Unravelling the progression of the zebrafish primary body axis with reconstructed spatiotemporal transcriptomics

  1. Yang Dong
  2. Tao Cheng
  3. Xiang Liu
  4. Xin-Xin Fu
  5. Yang Hu
  6. Xian-Fa Yang
  7. Ling-En Yang
  8. Hao-Ran Li
  9. Zhi-Wen Bian
  10. Naihe Jing
  11. Jie Liao
  12. Xiaohui Fan
  13. Peng-Fei Xu

Reviewed by

Read the full article See related articles

Discuss this preprint

Start a discussion What are Sciety discussions?

Listed in

Log in to save this article

Abstract

Elucidating the spatiotemporal dynamics of gene expression is essential for understanding complex physiological and pathological processes. Traditional technologies like in situ hybridization (ISH) and immunostaining have been restricted to analyzing expression patterns of a limited number of genes. Spatial transcriptomics (ST) has emerged as a robust alternative, enabling the investigation of spatial patterns of thousands of genes simultaneously. However, current ST methods are hindered by low read depths and limited gene detection capabilities. Here, we introduce Palette, a pipeline that infers detailed spatial gene expression patterns from bulk RNA-seq data, utilizing existing ST data as only reference. This method identifies more precise expression patterns by smoothing, imputing and adjusting gene expressions. We applied Palette to construct the D anio re rio S patio T emporal E xpression P rofiles ( Dre STEP) by integrating 53-slice serial bulk RNA-seq data from three developmental stages with existing ST references and 3D zebrafish embryo images. Dre STEP provides a comprehensive cartographic resource for examining gene expression and spatial cell-cell interactions within zebrafish embryos. Utilizing machine learning-based screening, we identified key morphogens and transcription factors (TFs) essential for anteroposterior (AP) axis development and characterized their dynamic distribution throughout embryogenesis. In addition, among these TFs, Hox family genes were found to be pivotal in AP axis refinement. Their expression was closely correlated with cellular AP identities, and hoxb genes may act as central regulators in this process.

Article activity feed

  1. Review Commons

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

    Learn more at Review Commons

    Reply to the reviewers

    Reviewer #1 (Evidence, reproducibility and clarity):

    Summary:

    The manuscript titled "Unravelling the Progression of the Zebrafish Primary Body Axis with Reconstructed Spatiotemporal Transcriptomics" presents a comprehensive analysis of the development of the primary body axis in zebrafish by integrating bulk RNA-seq, 3D images, and Stereo-Seq. The authors first clearly demonstrate the application of Palette for integrating RNA-seq and Stereo-Seq using published spatial transcriptomics data of Drosophila embryos. Subsequently, they produced serial bulk RNA-seq data for certain developmental stages of Danio rerio embryos and utilized published Stereo-Seq data. Through robust validation, the authors observe the molecular network involved in AP axis formation. While the authors show that integrating bulk RNA-seq data with Stereo-Seq improves spatial resolution, additional proof is required to demonstrate the extent of this improvement.

    Response: We thank the reviewer for the positive feedback on our Palette pipeline, zSTEP construction and analysis of primary body axis development. We appreciate the constructive suggestions provided, which we can implement to improve our manuscript. As pointed out by the reviewer, some analysis procedures were not described in sufficient detail. To address this, we have added more explanatory texts and additional schematic diagrams to make the methods clearer and more understandable. We also thank the reviewer for the meticulous reading and for reminding us to include parameters, references and essential texts, which significantly improve the manuscript quality and make the manuscript more rigorous. Furthermore, as suggested by the reviewer, the extent of the improvement on the spatial resolution was not clearly demonstrated in the manuscript. Therefore, we have provided an additional figure to show the original expression on the stacked Stereo-seq slices and 3D live image compared to the expression from zSTEP, and the results indicate that zSTEP provides better, more continuous expression patterns. We still have two remaining tasks that are expected to be completed within the next month. We hope our responses have address the concerns raised by the reviewer, and we are pleased to provide any additional proof as needed.

    Major Comments:

    1. Lines 66-68: Discuss the limitations of existing tools and explicitly state the advantages of using Palette.

    Response: We thank the reviewer for the valuable suggestion. We have added the following new texts after line 68 to emphasize the features and advantages of Palette.

    "Newly developed tools are committed to integrating bulk and/or scRNA-seq data with ST data to enhance spatial resolution, focusing on expression at the spot level. However, gene expression patterns are closely correlated to the biological functions and are more critical for understanding biological processes. Therefore, a tool focusing on inferring spatial gene expression patterns would be desirable."

    1. Body Pattern Genes Analysis: For both Drosophila and Danio rerio, it would be valuable to examine body pattern genes in Stereo-Seq and apply Palette to determine if the resolution of the segments improves or merges. The resolution of the A-P axis is convincing, but further evidence for other segments would be beneficial.

    Response: We thank the reviewer for the suggestions. For the Drosophila data, we only used two adjacent slices for Palette performance assessment, and thus were only able to evaluate the expression patterns within the slice.

    For the zebrafish data, although we have construct zSTEP as a 3D transcriptomic atlas, we have to admit that the left-right (LR) and dorsal-ventral (DV) patterning is not satisfactory enough. Here we show a section from the dorsal part of 16 hpf zSTEP that displays a relatively well-defined left-right pattern (Fig. 2). Along the left-right axis, the notochord cells are centrally located, flanked by somite cells on either side, with the outermost cells being pronephros.

    One reason for the limited LR and DV patterning is that the original annotation of the ST data does not clearly distinguish all the cell types. Another reason is likely due to the disordered cell positions when stacking ST slices. Thus, our zSTEP is most suitable for investigating the AP patterns, while the performances on LR and DV patterns may not achieve the same level of accuracy.

    See response letter for the figure.

    1. Figure 2d: Include the A-P line for which the intensity profile was plotted in the main figure, rather than just in the supplementary material. Additionally, consider simplifying the plot by not combining three lines into one, as it complicates the interpretation of observations.

    Response: We thank the reviewer for the helpful suggestions. We have updated Figure 2d and Figure S1b by adding a A-P line on each subfigure (Fig. 3). Additionally, as the reviewer suggested, we have separated the intensity plots so that each subfigure now includes a dedicated intensity plot along A-P axis.

    See response letter for the figure.

    1. Drosophila Data Analysis: While the alignment and validation of Danio rerio sections are clearly explained, the analysis and validation of Drosophila data are insufficiently detailed. Provide a more thorough explanation of how the intensity profiles between BDGP in situ data and Stereo-Seq data are adjusted.

    Response: We thank the reviewer for raising this issue. To make the analysis procedure clearer, we have updated Figure 2a (Fig. 4) and added explanatory texts in the figure legends to describe the processing procedure for the Drosophila ST data.

    See response letter for the figure.

    Additionally, the following sentences have been added into the Methods section to describe the generation of the intensity profiles.

    "The intensity plot profiles along AP axis were generated through the following steps: The expression pattern plot images or in situ hybridization images were imported into ImageJ and converted to grayscale. The colour was then inverted, and a line of a certain width (here set as 10) was drawn across from the anterior part to the posterior part (Fig. S1a). The signal intensities along the width of the line were measured and imported into R for generating intensity plots."

    1. Figure 3d: Present a plot with the expected expression profiles of the three genes if the embryo is aligned as anticipated.

    Response: We thank the reviewer for this helpful suggestion, which improves the clarity of our manuscript. We have added the following subfigure in as Figure 3d (Fig. 5) to show the expected expression profiles of the three midline genes along left-right axis.

