in an order that makes sense; * produces a tree that represents the syntax of the original code.

A compiler: * accepts a tree as input; * traverses the tree to produce bytecode operations.

An interpreter: * accepts bytecode as input; * traverses the bytecode and performs the operation that each one represents; * uses a stack to help with the computations.

The tokenizer is the part of your program that accepts the source code and produces a linear sequence of tokens – bits of source code that you identify as being relevant.

• Tokenizer takes source code as input and produces tokens;
• Parser takes tokens as input and produces an AST;
• Compiler takes an AST as input and produces bytecode;
• Interpreter takes bytecode as input and produces program results.

A lot of various sources being used since they really do not know much that is going on.

Reviewer #1 (Public Review):

The brain's code is not static. Neuronal activity patterns change as a result of learning, aging, and disease. Reliable tracking of activity from individual neurons across long time periods would enable detailed studies of these important dynamics. For this reason, the authors' efforts to track electrophysiological activity across days without relying on matching neural receptive fields (which can change due to learning, aging, and disease) are very important.

By utilizing the tightly-spaced electrodes on Neuropixels probes, they are able to measure the physical distance and the waveform shape 'distance' between sorted units recorded on different days. To tune the matching algorithm and validate the results, they used the visual receptive fields of neurons in the mouse visual cortex (which tend to change little over time) as ground truth. Their approach performs quite well, with a high proportion of neurons accurately matched across multiple weeks. This suggests that the method may be useable in other cases where the receptive fields can't be used as ground truth to validate the tracking. This potential extendibility to tougher applications is where this approach holds the most promise.

The main caveat (and disappointment) is that this paper does not address generalizability to other experimental conditions. Because it only looks at one brain area (visual cortex), in one species (mouse), using one type of spike sorter (Kilosort), and one type of behavioral prep (head-fixed), it is not clear if this approach is overfit to those conditions or if it will perform equally well in other conditions. Most importantly, in brain areas where neuronal receptive fields are more dynamic and can't be used as a ground truth diagnostic, it isn't clear how to apply the technique outlined in this study, since many of the parameters are tuned to a very specific set of conditions using visual receptive fields as ground truth.

If you want both the benefits of caching, and to get security updates within a reasonable amount of time, you will need two build processes:

1. The normal image build process that happens whenever you release new code.
2. Once a week, or every night, rebuild your Docker image from scratch using docker build --pull --no-cache to ensure you have security updates.

This might be a good place to put two notes on how the way we encounter these systems is changin One trend is toward integrating AI into existing software like Google Doc or Word or the Office Suite. Another trend is toward multimodal, more agentlike systems that "decide" to do multiple kinds of things, so not a discrete text generator app but a system that will run code, search the web, generate and analyze images as well as generating text. ChatGPT Premium is already like this.

Whether using sub interpreters or multiprocessing you cannot simply send existing Python objects to worker processes.

Multiprocessing uses pickle by default. When you start a process or use a process pool, you can use pipes, queues and shared memory as mechanisms to sending data to/from the workers and the main process. These mechanisms revolve around pickling. Pickling is the builtin serialization library for Python that can convert most Python objects into a byte string and back into a Python object.

Pickle is very flexible. You can serialize a lot of different types of Python objects (but not all) and Python objects can even define a method for how they can be serialized. It also handles nested objects and properties. However, with that flexibility comes a performance hit. Pickle is slow. So if you have a worker model that relies upon continuous inter-worker communication of complex pickled data you’ll likely see a bottleneck.

Sub interpreters can accept pickled data. They also have a second mechanism called shared data. Shared data is a high-speed shared memory space that interpreters can write to and share data with other interpreters. It supports only immutable types, those are:

• Strings
• Byte Strings
• Integers and Floats
• Boolean and None
• Tuples (and tuples of tuples)

To share data with an interpreter, you can either set it as initialization data or you can send it through a channel.

• A Python process, which contains one or more interpreters
• An interpreter, which contains a lock (the GIL) and one or more Python threads
• A thread, which contains information about the currently executing code.

This was a great way of introducing an outside audience to the technical terms. You made them seem more down to earth and familiar to the unfamiliar reader.

Getters and setters are typically named using the following conventions: * Getters start with the word "get" followed by the name of the variable. * Setters start with the word "set" followed by the name of the variable. For example, if you have a private variable named "name", you would create a getter method named "getName()" and a setter method named "setName()".

Strategies for Addressing Freedom and Autonomy * Ensure humans are in control of decisions * Allow clinicians control of the tech * Universal code for patient-clinician relationship * Make AI Info comprehensible to patients

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

Learn more at Review Commons

Reply to the reviewers

1. General Statements

We express our gratitude to the reviewers for their time and insightful comments, which have significantly contributed to the enhancement of our manuscript. We believe that the thoughtful critiques and suggestions have substantially improved the overall quality of our work. The changes made in the revised manuscript were highlighted in red. Below, we provide a point-by-point response to each comment, addressing the concerns raised by the reviewers.

2. Point-by-point description of the revisions

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

*Summary: *

*In the current study, Li et al investigated how TGF-beta signaling is controlled by protein abundances. Computational modeling and experiments indicated that the abundance of TGFBR1 and TGFBR2 affects the signaling, and those with lower abundance affect the signaling more, resembling Liebig's law of the minimum. Specifically, they showed that by using multiple cell lines with a different abundance of receptors, modulation of expression of the less abundant receptor impacts the signaling, which is measured by SMAD2 nuclear-to-cytosol ratio and/or relative phospho-SMAD2 level. Also, by using a light-induced interaction system, they showed that the signaling is dependent on the concentration of receptor complex when both receptors are expressed at similar amounts. *

*Major comments: *

*Computational predictions support the authors' idea. The computation and the experiments are well-documented. And it would gain substantially if the authors fill the gap between the predictions and the experiments as follows. *

*In Figure 4, the authors showed that perturbation on receptors with lower expression levels in each cell line changes the phospho-SMAD2 level. Although the data looks consistent with their claim, the result is only qualitative. The authors established a computational model in the former sections, thus it would be of great interest to assess if the experimental results quantitatively match the computational prediction. *

Response: The reviewer suggests that our work could benefit from a quantitative comparison between computational predictions and experimental data shown in Figure 4. We appreciate this suggestion. Given the challenges in obtaining precise quantification of TGFBR1 protein due to antibody issues (see the response to comment #2 from reviewer 2), a direct quantitative comparison between model predictions and experimental results is difficult. Our model predictions about the control principle with Liebig's law of the minimum should be interpreted qualitatively, rather than a strict quantitative law. We have explicitly indicated in the revised manuscript that our siRNA knockdown experiments are to qualitatively test our model predictions.

*In Figure 5, the authors computationally predicted that the expression level of receptors is correlated with SMAD2 N2C levels 1 hour after stimulation, and the strength of negative feedback with SMAD2 N2C levels 8 hours after stimulation. Because the authors employed iRFP-SMAD2 system, the prediction could be verified experimentally, at least the prediction on SMAD2 N2C 1 hour after stimulation could be checked. (In a sense, this is partially verified by the data in Figure 7, where both receptors are expressed at similar levels). It would gain substantially if the authors could verify the computational prediction in Figure 6. Since the authors stated in the introduction that "The same TGF-beta ligand can initiate different signaling responses depending on the cellular context, but the underlying control principle remains unclear...Together, these results revealed an effect of the minimum control in the TGF-beta pathway, which may be an important principle of control in signaling pathways with context-dependent outputs.", experimental verification of the prediction done in Figures 4-6 will be very important. Or the authors should stress that these points are only predicted by computational models. *

__Response: __The reviewer recommends verifying the model predictions in Figure 6 experimentally, particularly regarding SMAD2 N2C levels 1 hour after stimulation. We appreciate this valuable suggestion, which was also raised by reviewer 2. In response, we conducted experiments as recommended by reviewer #2, in which imbalanced expression of TGFBR1 and TGFBR2 was achieved by transfecting optoTGFBR1 or optoTGFBR2 plasmids into optoTGFBRs-HeLa cells, which initially expressed similar levels of both receptors. Western blot analysis confirmed the desired imbalance (Figure S13).

Consistent with the model predictions (Figure 6), the strong correlation between SMAD2 N2C fold change response at 1h and optoTGFBR2-tdTomato expression levels persisted in single cells when optoTGFBR1 was overexpressed (Figure 8A). Conversely, the high correlation between nuclear SMAD2 signaling and optoTGFBR2-tdTomato expression levels vanished at single cell level when optoTGFBR2 was overexpressed (Figure 8B). These experimental results validate our model predictions, confirming that the SMAD2 signaling is determined by the low abundance TGF-beta receptor in single cells. Incorporating these experimental validations enhances the quantitative support for our model predictions and clarifies the relationship between TGF-beta receptor abundance and signaling outcomes in single cells.

*As written in the below "Significance" section, the result is, in a sense, obvious. It should be stated that because the study utilized a slightly high concentration of TGF-beta in the experiments, it might be natural that the low-abundance receptor becomes a bottleneck of the signaling. It would gain to assess how receptor abundance affects signaling with the stimulation of lower concentrations of TGF-beta, or to examine the computational model if the low abundance of a receptor becomes a bottleneck of signaling because of saturation. Also, it is highly recommended to discuss the physiological implication of the current study, taking into account the experimental conditions used. *

Response: We appreciate the reviewer's insightful comments regarding the concentration of TGF-beta used in our experiments and the potential influence on the model predictions. In our experiments and model simulations, we utilized 100 pM TGF-beta, equivalent to 2.5 ng/mL (not 4.4 ng/mL as calculated by the reviewer). This concentration is a widely used dose in TGF-beta signaling studies. The reviewer's suggestion to explore how varying TGF-beta concentrations might influence the minimum control concept prompted us to extend our computational simulations. We used the extended model to perform simulations with lower TGF-beta concentrations (25 pM, equivalent to 0.625 ng/mL, and 10 pM, equivalent to 0.25 ng/mL). The results, depicted in Figure S7 of the revised manuscript, reaffirm that even at lower TGF-beta stimulations, a low abundance of a TGF-beta receptor acts as a bottleneck for SMAD2 signaling.

Following the reviewer’s suggestion, we have incorporated additional paragraphs to discuss the physiological implications and potential limitations of our study (Page 16-17 in the Main text).

It is pertinent to note that while the concept of TGF-beta signaling response being dictated by the minimum abundance of TGF-beta receptors may seem intuitive or even obvious, theoretical and experimental validations are crucial. As demonstrated in Figure S1B, our new simulation results from the minimal model illustrate similar response profiles when a high binding affinity (K1) is set for ligand-receptor interactions (Figure S1A). However, with a small binding affinity (K1), the minimal model indicates that TGF-beta signal response remains proportional to the product of TGFBR1 and TGFBR2 abundance and can be sensitive to the change of high abundance receptor in some region (Figure S1B). This highlights that the observed response patterns aligning with Liebig's law of the minimum depend on the binding affinity of ligand-receptor interactions in our minimal model. Consequently, the intuitive idea about Liebig's law of the minimum is not necessarily true theoretically. Moreover, given the non-linearity of the TGF-beta network, this complexity introduces an additional layer of uncertainty regarding the applicability of the minimum control principle to TGF-beta responses. This uncertainty led us to develop an extended model, with parameter values either experimentally measured or estimated from time course experimental data. The extended model predicted a similar minimum control principle at the TGF-beta receptor level, inspiring us to validate this prediction through diverse experiments. While we acknowledge the intuitive nature of our findings, we believe it is important for the field to prove this expectation, as emphasized by reviewer 4.

