A full-body transcription factor expression atlas with completely resolved cell identities in C. elegans

Invariant cell lineage in C. elegans enables spatiotemporal resolution of transcriptional regulatory mechanisms controlling the fate of each cell. Here, we develop RAPCAT (Robust-point-matching- And Piecewise-affine-based Cell Annotation Tool) to automate cell identity assignment in three-dimensional image stacks of L1 larvae and profile reporter expression of 620 transcription factors in every cell. Transcription factor profile-based clustering analysis defines 80 cell types distinct from conventional phenotypic cell types and identifies three general phenotypic modalities related to these classifications. First, transcription factors are broadly downregulated in quiescent stage Hermaphrodite Specific Neurons, suggesting stage- and cell type-specific variation in transcriptome size. Second, transcription factor expression is more closely associated with morphology than other phenotypic modalities in different pre- and post-differentiation developmental stages. Finally, embryonic cell lineages can be associated with specific transcription factor expression patterns and functions that persist throughout postembryonic life. This study presents a comprehensive transcription factor atlas for investigation of intra-cell type heterogeneity.

Problem #2: While the reporter collection established by the authors is potentially useful, it is a very significant shortcoming that the vast majority of reporters used are mere promoter fusions that very often do not capture the full expression of the respective genes.This is a fundamental problem in the interpretation of the expression patterns.The authors are in a quite unique position to address such discrepancies, by again making use of the homeobox reporter alleles mentioned above, and comparing their published expression patterns with the many homoeobox promoter gene fusions that they use in the paper.The comparison will give a sense of how accurate a reflection of real expression patterns these reporters are.Again, such comparison have to be done with proper sample sizes, e.g. the entire homeobox family down by Reilly et al., with the promoter-based expression patterns of homeobox genes by the present authors.One or two genes are not enough.Same random sampling of patterns that I did manually make me worry about quite substantial disparities, but this may be biased.If there are indeed substantial discrepancy between the expression patterns, the author will have to dampen a large number of their conclusions.

Two other comments:
-For the reporter data, the authors should not just give primer sequence, but coordinates relative to start codon of the gene so that it is easier to grasp how much genomic region is covered by the reporter.
-one of the greatest values of this paper is the reporter resource.I expect that according to journal policy, these strains are deposited in a strain repository (CGC).

Reviewer #2 (Remarks to the Author):
Li et al reports a TF expression atlas in C. elegans with single-cell resolution, covering 612 TFs whose reporter expression is mapped to each of the 558 cells in the newly hatched L1 larvae.It is an impressive body of work with unique values complementing the more commonly used scRNA atlases.The authors offered a systemic analysis of the expression patterns in terms of terminal differentiation, lineage history and body axes, and identified an interesting case of developmental quiescence.The study also established a computational approach to reliably identity individual cells based on nuclear positions, which is still difficult to do in C. elegans.
The text is well written and easy to follow.However, the figures are hard to read.The labels are too small to read.On the other hand, some figures (eg, 4 and 5) show excessive images --it would be more effective to show the quantification with a representative image while moving the entire montage to a sup figure.This way it might also be possible to merge figures 4 and 5 into one, and split Fig 3 to two for the real estate it needs.
Most of the comments are on the computational method of template matching, which lacks crucial details: The presentation of the method in the main text is a bit unclear about how the full set of examples is used in scoring, this could and should be included in overview in the main text.
The choice of primary alignment template for the atlas appears to have an, at first glance, unusually large influence on the quality of results.It would be helpful to give some intuitions why this is, and what characteristics the other 4 sub-optimal templates chosen have (from fig S2A they appear to have perhaps been chosen for their unusually good -worst caseperformance, though the caption suggests they are the runners up in fig S2b, this should be clarified a bit more).Regardless of method what does the choice of 5 and templates represent?Major modes of position?Could a different metric for co-alignment of the 100 examples (rather than least squares fit to one chosen example) create a single reference atlas with different/better properties.What are the intuitions that argue for this approach rather than doing RPM 100 times and voting over these answers.
It's valuable that the authors quantified the error rate of the overall annotation pipeline including screening and correction for low confidence matches.Can they provide some insight about what influence the residual error of around 3% might have on results, and what minimal accuracy is needed?Is the 95% cited for raw alignment sufficient?How long does exhaustive curation take and how long does curating low confidence matches take?