    See response letter for the figure.

    1. Analysis Without Palette: Between lines 277-438, the outcome of using Palette with bulk RNA-seq and Stereo-Seq is convincing. However, consider the following:

    o What would be the observations if the analysis were conducted solely with Stereo-Seq data, without incorporating bulk RNA-seq data and employing Palette?

    Response: We thank the reviewer for raising this important question. Here we show the comparison of ST expression on stacked Stereo-seq slices, ST expression projected on 3D live images, and the Palette-inferred expression (Fig. 6). The stacked ST slices do not fully reflect the zebrafish morphology, and the gene expression appears sparse, making it look massive (the first row). While after projecting ST expression onto the live image, the expression patterns can be observed on zebrafish morphology, but the expression is still sparsely distributed in spots (the second row). However, the expression patterns captured by Palette in zSTEP show more continuous expression patterns (the third row), which are more similar to the observations in in situ hybridization images (the fourth row). We are considering put these analyses into the supplementary figure.

    See response letter for the figure.

    o This study uses only Stereo-Seq as the spatial transcriptomics reference. It would strengthen the argument to use at least one other spatial transcriptomics method, such as Visium or MERFISH, in conjunction with bulk RNA-seq and Palette, to demonstrate whether Palette consistently improves gene expression resolution.

    Response: We thank the reviewer for raising this professional question. To demonstrate a broad application of Palette, it would be necessary to test Palette performance using different types of ST references. We plan to perform extra analyses to evaluate Palette performance using Visium and MERFISH data as ST references, respectively. Additionally, our Palette pipeline only takes the overlapped genes for inference. As only hundreds of genes can be detected by MERFISH, Palette can only infer the expression patterns of these genes. As mentioned in the work of Liu et al. (2023), MERFISH can independently resolve distinct cell types and spatial structures, and thus we believe Palette will also show great performance when using MERFISH as ST reference. We've already started the analyses and expect to accomplish it within the next month. And we will update the analyses as separated tutorials to the GitHub repository.

    Reference:

    Liu, J. et al. Concordance of MERFISH spatial transcriptomics with bulk and single-cell RNA sequencing. Life Sci Alliance 6 (2023).

    1. PDAC Data Analysis: Provide a more detailed explanation of the PDAC data analysis and use appropriate colors in the tissue images to clearly distinguish cell types.

    Response: We thank the reviewer for the suggestions. We have updated the colours used in the tissue images to be consistent to the colours in tissue clustering analysis. Additionally, we have added an additional subfigure in supplementary figure (Fig. 7) with more explanatory texts in the figure legends to provide a more thorough explanation for the analysis.

    See response letter for the figure.

    1. Comparison with Other Methods: State the limitations of not using STitch3D and Spateo for alignment and explain why these methods were not employed.

    Response: We thank the reviewer for raising this constructive comment. We fully agree with you that the introduction of published alignment algorithms would be helpful in our analysis. Currently, the slice alignment is adjusted manually, and thus the main limitation of not using these tools is that manual operation may induce bias compared to the alignment generated by computational algorithm. Unfortunately, STitch3D and Spateo are not included in this study because of two reasons. First, these two newly developed tools have been recently posted, and our analyses were largely completed before that. Therefore, we only mentioned these tools in the Discussion section. Second, we do not want to embed too many external tools into our analysis, which may increase the difficulties for researchers' operation. Specifically, STitch3D and Spateo are configured to run in Python environment, while Palette is based on R packages. Moreover, without these tools, our current manual alignment also achieves desired performance. However, we value this enlightening suggestion by the reviewer and therefore plan to further compare the performance of manual alignment versus the mentioned two alignment tools. At present, we have a preliminary comparison scheme and collected relevant datasets. Hopefully, we will complete this analysis within the next 1 to 2 weeks.

    Minor Comments:

    1. References: Add references to the statements in lines 51-53.

    Response: We thank the reviewer for reminding us of the missing references. We have added the works of Junker et al. (2014), Liu et al. (2022), Chen et al. (2022), Wang et al. (2022), Shi et al. (2023) and Satija et al. (2015) as references in line 53 as follows.

    "Thus, great efforts are ongoing to construct gene expression maps of these models with higher resolution, depth, and comprehensiveness1-6."

    References:

    1. Junker, J.P. et al. Genome-wide RNA Tomography in the zebrafish embryo. Cell 159, 662-675 (2014).
    2. Liu, C. et al. Spatiotemporal mapping of gene expression landscapes and developmental trajectories during zebrafish embryogenesis. Dev Cell 57, 1284-1298 e1285 (2022).
    3. Chen, A. et al. Spatiotemporal transcriptomic atlas of mouse organogenesis using DNA nanoball-patterned arrays. Cell 185, 1777-1792 e1721 (2022).
    4. Wang, M. et al. High-resolution 3D spatiotemporal transcriptomic maps of developing Drosophila embryos and larvae. Dev Cell 57, 1271-1283 e1274 (2022).
    5. Shi, H. et al. Spatial atlas of the mouse central nervous system at molecular resolution. Nature 622, 552-561 (2023).
    6. Satija, R. et al. Spatial reconstruction of single-cell gene expression data. Nature biotechnology 33, 495-502 (2015)
    1. Scientific Name Consistency: Ensure consistency in using either "Danio rerio" or "zebrafish" throughout the manuscript.

    Response: We thank the reviewer for this suggestion. We have changed "Danio rerio" to "zebrafish" to make "zebrafish" consistent throughout the manuscript.

    1. Related References: Include the following relevant references:

    o https://academic.oup.com/bib/article/25/4/bbae316/7705532

    o https://www.life-science-alliance.org/content/6/1/e202201701

    Response: We thank the reviewer for bringing these two relevant works to us. Baul et al. (2024) presented STGAT leveraging Graph Attention Networks for integrating spatial transcriptomics and bulk RNA-seq, and Liu et al. (2023) demonstrated the concordance of MERFISH ST with bulk and single-cell RNA-seq. Both are excellent works and relevant to our work. We have added these two references in line 61 and line 68, respectively.

    References:

    Baul, S. et al. Integrating spatial transcriptomics and bulk RNA-seq: predicting gene expression with enhanced resolution through graph attention networks. Brief Bioinform 25 (2024).

    Liu, J. et al. Concordance of MERFISH spatial transcriptomics with bulk and single-cell RNA sequencing. Life Sci Alliance 6 (2023).

    1. Figure 1a: In the Venn diagram, include the number of genes in the bulk and Stereo-Seq datasets, as well as the number of overlapping genes.

    Response: We thank the reviewer reminding us to include these important numbers. And in our current manuscript, we have added the following sentences in the Methods section to provide the gene numbers (Fig. 8). While the Venn diagram in Figure 1a serves as a schematic representation, so we did not include the gene numbers, as these may vary depending on the actual data.

    "Palette was performed on the aligned slices using the overlapped genes. For the 10 hpf embryo, there were 24,658 genes in the bulk data, 18,698 genes in the Stereo-seq data, and 16,601 overlapped genes. For the 12 hpf embryo, there were 23,018 genes in the bulk data, 18,948 genes in the Stereo-seq data, and 16,401 overlapped genes. For the 16 hpf embryo, there were 24,357 genes in the bulk data, 23,110 genes in the Stereo-seq data, and 19,539 overlapped genes."

    See response letter for the figure.

    1. Figure 1 Improvement: Enlarge Figure 1 and reduce repetitive elements, such as parts of the deconvolution and Figure 1b.