Reviewer #1 (Significance (Required)):

*TGF-beta signaling is one of the most rigorously studied pathways both computationally and experimentally. As written in the introduction of the manuscript, it is still unknown how the variability of responses arises not only between cell types but also differences among cells of single cell type. Studies showed that protein abundance accounts at least partly for a source of cell variability in TGF-beta signaling. While former studies examined the variability in SMAD protein abundance, the uniqueness of this study is that it focused on the abundance of TGF-beta receptors. *

*Given that both TGFBR1 and TGFBR2 are involved in the signaling, however, it's not difficult to imagine that a less abundant receptor affects the signaling more than the other, and serves as a bottleneck for the signaling. Specifically, because a slightly high concentration (100pM = 4.4 ng/mL of TGF-beta; other studies used much lower conc., e.g. 0, 0.03, 0.04, 0.07, and 2.4 ng/mL in Frick et al, PNAS, 2017, and 0, 1, 2.5, 5, 25, and 100 pM in Strasen et al, Mol Syst Biol, 2017) is used throughout the experiments to check cell-cell variability and the effect of receptor abundance in the current study, the formation of the receptor-ligand complex may be quite fast and be saturated at the level where the receptor with lower abundance is exhausted. In the reviewer's humble opinion, the authors' statement that this is Liebig's law of the minimum sounds a bit exaggerated. *

Nevertheless, the study is of some value because it utilized both computational and experimental analysis to show it is indeed the case. Of note, the current study showed that the variability in the different proteins leads to the variability in different time points, namely, the variability in the receptor abundance leads to the variability 1 hour after stimulation, while that in negative feedback strength leads to the variability 8 hours after stimulation. If the authors fill a small gap between their computational analysis and experimental verification, the study will be of interest to the specialist in the field.

__Response: __We are grateful for the valuable feedback provided by the reviewer. The concerns related to the TGF-beta dose have been thoroughly addressed in our responses to previous comments. Regarding the observation that the term "Liebig's law of the minimum" may sound a bit exaggerated, we acknowledge this consideration. We have refined the title to "Liebig’s Law of the Minimum in the TGF-β/SMAD Pathway," specifying its relevance to SMAD signaling exclusively, as non-SMAD signaling was not within the scope of this study. We appreciate the reviewer's constructive feedback and hope these adjustments enhance the specificity and accuracy of our manuscript.

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

Li et al. present an interesting and intuitive concept for the sensitivity and heterogeneity of biological networks: When two or more proteins form a functional complex, it is the limiting component with the lowest concentration that is most sensitive to perturbations and whose fluctuations dictate cell-to-cell variability of complex function. The authors apply this concept to the TGFb pathway and discuss sensitivity of SMAD signaling towards TGFb receptor I and II fluctuations. The paper is clearly written and convincing, but some improvements in the experimental validation would be beneficial as detailed further below.

1) The authors claim that the ratio of TGFb receptor I and II is very different across cell lines (Fig. 1) and use this observation for the validation of their model in Fig. 4. However, the relative expression TGFb receptor levels are purely based on RNAseq data which does not necessarily imply similar behavior at the protein level, especially on the cell surface. To address this issue, the authors should ideally provide absolute Western blot measurements of TGFbRI at the protein level to complement their absolute quantification of TGFbRII (Fig. S2). At the very least they should show that the observed relative expression levels of TGFbRI and II at the protein level (Figure S7) are correlated to differences in RNA levels (Fig. 1) using protein quantification. They should also confirm that similar receptor ratios for these receptors at the RNA level are observed in other published RNAseq datasets of the same cell lines (e.g., ENCODE for HepG2 and published RNAseq studies in HaCaT). Furthermore, they might take into account published mass spec datasets for quantifications of TGFbR protein levels.

Response: We appreciate the reviewer's thorough evaluation and constructive suggestions.

(A) Absolute quantification of TGFBR1: We acknowledge the importance of obtaining absolute quantification of TGFBR1 protein similar as what we have done for TGFBR2 protein (Figure S2). Despite significant efforts, our attempts to achieve this were hindered by challenges with available TGFBR1 antibodies and recombinant TGFBR1 proteins. Many commercial antibodies failed negative controls with TGFBR1 knockdown samples, while others validated TGFBR1 antibodies could not recognize the available recombinant TGFBR1 protein standards.

Although many mass spectrometry proteomics data available for different cell lines, it is difficult to convert these MS quantitative values to absolute protein abundance as mentioned in a recent publication (Nusinow et al.,bioRxiv 2020.02.03.932384): “Importantly, these values are all relative values to the other values for that same protein and not absolute values. This means that comparing the levels of different proteins to each other without using something like a correlation to standardize values won’t produce meaningful results.”

We share the reviewer's concern and fully agree that obtaining this absolute quantification is crucial. However, at the present stage, technical limitations prevent us from providing this information for TGFBR1. We commit to pursuing this aspect when feasible in the future.

(B) Validation of relative TGF-beta receptor expression ratios: Following the reviewer's suggestion, we conducted additional analyses to validate the relative expression ratios of TGFBR1 and TGFBR2 using different RNA-Seq databases. The results, presented in Table S1, demonstrate consistent imbalances in TGFBR1-to-TGFBR2 ratios across HepG2 and RH30 cell lines from various data sources, reinforcing the reliability of our observations.

(C) Correlation between RNA and protein expression: We appreciate the reviewer highlighting the challenges associated with correlating RNA and protein expression. Indeed, the correlations between RNA and protein levels vary widely, and direct comparisons can be challenging. To address this, we referenced a recent study (Nusinow et al., Cell 2020, 180:387), which reported that the protein data of TGFBR1 and TGFBR2 were highly correlated with the corresponding RNA data from the same cell line (Spearman’s correlation: 0.672 for TGFBR1, 0.771 for TGFBR2) based on quantitative proteomics and RNA expression data from 375 cancer cell lines.

2) Figure 4: To better judge the reproducibility of the knockdown titration, it would be good to show the different siRNA concentrations as a color code- Alternatively, TGFBR expression could be plotted as a function of the siRNA concentration in a Supplemental Figure, showing the effects of individual replicates.

Response: We thank the reviewer for the suggestion to enhance the clarity of the knockdown titration data. In response, we have now presented the quantified experimental data from three replicates with different colors in Figure 4. Additionally, we have created Figure S9 that plots the expression levels of relative TGFBR1 and TGFBR2 as a function of siRNA concentration, providing a more detailed view of the effects across individual replicates.

3) The simulations in Figs. 5 and 6 show that SMAD signaling fluctuations are mainly determined by cell-to-cell variability of receptor levels when using the SMAD nucleocytoplasmic ratio as a readout, and this is especially true for early time points. For downstream cellular responses, the absolute concentration of phosphorylated SMAD (complexes) in the nucleus is likely more relevant. Based on the authors work and evidence from the literature, I expect that this quantity will likely be heavily be influenced by receptor levels as well, but fluctuations in SMAD expression will play an important role as well. The authors should discuss this issue, and clarify that normalized quantities like SMAD N2C and pSMAD/SMAD mostly characterize receptor-level fluctuations while filtering SMAD fluctuations.

__Response: __We acknowledge the importance of discussing the relevance of different readouts in our study. In the revised manuscript, we have incorporated a discussion addressing this issue. Specifically, we highlight that while the SMAD nucleocytoplasmic ratio is sensitive to cell-to-cell variability in low abundance receptor levels, the absolute concentration of phosphorylated SMAD in the nucleus may be more relevant for downstream cellular responses (e.g.: gene expression). We have cited the work by Lucarelli et al, which demonstrated that variations in SMAD abundance could modulate the balance of different SMAD complexes, thereby regulating heterogeneous gene expression in diverse cell types (Lucarelli et al., Cell Systems 2018).

4) The single-cell measurements in Fig. 7 are interesting, but can only partially be seen as a direct validation of the model predictions, as it seems expected that varying the total input by introducing co-fluctuations in both receptors heavily influence the SMAD level. Wouldn't it be possible to design more specific validation experiments, in which the receptor co-expression construct (Fig. 7C) is used for baseline optoTGFBR expression and combined with an individual expression construct for one of the opto-receptors? This way, the authors could establish different regimes, in which one of the two receptors becomes dominant, and the impact fluctuations could be analyzed in a larger receptor expression space. Of course, a full validation of all possible scenarios is not necessary, but it would, for instance, be valuable to see whether the strong dependency of SMAD signaling of TGFBR2 levels vanishes when TGFBR2 is expressed at a higher level than TGFBR1.

Response: We appreciate the insightful comments and suggestions provided by the reviewer. Based on these recommendations, we have conducted additional experiments to further validate our model predictions. Reviewer 1 also raised this point, we quote our aforementioned response here: “consistent with the model predictions (Figure 6), the strong correlation between SMAD2 N2C fold change response at 1h and optoTGFBR2-tdTomato expression levels persisted in single cells when optoTGFBR1 was overexpressed (Figure 8A). Conversely, the high correlation between nuclear SMAD2 signaling and optoTGFBR2 expression levels vanished at single cell level when optoTGFBR2 was overexpressed (Figure 8B). These experimental results validate our model predictions, confirming that the SMAD2 signaling is determined by the low abundance TGF-beta receptor in single cells. Incorporating these experimental validations enhances the quantitative support for our model predictions and clarifies the relationship between TGF-beta receptor abundance and signaling outcomes in single cells.”

**Referees cross-commenting**

Comments from R2: I agree with most comments of the other reviewers, and highlight the most important overlaps with my comments below.

I agree with R1 that the model validation in Fig. 7 is incomplete and think that this will be a key point to improve the quality of the manuscript (see also my reviewer comment 4)

In line with R3 and R4, I think that the SMAD N/C simulations do not necessarily imply effects on TGFb target gene expression, cell fate decisions or human pathologies. The significance of the results for cellular behavior should be discussed (see also my comment 3)

__Response: __We are grateful for the reviewer's thoughtful comments. These comments have been now addressed (see our responses to the corresponding comments).

Reviewer #2 (Significance (Required)):

The manuscript presents an interesting and intuitive concept for the sensitivity and heterogeneity of biological networks. The authors apply this concept to the TGFb pathway and discuss sensitivity of SMAD signaling towards TGFb receptor I and II fluctuations.

Reviewer #3 (Evidence, reproducibility and clarity (Required)):

*Summary: *

*This is an interesting study that examines the output of the TGF-Beta pathway and how abundance/dosage can determine the signaling response in single cells across multiple cell types. The study is primarily mathematical. The focus is on the Type 1 and 2 TGF-Beta receptors driving nuclear SMAD2 expression. The authors observe that SMAD2 phosphorylation is sensitive to variations in the lower levels of either receptor but robust at variations of high abundance of the receptor reflected through SiRNA experiments shown in Figure 4. Their conclusion is that the feature is consistent with Liebig's law of the minimum- where in this case- a low abundance of the receptor serves as the rate-limiting step in signaling for this pathway. *

*Major comments: *

*- While the data as presented are interesting, it is unclear as to whether the abundance regulates biological function. SMAD2 phosphorylation is shown with some nuclear translocation. However, TGF-Beta target gene activation is not shown, and this needs to be completed. *

Response: We appreciate the reviewer's constructive comment. We have conducted new experiments and included quantitative real-time PCR data in the revised manuscript to evaluate the impact of TGFBR1 and TGFBR2 knockdown on the expression of TGF-beta target genes, such as SMAD7, PAI1, and JUNB. The results, presented in Figure S11, demonstrate differential sensitivity of these genes to the downregulation of TGFBR1 and TGFBR2 in various cell lines (HaCaT, HepG2, and RH30). Specifically, the expression of SMAD7, PAI1, and JUNB is sensitive to TGFBR2 knockdown in RH30 cells, while it is sensitive to TGFBR1 knockdown in HepG2 cells. HaCaT cells, expressing similar levels of both receptors, show comparable sensitivities to reductions in both TGFBR1 and TGFBR2. These findings provide additional insights into the regulatory role of TGF-beta receptor abundance on downstream target gene activation, complementing our study's focus on SMAD2 phosphorylation and nuclear translocation.

*- In addition, it is unclear as to what happens to SMAD3 and SMAD4 which are expressed endogenously in this setting. How are these other TGF-Beta signaling molecules addressed by these observations? *

__Response: __Thank you for bringing up this important point. In our study, the expression levels of endogenous SMAD2 and SMAD4 were found to be similar across HaCaT, RH30, and HepG2 cells. However, SMAD3 expression was notably lower in RH30 and HepG2 compared to HaCaT cells. The central conclusion of our study is based on the observed common control principle, which hinges on the relative expression levels of TGFBR1 and TGFBR2. Consequently, the applicability of this principle is more pertinent when comparing signal responses within the same cell type.