Reviewer #3 (Remarks to the Author):
In this herculean work, Li and co-authors have generated an extensive collection of promoter-driven fluorescent reporter lines of transcription factors (TFs) and presented an expression atlas of TF reporters that cover 65% of all predicted TFs in every lineageresolved single cell of the L1 C. elegans larvae.Although scRNA-seq approaches have been previously used to determine high-resolution transcriptomes during C. elegans embryonic (e.g., Packer et al. Science 2019) and post-embryonic development (e.g., Cao et al. Science 2017and Taylor et al. Cell 2021), only a small portion of the transcriptomes can be unequivocally linked to a specific cell.This uncertainty undermines the analysis of cell typecell state relationships and state heterogeneity among similar/identical cell types.In contrast, this study takes advantage of the invariant cell lineage and stereotypic cell positioning in C. elegans development, using an imaging-based approach that allows for the quantification of the expression of a gene of interest in every annotated single cell and the synthesis of individual expression patterns into an atlas.The key advantage of this dataset is the clarity of cell annotation and complete cell coverage.Given the functional importance and expression informativeness of TFs, this dataset provides an opportunity to define the molecular/regulatory state of every cell at a developmental stage.Although Imaging-based single-cell analysis of TF expression at comparable resolution and scale has been conducted in C. elegans embryos (Murray et al. Genome Research 2012 andMa et al. Nature Methods 2021), this study stands out as the first to generate a comprehensive TF expression atlas at the L1 stage, with a greater number of TFs and higher cell coverage.
Using the atlas, the authors tackle an important yet unresolved question regarding the complex relationships between cell types and states.They grouped cells based on their TF expression (molecular state) and compared them to classic classifications based on the morphology and function of cells, reporting both consistencies and discrepancies between the two classifications.The authors described three cases to demonstrate the usefulness of the atlas in uncovering new cell states and state-type relationships.The first case describes a previously unrecognized quiescent state in newborn HSN neurons, characterized by gene repression.The second case concerns the state heterogeneity between P and G2/W neuroblasts, in which the authors characterized the regulatory differences between the otherwise similar cell types.Finally, the authors analyzed the incomplete convergence of cell states between cells originating from unrelated cell lineages but differentiating into identical/similar cell types, a phenomenon previously described during embryogenesis.Here, the authors verified and extended previous findings, showing that the lineage-dependent differences in cell states (TF expression) persist at the L1 stage when the terminal differentiation of many cells is completed.Furthermore, they focused on muscle cells and identified that earlier expression TFs (unc-39 and hnd-1) are responsible for lineagedependent state heterogeneity through asymmetric functions in different lineages.
This study presents a valuable functional genomic resource for studying multicellular development.The over 900 reporter lines (including ~700 transgenic lines generated in this study) and associated cellular expression patterns will be valuable genomic resources for investigating TFs, cell fates, and cellular states.The experiments are generally well executed, and the results appear to be of high quality.The comparisons of cell states defined by TF expression to cell types are useful, and the case studies are interesting.However, the presentation of the results in many sections is not well organized and presented, making some parts of the manuscript hard to follow and unclear the key message the authors want to deliver.Additionally, the quality assessments of the expression data are somewhat insufficient, and the authors tend to overlook the caveats of using fluorescent reports to indicate gene transcription, which should be explicitly stated and discussed.Finally, while analysis of cellular states using the TF atlas is definitely a good direction, the authors missed the rich opportunity of using this atlas to prioritize and characterize new regulators of cell types.Including this analysis could further enhance the impact of this study.To improve the manuscript, I suggest the following revisions.
Major comments: 1. RAPCAT methods.The authors provided a detailed description of the development of the RAPCAT algorithm and its ability to automatically assign cell identities based on stereotypic relative cell positions using training datasets.However, the accuracy of cell segmentation is equally important as well, which is not mentioned.Therefore, the authors should provide more information on how they dealt with the errors in nuclei segmentation, which inevitably affects the robustness of cell matching.
2. Quality assessment of expression atlas.To increase the utility of the atlas, it is critical to provide a thorough quality evaluation of the data to give readers a general idea of its quality.Although the authors show that the technical reproducibility of expression patterns is reasonably high (R> 0.81 between replicates), this result does not necessarily validate the correctness of the expression.Indeed, the correlation between different reporters of the same TF is not high (Figure 2B), raising concerns about to what extent the reporter expression recapitulates the endogenous expression.Since endogenous fluorescent fusion lines are available for many of the TFs analyzed in this study, the authors should perform a small-scale comparative analysis to determine the consistency.In addition, the authors should compare their expression pattern to benchmark genes whose expression and function are relatively well-documented.Although the single-cell expression at the L1 stage is not available for most TFs, tissue-level analysis is feasible.