    Response****: We thank the reviewer for the helpful suggestion. We agree with the reviewer that the deconvolution sections appear repetitive. We have updated Figure 1 (Fig. 9) by replacing these repetitive elements with a clearer and simpler diagram.

    See response letter for the figure.

    1. Figure 3f: Explain the black discontinuous line in the plot.

    Response: We thank the reviewer for the reminder. We are sorry about the lack of the explanation. We have added the below explanation for the black discontinuous line in the legend of Figure 3 (Fig. 10) as follows.

    See response letter for the figure.

    1. Line 610: State the percentage of unpaired imaging spots.

    Response: We thank the review for the reminder. We are sorry about not including the paired and unpaired spot number. We have added the number of paired spots with the percentage in the total spots in the Method section as follows.

    "The numbers of mapped spots for the 10 hpf, 12 hpf and 16 hpf embryos are 15,379 (69.4% of the total spots), 14,697 (70.5% of the total spots) and 21,605 (77.2% of the total spots), respectively."

    1. Lines 616-618: Specify the unit for the spot diameter.

    Response: We thank the reviewer for the reminder. Again, we are sorry about not including the spot diameter information in our previous version of manuscript. We have added the spot diameter in Method section as follows.

    "In the Stereo-seq data, each spot contained 15 × 15 DNA nanoball (DNB) spots (The diameter of each spot is near 10 μm)."

    Reviewer #1 (Significance):

    This algorithm will be useful not only for the field of developmental biology but also for wider applications in spatial omics. Although I have expertise in spatial omics technology development, my understanding of computational biology is limited, which restricts my ability to fully evaluate the Palette algorithm presented in this paper.

    Response: We thank the reviewer for recognizing our work, and we greatly appreciate the constructive suggestions from the reviewer. Although the reviewer acknowledged limited expertise in computational biology, the comments from the reviewer are highly professional and valuable. Following the suggestions from the reviewer, we have not only included more explanatory texts and figures to make the analysis procedures clearer and more understandable, but also supplemented the important parameters that were missing in our previous manuscript. We also provided extra figure to demonstrate the improvements of zSTEP on gene expression patterns. We believe that our work is now more scientific and more understandable, and we will continue working to solve the remaining issues as planned. We express our thanks for the reviewer again.

    Reviewer #2 (Evidence, reproducibility and clarity):

    The authors of the study introduce the Palette method, a novel approach designed to infer spatial gene expression patterns from bulk RNA-sequencing (RNA-seq) data. This method is complemented by the development of the DreSTEP 3D spatial gene expression atlas of zebrafish embryos, establishing a comprehensive resource for visualizing gene expression and investigating spatial cell-cell interactions in developmental biology.

    Response: We sincerely appreciate the reviewer's positive feedback on our Palette pipeline and the zSTEP 3D spatial expression atlas of zebrafish embryos. We also thank the reviewer for the professional comments and constructive suggestions. The reviewer raised the concerns from the aspect of algorithm design and computational biology, which we did not address well in our previous manuscript. We agree with the reviewer that we did not clarify the selection criteria of the parameters in detail, and we are now working on the additional analyses to address this issue.

    We also agree with the reviewer that we did not provide enough discussion of the strategies used in the pipeline, the features of Palette and the application scenarios of Palette and zSTEP. For wide use of our tools, it is significantly important to state these aspects. In this revised version, we have added more paragraphs in the Discussion section to address this issue. Additionally, we acknowledge that we did not adequately demonstrate the computational efficacy and computational requirements, which are important for researchers. We are also working on the additional analyses to address this issue.

    Finally, we thank the reviewer again for the professional and constructive suggestions. These suggestions are addressable, and by following them, we believe our manuscript will see a significant improvement, especially in the Palette pipeline part, making the pipeline more rigorous and easier to access. We are confident that we can complete the planned additional tasks within the next 1-2 months.

    1. The efficacy of the Palette method may be compromised by its dependency on the quality of the reference spatial transcriptomics data. As highlighted in the study, variations in data quality can lead to significant challenges in reconstructing accurate spatial expression patterns from bulk data. This underscores the necessity of evaluating quality parameters, such as the number of gene detections and spatial resolution, to ensure reliable outcomes. Additional studies should rigorously assess how these quality factors influence the accuracy and efficiency of the algorithm in various data contexts, particularly under diverse conditions of gene detection.

    Response: We thank the reviewer for this valuable suggestion. We agree with the reviewer that the quality of the reference ST data may greatly influence the performance and efficacy of the Palette, and we have added paragraphs in the Discussion section to further discuss the impact of ST data quality on Palette performance. As mentioned by the reviewer, gene detections and spatial resolution are two important parameters that can influence the Palette performance. Low gene detection may impact the clustering process, making the cell types of spots not distinguished well. To evaluate the performance of Palette when ST data shows low gene detection, we plan to applied Palette using MERFISH data as the ST reference, which only captures hundreds of genes. Moreover, we will also investigate the impact of spatial resolution on Palette performance by merging ST spots to simulate lower resolution scenarios, as well as the impact of gene detection by randomly reducing detected genes. Through the comparison among the inferred expression patterns with ST data of different spatial resolutions or different numbers of detected genes, we can better access the performance of Palette and provide guidance to researchers on the appropriate ST data requirements for optimal performance. These analyses will take another one month to accomplish after this round of revision due to the limited response time.

    1. The methodology raises pertinent questions regarding how the clustering results from different algorithms may affect the reconstructions by the Palette method. The authors would better provide a detailed discussion/comparison of clustering processes that optimize the reconstruction of spatial patterns, ensuring precision in the downstream analyses.

    Response: We thank the reviewer for the constructive comments. We agree with the reviewer that the differences in clustering results would impact the inference of the Palette. In our Palette pipeline, rather than develop a new methodology for clustering, we employ the BayesSpace for spot clustering, which considers both spot transcriptional similarity and neighbouring structure for clustering. In this case, researchers may adjust the parameters in the BayesSpace package to achieve optimal clustering results. Actually, in most cases, the spot identities were achieved through UMAP analysis, which only considers the transcriptional differences but does not consider the spatial information. This kind of clustering strategy will potentially lead to an intricate arrangement of spots belonging to different clusters, and may result in sparse gene expression in Palette outcome, which is different from the patterns in bona fide tissues. Therefore, a suitable clustering strategy will definitely help capture the local patterns.

    Moreover, our Palette pipeline also can use the clustering results from the tissue histomorphology. Using tissue histomorphology for clustering would be a good choice, as it is closer to the real case. The following Figure (Fig. 11) displays the Palette performance on PDAC datasets using both spatial clustering and histomorphology clustering strategies. The result using histomorphology clustering captures the weak pattern (indicated by the red circle) that were missed when using the spatial clustering (Fig. 11d).

    See response letter for the figure.

    1. The choice to utilize only highly expressed genes in the initial stages of the Palette algorithm also warrants further exploration. Addressing the criteria for determining which genes qualify as "highly expressed" and outlining robust cutoff will enhance the algorithm's rigor and applicability. Similarly, in the iterative estimation of gene expression across spatial spots, establishing optimal iteration conditions is crucial. Implementing a loss function may offer a systematic method for concluding iterations, thus refining computational efficiency.