We acknowledge the relevance of endogenous SMAD proteins, and in the revised manuscript, we have expanded our discussion on how differences in SMAD protein expression levels and potential mutations (page 16 in main text), as observed in certain cancers, could influence the formation of homo- and hetero-oligomeric SMAD complexes. These considerations contribute to a more comprehensive understanding of downstream gene expression responses, as discussed in the work of Lucarelli et al. (Cell Systems 2018).

*-Specific biological readouts- cell differentiation etc. are not examined and would need to be provided and discussed. Therefore, the claims put forward while interesting require additional experiments examining SMAD2 target gene activation and biological readouts. *

__Response: __We appreciate this valuable suggestion. While we acknowledge the importance of exploring long-term biological responses, including cell differentiation, it is crucial to note that specific biological readouts are not solely dependent on SMAD signaling; they also involve other non-SMAD signaling pathways. Additionally, these responses are highly cell type-specific. Undertaking extensive investigations into these responses would extend beyond the current scope of our work. Nevertheless, we have discussed this topic in the revised manuscript (page 16 in main text).

Following the reviewers’ suggestion on examining TGF-beta target genes, we have performed experiments examining the expression of SMAD7, PAI1, and JUNB with respect to the changes of TGFBR1 and TGFBR2, respectively (see our response to the first major comment of this reviewer).

*- Lastly, statistical analyses are not provided and would need to be provided. For instance, in Figure 4, how many experiments were replicated and statistical analysis performed for this Figure? *

__Response: __In addressing this concern, we conducted three siRNA knockdown titration experiments for each cell line, as detailed in the figure legend. Due to batch effects, different percentages of TGF-beta receptors were knocked down in different experiments using the same concentration of siRNA. To transparently present the data, we utilized a scatter plot. Following the suggestion from reviewer 2, we have further enhanced the clarity of our data presentation by labeling the results of different experiments with a color code. In addition, we have performed statistical analysis of TGF-β receptor fold-change effects leading to a 50% reduction in the P-Smad2 response compared to that in the non-targeting siRNA control group (EC50) during siRNA knockdown experiments (Figure S10). The results of this analysis unveil significant differences in the sensitivities of pSMAD2 responses to variations in TGFBR1 and TGFBR2 within RH30 and HepG2 cells.

Reviewer #3 (Significance (Required)):

*- Conceptually this is an important study because dosage is a prominent issue in TGF-Beta signaling. For instance, in my field of expertise- mouse models of TGF-beta signaling e.g. SMAD2 knockouts- the cancer phenotypes are evident in haploid animals. Yet how and why dosage plays such a large role in tumorigenesis remains unclear. *

__Response: __We sincerely appreciate your recognition of the conceptual importance of our study in addressing the dosage-related complexities of TGF-beta signaling. Your insights into dosage effects in mouse models, particularly in haploid animals, highlight the relevance of our work underlying tumorigenesis. We have incorporated relevant citations and expanded our discussion in the revised manuscript, providing additional context to the importance of dosage in tumorigenesis (page 18 in main text).

Reviewer #4 (Evidence, reproducibility and clarity (Required)):

Summary: In this study, Li and co-workers combined computational modeling and experimental analysis to study the dependence of the output of the TGF-beta pathway on the abundance of signaling molecules in the pathway, mainly the most upstream regulators of SMAD2, TGFbeta type I and type II receptors. They showed by a combination of biochemical studies (mainly pSmad2 WB and type I/II receptor expression profiling) in HaCaT and HeLa cells as well as stable optogenetical receptor variants expressed by those cell lines, that TGF-beta receptor abundance influences signaling outputs using the concept of Liebigs law of the minimum, meaning that the output-modifying factor is the signaling protein that is most limited, to determine signaling responses across cell types and in single cells.

*Major comments: *

The study is very interesting, the combination of biochemistry and computational modeling to better understand the compexity of the TGFbeta pathway is very much required in the field and should stimulate others to further expand this approach.

__Response: __Thank you for the positive evaluation of this work.

*However, the authors must further explain that the model depicted here to explain pathway kinetics and dynamics lacks multiple crossroads and feedbacks and is until now oversimplified in the manuscript. They have mentioned receptor internalization and recycling, nuclear import and export of SMAD protein, and the feedback regulations e.g. by SMADs regulating receptor expression. Beyond, there is non- SMAD signaling (Derynck et al.; SMAD Linker regulation, deRobertis et al.), different receptor oligomerization modes (Ehrlich/Henis et al.) and heteromeric receptor complexes of TGFbeta receptors known (Hill et al.), that further diversify beyond these mentioned mechanisms. It is understandable that the mathematical model cannot include those considerations to date, however, they must be further explained and commented on to allow that this model can be expanded in the future. *

Response: We acknowledge that there are multiple crossroads and feedbacks that exist in the TGF-beta signaling pathway that have not been explicitly incorporated into our model. We appreciate the reviewer's understanding that current model cannot include these considerations and his/her suggestions for potential future extensions. In the revised manuscript, we have mentioned one of the limitations of our model: non-Smad signaling and crosstalk with other signaling pathways were not considered for simplicity. We have also discussed how to expand this model by including these regulations when more quantitative data are available in the future (page 16-17 in main text).

*A myriad of research labs focus on these intricate fine tuning ot the TGFbeta pathway by those mechanisms which makes the difference between "good" TGFbeta signaling and "bad" TGFbeta signaling in different context and this complexity must be acknowledged by more introduction and discussion. *

Response: In the revised manuscript, we have added an introduction and discussion about the dual role of TGF-beta signaling (page 4 and page 18 in main text).

*The model here will be important to explain *

*A: the mode of heterooligomeric TGFbeta/BMP receptor assemblies as e.g. found in pathological conditions and *

B: Can maybe explain the formation of mixed SMAD complexes as activated by lateral signaling comprising TGFbeta *and BMP receptors once one receptor is of lower abundance to form a high affinity complex. *

*It is therefore required to comment on these aspects at multiple points in the manuscript. *

*It is very important that the visual model used in this manuscript depicts on the possibility, that a TGFbeta type I receptor can team up with e.g. another TGFbeta type I receptor together with two TGFbeta type II receptors but also with an activin type II receptor or that a BMP type I receptor (e.g. ALK1) can form heterooligomeric complexes with ALK5 (TGFbeta type I). *

__Response: __Thank you for this comment. We cited the relevant work (Ramachandran et al, eLife 2018; Szilagyi et al, BMC Biology 2022) and added a discussion about the complexity of the mode of heterooligomeric TGFbeta/BMP receptor assemblies and its effect on the induction of mixed SMAD complexes (page 17 in the main text).

*While the use of optogenetical TGFbeta receptor biosensors is highly interesting, their mode of oligomerization is not yet fully described. It is not known if those biosensors behave like wt receptors in terms of oligomerization and ligand binding. This should be mentioned somewehere. For this reason, the authors should also consider to draw the TGFbeta receptor complex in the cartoons with more detail towards the heterooligomeric assembly that is standard to the field. *

__Response: __The reviewer is correct that the optogenetic TGF-beta receptors might behave differently from the natural TGF-beta receptor system in terms of ligand binding. We have added this point in the Discussion part to highlight the potential difference between the optogenetic TGF-beta systems and the wild-type system (page 16 in the main text).

*While the general finding is not surprising (manipulating the receptor with the lowest abundancy has the biggest impact on signaling output) the methods and models used here are very important to the field to proof that this expectation is actually true and can be experimentally addressed by a combination of bioinformatics and biochemistry. The model developed will be valuable to expand to much more complex and interesting questions in TGFbeta signaling and possibly also BMP signaling e.g. in pathological context (see below). *

*Minor comments: *

*The authors should discuss their findings in the context of: *

• non-Smad signaling outputs (similar or different to the observations on pSMAD2)*
• What do these findings mean for e.g. human pathologies, where type I or type II receptor expression is altered? *
• Can those findings integrate into the "switch" in TGFbeta signaling? *
• How do these findings translate towards BMP SMAD 1/5/9 signaling? * Response: First, we sincerely appreciate the reviewer’s recognition that our work is very important to the field in proving that manipulating the receptor with the lowest abundance has the biggest impact on signaling output. The reviewer’s suggestions about discussing our work in the context of non-Smad signaling, BMP SMAD1/5/9 branch, and the relevance to the dual role of TGF-beta signaling are all constructive. We have incorporated these suggestions and discussed them in the revised manuscript (page 17 in the main text).

Reviewer #4 (Significance (Required)):

*The manuscript is novel and interesting, partiular the combination of bioinformatical and biochemical approaches. The use of optogenetics is state-of-art while some more care should be given to interpretation of results with optogenetical TGfbeta receptor biosensors, is is not known if they really behave similar in terms of receptor oligomerization and signaling. Also it is not shown how their interactome in terms of effector proteins looks like that can potentially influence SMAD signaling output (e.g. Phosphathases to SMADs known to interact with wt receptors). *

*The models drawn need to depict more accurately on the nature of type I and type II receptor complexes (heterotetrameric) and high affinity towards the ligand. The current versions are too oversimplified at this stage. The pathway crosstalks and feedbacks need to be more visible, in order for non experts to not draw too simple conclusions from the visual representations presented in this MS. Particularly the work by Hill and co-workers on receptor oligomerization and SMAD shuttling and feedback need to be included. *

Overall, the manuscript is very significant to the field.

__Response: __We would like to thank the reviewer again for his/her positive evaluation of the novelty and significance of our work. We have taken the reviewer's comments into consideration and made revisions to the manuscript. We now provide more information on the limitations of our current model and the optogenetic TGF-beta receptor biosensors in the Discussion section. We have also included more details about the receptor complex nature and the high affinity towards the ligand. The ligand receptor complex in the model is now drawn as heterotetrametric complex (1 ligand dimer with two TGFBR1s and two TGFBR2s). Additionally, we have incorporated information about pathway crosstalks and feedbacks, giving a more comprehensive view for non-experts. The work by Hill and co-workers on receptor oligomerization, SMAD shuttling, and feedback has been included in the revised manuscript to provide a more complete and accurate representation of the current knowledge in the field.

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

Li et al. present an interesting and intuitive concept for the sensitivity and heterogeneity of biological networks: When two or more proteins form a functional complex, it is the limiting component with the lowest concentration that is most sensitive to perturbations and whose fluctuations dictate cell-to-cell variability of complex function. The authors apply this concept to the TGFb pathway and discuss sensitivity of SMAD signaling towards TGFb receptor I and II fluctuations. The paper is clearly written and convincing, but some improvements in the experimental validation would be beneficial as detailed further below.

1. The authors claim that the ratio of TGFb receptor I and II is very different across cell lines (Fig. 1) and use this observation for the validation of their model in Fig. 4. However, the relative expression TGFb receptor levels are purely based on RNAseq data which does not necessarily imply similar behavior at the protein level, especially on the cell surface. To address this issue, the authors should ideally provide absolute Western blot measurements of TGFbRI at the protein level to complement their absolute quantification of TGFbRII (Fig. S2). At the very least they should show that the observed relative expression levels of TGFbRI and II at the protein level (Figure S7) are correlated to differences in RNA levels (Fig. 1) using protein quantification. They should also confirm that similar receptor ratios for these receptors at the RNA level are observed in other published RNAseq datasets of the same cell lines(e.g., ENCODE for HepG2 and published RNAseq studies in HaCaT). Furthermore, they might take into account published mass spec datasets for quantifications of TGFbR protein levels.
2. Figure 4: To better judge the reproducibility of the knockdown titration, it would be good to show the different siRNA concentrations as a color code- Alternatively, TGFBR expression could be plotted as a function of the siRNA concentration in a Supplemental Figure, showing the effects of individual replicates.
3. The simulations in Figs. 5 and 6 show that SMAD signaling fluctuations are mainly determined by cell-to-cell variability of receptor levels when using the SMAD nucleocytoplasmic ratio as a readout, and this is especially true for early time points. For downstream cellular responses, the absolute concentration of phosphorylated SMAD (complexes) in the nucleus is likely more relevant. Based on the authors work and evidence from the literature, I expect that this quantity will likely be heavily be influenced by receptor levels as well, but fluctuations in SMAD expression will play an important role as well. The authors should discuss this issue, and clarify that normalized quantitites like SMAD N2C and pSMAD/SMAD mostly characterize receptor-level fluctuations while filtering SMAD fluctuations.
4. The single-cell measurements in Fig. 7 are interesting, but can only partially be seen as a direct validation of the model predictions, as it seems expected that varying the total input by introducing co-fluctuations in both receptors heavily influence the SMAD level. Wouldn't it be possible to design more specific validation experiments, in which the receptor co-expression construct (Fig. 7C) is used for baseline optoTGFBR expression and combined with an individual expression construct for one of the opto-receptors? This way, the authors could establish different regimes, in which one of the two receptors becomes dominant, and the impact fluctuations could be analyzed in a larger receptor expression space. Of course, a full validation of all possible scenarios is not necessary, but it would, for instance, be valuable to see whether the strong dependency of SMAD signaling of TGFBR2 levels vanishes when TGFBR2 is expressed at a higher level than TGFBR1.