3. Caveats of reporter-based expression profiling.The authors should explicitly mention and discuss the limitations of the fluorescent reporter-based strategy.Although this approach is a reasonable choice, it has several caveats that compromise the accuracy of the expression data.These include but are not limited to (1) the inability to fully recapitulate endogenous expression due to the relatively short promoter sequence included in the reporter and missing critical regulatory elements, (2) false negatives due to low detection sensitivity of imaging, (3) the potential influence of the integration site on expression, and (4) some of the expression signals are not from the subject cell but inherited from earlier cells through the cell lineage due to long half-life of the fluorescent protein.Importantly, some of the caveats mentioned above may also affect the reliability of TF-based cell clustering results.For example, the long half-life of fluorescent protein would cause an overestimation of the influence of cell lineage on cellular states, which should be taken into account when interpreting the results (see below).

Cell clustering by TF expression.
The key finding of this study is presented in Figure 3, but the results are not well-organized, making it difficult to follow.The use of many different colors in the figure makes it impossible to understand the exact relationships between different cell classifications.Additionally, the presentation of the different cell classifications in Table S4 is unclear.In particular, organizing the classifications into a matrix is highly unintuitive.A more effective way to present the information would be to show each cell and then different clustering or classifications as numbers or text.
5. Gene-level analysis.The cellular expression patterns of individual TFs are informative in revealing regulators of specific cell fates, types, or states.However, the authors did not perform any analysis to characterize TFs with interesting yet uncharacterized expression patterns.At the very least, the authors should identify and discuss individual TFs exhibiting cell-type or cell-state-specific expression patterns.This analysis is critical because it helps to demonstrate the value of high-resolution expression analysis to identify novel expression enrichment/functions of TFs.Furthermore, it is also an integral part of quality control to show how many known patterns are recapitulated (see Major comment 2).
6. Quiescent HSN state.TF expression status in the sister cell of HSNs (PHBs) should be included as an internal control to demonstrate gene repression in the quiescent HSN state.As the authors use the fluorescent signal to monitor gene transcription, protein turnover is likely involved in addition to repressing gene expression, and this possibility should be discussed.Moreover, if possible, the authors should use more endogenously-tagging reporters to verify the proposed gene repression in HSN neurons.Furthermore, since HSNs are hermaphrodite-specific neurons (the equivalent cells undergo programmed cell death in males), it is worth discussing whether apoptosis-related pathways play a role in gene repression.
7. P and G2/W states.This section is difficult to follow, and it is unclear what the authors intended to highlight.If I understand correctly, the authors identified P and G2/W as two subtypes of neuroblasts.They then conducted follow-up experiments and characterized the shared and distinct molecular regulation of the two states by several known regulators (ref-2, vab-15, and tlp-1).At first glance, the authors seem to highlight the molecular distinction of the subtypes.However, the authors concluded the findings as "… revealed a TF battery (ref-2, vab-15, and tlp-1) shared between the TF-based G2/W and P neuroblast subtypes.".In addition, it is unclear what hypothesis the authors are trying to form in this long sentence: "So we speculated that two neuroblast subtypes specified by similar pro-neural TF batteries would have smaller inter-subtype difference in TF expression normalized by their crosssister clade difference than that between subtypes specified by different TF batteries (Figure S5D)."This section should be significantly revised to make it clear what the hypothesis is and how each finding supports it.
8. Lienage-dependent heterogeneity.As mentioned in Major Comment 3, the long half-life of fluorescent proteins will make early expressed proteins persist in descendants through the cell lineage even though the endogenous transcripts or proteins no longer exist.As a result, the lineage effect may be overestimated since the propagation of protein is through mitotic divisions in each lineage.To address this concern, the authors should validate their findings by focusing on reporters that are expressed since the L1 stage, thereby avoiding this caveat.
Minor Comments: Line 83: "Moreover, a significant source of this intra-type molecular heterogeneity was convergence of different lineages into the same cell fate, i.e., multiple cell lineages could produce cells of the same phenotype and function."-A complete convergence should not produce heterogeneity.
Line 207-211: Is the integration site the only difference between these strains?Line 218: The correlation coefficient should be provided instead of being displayed as a color gradient.
Line 225: Figure 3E-G should be Figure 2E-G, also Figure 2H is not cited before Figure 3 Line 257: sister-clade cells, To avoid confusion with sister cells, I would suggest using a term like "neighbor-clade cells" or something similar Lin 1190-1192: "The mean expression profile across all image stacks of a strain was used as the strain expression profile.The mean expression profile across all strains of a TF was used as expression profile of the TF." -While it is reasonable to average expression across experimental replicates, it is uncertain whether it is valid to average expression between different reporters for the same transcription factor, especially if the expression levels differ significantly Figure 3A: Rather than organizing the hierarchical clustering into a circular-shaped plot, a classic horizontal view would be more intuitive and effective.Additionally, the current plot does not allow for easy visualization of which cells belong to the same clade Line 324: in the paragraph, Figure 6 should be Figure 5.