    Response: We thank the reviewer for the professional suggestions. As pointed out by the reviewer, the selection of highly expressed genes and the iteration times are two important parameters in our pipeline. The definition of highly expressed genes and the number of highly expressed genes are important for achieving a satisfactory clustering performance. We tested the impact of different numbers of highly expressed genes on cluster performance in our preliminary analyses, while we did not summarize these tests and specify the parameters. Therefore, we plan to include a supplementary figure showing the clustering performances under different definitions of highly expressed genes and different numbers of highly expressed genes. Additionally, for the iteration conditions, we have tested different iteration numbers to find out a suitable iteration number to achieve a stable expression in each spot. The following figure (Fig. 1) shows the results after performing Palette with different iteration times. We randomly selected 20 cells and compared their expression across tests with varying iteration times. The results indicate that for a ST dataset with 819 spots, the expression in each spot becomes nearly stable after 5000 iteration times. We previously did not consider the computational efficiency, while here the reviewer raises a valuable and professional suggestion to implement a loss function to determine the optimal number of iterations. We greatly appreciate this suggestion, and plan to apply a loss function to summarize the optimal iteration times for ST datasets of different sizes. This will provide guidance for potential researchers in selecting iteration times and enhance computational efficiency.

    See response letter for the figure.

    1. Performance metrics relating to processing speed and computational demands remain inadequately addressed in the current framework. Understanding how the Palette method scales across varying gene counts and bulk RNA-seq datasets will be essential for potential applications in larger biological contexts. Notably, the quantitative demands of analyzing 20,000 genes when processing 10, 100, or 1,000 bulk RNA profiles must be articulated to guide researchers in planning accordingly.

    Response: We thank the reviewer for this valuable and professional suggestion. In our previous analyses, we did not consider the computation efficiency, processing speed and computational demands, which are important information for potential researchers. To address this issue, we will list our computer configuration first. And under this configuration, we plan to run Palette on datasets with different numbers of overlapped genes or ST references with varying spot numbers, and then summarize the running times into a metrics table. This will help researchers estimate the running time for their datasets and guide them in planning the analyses. We will begin the analyses soon and expect to complete the analysis within the next 1 to 2 months.

    Minor opinions:

    1. Despite the promising advances offered by the zebrafish 3D reconstruction, there is a lack of details regarding numbers of the spatial transcriptomics (ST) data utilized, and the number of bulk RNA-seq data employed in the analyses. These parameters need to be clarified.

    Response: We thank the reviewer for reminding us of these parameters. We are sorry for not including these parameters in our previous manuscript. We have now included the numbers of bulk, ST and overlap genes in the Methods section as follows (Fig. 12).

    "Palette was performed on the aligned slices using the overlapped genes. For the 10 hpf embryo, there were 24,658 genes in the bulk data, 18,698 genes in the Stereo-seq data, and 16,601 overlapped genes. For the 12 hpf embryo, there were 23,018 genes in the bulk data, 18,948 genes in the Stereo-seq data, and 16,401 overlapped genes. For the 16 hpf embryo, there were 24,357 genes in the bulk data, 23,110 genes in the Stereo-seq data, and 19,539 overlapped genes."

    See response letter for the figure.

    1. Issues regarding spatial cell-cell communication, especially concerning interactions over longer distances, necessitate careful consideration. Introducing spatial distance constraints could help formulate more realistic models of cellular interactions, a vital aspect of embryonic development.

    Response: We thank the reviewer for this essential comment. We agree with the reviewer that the spatial distance is an essential factor to investigate in vivo cell-cell communication during embryonic development. Therefore, in our analyses, we employed CellChat for spatial cell-cell communication analysis, which can be used to infer and visualize spatial cell-cell communication network for ST datasets, considering the spatial distance as constrains of the computed communication probability. However, during our analyses, we observed that there were interactions between cell types over longer distances, as mentioned by the reviewer. We then investigated how these interactions of longer distances occurred. Here, we show the FGF interaction between tail bud and neural crest cells from our spatial cell-cell analysis as an example, and the distance between these two cell types appears quite significant (Fig. 13). We labelled tail bud cells and neural crest cells on the selected midline section and observed that, although most neural crest cells are distributed anteriorly, a small number of neural crest cells are located at tail, close to the tail bud cells. Therefore, the observed interaction between tail bud and neural crest cells is likely due to their adjacent distribution in the tail region, while the anteriorly distributed of neural crest spot in spatial cell-cell communication analysis reflects the anterior positioning of most neural crest cells. As a result, the distances shown on the spatial cell-cell communication analysis are not the real distance between two cell types.

    In most cases in our spatial cell-cell communication analyses, the observed interactions over longer distances are likely influenced by this visualization strategy. Additionally, pre-processing the dataset may enhance the performance of the analyses. Here we performed systematic analyses of the entire embryo, which can make the interactions between cell types appear massive. To investigate specific biological questions, researchers can subset cell types of interest or categorize them into different subtypes based on their positions.

    See response letter for the figure.

    1. Evaluation metrics such as the Adjusted Rand Index (ARI) and Root Mean Square Error (RMSE) represent critical tools for systematically measuring the similarity of inferred spatial patterns, yet their specific application within this context should be elaborated.

    Response: We thank the reviewer for recommending these two tools. We have applied them to evaluate the similarity between the expression patterns (Fig. 14). The inclusion of these statistical values makes our comparisons of expression patterns more scientific and convincing. And we have added the following texts in the Methods section to describe the calculation of these two values.

    "The Adjusted Rand Index (ARI) and Root Mean Square Error (RMSE) were used to evaluate the similarity of the expression patterns. The expression patterns of in situ hybridization images were considered as the expected values, and the expression patterns of ST data and inferred expression patterns were compared to the expected values. Common positions along the AP axis within all three expression profiles were used, and the RMSE were calculated based on the scaled intensity of these positions. Values greater than the threshold were set to 1; otherwise, they were set to 0, and the ARI was then calculated based on the intensity category. Higher ARI and lower RMSE indicate greater similarity."

    See response letter for the figure.

    1. The study's limitations surrounding ST data quality cannot be overstated. Discussing scenarios where only limited or poor-quality ST data are available will be crucial for guiding future studies. Furthermore, a clear explanation of how enhanced specificity and accuracy translate into tangible biological insights is essential for demystifying the underlying mechanisms driving developmental processes.

    Response: We thank the reviewer for raising this essential suggestion. We have realized that in our previous manuscript, our discussion on the advantages and limitations of Palette and zSTEP was neither broad nor detailed enough.

    Therefore, in our revised manuscript, we have added the following paragraphs to further discuss the advantages and limitations of Palette and zSTEP, as well as the potential application of zSTEP in developmental biology.

    In this section, we have emphasized again the impact of ST data quality on the performance of Palette and zSTEP, and then compared Palette with the strategy that uses well-established marker genes to infer spatial information. We demonstrated that although Palette cannot achieve single cell resolution, it captures the major expression patterns, which are closely correlated to biological functions and critical for embryonic development. Furthermore, we further discussed that zSTEP is not only a valuable tool for investigating gene expression patterns, but also has the potential in evaluating the reaction-diffusion model to investigate the complicated and well-choreographed pattern formation during embryonic development.

    As here we have provided a more comprehensive discussion about Palette and zSTEP, we think that the potential researchers will better understand the application scenarios of our inference pipeline and our datasets. We hope our study can assist and inspire further research in the field of spatial transcriptomics and developmental biology.

    "Thirdly, the performance of Palette and zSTEP heavily relied on the quality of ST data. If the quality of ST data is not of sufficient quality, the low-expression genes may not be detected or only appear in very few scattered spots, and the performance of spot clustering could also be affected. Moreover, in this study, for example, the Stereo-seq data of 12 hpf zebrafish embryo had fewer slices on the right side (Fig. S3b), resulting in more blank spots in the right part of zSTEP for the 12 hpf embryo. However, with the ongoing advancements in spatial resolution and data quality, the performance of Palette is expected to be enhanced and demonstrate even greater potential for analysing spatiotemporal gene expression.