Referees cross-commenting

Comments from R2: I agree with most comments of the other reviewers, and highlight the most important overlaps with my comments below.

I agree with R1 that the model validation in Fig. 7 is incomplete and think that this will be a key point to improve the quality of the manuscript (see also my reviewer comment 4)

In line with R3 and R4, I think that the SMAD N/C simulations do not necessarily imply effects on TGFb target gene expression, cell fate decisions or human pathologies. The significance of the results for cellular behavior should be discussed (see also my comment 3)

Significance

The manuscript presents an interesting and intuitive concept for the sensitivity and heterogeneity of biological networks. The authors apply this concept to the TGFb pathway and discuss sensitivity of SMAD signaling towards TGFb receptor I and II fluctuations.

Author Response

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

We are grateful to the reviewers for their remarks, which significantly improved the paper. We repeated the biochemical assay concerning SIRT6 activity on H3-K27Ac and quantified the results as requested. Please find our detailed answers bellow each recommendation of the reviewers.

Major recommendations:

1. Grammatical errors are still common; the authors may need to consider an external editing service if they intend to fix the problems as they indicate that they believe the errors have been removed. The Results section is relatively clean, but parts of the Abstract, Introduction, and Discussion are more difficult to understand, and errors are especially common in the Methods section and those parts of the manuscript that are new in this revision.

We corrected the grammatical errors.

1. The introduction doesn't mention the other structures published; this is considered to be a serious deficiency as it prevents the reader from understanding the context for the contributions described here. Withholding the comparison with (or mention of) the previously published work to the last sentence of the Discussion seems misleading and does not give the reader adequate ability to judge the novelty of the results presented in this manuscript.

A paragraph comparing our paper to the other structures published appear at the end of the discussion. We feel this is still the right place for such a paragraph.

1. The addition of the assay for deacetylation is a significant improvement over the initial submission. This is important both for validating the importance of the acidic patch contacts and for helping to resolve the conflicting reports regarding activity on H3-K27Ac. Given the importance of this assay for the impact of the manuscript, it is not clear why the authors chose to 1) put the data in the supplement instead of in the main manuscript, and 2) provide only single samples without quantitation. These both seem to be significant limitations.

We repeated the experiment and provided quantification of the results. We placed the figure in the main manuscript.

1. The authors should add text or a table to the Methods section explaining which maps were used for each figure. By our count, there are 8 maps and 5 models (plus MD models) based on two datasets, but the relationships among them are not clearly stated, and the names of the maps (such as "Zn-finger focused" and "Rossman-Fold-Focused") might be changed to be more helpful to the reader (for example, the latter includes more than the Rossman fold and might be renamed "Sirt6-focused"). The authors should also explain how the maps were validated, which data were deposited in public repositories, and why some data were not deposited. For example, no statistics or methods regarding how particles were separated into integrated vs. non-integrated motion are provided for the CryoDRGN models. Further, the "two principle movements" described are depicted in 4 maps from two CryoDRGN runs using two separate sets of particles, but the relationships among them are not defined clearly. Finally, the connectivity of densities in Fig 8 are not obvious in the submitted maps. Until these points are addressed, the work is considered incomplete.

AND

1. The PDB model provided for review and submitted to the PDB database shows loosely bound DNA at the nucleosomal entry/exit points near the binding site of SIRT6, but the maps provided for review and submitted to the EMDB show stronger density for the canonical location of the DNA expected at these sites. The CryoDRGN maps support a more extended conformation, but these maps were not deposited or provided for review so their validity cannot be assessed.

We added a section to the methods listing the different maps used for the figures. We deposited the map we used to trance the H2A N-terminal tail (EMD-18497). Unfortunately, we couldn’t deposit the cryoDRGN maps as the deposition system either accepts composite maps, where the consensus should be deposited too or experimental maps, where the deposition of half maps are mandatory. Nevertheless, the cryoDRGN maps are available upon request. We also added a supplementary figure (Supplementary Fig 6) to show how the cryoDRGN analyses were performed.

1. The orientation, angle and threshold used in Fig 1 make it difficult to see the multiple DNA orientations that are visible in the deposited consensus map. Examination of the map suggests that the DNA model submitted to PDB corresponds to a weaker DNA conformation than is present in the map where both DNA conformations are visible. The authors should consider modeling both conformations in their deposited model to provide a more complete, accurate representation of the data. It is concerning that a key conclusion of the manuscript is that the DNA conformation changes upon SIRT6 binding, but density for the canonical position is observable in Fig 8a.

Figure 1 is showing the overall representation of the SIRT6 bound nucleosome structure. We show the DNA linker orientations in the subsequent figure. Figure 8 (now Figure 9) shows the rearrangement of the SIRT6 Rossmann fold domain not the DNA linker.

1. Figure 4 needs a more complete legend, indicating that it is a hybrid of the consensus structure (one color) and the MD simulations (another color). In general, the colors used in the figure should be changed to make the main points more accessible.

As there is a color code for the histones, changing colors might be confusing. The figure legend mentions that panels c, d and e are from MD simulations.

Minor recommendations:

1. Figures 2c, e, and f are not referenced in the text.

We now referenced all figure panels in the text.

1. Consider moving Supp. 5C to Fig. 2 as the models in that figure come from the CryoDRGN maps and not the consensus map.

Supplemental Figure 5c show the DNA linker deviation upon SIRT6 binding from another angle. We prefer to keep it there.

1.) Supp Fig 3 is labeled "ZnF-nucleosome" refinement, but this appears to come from Data Set #2 processing. The map might be labeled ZnF-nucleosome but then a mask should be shown that excludes the Rossman Fold. It is not clear if this is a focused refinement or just a 2.9 A map that was merged with the "Rossman-fold" map.

We changed both supplemental figures accordingly.

1. The orientation of Fig 2 b and e do not show the differences in these models as well as panels c and f. Panels b and e could be replaced with the 4 CryoDRGN maps.

The models reflect the cryoDRGN maps and panels c and f were added to clarify the movement.

1. The MD description should emphasize that the H3 tails are moving with respect to the active site, as it currently suggests the active site is moving.

In the results and in the discussion section we mention that we observe new conformations of the H3 tail, not of the active site.

1. The authors refer to the "flexibility of the Rossmann fold domain," but the Rossman Fold domain isn't flexible, the linkage to the ZnF is flexible. Perhaps "observed conformational space" or "dynamic Rossman-fold domain position" are meant.

The text was changed accordingly.

1. The H2A C-terminal tail present in Fig 1 (bottom right) and Figure 3e is not present in the model in Fig 4a,b.

The H2A tails conformation was not resolved in the cryoDRGN maps so we didn’t model it.

1. The crosslinking agent used is not specified.

The crosslinking agent used is specified more clearly in the methods.

1. Supp Table 1 and EM methods do not agree on the magnification for Dataset #1. Verify nominal versus binned magnification and reported pixel size.<br /> The magnification in the methods was changed.

2. Fig 3F showing the difference between affinity for H2A and H2A.Z-containing nucleosomes would be more convincing with a titration rather than the current comparison of a single concentration.

We agree with this remark however, we find single concentration comparison is convincing enough for the purposes of this paper as it is not a central finding.

1. Fig S1 legend; both the Zn-finger and helix bundle are stated to be shown in green.

Figure S1 legend was changed.

The names of the IMD data files in the code block near "Let us load the IMD data first:" do not resemble names in the IMD folder in the zipped file.

100% true the real impact that can be made is on a people scale, with people putting in better policy's and such this CAN be prevented.

I wonder what this code is and how easily it could be fixed to detect everyone's face. Also, what a coincidence.

Big metaphor for the implications of AI bias which is what this video is on, interesting for her very first sentence to be a metaphor like this and I wonder if there is a good reason for it like it being a hook or something. Additionally, I will say that it is kind of funny that her name is "Joy" and throughout the whole video she is happy even though she is talking about a serious issue.

This is important as it is one of the main reasons for the poor, biased, decisions of the ai code. It is also instutuional oppression as it is the comapny deciding who to test for their ai data, even if it is an accident. i think with a bigger diversiety of testers, ai will be mroe ethical and omptimized for everyone to use. i feel hopeful

I remembered that happen to me, but it involve some python code and Telegram Bot API worked for me.

rat race

Author Response

The authors wish to thank the Reviewers for valuable and constructive comments that will help up improve the paper’s quality.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

This manuscript builds upon the authors' previous work on the cross-talk between transcription initiation and post-transcriptional events in yeast gene expression. These prior studies identified an mRNA 'imprinting' phenomenon linked to genes activated by the Rap1 transcription factor (TF), a surprising role for the Sfp1 TF in promoting RNA polymerase II (RNAPII) backtracking, and a role for the non-essential RNAPII subunits Rpb4/7 in the regulation of mRNA decay and translation. Here the authors aimed to extend these observations to provide a more coherent picture of the role of Sfp1 in transcription initiation and subsequent steps in gene expression. They provide evidence for (1) a physical interaction between Sfp1 and Rpb4, (2) Sfp1 binding and stabilization of mRNAs derived from genes whose promoters are bound by both Rap1 and Sfp1 and (3) an effect of Sfp1 on Rpb4 binding or conformation during transcription elongation.

Strengths:

This study provides evidence that a TF (yeast Sfp1), in addition to stimulating transcription initiation, can at some target genes interact with their mRNA transcripts and promote their stability. Sfp1 thus has a positive effect on two distinct regulatory steps. Furthermore, evidence is presented indicating that strong Sfp1 mRNA association requires both Rap1 and Sfp1 promoter binding and is increased at a sequence motif near the polyA track of many target mRNAs. Finally, they provide compelling evidence that Sfp1-bound mRNAs have higher levels of RNAPII backtracking and altered Rpb4 association or conformation compared to those not bound by Sfp1.

Weaknesses:

The Sfp1-Rpb4 association is supported only by a two-hybrid assay that is poorly described and lacks an important control. Furthermore, there is no evidence that this interaction is direct, nor are the interaction domains on either protein identified (or mutated to address function).

Indeed, our two hybrid, immunoprecipitation and imaging results do not allow us to conclusively discern whether the interaction between Rpb4 and Sfp1 is direct or indirect. While the interaction holds significance, we consider the direct versus indirect distinction to be of secondary importance in the context of this paper. We intend to give more attention to this matter in our revised paper. In addition, we will make an effort to investigate an in vitro interaction between Sfp1 and Rpb4 by employing purified Sfp1 and Rpb4 proteins.

The contention that Sfp1 nuclear export to the cytoplasm is transcription-dependent is not well supported by the experiments shown, which are not properly described in the text and are not accompanied by any primary data.

We note that this assay has been developed and published in prior research by Lee, M. S., M. Henry, and P. A. Silver. (G&D, 1996) and was reported in a number of subsequent papers. Reassuringly, our conclusion is supported by the observation that Sfp1 binds to Pol II transcripts co-transcriptionally suggesting that Sfp1 is exported in the context of the mRNA.

The presence of Sfp1 in P-bodies is of unclear relevance and the authors do not ask whether Sfp1-bound mRNAs are also present in these condensates.

In the revised paper, we will indicate that we do not know whether RP mRNAs are present in the actual foci shown in Fig. 1B.

Further analysis of Sfp1-bound mRNAs would be of interest, particularly to address the question of whether those from ribosomal protein genes and other growth-related genes that are known to display Sfp1 binding in their promoters are regulated (either stabilized or destabilized) by Sfp1.

Fig. 4A, C and D show that RP mRNAs become destabilized in sfp1Δ cells.