Line 370: "It is well-established that 97 of these 159 cell pairs are symmetric in cell lineage, while the other 62 pairs are converged cell pairs, whose left and right cell mates have an asymmetric cell lineage history."-To clarify the meaning of "symmetric" and "asymmetric," an example should be provided.The paper has great potential for three reasons: (1) the RAPCAT tool (2) the expression patterns serve as great hypothesis building tools for future functional analysis (3) the reporter transgenes serve as value cell fate tools This being said, there are substantial potential problems with the paper that the authors need to address before I can make a suggestion on suitability for publication in Nature Comm.
Problem #1: the cell ID tool RAPCAT has not been properly validated for whether it can indeed reliably identify cells, particularly in the densely packed brain of the worm.This is absolutely essential.The authors need to do such validation with reporters whose expression has been unambiguously and correctly identified.One set of reporters that comes to mind are the homeodomain TF reporter alleles published by Reilly et al. in 2020.Many dozens show very selective expression in the nervous system.These expression pattern have been identified in L4/young adults, but there are no differences in expression at the L1 stage.The author need to see whether RAPCAP can properly identify these patterns.The authors shall not limit this to a single reporter, but to a substantial number of reporters that in aggregrate cover most of the nervous system.
We agree that homeodomain TF reporter alleles published by Reilly et al. in 2020 could serve as reliable and well-defined positive controls to validate the annotation accuracy of RAPCAT since their neuronal expression patterns were determined through wellestablished markers.As recommended, we obtained 11 of these TF reporters from colleagues in the worm research community.Among the 222 neurons already present in L1 larvae, these reporters were expressed in 1 -108 neurons of young adults (Supplementary Figure 4E).We generated image stacks of the 11 reporter strains and annotated them using RAPCAT without the manual EPC step (i.e., the stacks were annotated automatically).One stack, capturing hmbx-1 expression, did not pass the 0.995 WAC threshold and was therefore excluded from the validation of automated RAPCAT annotation accuracy.For the remaining 10 reporters, RAPCAT annotation indicated that a TF was simultaneously expressed in at least 1 and as many as 80 neurons.Among the 222 total neurons, 140 showed detectable GFP expression by at least one reporter at the L1 stage in our 10-TF profile.Among the 140 GFP-expressing neurons in our results, 128 were also positive for expression of all corresponding reporters in young adults (Supplementary Figure 4E), suggesting 91.4% consistency between the automated neuron annotation by RAPCAT and marker-based identity assignment by NeuroPAL.
Transcriptomic changes have been previously reported in the post-mitotic nervous system during larval development (Sun and Hobert, 2021).It is thus reasonable to speculate that some discrepancies between RAPCAT annotation based on L1 expression profile and the adult profile reported by Reilly may be due to temporal changes in gene expression rather than error in RAPCAT annotation.We conducted a manual examination of the RAPCAT annotation outcomes for the 12 neurons that exhibited detectable GFP expression exclusively in L1 larvae, without the same expression observed in young adults.Based on our thorough inspection, it was determined that seven out of these 12 neurons had received accurate cell identity assignment through the RAPCAT annotation process.In other words, automatic RAPCAT annotation errors were present in only five neurons.These included a tail neuron PHB, a pharyngeal neuron I6, and three neurons (ASGL, AVDL, and AVHR) located around the nerve ring (Supplementary Figure 4E).Collectively, the automatic RAPCAT annotation method effectively and unambiguously identifies neurons in L1 larvae, achieving an accuracy ranging from 91.4% to 96.4%, encompassing even the densely packed neurons in the brain.
We have added these new data to Lines 165 -169 of the revised manuscript and Lines 51 -88 of the revised Supplementary Information.We appreciate this highly constructive suggestion.It is well-established that promoterfusion reporters may not always fully recapitulate the expression patterns of their corresponding genes because the necessary regulatory elements are located outside the upstream sequence used in the construct, such as those in introns (Bradnam et al., 2008;Fuxman Bass et al., 2014) or distal enhancers in long intergenic sequence (Okkema and Krause, 2005).Similarly, the transcription of downstream genes in an operon is usually controlled by a promoter outside of its direct upstream intergenic sequence (Okkema and Krause, 2005), and is consequently excluded from screens of the immediate upstream region, such as our promoter fusions.We therefore tested the extent of their influence by comparing the profiles of 84 TFs shared between our L1 profile and the adult profiles reported by Reilly and colleagues (2020).This set included 58 promoter-fusion reporters, 22 fosmid-based reporters, and 4 knock-in reporters (Supplementary Data 1), with fosmid and knock-in reporters serving as gold standards because they contained the full genomic context of the respective genes.