    On the other hand, compared to the brilliant strategy that infers spatial information of scRNA-seq data from well-established genes, our Palette pipeline cannot achieve single cell resolution. However, our Palette pipeline is based on the ST reference, and thus preserves the real positional relationships between spots. Furthermore, the focus of our pipeline is to infer the gene expression patterns, which are closely correlated to biological functions and critical for embryonic development, rather than the sparse expression within individual spots. In this regard, our Palette pipeline can be advantageous, as it allows for reconstruction of the major expression profiles, which are often more relevant for understanding developmental processes. Additionally, our Palette can be applied to serial sections, enabling the construction of 3D ST atlas.

    Finally, while the current analyses demonstrated that zSTEP can serve as a valuable tool for identifying genes having specific patterns at certain developmental stages, the exploration of zSTEP is still limited. During animal development, pattern formation is always one of the most important developmental issues. As demonstrated by the reaction-diffusion (RD) model, morphogen molecules are produced at specific regions of the embryo, forming morphogen gradients to guide cell specification, while interactions between different morphogens instruct more complicated and well-choreographed pattern formation. Our Palette constructed zSTEP, as a comprehensive transcriptomic expression pattern during development, could be leveraged to evaluate and prove the RD model during development, including AP patterning. Moreover, the investigation of gene expression patterns should not be limited to morphogens and TFs, and further investigation of their roles in AP patterning is desirable. Additionally, here a random forest model may be sufficient for investigating the most essential morphogens and TFs for AP axis refinement, while more sophisticated machine learning models may be required for addressing more specific biological questions."

    Reviewer #2 (Significance):

    The Palette pipeline demonstrates a marked improvement in specificity and accuracy when predicting spatial gene expression patterns. Evaluative studies on Drosophila and zebrafish datasets affirm its enhanced performance compared to existing methodologies. By effectively reconstructing spatial information from bulk transcriptomic data, the Palette method innovatively merges the philosophy of leveraging single-cell transcriptomic data for deconvolution analyses. This integration is pivotal, advancing traditional bulk RNA-seq approaches while laying the groundwork for future research.

    One of the notable achievements in this work is the construction of the DreSTEP atlas, which integrates serial bulk RNA-seq data with advanced 3D imaging techniques. This resource grants researchers unprecedented access to the visualization of gene expression patterns across the zebrafish embryo, facilitating the investigation of spatial relationships and cell-cell interactions critical for developmental processes. Such capabilities are invaluable for understanding the intricate dynamics of embryogenesis and the distinct roles of individual cell types.

    Response****: We thank the reviewer for the positive evaluation of our work, either the Palette pipeline or zSTEP. The reviewer has strong expertise in algorithm development and computational biology, and the concerns and suggestions from the reviewer are significantly precious and valuable for us. Regarding the bioinformatics tool development, we did not have extensive experiences, and thus we did not thoroughly address the selection criteria or clarify the parameters used in the pipeline, which may influence the application by other researchers. Therefore, we sincerely appreciate the professional suggestions from the reviewer, which we can follow to address these issues, improve our manuscript and make our work more impactful for researchers. Additionally, we did not consider computation efficiency, processing speed and computational demands, which would be important factors for other researchers to use Palette. We would like to add extra analyses to address these aspects.

    Currently, based on the suggestions from the reviewer, we have added extra texts discussing the clustering strategy in Palette pipeline, the advantages and limitations of Palette, and the potential application of zSTEP in developmental biology. We believe that readers will now have a clearer understanding of the performance of Palette and the application scenarios of both Palette and zSTEP. We have not fully addressed the comments raised by the reviewer yet, while we are working on the planned additional analyses and expect to complete all these tasks within the next 1-2 months. We sincerely thank the reviewer for the professional and valuable suggestions, which definitely improve our work and will make it accessible for a wide range of researchers.

    Finally, through this review process, we have learned a lot about the important considerations and requirements when designing bioinformatics tools, and we benefit a lot from the thoughtful guidance. We express our thanks to the reviewer again for the guidance, and we will try our best to address the remaining issues to further improve our manuscript.

    Reviewer #3 (Evidence, reproducibility and clarity):

    Evidence, reproducibility and clarity

    In this study, Dong and colleagues developed a computational pipeline to use spatial transcriptomics (ST) datasets as a reference to infer the spatial patterns of gene expression from bulk RNA sequencing data. This approach aims to overcome the low read depth and limited gene detection capabilities in current ST datasets, while exploiting its ability to provide highly resolved spatial information. By combining bulk RNA-seq datasets from 3 developmental stages during early zebrafish development with previously available ST and imaging datasets, the authors build DreSTEP (Danio rerio spatiotemporal expression profiles). Using this approach, they go on to identify the morphogens and transcription factors involved in anteroposterior patterning.

    The paper is well written, and the pipeline presented in this study is likely to be useful beyond the case studies included in this study. There are a few questions that, in my view, would be important to clarify to increase the impact of this work:

    Response: We sincerely appreciate the positive feedback from the reviewer on the Palette pipeline and zebrafish spatiotemporal expression profiles zSTEP. We thank the reviewer for the constructive suggestions, which have inspired us to think deeply about application and advantages of Palette and zSTEP for future studies.

    We fully agree with the reviewer that we do not sufficiently clarify the advantages and limitations of our inference pipeline in the original manuscript. The questions raised by the reviewer are very insightful. For example, while the inference expression patterns may closely resemble the in situ hybridization observation, which we consider as good performance, the reviewer pointed out that we should consider whether weak, yet real expression may have been removed. These questions have motivated us to think more deeply about the underlying principles and assumptions of our inference pipeline. Following the reviewer's questions, we have expanded our discussion on the application of zSTEP in developmental biology and the features of Palette compared to the existing strategies.

    We believe that after incorporating the revisions, our current manuscript now demonstrates the application scenario of Palette clearer and suggested the application of zSTEP for investigating biological questions in developmental biology. We are grateful for the reviewer's guidance, which helps us increase the impact of our work.

    1. The authors mention that they used a variable factor to adjust expression differences between the ST and bulk RNA-seq datasets. It would be important for the authors to comment on how much overlap in gene expression is necessary between the datasets for an accurate calculation of this variable factor? Can this be directly tested, for instance, by testing how their conclusions vary if expression is adjusted by a variable factor calculated from only a smaller set of genes?

    Response: We thank the reviewer for the professional questions. We are sorry about not including the gene numbers in our previous manuscript. And now we have provided the numbers of genes in bulk and ST data and the numbers of the overlapped genes (Fig. 15).

    "Palette was performed on the aligned slices using the overlapped genes. For the 10 hpf embryo, there were 24,658 genes in the bulk data, 18,698 genes in the Stereo-seq data, and 16,601 overlapped genes. For the 12 hpf embryo, there were 23,018 genes in the bulk data, 18,948 genes in the Stereo-seq data, and 16,401 overlapped genes. For the 16 hpf embryo, there were 24,357 genes in the bulk data, 23,110 genes in the Stereo-seq data, and 19,539 overlapped genes."

    See response letter for the figure.