The authors need to discuss, and ideally address, the apparent paradox that their previous findings showed that Rap1 acts to destabilize its downstream transcripts, i.e. that it has the opposite effect of Sfp1 shown here.

We would like to thank Reviewer 1 for this valuable comment. In the revised paper, we will delve into our hypothesis suggesting that Rap1 is likely responsible for regulating the imprinting of other proteins, that, in turn, lead to the destabilization of mRNAs, such as Rpb4.

Finally, recent studies indicate that the drugs used here to measure mRNA stability induce a strong stress response accompanied by rapid and complex effects on transcription. Their relevance to mRNA stability in unstressed cells is questionable.

Half-lives were determined mainly by the GRO analysis of optimally proliferating cells. This method does not requires any drug or stressful treatment. The results obtained by this method were consistent with the those obtained after thiolutin addition. Nevertheless, in our revised manuscript, we plan to supplement the half-life data with results obtained by subjecting cells to a temperature shift to 42°C, a natural method to block transcription in wild-type (WT) cells. This approach to determine half-lives has been previously reported in our publications, such as Lotan et al. (2005, 2007) and Goler Baron et al. (2008). This may rule out effects of the drug on halfe-life.

Reviewer #2 (Public Review):

Summary:

The manuscript by Kelbert et al. presents results on the involvement of the yeast transcription factor Sfp1 in the stabilisation of transcripts whose synthesis it stimulates. Sfp1 is known to affect the synthesis of a number of important cellular transcripts, such as many of those that code for ribosomal proteins. The hypothesis that a transcription factor can remain bound to the nascent transcript and affect its cytoplasmic half-life is attractive, but the methods used to demonstrate the half-life effects and the association of Sfp1 with cytoplasmic transcripts remain to be fully validated, as explained in my comments on the results below:

Comments on methodology and results:

1. A two-hybrid-based assay for protein-protein interactions identified Sfp1, a transcription factor known for its effects on ribosomal protein gene expression, as interacting with Rpb4, a subunit of RNA polymerase II. Classical two-hybrid experiments depend on the presence of the tested proteins in the nucleus of yeast cells, suggesting that the observed interaction occurs in the nucleus. Unfortunately, the two-hybrid method cannot determine whether the interaction is direct or mediated by nucleic acids.

Please see our response to comment 1 of Reviewer 1.

1. Inactivation of nup49, a component of the nuclear pore complex, resulted in the redistribution of GFP-Sfp1 into the cytoplasm at the temperature non-permissive for the nup49-313 strain, suggesting that GFP-Sfp1 is a nucleo-cytoplasmic shuttling protein. This observation confirmed the dynamic nature of the nucleo-cytoplasmic distribution of Sfp1. For example, a similar redistribution to the cytoplasm was previously reported following rapamycin treatment and under starvation (Marion et al., PNAS 2004). In conjunction with the observation of an interaction with Rpb4, the authors observed slower nuclear import kinetics for GFP-Sfp1 in the absence of Rpb4 when cells were transferred to a glucose-containing medium after a period of starvation. Since the redistribution of GFP-Sfp1 was abolished in an rpb1-1/nup49-313 double mutant, the authors concluded that Sfp1 localisation to the cytoplasm depends on transcription. The double mutant yeast cells may show a variety of non-specific effects at the restrictive temperature, and whether transcription is required for Sfp1 cytoplasmic localisation remains incompletely demonstrated.

We concur with Reviewer 2 that any heat inactivation of a temperature-sensitive (ts) protein can result in non-specific effects. In the instance of rpb1-1, these non-specific effects are anticipated because of the transcriptional arrest, which can eventually lead to a reduction in protein content. However, it is worth noting that this process takes some time, whereas the impact on export is more rapid. We note that that this assay has been developed and published in prior research by Pam Silver (op. cit.) and was reported in a number of subsequent papers. Reassuringly, our conclusion is supported by the observation that Sfp1 binds to Pol II transcripts co-transcriptionally.

1. Under starvation conditions, which led to the presence of Sfp1 in the cytoplasm and have previously been correlated with a decrease in the transcription of Sfp1 target genes, the authors observed that a plasmid-based expressed GFP-Sfp1 accumulated in cytoplasmic foci. These foci were also labelled by P-body markers such as Dcp2 and Lsm1. The quality of the microscopic images provided does not allow to determine whether Rpb4-RFP colocalises with GFP-Sfp1.

The submitted PDF figure is of low quality. We believe that high quality figure will be convincing.

1. To understand to which RNA Sfp1 might bind, the authors used an N-terminally tagged fusion protein in a cross-linking and purification experiment. This method identified 264 transcripts for which the CRAC signal was considered positive and which mostly correspond to abundant mRNAs, including 74 ribosomal protein mRNAs or metabolic enzyme-abundant mRNAs such as PGK1. The authors did not provide evidence for the specificity of the observed CRAC signal, in particular, what would be the background of a similar experiment performed without UV cross-linking. In a validation experiment, the presence of several mRNAs in a purified SFP1 fraction was measured at levels that reflect the relative levels of RNA in a total RNA extract. Negative controls showing that abundant mRNAs not found in the CRAC experiment were clearly depleted from the purified fraction with Sfp1 would be crucial to assessing the specificity of the observed protein-RNA interactions. The CRAC-selected mRNAs were enriched for genes whose expression was previously shown to be upregulated upon Sfp1 overexpression (Albert et al., 2019). The presence of unspliced RPL30 pre-mRNA in the Sfp1 purification was interpreted as a sign of co-transcriptional assembly of Sfp1 into mRNA, but in the absence of valid negative controls, this hypothesis would require further experimental validation.

We argue that the 264 CRAC+ genes represent a distinct group with many unique features. Moreover, many CRAC+ genes do not fall into the category of highly transcribed genes.

The biological significance of the 264 CRAC+ mRNAs was demonstrated by various experiments; all are inconsistent with technical flaws. Some examples are:

1. Fig. 2a and B show that most reads of CRAC+ mRNA were mapped to specific location – close the pA sites.

2. Fig. 2C shows that most reads of CRAC+ mRNA were mapped to specific RNA motif.

3. Most RiBi CRAC+ promoter contain Rap1 binding sites (p= 1.9x10-22), whereas the vast majority of RiBi CRAC- promoters do not contain Rap1 binding site. (Fig. 3C).

4. Fig. 4A shows that RiBi CRAC+ mRNAs become destabilized due to Sfp1 deletion, whereas RiBi CRAC- mRNAs do not. Fig. 4B shows similar results due to

5. Fig. 6B shows that the impact of Sfp1 on backtracking is substantially higher for CRAC+ than for CRAC- genes. This is most clearly visible in RiBi genes.

6. Fig. 7A shows that the Sfp1-dependent changes along the transcription units is substantially more rigorous for CRAC+ than for CRAC-.

7. Fig. S4B Shows that chromatin binding profile of Sfp1 is different for CRAC+ and CRAC- genes

Moreover, only a portion of the RiBi mRNAs binds Sfp1, despite similar expression of all RiBi.

Most importantly, these genes do not all fall into the category of highly transcribed genes. On the contrary, as depicted in Figure 6A (green dots), it is evident that CRAC+ genes exhibit a diverse range of Rpb3 ChIP and GRO signals. Furthermore, as illustrated in Figure 7A, when comparing CRAC+ to Q1 (the most highly transcribed genes), it becomes evident that the Rpb4/Rpb3 profile of CRAC+ genes is not a result of high transcription levels. In our revised paper, we will give increased attention to this matter in the Discussion section.

1. To address the important question of whether co-transcriptional assembly of Spf1 with transcripts could alter their stability, the authors first used a reporter system in which the RPL30 transcription unit is transferred to vectors under different transcriptional contexts, as previously described by the Choder laboratory (Bregman et al. 2011). While RPL30 expressed under an ACT1 promoter was barely detectable, the highest levels of RNA were observed in the context of the native upstream RPL30 sequence when Rap1 binding sites were also present. Sfp1 showed better association with reporter mRNAs containing Rap1 binding sites in the promoter region. However, removal of the Rap1 binding sites from the reporter vector also led to a drastic decrease in reporter mRNA levels. Whether the fraction of co-purified RNA is nuclear and co-transcriptional or not cannot be inferred from these results.

The proposed co-transcriptional binding of Sfp1 is based on the findings presented in Figure 5C and Figure S2D, as well as the observed binding of Sfp1 to transcripts containing introns, as shown in Figures 2D and 3B. Our conclusion, which we still uphold, was drawn from the results presented in Figure 3. These results led us to the assertion that the "RNA-binding capacity of Sfp1 is regulated by Rap1-binding sites located at the promoter." We maintain our stance on this conclusion. Indeed, the Rap1 binding site does impact mRNA levels, as highlighted by Reviewer 2. However, "construct E," which possesses a promoter with a Rap1 binding site, exhibits lower transcript levels compared to "construct F," which lacks such a binding site in its promoter. Despite this difference in transcript levels, Sfp1 was able to pull down the former transcript but not the latter, even though expression of the former gene is relatively low. Thus, the results appear to be more reliant on the specific capacity of Sfp1 to interact with the transcript rather than on the transcript's expression level.

1. To complement the biochemical data presented in the first part of the manuscript, the authors turned to the deletion or rapid depletion of SFP1 and used labelling experiments to assess changes in the rate of synthesis, abundance, and decay of mRNAs under these conditions. An important observation was that in the absence of Sfp1, mRNAs encoding ribosomal protein genes not only had a reduced synthesis rate but also an increased degradation rate. This important observation needs careful validation, as genomic run-on experiments were used to measure half-lives, and this particular method was found to give results that correlated poorly with other measures of half-life in yeast (e.g. Chappelboim et al., 2022 for a comparison). Similarly, the use of thiolutin to block transcription as a method of assessing mRNA half-life has been reported to be problematic, as thiolutin can specifically inhibit the degradation of ribosomal protein mRNA (Pelechano & Perez-Ortin, 2008). Specific repressible reporters, such as those used by Baudrimont et al. (2017), would need to be tested to validate the effect of Sfp1 on the half-life of specific mRNAs. Also, it would be very difficult to infer from the images presented whether the rate of deadenylation is altered by Sfp1.

Various methods exist for assessing mRNA half-lives (HLs), and each of them carries its own set of challenges and biases. Consequently, it becomes problematic to directly compare HL values of a specific mRNA when different methods are employed. The superiority of one particular method over others remains unclear. However, they all exhibit a high degree of reliability when it comes to comparing different strains under the identical conditions using a single method.

Estimating half-lives through the GRO approach is a non-invasive method, applied on optimally proliferating cells, which has been employed in numerous publications. While no method is without its limitations, we consider this approach to be among the most dependable. Our HL determination using thiolutin to block transcription provided results that were consistent with the values obtained by the GRO approach.

Nevertheless, in our revised manuscript, we plan to supplement the HL data, obtain by thiolutin, with results obtained by subjecting cells to a temperature shift to 42°C, a natural method to block transcription in wild-type (WT) cells. This approach to determine HLs has been previously reported in our publications, such as Lotan et al. (2005, 2007) and Goler Baron et al. (2008).

1. The effects of SFP1 on transcription were investigated by chromatin purification with Rpb3, a subunit of RNA polymerase, and the results were compared with synthesis rates determined by genomic run-on experiments. The decrease in polII presence on transcripts in the absence of SFP1 was not accompanied by a marked decrease in transcript output, suggesting an effect of Sfp1 in ensuring robust transcription and avoiding RNA polymerase backtracking. To further investigate the phenotypes associated with the depletion or absence of Sfp1, the authors examined the presence of Rpb4 along transcription units compared to Rpb3. One effect of spf1 deficiency was that this ratio, which decreased from the start of transcription towards the end of transcripts, increased slightly. The results presented are largely correlative and could arise from the focus on very specific types of mRNAs, such as those of ribosomal protein genes, which are sensitive to stress and are targeted by very active RNA degradation mechanisms activated, for example, under heat stress (Bresson et al., 2020).