Target genes were then classified into four types, i.e., upstream intergenic sequence >7kb, upstream intergenic sequence <1kb, genes with >1kb intron following the start codon, and all other genes.The cloned promoter regions of genes categorized in the fourth group were presumed to contain the necessary and sufficient regulatory elements for driving their expression, termed as high context promoter reporters.
Substantiating this proposition, the L1 profiles of these high context promoter reporters displayed strong correlations (median of ROCAUC = 0.92) with corresponding adult homeobox TF profiles as reported by Reilly and colleagues.This correlation level resembled the relationship observed between the gold standard fosmid/knock-in reporters and the adult profiles (median of ROCAUC = 0.90) (Figure 2C).In our study, these closely correlated promoter reporters constituted a total of 296 TFs.In contrast, promoter fusions belonging to the first three categories, referred to as low context promoter reporters, displayed weak correlations with adult stage profiles by Reilly et al. (Figure 2C).These low-context promoter reporters encompassed 263 TFs (Supplementary Data 1).
These data have been added to lines 251 -277 of the revised manuscript.
If there are indeed substantial discrepancy between the expression patterns, the author will have to dampen a large number of their conclusions.
We are very grateful for the reviewer's advice.
As explained in the initial section of Question #2, we categorized the reporters of the 620 profiled TFs into three distinct types based on the extent to which their reporter constructs encompassed regulatory elements (Figure 2C).These categorized comprised 61 fosmid/knock-in reporters, 296 high-context promoter reporters, and 263 lowcontext promoter reporters (Supplementary Data 1).Upon conducting further analysis and comparing the results with the adult stage homeobox TF profiles reported by Reilly and colleagues, it became evident that both fosmid/knock-in and high-context promoter reporters effectively captured the full expression of their respective genes.In contrast, low-context promoter reporters demonstrated relatively weaker correlations with the published adult stage profiles.
We conducted an analysis to measure the alignment between traditional cell classifications and TF-driven cell clustering, aiming to quantify the impact of discrepancies between reporter expression and endogenous gene expression on our interpretations of biological function.It is noteworthy that all biological implications explored within this study were based on discrepancies between phenotype-based classification and molecular clustering.Our findings revealed that out of the total 119 multi-cellular phenotypic cell classes, fourteen classes were distributed across multiple clades in the dendrogram (Figure 3 and Supplementary Data 4).This distribution pointed to disparities between phenotype-based classification and molecular clustering, specifically relying on the profiles of all 620 TFs that were examined.Upon excluding the 263 low-context promoter reporters from our profiles, additional six cell classes exhibited discordance between phenotype-based classification and TF-based clustering (Supplementary Figure 6L).This suggests that low-context promoter reporters could potentially offer insights in the gene regulation program sensed by specific DNA sequences, even if they did not entirely encompass the expression patterns of their corresponding endogenous genes.
Notably, when we excluded reporters that were likely comprehensive in capturing full expression of their respective genes (specifically, all 61 fosmid/knock-in reporters and a random selection of 202 high-context promoter reporters), it is intriguing that the count of cell classes exhibiting discordance between phenotype-based classification and TF-based clustering remained comparable to that observed when profiles excluding the 263 low-context promoter reporters were considered (18 vs. 20) (Supplementary Figure 6L).Moreover, we found molecular clustering of the large set of 263 low-context reporters could recapitulate well-established phenotype-based cell type classifications to strikingly higher extent than a smaller set of 61 fosmid/knock-in reporters (Supplementary Figure 6K).This finding led us to conclude that the number of profiled genes affected TF-based cell type identification more strongly than full genomic context.While we have included our findings of low-and high-context reporter profiles in the revised manuscript, we respectfully maintain the high number of reporter profiles for reliable TF-based cell type annotation.
These data have been added to Lines 91 -99 and Lines 119 -147 of the revised Supplementary Information.

Two other comments:
-For the reporter data, the authors should not just give primer sequence, but coordinates relative to start codon of the gene so that it is easier to grasp how much genomic region is covered by the reporter.
Supplementary Data 1 has been updated to include the suggested information.
-one of the greatest values of this paper is the reporter resource.I expect that according to journal policy, these strains are deposited in a strain repository (CGC).
We completely agree.The reporter strains will indeed be available to the worm research community, and the set may be further expanded as the tool is refined.