    For Palette implementation, we took all the overlapped genes. To calculate the variable factor, we aggregated the expression of each gene in the ST data, and then used the expression of the bulk data to divide the aggregated expression for variable factor calculation. As a result, each overlapped gene was assigned a variable factor to adjust its expression, based on its difference between bulk and ST data. The rationale behind this approach is that by considering the ST data as a whole, we can effectively reduce the variations among individual spots. This allows the variable factors to provide reasonable adjustment to gene expression.

    Above all, the variable factors can be directly calculated. Currently Palette only can infer the expression patterns of overlapped genes. It means when the number of overlapped genes is small, such as MERFISH only detecting hundreds of genes, Palette can only infer the expression patterns of these genes. However, if the MERFISH data have good quality, which enable resolving distinct cell types, we believe Palette will also show good performance when using MERFISH as ST reference. Additionally, we plan to perform Palette using MERFISH as ST reference to further demonstrate its broad application when using different ST references.

    1. Palette gives rise to highly spatially precise patterns, which closely match those found in ISH. However, the smoothening of the expression can also remove weak, yet real, local expression patterns, as shown for idgf6 in Fig. 2a. Can the authors test this more extensively for other genes?

    Response: We thank the reviewer for this essential question. We agree with the reviewer that weak, yet real expression might be removed in our Palette inference pipeline. The weak, sparse expression may be due to the ST technique itself or the variations in samples. However, that sparse gene expression may not have biological meaning, and the focus of our pipeline in to capture the expression patterns, which are closely correlated with functions and crucial for embryonic development. Therefore, our algorithm considers spot characteristics and emphasize cluster-specific expression, resulting in spatial-specific expression patterns. In most cases, the main gene expression patterns can be captured, which can help understand gene functions and roles in embryonic development. We have updated Supplementary Figure S1a (Fig. 16) to include more gene patterns to demonstrate this point.

    See response letter for the figure.

    1. Using adjacent slices for ST and "bulk RNA-seq" may provide better results than those obtained when comparing two independent datasets. Could the authors also extend the analysis of Palette's functionalities by using separate, previously available but independent datasets, for ST and bulk RNA-seq in Drosophila as well?

    Response: We thank the reviewer for the valuable question. We agree with the reviewer that using adjacent slices may provide better results. The idea here is that the inferred spatial expression patterns from pseudo bulk RNA-seq can be used to compare with the real expression of ST to evaluate Palette performance. We have updated our Figure 2a (Fig. 17) to illustrate the analysis clearer.

    See response letter for the figure.

    To demonstrate the Palette's functionalities, we have used Palette to infer zebrafish bulk RNA-seq slice (Junker et al., 2014) using Stereo-seq slice (Liu et al., 2022) as ST reference, and these two datasets are separate and independent. We agree with the reviewer that it would be good to use separate datasets to test in Drosophila to further demonstrate the Palette's functionalities. However, unfortunately, we did not find the Drosophila serial bulk RNA-seq data along left-right axis of the corresponding stages, and thus we might be unable to perform the extra analyses using independent Drosophila datasets.

    References:

    Junker, J.P. et al. Genome-wide RNA Tomography in the zebrafish embryo. Cell 159, 662-675 (2014).

    Liu, C. et al. Spatiotemporal mapping of gene expression landscapes and developmental trajectories during zebrafish embryogenesis. Dev Cell 57, 1284-1298 e1285 (2022).

    1. The DreSTEP analysis in zebrafish embryos is interesting and validates well-established observations in the field. Can the authors also discuss whether and how their dataset allows them to refine our understanding of the spatial or temporal pattern of the morphogens and TFs involved in AP patterning? This would further validate their approach.

    Response: We appreciate the reviewer for recognition of our zSTEP and raising this valuable question, which has inspired us to think more deeply about the potential application of zSTEP in developmental biology. As the reviewer noted, our zSTEP analyses have validated well-established observations in the field. Rather than focusing on the sparse expression detected in ST data, zSTEP emphasizes the gene expression patterns that are closely correlated with biological functions and critical for embryonic development. Therefore, zSTEP can serve as a valuable tool for identifying the genes having specific patterns at certain developmental stages.

    Pattern formation is one of the most important developmental issues for all animals. The reaction-diffusion (RD) model is a widely recognized theoretical framework used to explain self-regulated pattern formation in developing animal embryos (Kondo & Miura, 2010). Morphogen molecules are produced at specific regions of the embryo, forming morphogen gradients to guide cell specification. Most importantly, interactions between different morphogens instruct more complicated and well-choreographed pattern formation. Our Palette-constructed zSTEP provides a comprehensive transcriptomic expression pattern, including all morphogens and TFs, across the whole embryo during development. These valuable resources, in our opinion, could be leveraged to evaluate and prove the RD model during development, including AP patterning. In our current zSTEP analyses, we have already identified genes that exhibit specific expression patterns along AP axis, some of which have not been fully characterized. These genes could be potential targets for further investigation into their roles in AP patterning, although they are not the primary focus of this study. Additionally, our analyses only focused on morphogens and TFs, but zSTEP can be used to investigate the expression patterns of other genes as well. Moreover, we employed a random forest model to investigate the most essential morphogens and TFs for AP axis refinement, which is one of the basic applications of zSTEP. To investigate specific biological questions of interest, it would be worth exploring the use of more sophisticated machine learning models.

    We have added the following paragraph in the Discussion section to discuss the potential application of zSTEP in future studies.

    "Finally, while the current analyses demonstrated that zSTEP can serve as a valuable tool for identifying genes having specific patterns at certain developmental stages, the exploration of zSTEP is still limited. During animal development, pattern formation is always one of the most important developmental issues. As demonstrated by the reaction-diffusion (RD) model, morphogen molecules are produced at specific regions of the embryo, forming morphogen gradients to guide cell specification, while interactions between different morphogens instruct more complicated and well-choreographed pattern formation. Our Palette constructed zSTEP, as a comprehensive transcriptomic expression pattern during development, could be leveraged to evaluate and prove the RD model during development, including AP patterning. Moreover, the investigation of gene expression patterns should not be limited to morphogens and TFs, and further investigation of their roles in AP patterning is desirable. Additionally, here a random forest model may be sufficient for investigating the most essential morphogens and TFs for AP axis refinement, while more sophisticated machine learning models may be required for addressing more specific biological questions."

    Reference

    Kondo, S. & Miura, T. Reaction-Diffusion model as a framework for understanding biological pattern formation. Science 329, 1616-1620 (2010).

    1. Can the authors comment on the limits of this inference pipeline? And how it performs as compared to single-cell RNA sequencing datasets where spatial information is inferred from well-established marker genes?

    Response: We appreciate the reviewer for this insightful question, which has inspired us to further explore the advantages and limitations of the Palette pipeline in comparison with other inference strategies. As mentioned in the Discussion section, a key limitation of the inference pipeline is its heavy reliance on the quality of ST data. It is obvious that if the quality of ST data is not of sufficient quality, the low-expression genes may not be detected or only appear in very few scattered spots. We think it is a common issue for any inference tools using ST data as the reference. However, with the ongoing advancements in spatial resolution and data quality, the performance of Palette is expected to be improved.

    As a comparison, the single-cell RNA sequencing datasets where spatial information is inferred from well-established marker genes do not face this limitation. The ground-breaking work by Satija et al. (2015) used such a strategy that combined scRNA-seq and in situ hybridizations of well-established marker genes to infer spatial location, enabling single cell resolution, as it maintains the high read depth and gene detection. One advantages of this scRNA-seq-based strategy is that it provides the transcriptomics of individual cells, rather than a combination of cell within a ST spot, although the positional relationships between cells are not real.