Figure 7A illustrates a significant reduction in Rpb4/Rpb3 ratios along the transcription unit in WT cells. This reduction is notably more pronounced in CRAC+ genes compared to the highly transcribed quartile (Q1), which includes all ribosomal protein (RP) genes, and it is completely absent in sfp1∆ cells. Furthermore, it's important to highlight that the CRAC+ gene group displays a wide range of transcription rates, as measured by either Rpb3 ChIP or GRO (Figure 6A). Given these observations, it is challenging to reconcile how the heightened sensitivity of RP mRNA degradation in response to stress could account for the more pronounced differences in the configuration of the Pol II elongation complex that are detected in CRAC+ genes under standard culture conditions in wt cells.

Correlative studies are particularly informative when a gene mutation eliminates a correlation, and this is precisely the type of study depicted in Figure 7B-C. The configuration of elongating Pol II (as reflected by Rpb4/Rpb3 ratios) and the backtracking index are both transcriptional outputs. It is difficult to envision how stress-induced destabilization of RP mRNAs could explain the twofold higher correlation between these two parameters observed in CRAC+ genes under non-stressful conditions in WT cells (Figure 7B).

Furthermore, it's worth noting that in WT cells, CRAC+ genes did not display any apparent unusual destabilization, but rather exhibited higher (not lower) mRNA stability compared to CRAC- genes (Figure 7C).

Strengths: - Diversity of experimental approaches used - Validation of large-scale results with appropriate reporters

Weaknesses: - Choice of evaluation method to test mRNA half-life - Lack of controls for the CRAC results

Reviewer #1 (Public Review):

The manuscript investigates the binding of PHD-BD, a tandem of reader domains in the C-terminus of BPTF, to modified histone tail peptides and nucleosomes. It focuses on the differences in binding affinity between peptide and nucleosome substrates for BPTF PHD-BD. Using the dCypher approach, they find that multi-modified peptide substrates (both acetylation and methylation) do not increase PHD-BD binding affinity. They argue that histone peptide substrates do not support the histone code model, which champions that multivalent engagement by PHD-BD with a multi-modified substrate would lead to stronger binding when compared to the engagement of each domain alone. In contrast, when using nucleosome substrates, even though the overall affinity is reduced, the affinity for H3K4me3triac (double modification) is tighter than either modification on its own. This is consistent with the histone code model.

A strength of the manuscript is that it further delineates the contribution of each domain by again using dCypher to compare peptide and nucleosome binding of the PHD and BD domains alone, as well as tandem domain constructs where each domain has been inactivated by a point mutation (W2891A for the PHD and N3007A for the BD). PHD alone had a lower affinity for nucleosomes than peptides overall. With peptide substrates, PHD had the highest affinity for H3K4me3 and reduced affinity for H3K4me3triac; while with nucleosomes this trend was reversed. BD alone showed an affinity for acetylated H3 and H4 peptides but surprisingly was unable to bind nucleosomes. PHD requires the combination of H3K4 methylation and H3 tail acetylation for binding, and when partnered with BD, which is not able to bind nucleosomes alone, interestingly confers specificity for K14ac and K18ac. The in vivo relevance is argued using CUT&RUN analysis.

NMR spectroscopy is further used to show that PHD-BD binds acetylated H3 in a multivalent manner while forming a unique complex with H3K4me3triac. Deleting the N-terminal A1 region of H3 abolishes the binding of PHD-BD, implying its importance for recognition. The authors also discuss a "fuzzy complex" that forms between H3 and DNA, as well as H4 and DNA, which explains the occlusion of histone tail accessibility in the nucleosome. By changing the sidechain charge, such as with PTMs, this interaction can be weakened and allow PHD in this case to bind to the modified H3 tail. Comparisons between spectra of the H4 tail, H4 tail with DNA, and the H4 tail in the nucleosome are made and used to argue for H4-DNA interactions in the nucleosome.

The conclusions of the manuscript are very well-supported by the data and reveal a lot of insight into how the two reader domains of BPTF interact with modified nucleosomes. In many places, however, the manuscript is written more generally as if the conclusions apply in all cases (e.g. the title, abstract, and introduction) and this remains to be determined. It is also overstated that there is a belief that peptides perfectly recapitulate nucleosomes. It should also be pointed out that the nucleosomes are multi-valent and the data cannot discriminate binding of a single PHD-BD to single or multiple tails, and that the work is limited as it is using a construct of BPTF and in fact, there is at least one other reader domain involved.

Reviewer #2 (Public Review):

This manuscript by Musselman and coworkers uses a commercial library of modified histone peptides and mononucleosomes to probe the substrate specificity of the PHD-bromodomain combination of the BPTF protein. They arrive at the conclusion that BPTF preferably binds H3K4me3 and H3K18ac in the H3 tail. By using NMR with lableled H4 protein in nucleosomes they show that the H4 tail interacts with DNA, which may limit its ability to interact with BPTF. Finally, experiments in cells demonstrate that BPTF, H3K4me3, and H3K18ac occupy overlapping regions of chromatin. The authors suggest that recruitment of BPTF to specific regions of chromatin is driven by the co-binding of H3K4me3 and H3K18ac by BPTF. This study is of interest to readers interested in understanding the functions of the BPTF protein in cells.

In this reviewer's opinion, the manuscript needs some revision and the inclusion of some missing information.

1) The authors seem to have overlooked the fact that mononucleosome substrates have been in use for determining the substrate specificity and mechanisms of quite a few enzymes that simply do not act on peptide substrates. For example, Dot1L doesn't do anything with peptides nor does COMPASS/Set1, both of which require intact nucleosomal substrates to measure their activity in response to ubiquitylated H2B. Thus, the authors' refinement of the "histone code hypothesis" is unnecessary and overdone. I would suggest that they instead cite examples where nucleosome substrates have provided answers that cannot be obtained from peptide substrates alone. For example, extensive work from the Muir and Allis labs.

2) Ruthenburg and Allis in Cell 2011 conducted similar experimentation and concluded that H3K4me3-H4K16ac is a modification state bound by BPTF in cells. They also showed co-localization in ChIP-seq experiments and demonstrated preferential pulldowns with BPTF and semisynthetic methylated and acetylated nucleosomes. The authors have entirely ignored these previous results in their own discussions. Readers would benefit from a side-by-side comparison of the two acetylation states to get a sense of which is a stronger interaction and why both seemingly correlate in CUTnRUN or ChIP-seq.

3) The idea that electrostatics may modulate tail accessibility was reported by Musselman and coworkers for the H3 tail in eLife 2018. Yet the PHD domain of BPTF clearly binds H3K4me3 in nucleosomes. In light of this prior observation, the NMR experiments now with H4 tail seem repetitive and not informative regarding BPTF's bromodomain binding. Also, missing is the effect of H4K16acetylation on H4 tail dynamics, which would be pertinent to addressing the hypothesis regarding the BPTF bromodomain binding H4K16ac

4) The NMR experiments are all undertaken with 150mM KCl with no NaCl present. While NMR experimental constraints are understandable, the authors should avoid sweeping statements from NMR experiments regarding the dynamism of histone tails in chromatin, unless specific experiments are cited/conducted to demonstrate the same in cells. Many factors may contribute to the exclusion of BPTF from modified histone tails in cells, including the binding of other reader proteins, and the precise genomic localization of these modifications vis-a-vis BPTF. The important role of anchoring proteins must also be taken into account when considering binding/non-binding of substrates by CAPs. Thus, the NMR experiments presented in the manuscript do not report on whether BPTF binds H4K16ac in cells or indeed in vitro. If the PHD domain is capable of ultimately binding the H3 tail despite the tail's fuzzy interaction with DNA, the question remains as to why the bromodomain may not do so for acetylated H4 tails?

This manuscript reports several interesting elements regarding BPTF regulation, but as presented it is missing some key comparisons with prior information that makes it hard for readers to assess the relevance of the results presented.

• for: stone age code, similar to - Ronald Wright - computer metaphor, evolutionary psychology - examples, evolutionary paradox of modernity, evolution - last mile link, major evolutionary transition - full spectrum in modern humans, example - MET - full spectrum embedded in modern humans

• comment

• insights

  • evolutionary paradox of modernity
    • modern humans , like all the living species we share the world with, are the last mile link of the evolution of life we've made it to the present, so all species of the present are, in an evolutionary sense, winners of their respective evolutionary game
    • this means that all our present behaviors contain the full spectrum of the evolutionary history of 4 billion years of life
    • the modern human embodies all major evolutionary transitions of the past
    • so our behavior, at all levels of our being is a complex and heterogenous mixture of evolutionary adaptations from different time periods of the 4 billion years that life has taken to evolve.
    • Some behaviors may have originated billions of years ago, and others hundred thousand years ago.

• Examples: humans embody full spectrum of METs in our evolutionary past

  • fight and flight response
    • early hominids on African Savannah hundreds of thousands to millions of years ago when hominids were predated upon by wild predators
  • cancer
    • normative intercell communication breaks down and reverts to individual cell behavior from billions of years ago
      • see Michael Levin's research on how to make metastatic cancer cells return to normative collective, cooperative behavior
  • children afraid to sleep in the dark
    • evolutionary adaptation against dangerous animals that might have hid in the dark - dangerous insiects, snakes, etc, which in the past may have resulted in human fatalities
  • obesity
    • hunter gatherer hominid attraction to rich sources of fruit. Eating as much of it as we can and maybe harvesting as much as we can and carrying that with us.
      • like squirrels storing away for the winter.

Gilad Bracha's criticism:

Semi-tangent: source code seems the most obvious thing in the world; but traditional Smalltalk’s have no real syntax above the method level! Classes are defined via the evaluation of reflective expressions, which rely on the reflective API. This is very problematic: the API often varies from one implementation to another. By the way, this is one of the ways Newspeak differs from almost every Smalltalk (the late, great Resilient being the only exception I can recall). Newspeak has a true syntax. Furthermore, because Newspeak module declarations are fully parametric in all their external dependencies, they can be compiled at any time in any order - unlike code in most languages (say Java packages) where there are numerous constraints on compilation order (e.g., imports must be defined).

See the full blog post An Image Problem.

Change this language

That resonates with me deeply. While English is not my native language, I've come to appreciate the wisdom behind programming being conducted exclusively in English. It establishes a universal standard, allowing coders worldwide to share a common language. This not only promotes convenience but ensures that even if someone's English proficiency is not perfect, their code remains understandable to English speakers.

I think that even with people who speak in other languages, the "for" and "if" coding terms are just symbols to them that they interpret in their brains as something necessary to code. I think the universal code makes it easier to handle.

As a child, I initially thought programming languages were translations of other languages, but I soon discovered that this wasn't the case when I began learning how to code. Given that English was the language of choice in the early development of programming languages, it makes sense that many programs were designed using the same foundation. However, I believe that to promote global learning and ensure efficient access to this skill, we should explore ways to make coding more accessible in languages other than English.

N

IPFS Connect Istanbul 2023: IPVM - Bringing Wasm-Based Edge Compute to IPFS

[docdrop](IPFS Connect Istanbul 2023: IPVM - Bringing Wasm-Based Edge Compute to IPFS

15 views 22 Nov 2023 The advent of TCP/IP and the web produced an explosion of innovation by radically lowering the barrier to entry to connect over the network. Thanks to new technical and social innovations, we now have the building blocks for the next generation of open services: location-free verifiable data and computation.

Verifiable computation opens the door to content-addressed function invocations, results, and workflows. This radically lowers the complexity of historical architectures (e.g. the LAMP stack), networked, and distributed systems. Not only is this easier to reason about, it also (paradoxically) enables superlinear scaling: the more it gets used, the more efficient it becomes!

This talk presents UCAN Invocations and the Interplanetary VM (IPVM). Code in this model can run anywhere (even offline), respects data privacy, and services interoperate seamlessly out of the box without pre-negotiation. Since computation doesn't happen in a vacuum, we will also describe how the workflow planner interacts with existing services and how to lift them into this seamless paradigm.) 15 views 22 Nov 2023 The advent of TCP/IP and the web produced an explosion of innovation by radically lowering the barrier to entry to connect over the network. Thanks to new technical and social innovations, we now have the building blocks for the next generation of open services: location-free verifiable data and computation.

Verifiable computation opens the door to content-addressed function invocations, results, and workflows. This radically lowers the complexity of historical architectures (e.g. the LAMP stack), networked, and distributed systems. Not only is this easier to reason about, it also (paradoxically) enables superlinear scaling: the more it gets used, the more efficient it becomes!