We would like to thank Reviewer #1 for their time and careful reading of our paper, and for their highly constructive remarks, which have helped us to bolster the rigor of our analyses and ultimately strengthen our conclusions.As suggested, Figure 3 was split into Figures 3 and 4 in the revised manuscript, which has allowed us to enlarge the labels of cell clusters.Original Figure 4 was moved to Supplementary Figures 5A-5I and HSN expression is shown as dot plots in Figure 4I.
Most of the comments are on the computational method of template matching, which lacks crucial details: The presentation of the method in the main text is a bit unclear about how the full set of examples is used in scoring, this could and should be included in overview in the main text.
We are grateful for this constructive suggestion.How the full set of 100 training image stacks was used to build 100 templates is summarized in the original caption of Supplementary Figure 2.This content has been moved to the main text with expanded details.Specifically, one training worm was selected as the initial target and all straightened worm stacks were registered to the target using global affine transformation, resulting in a template L1 cell organization (Supplementary Figures 1D-1G).This process was repeated 100 times, with each training stack used once as the initial target for template generation, resulting in 100 digital templates of cell organization for L1 worms.
These data have been added to Lines 92 -97 of the revised manuscript.
The choice of primary alignment template for the atlas appears to have an, at first glance, unusually large influence on the quality of results.It would be helpful to give some intuitions why this is, and what characteristics the other 4 sub-optimal templates chosen have We appreciate this valuable question.We selected multiple templates instead of a single reference atlas because the distribution of cell nuclei does not have a strict stereotype (i.e., can vary among individual worms).For instance, variations in developmental program have been well-documented to result in slight variations in the exact location of cell nuclei (Sulston and Horvitz, 1977;Sulston et. al., 1983;Yemini et. al., 2021).Furthermore, cell identity can determine potential variability of developmental program for individual cells.For example, a ring of four nuclei is invariably found surrounding the lumen in the anterior-most region of the intestine (Sulston and Horvitz, 1977).By contrast, the arrangement of neurons anterior to the intestine shows relatively high variability among newly hatched L1 larvae (Sulston et. al., 1983).In addition, worms adopt different curvatures at the time of fixation for confocal imaging.As a result, their 3D image stacks require computational straightening so that they can be annotated by a canonical coordinate system.The straightening process could potentially deform relative cell locations (Peng et. al., 2008).In short, several factors, such as cell identityrelated differences in developmental program, and artifacts of the image straightening process for curved worm bodies can collectively contribute slight differences in cell positions among stereotype templates.
In practice, the 5-template-based annotation showed higher accuracy and more robust annotations compared to the optimal-template-based results in tests of 100 image stacks (Figure 1F), thus supporting our intuition that multiple templates can better capture inter-prototype variation in the positions of cell nuclei.
These data have been added to Lines 24 -44 in the revised Supplementary Information.We thank the reviewer for pointing out this unclear language.These templates were indeed selected based on their strong performance in annotating stacks that were 'worst case' examples for the optimal template.To avoid confusion, we have revised the caption of Supplementary Figure 2B as follows: These templates that provided the best performance in annotating worm stacks that could not be readily annotated by the optimal template (blue) were then adopted as the alternative templates.(See Supplementary Figure 2B caption, Lines 243 -246 in the revised Supplementary Information) Regardless of method what does the choice of 5 and templates represent?
Major modes of position?
As explained above, we hypothesized that computationally straightened image stacks could have slightly different stereotypical cell positions and that 5 templates could accommodate the range of variability in nuclei positions resulting from developmental differences or the fixation process.It is important to highlight that the disparity in mean and variance of cell positions among these five templates was exceedingly simple (Supplementary Figures 2E and S2F).Nonetheless, this subtle difference yielded distinct matching scores (Supplementary Figures 2G and 2H), underscoring the necessity for separate templates to effectively assign cell identities.
These data have been added to Lines 44 -49 in the revised Supplementary Information.
Could a different metric for co-alignment of the 100 examples (rather than least squares fit to one chosen example) create a single reference atlas with different/better properties.
We thank the reviewer for this thoughtful comment.Least-squares matching is a welldeveloped image registration method (Pattern recognition 32: 1783(Pattern recognition 32: , 1999;;International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences 38:86, 2010;Automatica 73: 155, 2016, andJournal of WSCG 25:21, 2017).Moreover, global affine transformation using least squares has been established as a reliable method for accurate identity assignment of trunk and tail cells in worm image stacks (Nature Methods 6:667, 2009).We therefore decided to build upon the least-squares matching method in the current study.
However, we agree with the reviewer's suggestion that a single, consensus reference atlas with minimal position bias towards a specific worm image stack could be valuable for cell annotation.We therefore followed a 3-step process in which each worm training stack was first aligned to the optimal template (i.e., template #2) by global affine transformation.Subsequently, we computed the average deformation field from these transformed stacks and inverted it.This inverted parameter was then utilized to deform the optimal template, resulting in a new template.Finally, step one was reiterated and the new template replaced the optimal template.This iterative process continued until the template converged into a stable template, which was designated as the consensus template.