    However, compared to the inference from ST data, the positional relationships between cells are not directly captured. On the other hand, as the embryonic development progresses, more cell types will be specified, and the body patterning becomes more complex. In this scenario, using well-established marker gene to infer spatial information would be much more challenging. Additionally, there are not many scRNA-seq datasets of serial sections, and thus this strategy may not be used to construct 3D ST atlas.

    In contrast, our Palette inference pipeline is based on the ST data, which preserves the real positional relationships between spots. Although our inference pipeline cannot achieve single cell resolution, it focuses on the gene expression patterns rather than the sparse expression within individual spots. By applying Palette to paired serial sections, we were able to generated a 3D spatial expression atlas of zebrafish embryos, which has showed promising performance for investigating gene expression patterns and their involvement in AP patterning.

    Reference

    Satija, R. et al. Spatial reconstruction of single-cell gene expression data. Nature biotechnology 33, 495-502 (2015)

    We have updated the following paragraphs to further demonstrating the limitation of the inference pipeline in details in the Discussion section.

    "Thirdly, the performance of Palette and zSTEP heavily relied on the quality of ST data. If the quality of ST data is not of sufficient quality, the low-expression genes may not be detected or only appear in very few scattered spots, and the performance of spot clustering could also be affected. Moreover, in this study, for example, the Stereo-seq data of 12 hpf zebrafish embryo had fewer slices on the right side (Fig. S3b), resulting in more blank spots in the right part of zSTEP for the 12 hpf embryo. However, with the ongoing advancements in spatial resolution and data quality, the performance of Palette is expected to be enhanced and demonstrate even greater potential for analysing spatiotemporal gene expression.

    On the other hand, compared to the brilliant strategy that infers spatial information of scRNA-seq data from well-established genes, our Palette pipeline cannot achieve single cell resolution. However, our Palette pipeline is based on the ST reference, and thus preserves the real positional relationships between spots. Furthermore, the focus of our pipeline is to infer the gene expression patterns, which are closely correlated to biological functions and critical for embryonic development, rather than the sparse expression within individual spots. In this regard, our Palette pipeline can be advantageous, as it allows for reconstruction of the major expression profiles, which are often more relevant for understanding developmental processes. Additionally, our Palette can be applied to serial sections, enabling the construction of 3D ST atlas."

    Reviewer #3 (Significance):

    This study tackles an important challenge in biology - the difficult to resolve gene expression patterns with high spatial precision and in a high-throughput manner. By integrating sequencing datasets from previously published studies, as well as newly-generated datasets, the authors provide evidence that their novel inference pipeline enables them to obtain high-quality spatial information simply from bulk RNA-seq datasets, using ST as a reference. The development of this pipeline - Palette - is a major part of this manuscript and its applicability is validated using datasets from Drosophila and zebrafish embryos. This in an important advance for the field, but it would be nice for the authors to further comment on i) the validity of some of their approaches and how they may influence the quality of their inference, as well as, ii) potential pitfalls/limitations of this approach as compared to others available in the field. This would synthetize both previous and current findings into a conceptual and technological framework that would have a strong impact well beyond cell and developmental biology.

    Audience: This study would be relevant for a broad audience of biologists, interested in morphogen signaling, gene regulatory networks and cell fate specification.

    Expertise in zebrafish development, gastrulation, morphogen signaling and morphogenesis.

    Response: We thank the reviewer for providing the positive feedback, arising these valuable questions, which have motivated us to deeply consider the design concept and further application of Palette and zSTEP. Based on the insightful questions from the reviewer, we have added two extra paragraphs in the Discussion section to further discuss the potential application of zSTEP in developmental biology and application scenarios of the Palette pipeline. Specially, we have demonstrated that the performance of the inference pipeline relies on the spatial resolution and data quality of the ST data. We have then compared the advantages and limitations of Palette with the existing brilliant spatial inference strategy, which infers spatial information of scRNA-seq from well-established marker genes. Although our inference pipeline cannot achieve single cell resolution, it can capture the major expression patterns, which are closely correlated to functions and critical for embryonic development. We believe this will help readers gain a clearer understanding of the advantage and limitations of our pipeline compared to other tools, as well as the tasks for which Palette and our constructed zSTEP can be utilized. We express our thanks to the reviewer again for the valuable comments.

  2. Review Commons

    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

    In this study, Dong and colleagues developed a computational pipeline to use spatial transcriptomics (ST) datasets as a reference to infer the spatial patterns of gene expression from bulk RNA sequencing data. This approach aims to overcome the low read depth and limited gene detection capabilities in current ST datasets, while exploiting its ability to provide highly resolved spatial information. By combining bulk RNAseq datasets from 3 developmental stages during early zebrafish development with previously available ST and imaging datasets, the authors build DreSTEP (Danio rerio spatiotemporal expression profiles). Using this approach, they go on to identify the morphogens and transcription factors involved in anteroposterior patterning.

    The paper is well written, and the pipeline presented in this study is likely to be useful beyond the case studies included in this study. There are a few questions that, in my view, would be important to clarify to increase the impact of this work:

    1. The authors mention that they used a variable factor to adjust expression differences between the ST and bulk RNAseq datasets. It would be important for the authors to comment on how much overlap in gene expression is necessary between the datasets for an accurate calculation of this variable factor? Can this be directly tested, for instance, by testing how their conclusions vary if expression is adjusted by a variable factor calculated from only a smaller set of genes?
    2. Palette gives rise to highly spatially precise patterns, which closely match those found in ISH. However, the smoothening of the expression can also remove weak, yet real, local expression patterns, as shown for idgf6 in Fig. 2a. Can the authors test this more extensively for other genes?
    3. Using adjacent slices for ST and "bulk RNAseq" may provide better results than those obtained when comparing two independent datasets. Could the authors also extend the analysis of Palette's functionalities by using separate, previously available but independent datasets, for ST and bulk RNAseq in Drosophila as well?
    4. The DreSTEP analysis in zebrafish embryos is interesting and validates well-established observations in the field. Can the authors also discuss whether and how their dataset allows them to refine our understanding of the spatial or temporal pattern of the morphogens and TFs involved in AP patterning? This would further validate their approach.
    5. Can the authors comment on the limits of this inference pipeline? And how it performs as compared to single-cell RNA sequencing datasets where spatial information is inferred from well-established marker genes?

    Significance

    This study tackles an important challenge in biology - the difficult to resolve gene expression patterns with high spatial precision and in a high-throughput manner. By integrating sequencing datasets from previously published studies, as well as newly-generated datasets, the authors provide evidence that their novel inference pipeline enables them to obtain high-quality spatial information simply from bulk RNAseq datasets, using ST as a reference. The development of this pipeline - Palette - is a major part of this manuscript and its applicability is validated using datasets from Drosophila and zebrafish embryos. This in an important advance for the field, but it would be nice for the authors to further comment on i) the validity of some of their approaches and how they may influence the quality of their inference, as well as, ii) potential pitfalls/limitations of this approach as compared to others available in the field. This would synthetize both previous and current findings into a conceptual and technological framework that would have a strong impact well beyond cell and developmental biology.

    Audience: This study would be relevant for a broad audience of biologists, interested in morphogen signaling, gene regulatory networks and cell fate specification.

    Expertise in zebrafish development, gastrulation, morphogen signaling and morphogenesis.

  3. Review Commons

    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

    The authors of the study introduce the Palette method, a novel approach designed to infer spatial gene expression patterns from bulk RNA-sequencing (RNA-seq) data. This method is complemented by the development of the DreSTEP 3D spatial gene expression atlas of zebrafish embryos, establishing a comprehensive resource for visualizing gene expression and investigating spatial cell-cell interactions in developmental biology.