This talk presents UCAN Invocations and the Interplanetary VM (IPVM). Code in this model can run anywhere (even offline), respects data privacy, and services interoperate seamlessly out of the box without pre-negotiation. Since computation doesn't happen in a vacuum, we will also describe how the workflow planner interacts with existing services and how to lift them into this seamless paradigm.

IPFS Connect Istanbul 2023: IPVM - Bringing Wasm-Based Edge Compute to IPFS

15 views 22 Nov 2023 The advent of TCP/IP and the web produced an explosion of innovation by radically lowering the barrier to entry to connect over the network. Thanks to new technical and social innovations, we now have the building blocks for the next generation of open services: location-free verifiable data and computation.

Verifiable computation opens the door to content-addressed function invocations, results, and workflows. This radically lowers the complexity of historical architectures (e.g. the LAMP stack), networked, and distributed systems. Not only is this easier to reason about, it also (paradoxically) enables superlinear scaling: the more it gets used, the more efficient it becomes!

This talk presents UCAN Invocations and the Interplanetary VM (IPVM). Code in this model can run anywhere (even offline), respects data privacy, and services interoperate seamlessly out of the box without pre-negotiation. Since computation doesn't happen in a vacuum, we will also describe how the workflow planner interacts with existing services and how to lift them into this seamless paradigm.

15 views 22 Nov 2023 The advent of TCP/IP and the web produced an explosion of innovation by radically lowering the barrier to entry to connect over the network. Thanks to new technical and social innovations, we now have the building blocks for the next generation of open services: location-free verifiable data and computation.

Verifiable computation opens the door to content-addressed function invocations, results, and workflows. This radically lowers the complexity of historical architectures (e.g. the LAMP stack), networked, and distributed systems. Not only is this easier to reason about, it also (paradoxically) enables superlinear scaling: the more it gets used, the more efficient it becomes!

This talk presents UCAN Invocations and the Interplanetary VM (IPVM). Code in this model can run anywhere (even offline), respects data privacy, and services interoperate seamlessly out of the box without pre-negotiation. Since computation doesn't happen in a vacuum, we will also describe how the workflow planner interacts with existing services and how to lift them into this seamless paradigm.

This talk presents UCAN Invocations and the Interplanetary VM (IPVM). Code in this model can run anywhere (even offline), respects data privacy, and services interoperate seamlessly out of the box without pre-negotiation. Since computation doesn't happen in a vacuum, we will also describe how the workflow planner interacts with existing services and how to lift them into this seamless paradigm.

I do believe this might be the culprit for all the problems I've had with suddenly/inconsistently unresponsive keyboard shortcuts recently...

​We must start with end-programmer programming - to achieve end-user programming,

- where users can

• spin up personalized apps
  • without knowing how to code.

With end-programmer programming, software engineers can build - folk applications, integrations, and mini-apps - to customize - their experience interacting with third-party software.

local ipfs

Machine-readable license specifiers are important in a setting where publishing the source code is the default.

代码不是写给自己一个人的

adaptive multi-factor authentication

Did they make their code available for others to do the test?

Suggesting changes (diff) instead of "auto-changing" would be more in the comfort zone for most, even in 2024 (??)

This a whole other project :) With a baggage of structure editing, edit-AST->code generation

Generating diagrams would be useful (e.g. mermaid), a la chatGPT plugins

Overlay UI * for missing docstrings, top level docstring for package/module/namespace * line level explainers / exploratory comments

General overview w/ jump to symbol links

Quote documentation from relavant packages w/ links

Needs: * accept all * accept granular changes * ability to ask for variants & iterate

Interesting, real-time-ish as you type HUD

(for that Also high-level summary of the changes)

Reviewer #1 (Public Review):

This work seeks to understand how behaviour-related information is represented in the neural activity of the primate motor cortex. To this end, a statistical model of neural activity is presented that enables a non-linear separation of behaviour-related from unrelated activity. As a generative model, it enables the separate analysis of these two activity modes, here primarily done by assessing the decoding performance of hand movements the monkeys perform in the experiments. Several lines of analysis are presented to show that while the neurons with significant tuning to movements strongly contribute to the behaviourally-relevant activity subspace, less or un-tuned neurons also carry decodable information. It is further shown that the discovered subspaces enable linear decoding, leading the authors to conclude that motor cortex read-out can be linear.

Strengths:

In my opinion, using an expressive generative model to analyse neural state spaces is an interesting approach to understanding neural population coding. While potentially sacrificing interpretability, this approach allows capturing both redundancies and synergies in the code as done in this paper. The model presented here is a natural non-linear extension of a previous linear model (PSID) and

Weaknesses:

First, the model in the paper is almost identical to an existing VAE model (TNDM) that makes use of weak supervision with behaviour in the same way [1]. This paper should at least be referenced. If the authors wish they could compare their model to TNDM, which combines a state space model with smoothing similar to LFADS. Given that TNDM achieves very good behaviour reconstructions, it may be on par with this model without the need for a Kalman filter (and hence may achieve better separation of behaviour-related and unrelated dynamics).

Second, in my opinion, the claims regarding identifiability are overstated - this matters as the results depend on this to some extent. Recent work shows that VAEs generally suffer from identifiability problems due to the Gaussian latent space [2]. This paper also hints that weak supervision may help to resolve such issues, so this model as well as TNDM and CEBRA may indeed benefit from this. In addition however, it appears that the relative weight of the KL Divergence in the VAE objective is chosen very small compared to the likelihood (0.1%), so the influence of the prior is weak and the model may essentially learn the average neural trajectories while underestimating the noise in the latent variables. This, in turn, could mean that the model will not autoencode neural activity as well as it should, note that an average R2 in this case will still be high (I could not see how this is actually computed). At the same time, the behaviour R2 will be large simply because the different movement trajectories are very distinct. Since the paper makes claims about the roles of different neurons, it would be important to understand how well their single trial activities are reconstructed, which can perhaps best be investigated by comparing the Poisson likelihood (LFADS is a good baseline model). Taken together, while it certainly makes sense that well-tuned neurons contribute more to behaviour decoding, I worry that the very interesting claim that neurons with weak tuning contain behavioural signals is not well supported.

Third, and relating to this issue, I could not entirely follow the reasoning in the section arguing that behavioural information can be inferred from neurons with weak selectivity, but that it is not linearly decodable. It is right to test if weak supervision signals bleed into the irrelevant subspace, but I could not follow the explanations. Why, for instance, is the ANN decoder on raw data (I assume this is a decoder trained fully supervised) not equal in performance to the revenant distilled signals? Should a well-trained non-linear decoder not simply yield a performance ceiling? Next, if I understand correctly, distilled signals were obtained from the full model. How does a model perform trained only on the weakly tuned neurons? Is it possible that the subspaces obtained with the model are just not optimally aligned for decoding? This could be a result of limited identifiability or model specifics that bias reconstruction to averages (a well-known problem of VAEs). I, therefore, think this analysis should be complemented with tests that do not depend on the model.

Finally, a more technical issue to note is related to the choice to learn a non-parametric prior instead of using a conventional Gaussian prior. How is this implemented? Is just a single sample taken during a forward pass? I worry this may be insufficient as this would not sample the prior well, and some other strategy such as importance sampling may be required (unless the prior is not relevant as it weakly contributed to the ELBO, in which case this choice seems not very relevant). Generally, it would be useful to see visualisations of the latent variables to see how information about behaviour is represented by the model.

Summary:

This paper presents a very interesting analysis, but I have several concerns as to well the analysis supports the main conclusions. I think the work could benefit from an additional complementary analysis that seeks to confirm with another method if weakly tuned neurons indeed show an encoding that differs qualitatively from the strongly tuned ones.

[1] Hurwitz, Cole, et al. "Targeted neural dynamical modeling." Advances in Neural Information Processing Systems 34 (2021): 29379-29392.<br /> [2] Hyvarinen, Aapo, Ilyes Khemakhem, and Hiroshi Morioka. "Nonlinear Independent Component Analysis for Principled Disentanglement in Unsupervised Deep Learning." arXiv preprint arXiv:2303.16535 (2023).

Use a Cascade to send several messages to the same receiver. Separate the messages with a semicolon. Put each message on its own line and indent one tab. Only use Cascades for messages with zero or one argument.

—Kent Beck, Smalltalk Best Practice Patterns

In the Spacewar! code snippet below, the parentheses aren't necessary but clarify that the addition happens first:

Lyle Bickley explains the PDP-1 (and we play the original Spacewar! at Curious Mark YouTube channel

From the description:

Lyle Bickley, of the PDP-1 restoration team, gives us a tour of this amazing, early scientific interactive computer at the Computer History Museum. The first machine built by DEC in 1959, it features a superb graphics screen. DEC gave one to MIT, and some very bright students went wild. Gems such as Spacewar!, Snowflake, 4-voice music programs were all developed by moonlighting MIT students, unencumbered by its measly 12kW memory and pokey 100,000 instructions per second. Along with much more serious debugging and programming languages of course. You can come and see the real machine for yourself at the Computer History Museum in Mountain View, California: http://www.computerhistory.org/

Also, Norbert Landsteiner made this incredible simulation of the PDP-1 that can run the original Spacewar! and Minkytron code in your browser: https://www.masswerk.at/spacewar/ https://www.masswerk.at/minskytron/ He also made a gate exact replica with Verilog code on github: https://www.youtube.com/watch?v=iymD9eysqXo

A website is both a hobby and a life's work

Scarlet Harris (profile at University of Cambridge)

Contact Information: sh2232@cam.ac.uk

Dr Scarlet Harris is a Teaching Associate in the department of Sociology.

She received her BA in Sociology from the University of Edinburgh and her MRes and PhD in Sociology from the University of Glasgow. Dr Harris has held various research posts at the University of Manchester, including with the Centre on the Dynamics of Ethnicity (CoDE).

She is currently writing a book based on her doctoral research, entitled 'Islamophobia, anti-racism and the British left', which will be published by Manchester University Press.

https://research.sociology.cam.ac.uk/profile/dr-scarlet-harris

accessed:: 2023-11-25 17:10

Les points avant et après la variable sont-ils utiles ou es-ce dans un souci de lisibilité du code ?

Are the people in the learning center for sure the best way to learn the code because in 120 I did a lot of work with my professor and that helped me a lot.

我无法理解为什么会有人觉得：对于一个已经以复杂著称的Rust来说，这种 Aysnc 实现还能称得上是一个好设计。 我曾尝试去理解它，但还是无法接受，它会污染所有跟它有关的代码。我很喜欢 Rust，同时它也是我最常用的编程语言，但每次我在编写异步密集的代码时，我牙关就会开始打颤、视线也变得模糊。

Reviewer #1 (Public Review):

Summary:

The goal of Pawel et al. is to provide a more rigorous and quantitative approach for judging whether or not an initial null finding (conventionally with p >= 0.05) has been replicated by a second similarly null finding. They discuss important objections to relying on the qualitative significant/non-significant dichotomy to make this judgement. They present two complementary methods (one frequentist and the other Bayesian) which provide a superior quantitative framework for assessing the replicability of null findings.

Strengths:

Clear presentation; illuminating examples drawn from the well-known Reproducibility Project: Cancer Biology data set; R-code that implements suggested analyses. Using both methods as suggested provides a superior procedure for judging the replicability of null findings.

Weaknesses:

The proposed frequentist and the Bayesian methods both rely on binary assessments of an original finding and its replication. I'm not sure if this is a weakness or is inherent to making binary decisions based on continuous data.

For the frequentist method, a null finding is considered replicated if the original and replication 90% confidence intervals for the effects both fall within the equivalence range. According to this approach, a null finding would be considered replicated if p-values of both equivalences tests (original and replication) were, say, 0.049, whereas would not be considered replicated if, for example, the equivalence test of the original study had a p-value of 0.051 and the replication had a p-value of 0.001. Intuitively, the evidence for replication would seem to be stronger in the second instance. The recommended Bayesian approach similarly relies on a dichotomy (e.g. Bayes factor > 1).

Reviewer #2 (Public Review):

Summary:

The study demonstrates how inconclusive replications of studies initially with p > 0.05 can be and employs equivalence tests and Bayesian factor approaches to illustrate this concept. Interestingly, the study reveals that achieving a success rate of 11 out of 15, or 73%, as was accomplished with the non-significance criterion from the RPCB (Reproducibility Project: Cancer Biology), requires unrealistic margins of Δ > 2 for equivalence testing.