Unexpectedly, the consensus template showed lower accuracy than the 5-template set How long does exhaustive curation take and how long does curating low confidence matches take?
For an image stack passing the 0.995 WAC threshold, EPC curation of low confidence matches typically requires about a quarter of an hour for a well-trained annotator.For image stacks that do not pass the WAC threshold, manual annotation of all 558 cells is necessary, often requiring more than two hours.
These data have been added to Lines 197 -199 of the revised manuscript.
We would like to thank the reviewer for his careful attention to detail and insightful questions that have given us much to consider and helped strengthen our conclusions.Using the atlas, the authors tackle an important yet unresolved question regarding the complex relationships between cell types and states.They grouped cells based on their TF expression (molecular state) and compared them to classic classifications based on the morphology and function of cells, reporting both consistencies and discrepancies between the two classifications.The authors described three cases to demonstrate the usefulness of the atlas in uncovering new cell states and state-type relationships.The first case describes a previously unrecognized quiescent state in newborn HSN neurons, characterized by gene repression.The second case concerns the state heterogeneity between P and G2/W neuroblasts, in which the authors characterized the regulatory differences between the otherwise similar cell types.Finally, the authors analyzed the incomplete convergence of cell states between cells originating from unrelated cell lineages but differentiating into identical/similar cell types, a phenomenon previously described during embryogenesis.Here, the authors verified and extended previous findings, showing that the lineage-dependent differences in cell states (TF expression) persist at the L1 stage when the terminal differentiation of many cells is completed.Furthermore, they focused on muscle cells and identified that earlier expression TFs (unc-39 and hnd-1) are responsible for lineage-dependent state heterogeneity through asymmetric functions in different lineages.
This study presents a valuable functional genomic resource for studying multicellular development.The over 900 reporter lines (including ~700 transgenic lines generated in this study) and associated cellular expression patterns will be valuable genomic resources for investigating TFs, cell fates, and cellular states.The experiments are generally well executed, and the results appear to be of high quality.The comparisons of cell states defined by TF expression to cell types are useful, and the case studies are interesting.
However, the presentation of the results in many sections is not well organized and presented, making some parts of the manuscript hard to follow and unclear the key message the authors want to deliver.Additionally, the quality assessments of the expression data are somewhat insufficient, and the authors tend to overlook the caveats of using fluorescent reports to indicate gene transcription, which should be explicitly stated and discussed.Finally, while analysis of cellular states using the TF atlas is definitely a good direction, the authors missed the rich opportunity of using this atlas to prioritize and characterize new regulators of cell types.Including this analysis could further enhance the impact of this study.To improve the manuscript, I suggest the following revisions.
Major comments: 1. RAPCAT methods.The authors provided a detailed description of the development of the RAPCAT algorithm and its ability to automatically assign cell identities based on stereotypic relative cell positions using training datasets.However, the accuracy of cell segmentation is equally important as well, which is not mentioned.Therefore, the authors should provide more information on how they dealt with the errors in nuclei segmentation, which inevitably affects the robustness of cell matching.
We are grateful for the Reviewer's supportive comments regarding our study and for their astute suggestions for improvement.We agree that correct segmentation is indeed essential for accurate cell annotation.Since the segmentation software was designed for the trunk and tail of L1 larvae (Long et. al. 2009), there are typically numerous segmentation errors in the densely packed brain region of the worm that require manual curation to resolve.Using VANO, a well-trained worm biologist may require two hours to identify and curate all segmentation errors.An image stack without segmentation errors has around 558 nuclear masks, depending on whether there are more or less than 20 intestinal nuclei.
These data have been added to Lines 808 -815 of the revised manuscript.
2. Quality assessment of expression atlas.To increase the utility of the atlas, it is critical to provide a thorough quality evaluation of the data to give readers a general idea of its quality.Although the authors show that the technical reproducibility of expression patterns is reasonably high (R> 0.81 between replicates), this result does not necessarily validate the correctness of the expression.Indeed, the correlation between different reporters of the same TF is not high (Figure 2B), raising concerns about to what extent the reporter expression recapitulates the endogenous expression.Since endogenous fluorescent fusion lines are available for many of the TFs analyzed in this study, the authors should perform a small-scale comparative analysis to determine the consistency.In addition, the authors should compare their expression pattern to benchmark genes whose expression and function are relatively welldocumented.Although the single-cell expression at the L1 stage is not available for most TFs, tissue-level analysis is feasible.