    Major concerns:

    1. The efficacy of the Palette method may be compromised by its dependency on the quality of the reference spatial transcriptomics data. As highlighted in the study, variations in data quality can lead to significant challenges in reconstructing accurate spatial expression patterns from bulk data. This underscores the necessity of evaluating quality parameters, such as the number of gene detections and spatial resolution, to ensure reliable outcomes. Additional studies should rigorously assess how these quality factors influence the accuracy and efficiency of the algorithm in various data contexts, particularly under diverse conditions of gene detection.
    2. The methodology raises pertinent questions regarding how the clustering results from different algorithms may affect the reconstructions by the Palette method. The authors would better provide a detailed discussion/comparison of clustering processes that optimize the reconstruction of spatial patterns, ensuring precision in the downstream analyses.
    3. The choice to utilize only highly expressed genes in the initial stages of the Palette algorithm also warrants further exploration. Addressing the criteria for determining which genes qualify as "highly expressed" and outlining robust cutoff will enhance the algorithm's rigor and applicability. Similarly, in the iterative estimation of gene expression across spatial spots, establishing optimal iteration conditions is crucial. Implementing a loss function may offer a systematic method for concluding iterations, thus refining computational efficiency.
    4. Performance metrics relating to processing speed and computational demands remain inadequately addressed in the current framework. Understanding how the Palette method scales across varying gene counts and bulk RNA-seq datasets will be essential for potential applications in larger biological contexts. Notably, the quantitative demands of analyzing 20,000 genes when processing 10, 100, or 1,000 bulk RNA profiles must be articulated to guide researchers in planning accordingly.

    Minor opinions:

    1. Despite the promising advances offered by the zebrafish 3D reconstruction, there is a lack of details regarding numbers of the spatial transcriptomics (ST) data utilized, and the number of bulk RNA-seq data employed in the analyses. These parameters need to be clarified.
    2. Issues regarding spatial cell-cell communication, especially concerning interactions over longer distances, necessitate careful consideration. Introducing spatial distance constraints could help formulate more realistic models of cellular interactions, a vital aspect of embryonic development.
    3. Evaluation metrics such as the Adjusted Rand Index (ARI) and Root Mean Square Error (RMSE) represent critical tools for systematically measuring the similarity of inferred spatial patterns, yet their specific application within this context should be elaborated.
    4. The study's limitations surrounding ST data quality cannot be overstated. Discussing scenarios where only limited or poor-quality ST data are available will be crucial for guiding future studies. Furthermore, a clear explanation of how enhanced specificity and accuracy translate into tangible biological insights is essential for demystifying the underlying mechanisms driving developmental processes.

    Significance

    The Palette pipeline demonstrates a marked improvement in specificity and accuracy when predicting spatial gene expression patterns. Evaluative studies on Drosophila and zebrafish datasets affirm its enhanced performance compared to existing methodologies. By effectively reconstructing spatial information from bulk transcriptomic data, the Palette method innovatively merges the philosophy of leveraging single-cell transcriptomic data for deconvolution analyses. This integration is pivotal, advancing traditional bulk RNA-seq approaches while laying the groundwork for future research.

    One of the notable achievements in this work is the construction of the DreSTEP atlas, which integrates serial bulk RNA-seq data with advanced 3D imaging techniques. This resource grants researchers unprecedented access to the visualization of gene expression patterns across the zebrafish embryo, facilitating the investigation of spatial relationships and cell-cell interactions critical for developmental processes. Such capabilities are invaluable for understanding the intricate dynamics of embryogenesis and the distinct roles of individual cell types.

  4. Review Commons

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

    Learn more at Review Commons

    Referee #1

    Evidence, reproducibility and clarity

    Summary:

    The manuscript titled "Unravelling the Progression of the Zebrafish Primary Body Axis with Reconstructed Spatiotemporal Transcriptomics" presents a comprehensive analysis of the development of the primary body axis in zebrafish by integrating bulk RNA-seq, 3D images, and Stereo-Seq. The authors first clearly demonstrate the application of Palette for integrating RNA-seq and Stereo-Seq using published spatial transcriptomics data of Drosophila embryos. Subsequently, they produced serial bulk RNA-seq data for certain developmental stages of Danio rerio embryos and utilized published Stereo-Seq data. Through robust validation, the authors observe the molecular network involved in AP axis formation. While the authors show that integrating bulk RNA-seq data with Stereo-Seq improves spatial resolution, additional proof is required to demonstrate the extent of this improvement.

    Major Comments:

    1. Lines 66-68: Discuss the limitations of existing tools and explicitly state the advantages of using Palette.
    2. Body Pattern Genes Analysis: For both Drosophila and Danio rerio, it would be valuable to examine body pattern genes in Stereo-Seq and apply Palette to determine if the resolution of the segments improves or merges. The resolution of the A-P axis is convincing, but further evidence for other segments would be beneficial.
    3. Figure 2d: Include the A-P line for which the intensity profile was plotted in the main figure, rather than just in the supplementary material. Additionally, consider simplifying the plot by not combining three lines into one, as it complicates the interpretation of observations.
    4. Drosophila Data Analysis: While the alignment and validation of Danio rerio sections are clearly explained, the analysis and validation of Drosophila data are insufficiently detailed. Provide a more thorough explanation of how the intensity profiles between BDGP in situ data and Stereo-Seq data are adjusted.
    5. Figure 3d: Present a plot with the expected expression profiles of the three genes if the embryo is aligned as anticipated.
    6. Analysis Without Palette: Between lines 277-438, the outcome of using Palette with bulk RNA-seq and Stereo-Seq is convincing. However, consider the following:
      o What would be the observations if the analysis were conducted solely with Stereo-Seq data, without incorporating bulk RNA-seq data and employing Palette?
      o This study uses only Stereo-Seq as the spatial transcriptomics reference. It would strengthen the argument to use at least one other spatial transcriptomics method, such as Visium or MERFISH, in conjunction with bulk RNA-seq and Palette, to demonstrate whether Palette consistently improves gene expression resolution.
    7. PDAC Data Analysis: Provide a more detailed explanation of the PDAC data analysis and use appropriate colors in the tissue images to clearly distinguish cell types.
    8. Comparison with Other Methods: State the limitations of not using STitch3D and Spateo for alignment and explain why these methods were not employed.

    Minor Comments:

    1. References: Add references to the statements in lines 51-53.
    2. Scientific Name Consistency: Ensure consistency in using either "Danio rerio" or "zebrafish" throughout the manuscript.
    3. Related References: Include the following relevant references:
    1. Figure 1a: In the Venn diagram, include the number of genes in the bulk and Stereo-Seq datasets, as well as the number of overlapping genes.
    2. Figure 1 Improvement: Enlarge Figure 1 and reduce repetitive elements, such as parts of the deconvolution and Figure 1b.
    3. Figure 3f: Explain the black discontinuous line in the plot.
    4. Line 610: State the percentage of unpaired imaging spots.
    5. Lines 616-618: Specify the unit for the spot diameter.

    Significance

    This algorithm will be useful not only for the field of developmental biology but also for wider applications in spatial omics. Although I have expertise in spatial omics technology development, my understanding of computational biology is limited, which restricts my ability to fully evaluate the Palette algorithm presented in this paper.

  5. Version published to 10.1101/2024.07.01.601472 on bioRxiv