Strengths:

The study uses reliable and sharable/open data to demonstrate its findings, sharing as well the code for statistical analysis. The study provides sensitivity analysis for different scenarios of equivalence margin and alfa level, as well as for different scenarios of standard deviations for the prior of Bayes factors and different thresholds to consider. All analysis and code of the work is open and can be replicated. As well, the study demonstrates on a case-by-case basis how the different criteria can diverge, regarding one sample of a field of science: preclinical cancer biology. It also explains clearly what Bayes factors and equivalence tests are.

Weaknesses:

It would be interesting to investigate whether using Bayes factors and equivalence tests in addition to p-values results in a clearer scenario when applied to replication data from other fields. As mentioned by the authors, the Reproducibility Project: Experimental Philosophy (RPEP) and the Reproducibility Project: Psychology (RPP) have data attempting to replicate some original studies with null results. While the RPCB analysis yielded a similar picture when using both criteria, it is worth exploring whether this holds true for RPP and RPEP. Considerations for further research in this direction are suggested. Even if the original null results were excluded in the calculation of an overall replicability rate based on significance, sensitivity analyses considering them could have been conducted. The present authors can demonstrate replication success using the significance criteria in these two projects with initially p < 0.05 studies, both positive and non-positive.

Other comments:

- Introduction: The study demonstrates how inconclusive replications of studies initially with p > 0.05 can be and employs equivalence tests and Bayesian factor approaches to illustrate this concept. Interestingly, the study reveals that achieving a success rate of 11 out of 15, or 73%, as was accomplished with the non-significance criterion from the RPCB (Reproducibility Project: Cancer Biology), requires unrealistic margins of Δ > 2 for equivalence testing.

- Overall picture vs. case-by-case scenario: An interesting finding is that the authors observe that in most cases, there is no substantial evidence for either the absence or the presence of an effect, as evidenced by the equivalence tests. Thus, using both suggested criteria results in a picture similar to the one initially raised by the paper itself. The work done by the authors highlights additional criteria that can be used to further analyze replication success on a case-by-case basis, and I believe that this is where the paper's main contributions lie. Despite not changing the overall picture much, I agree that the p-value criterion by itself does not distinguish between (1) a situation where the original study had low statistical power, resulting in a highly inconclusive non-significant result that does not provide evidence for the absence of an effect and (2) a scenario where the original study was adequately powered, and a non-significant result may indeed provide some evidence for the absence of an effect when analyzed with appropriate methods. Equivalence testing and Bayesian factor approaches are valuable tools in both cases.

Regarding the 0.05 threshold, the choice of the prior distribution for the SMD under the alternative 𝐻1 is debatable, and this also applies to the equivalence margin. Sensitivity analyses, as highlighted by the authors, are helpful in these scenarios.

What happens if I hit the endpoint multiple times?

class of bugs

They are not "used" in isolation though, as it is the context of these code snippets gleaned from tons of code that makes it valuable to the ML scene in comparison to static autocompletion of the prior generation

This assumption is flat-out incorrect. Virginia statute explicitly refers to "early voting" as "[a]bsentee voting in person." Va. Code Ann. § 24.2-701.1. If Mr. Dreyer had been familiar with the terminology used by election officials, the answer Mr. Dreyer should have provided the officer of election when asked whether he had voted absentee would have been "yes." That would have avoided all of the ensuing confusion. With this context, an "unsettling Election Day story," becomes nothing more than another example of the system working.

ChatGPT augments, not replaces

You have to know what you want to do before you can provide the prompt for the code generation.

I was just telling someone that OpenAI must decide who they wants to be. Do they want to be AWS/Azure or do they want to '"freely collaborate" with other institutions and researchers by making its patents and research open to the public'?

Well it looks like the choice was made for them. They just changed the name from "OpenAI" to... "Microsoft":

Two-factor authentication is important to teach students. The process of login in the website as well as the passwords or codes that can be sent through a text.

Working in IT, I can not tell you the sheer number of people who would do anything to avoid having to take the extra steps for security. Even when we explain to them the exact issues that come up, it seems there is a misunderstanding often about how dire a hacked account could become. This is part of why it is so necessary to make these security measures normalized as early as possible with students and technology!

eLife assessment

This work is a valuable presentation of sharp-wave-ripple reactivation of hippocampal neural ensemble activity recorded as animals explored two different environments. It attempts to use the fact that the ensemble code remaps between the two mazes to identify the best replay-detection procedures for analyzing this type of data. The reviewers found the evidence for a prescriptive conclusion inadequate, while still appreciating the concept of comparing maze-identity discrimination with replay.

Reviewer #1 (Public Review):

This work introduces a novel framework for evaluating the performance of statistical methods that identify replay events. This is challenging because hippocampal replay is a latent cognitive process, where the ground truth is inaccessible, so methods cannot be evaluated against a known answer. The framework consists of two elements:<br /> 1. A replay sequence p-value, evaluated against shuffled permutations of the data, such as radon line fitting, rank-order correlation, or weighted correlation. This element determines how trajectory-like the spiking representation is. The p-value threshold for all accepted replay events is adjusted based on an empirical shuffled distribution to control for the false discovery rate.<br /> 2. A trajectory discriminability score, also evaluated against shuffled permutations of the data. In this case, there are two different possible spatial environments that can be replayed, so the method compares the log odds of track 1 vs. track 2.

The authors then use this framework (accepted number of replay events and trajectory discriminability) to study the performance of replay identification methods. They conclude that sharp wave ripple power is not a necessary criterion for identifying replay event candidates during awake run behavior if you have high multiunit activity, a higher number of permutations is better for identifying replay events, linear Bayesian decoding methods outperform rank-order correlation, and there is no evidence for pre-play.

The authors tackle a difficult and important problem for those studying hippocampal replay (and indeed all latent cognitive processes in the brain) with spiking data: how do we understand how well our methods are doing when the ground truth is inaccessible? Additionally, systematically studying how the variety of methods for identifying replay perform, is important for understanding the sometimes contradictory conclusions from replay papers. It helps consolidate the field around particular methods, leading to better reproducibility in the future. The authors' framework is also simple to implement and understand and the code has been provided, making it accessible to other neuroscientists. Testing for track discriminability, as well as the sequentiality of the replay event, is a sensible additional data point to eliminate "spurious" replay events.

However, there are some concerns with the framework as well. The novelty of the framework is questionable as it consists of a log odds measure previously used in two prior papers (Carey et al. 2019 and the authors' own Tirole & Huelin Gorriz, et al., 2022) and a multiple comparisons correction, albeit a unique empirical multiple comparisons correction based on shuffled data.

With respect to the log odds measure itself, as presented, it is reliant on having only two options to test between, limiting its general applicability. Even in the data used for the paper, there are sometimes three tracks, which could influence the conclusions of the paper about the validity of replay methods. This also highlights a weakness of the method in that it assumes that the true model (spatial track environment) is present in the set of options being tested. Furthermore, the log odds measure itself is sensitive to the defined ripple or multiunit start and end times, because it marginalizes over both position and time, so any inclusion of place cells that fire for the animal's stationary position could influence the discriminability of the track. Multiple track representations during a candidate replay event would also limit track discriminability. Finally, the authors call this measure "trajectory discriminability", which seems a misnomer as the time and position information are integrated out, so there is no notion of trajectory.

The authors also fail to make the connection with the control of the false discovery rate via false positives on empirical shuffles with existing multiple comparison corrections that control for false discovery rates (such as the Benjamini and Hochberg procedure or Storey's q-value). Additionally, the particular type of shuffle used will influence the empirically determined p-value, making the procedure dependent on the defined null distribution. Shuffling the data is also considerably more computationally intensive than the existing multiple comparison corrections.

Overall, the authors make interesting conclusions with respect to hippocampal replay methods, but the utility of the method is limited in scope because of its reliance on having exactly two comparisons and having to specify the null distribution to control for the false discovery rate. This work will be of interest to electrophysiologists studying hippocampal replay in spiking data.

eLife assessment

This work is a valuable presentation of sharp-wave-ripple reactivation of hippocampal neural ensemble activity recorded as animals explored two different environments. It attempts to use the fact that the ensemble code remaps between the two mazes to identify the best replay-detection procedures for analyzing this type of data. The reviewers found the evidence for a prescriptive conclusion inadequate, while still appreciating the concept of comparing maze-identity discrimination with replay.

Reviewer #1 (Public Review):

This work introduces a novel framework for evaluating the performance of statistical methods that identify replay events. This is challenging because hippocampal replay is a latent cognitive process, where the ground truth is inaccessible, so methods cannot be evaluated against a known answer. The framework consists of two elements:<br /> 1. A replay sequence p-value, evaluated against shuffled permutations of the data, such as radon line fitting, rank-order correlation, or weighted correlation. This element determines how trajectory-like the spiking representation is. The p-value threshold for all accepted replay events is adjusted based on an empirical shuffled distribution to control for the false discovery rate.<br /> 2. A trajectory discriminability score, also evaluated against shuffled permutations of the data. In this case, there are two different possible spatial environments that can be replayed, so the method compares the log odds of track 1 vs. track 2.

The authors then use this framework (accepted number of replay events and trajectory discriminability) to study the performance of replay identification methods. They conclude that sharp wave ripple power is not a necessary criterion for identifying replay event candidates during awake run behavior if you have high multiunit activity, a higher number of permutations is better for identifying replay events, linear Bayesian decoding methods outperform rank-order correlation, and there is no evidence for pre-play.

The authors tackle a difficult and important problem for those studying hippocampal replay (and indeed all latent cognitive processes in the brain) with spiking data: how do we understand how well our methods are doing when the ground truth is inaccessible? Additionally, systematically studying how the variety of methods for identifying replay perform, is important for understanding the sometimes contradictory conclusions from replay papers. It helps consolidate the field around particular methods, leading to better reproducibility in the future. The authors' framework is also simple to implement and understand and the code has been provided, making it accessible to other neuroscientists. Testing for track discriminability, as well as the sequentiality of the replay event, is a sensible additional data point to eliminate "spurious" replay events.

However, there are some concerns with the framework as well. The novelty of the framework is questionable as it consists of a log odds measure previously used in two prior papers (Carey et al. 2019 and the authors' own Tirole & Huelin Gorriz, et al., 2022) and a multiple comparisons correction, albeit a unique empirical multiple comparisons correction based on shuffled data.

With respect to the log odds measure itself, as presented, it is reliant on having only two options to test between, limiting its general applicability. Even in the data used for the paper, there are sometimes three tracks, which could influence the conclusions of the paper about the validity of replay methods. This also highlights a weakness of the method in that it assumes that the true model (spatial track environment) is present in the set of options being tested. Furthermore, the log odds measure itself is sensitive to the defined ripple or multiunit start and end times, because it marginalizes over both position and time, so any inclusion of place cells that fire for the animal's stationary position could influence the discriminability of the track. Multiple track representations during a candidate replay event would also limit track discriminability. Finally, the authors call this measure "trajectory discriminability", which seems a misnomer as the time and position information are integrated out, so there is no notion of trajectory.

The authors also fail to make the connection with the control of the false discovery rate via false positives on empirical shuffles with existing multiple comparison corrections that control for false discovery rates (such as the Benjamini and Hochberg procedure or Storey's q-value). Additionally, the particular type of shuffle used will influence the empirically determined p-value, making the procedure dependent on the defined null distribution. Shuffling the data is also considerably more computationally intensive than the existing multiple comparison corrections.

Overall, the authors make interesting conclusions with respect to hippocampal replay methods, but the utility of the method is limited in scope because of its reliance on having exactly two comparisons and having to specify the null distribution to control for the false discovery rate. This work will be of interest to electrophysiologists studying hippocampal replay in spiking data.

I guess this explains what the interaction trace is about, and why there are interaction elements coming from verifier in the code base.

If you're going to write a plugin for an IDE, prepare for your hello-world to be days of learning and pages of code just to do the hello-world.

Now, I would love someday to do a plug-in for IntelliJ that understands all of the custom stuff for my game code. Right? You know, I would love to, but you know that's that's a project.