We thank the reviewer for this highly constructive suggestion, and we agree that the quality of our promoter-fusion reporter profiles should be evaluated by comparison with well-documented TFs.Previous work by Reilly and colleagues used fosmid reporters to generate expression profiles for 101 homeodomain TFs at single cell resolution in neurons of L4/young adults, several dozens of which showed highly selective expression in the nervous system (Reilly et al., 2020).We compared our L1 profile with the L4/young adult profile for all 84 homoeobox TFs shared by two datasets.This set included 58 promoter-fusion reporters, 22 fosmid-based reporters, and 4 knock-in reporters (Supplementary Data 1), with fosmid and knock-in reporters serving as gold standards because they contained the full genomic context of the respective genes.
Target genes were then classified into four types, i.e., upstream intergenic

Reviewer # 1 (
Remarks to the Author): Li et al. describe a tool, RAPCAT, to automatically identify cell position.The authors then go on to generate an atlas of expression of hundreds of C.elegans transcription factors.Analyzing these expression patterns, they describe a number of vignettes of potential biological significance, such as the identification of potentially quiescent stage of the HSN neurons.The authors also make a number of conclusions in regard to the relationship of TF expression and lineage.

Problem # 2 :
While the reporter collection established by the authors is potentially useful, it is a very significant shortcoming that the vast majority of reporters used are mere promoter fusions that very often do not capture the full expression of the respective genes.This is a fundamental problem in the interpretation of the expression patterns.The authors are in a quite unique position to address such discrepancies, by again making use of the homeobox reporter alleles mentioned above, and comparing their published expression patterns with the many homoeobox promoter gene fusions that they use in the paper.The comparison will give a sense of how accurate a reflection of real expression patterns these reporters are.Again, such comparison have to be done with proper sample sizes, e.g. the entire homeobox family down by Reilly et al., with the promoter-based expression patterns of homeobox genes by the present authors.One or two genes are not enough.Same random sampling of patterns that I did manually make me worry about quite substantial disparities, but this may be biased.

Reviewer # 2 (
Remarks to the Author): Li et al reports a TF expression atlas in C. elegans with single-cell resolution, covering 612 TFs whose reporter expression is mapped to each of the 558 cells in the newly hatched L1 larvae.It is an impressive body of work with unique values complementing the more commonly used scRNA atlases.The authors offered a systemic analysis of the expression patterns in terms of terminal differentiation, lineage history and body axes, and identified an interesting case of developmental quiescence.The study also established a computational approach to reliably identity individual cells based on nuclear positions, which is still difficult to do in C. elegans.The text is well written and easy to follow.However, the figures are hard to read.The labels are too small to read.On the other hand, some figures (eg, 4 and 5) show excessive images --it would be more effective to show the quantification with a representative image while moving the entire montage to a sup figure.This way it might also be possible to merge figures 4 and 5 into one, and split Fig 3 to two for the real estate it needs.

(
from fig S2A they appear to have perhaps been chosen for their unusually good -worst case-performance, though the caption suggests they are the runners up in fig S2b, this should be clarified a bit more).

Reviewer # 3 (
Remarks to the Author): In this herculean work, Li and co-authors have generated an extensive collection of promoter-driven fluorescent reporter lines of transcription factors (TFs) and presented an expression atlas of TF reporters that cover 65% of all predicted TFs in every lineage-resolved single cell of the L1 C. elegans larvae.Although scRNA-seq approaches have been previously used to determine high-resolution transcriptomes during C. elegans embryonic (e.g., Packer et al.Science 2019) and post-embryonic development (e.g., Cao et al.Science 2017 and Taylor et al. Cell 2021), only a small portion of the transcriptomes can be unequivocally linked to a specific cell.This uncertainty undermines the analysis of cell type-cell state relationships and state heterogeneity among similar/identical cell types.In contrast, this study takes advantage of the invariant cell lineage and stereotypic cell positioning in C. elegans development, using an imaging-based approach that allows for the quantification of the expression of a gene of interest in every annotated single cell and the synthesis of individual expression patterns into an atlas.The key advantage of this dataset is the clarity of cell annotation and complete cell coverage.Given the functional importance and expression informativeness of TFs, this dataset provides an opportunity to define the molecular/regulatory state of every cell at a developmental stage.Although Imaging-based singlecell analysis of TF expression at comparable resolution and scale has been conducted in C. elegans embryos (Murray et al.Genome Research 2012 and Ma et al.Nature Methods 2021), this study stands out as the first to generate a comprehensive TF expression atlas at the L1 stage, with a greater number of TFs and higher cell coverage.