Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

  • Brief Communication
  • Published:

scDesign3 generates realistic in silico data for multimodal single-cell and spatial omics

Nature Biotechnology volume 42, pages 247–252 (2024)Cite this article

Subjects

Abstract

We present a statistical simulator, scDesign3, to generate realistic single-cell and spatial omics data, including various cell states, experimental designs and feature modalities, by learning interpretable parameters from real data. Using a unified probabilistic model for single-cell and spatial omics data, scDesign3 infers biologically meaningful parameters; assesses the goodness-of-fit of inferred cell clusters, trajectories and spatial locations; and generates in silico negative and positive controls for benchmarking computational tools.

This is a preview of subscription content, access via your institution

Access options

Buy this article

  • Purchase on Springer Link
  • Instant access to full article PDF
Buy now

Prices may be subject to local taxes which are calculated during checkout

Fig. 1: scDesign3 generates realistic synthetic data of diverse single-cell and spatial omics technologies.
Fig. 2: scDesign3 enables comprehensive interpretation of real data.

Similar content being viewed by others

Data availability

All datasets used in the study are publicly available. Supplementary Table 2 lists the datasets from 17 published studies (sources included). The preprocessed datasets are available in the Zenodo repository at https://doi.org/10.5281/zenodo.711076152.

Code availability

The scDesign3 package is available at https://github.com/SONGDONGYUAN1994/scDesign3. The comprehensive tutorials are available at https://songdongyuan1994.github.io/scDesign3/docs/index.html. In the tutorials, we described the input and output formats, model parameters and exemplary datasets for each functionality of scDesign3. The source code for reproducing the results is available in the Zenodo repository at https://doi.org/10.5281/zenodo.711076152.

References

  1. Tang, F. et al. mRNA-seq whole-transcriptome analysis of a single cell. Nat. Methods 6, 377–382 (2009).

    Article  CAS  PubMed  Google Scholar 

  2. Cao, J. et al. The single-cell transcriptional landscape of mammalian organogenesis. Nature 566, 496–502 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  3. Buenrostro, J. D. et al. Single-cell chromatin accessibility reveals principles of regulatory variation. Nature 523, 486–490 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  4. Cusanovich, D. A. et al. Multiplex single-cell profiling of chromatin accessibility by combinatorial cellular indexing. Science 348, 910–914 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  5. Karemaker, I. D. & Vermeulen, M. Single-cell DNA methylation profiling: technologies and biological applications. Trends Biotechnol. 36, 952–965 (2018).

    Article  CAS  PubMed  Google Scholar 

  6. Bendall, S. C. et al. Single-cell mass cytometry of differential immune and drug responses across a human hematopoietic continuum. Science 332, 687–696 (2011).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  7. Chen, S., Lake, B. B. & Zhang, K. High-throughput sequencing of the transcriptome and chromatin accessibility in the same cell. Nat. Biotechnol. 37, 1452–1457 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  8. Stoeckius, M. et al. Simultaneous epitope and transcriptome measurement in single cells. Nat. Methods 14, 865–868 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  9. Rao, N., Clark, S. & Habern, O. Bridging genomics and tissue pathology: 10x genomics explores new frontiers with the visium spatial gene expression solution. Genet. Eng. Biotechnol. News 40, 50–51 (2020).

    Article  Google Scholar 

  10. Rodriques, S. G. et al. Slide-seq: a scalable technology for measuring genome-wide expression at high spatial resolution. Science 363, 1463–1467 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  11. Stickels, R. R. et al. Highly sensitive spatial transcriptomics at near-cellular resolution with Slide-seqV2. Nat. Biotechnol. 39, 313–319 (2021).

    Article  CAS  PubMed  Google Scholar 

  12. Moffitt, J. R. et al. Molecular, spatial, and functional single-cell profiling of the hypothalamic preoptic region. Science 362, eaau5324 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

  13. Efremova, M. & Teichmann, S. A. Computational methods for single-cell omics across modalities. Nat. Methods 17, 14–17 (2020).

    Article  CAS  PubMed  Google Scholar 

  14. Cao, Y., Yang, P. & Yang, J. Y. H. A benchmark study of simulation methods for single-cell RNA sequencing data. Nat. Commun. 12, 6911 (2021).

  15. Crowell, H. L., Morillo Leonardo, S. X., Soneson, C. & Robinson, M. D. The shaky foundations of simulating single-cell RNA sequencing data. Genome Biol. 24, 62 (2023).

  16. Sun, T., Song, D., Li, W. V. & Li, J. J. scDesign2: a transparent simulator that generates high-fidelity single-cell gene expression count data with gene correlations captured. Genome Biol. 22, 163 (2021).

  17. Risso, D., Perraudeau, F., Gribkova, S., Dudoit, S. & Vert, J.-P. A general and flexible method for signal extraction from single-cell RNA-seq data. Nat. Commun. 9, 284 (2018).

  18. Crowell, H. L. et al. Muscat detects subpopulation-specific state transitions from multi-sample multi-condition single-cell transcriptomics data. Nat. Commun. 11, 6077 (2020).

  19. Cannoodt, R., Saelens, W., Deconinck, L. & Saeys, Y. Spearheading future omics analyses using dyngen, a multi-modal simulator of single cells. Nat. Commun. 12, 3942 (2021).

  20. Dibaeinia, P. & Sinha, S. Sergio: a single-cell expression simulator guided by gene regulatory networks. Cell Syst. 11, 252–271 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  21. Papadopoulos, N., Gonzalo, P. R. & Söding, J. Prosstt: probabilistic simulation of single-cell RNA-seq data for complex differentiation processes. Bioinformatics 35, 3517–3519 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Tian, J., Wang, J. & Roeder, K. Esco: single cell expression simulation incorporating gene co-expression. Bioinformatics 37, 2374–2381 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  23. Navidi, Z., Zhang, L. & Wang, B. simATAC: a single-cell ATAC-seq simulation framework. Genome Biol. 22, 74 (2021).

  24. Li, W. V. & Li, J. J. A statistical simulator scDesign for rational scRNA-seq experimental design. Bioinformatics 35, i41–i50 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  25. Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat. Methods 16, 1289–1296 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  26. Marouf, M. et al. Realistic in silico generation and augmentation of single-cell RNA-seq data using generative adversarial networks. Nat. Commun. 11, 166 (2020).

  27. Ma, Y. & Zhou, X. Spatially informed cell-type deconvolution for spatial transcriptomics. Nat. Biotechnol. 40, 1349–1359 (2022).

  28. Cable, D. M. et al. Robust decomposition of cell type mixtures in spatial transcriptomics. Nat. Biotechnol. 40, 517–526 (2022).

    Article  CAS  PubMed  Google Scholar 

  29. Elosua-Bayes, M., Nieto, P., Mereu, E., Gut, I. & Heyn, H. SPOTlight: seeded NMF regression to deconvolute spatial transcriptomics spots with single-cell transcriptomes. Nucleic Acids Res. 49, e50 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  30. Yan, G. & Li, J. J. scReadSim: a single-cell multi-omics read simulator. Preprint at bioRxiv https://doi.org/10.1101/2022.05.29.493924 (2022).

  31. Cao, K., Hong, Y. & Wan, L. Manifold alignment for heterogeneous single-cell multi-omics data integration using Pamona. Bioinformatics 38, 211–219 (2022).

    Article  CAS  Google Scholar 

  32. Argelaguet, R., Cuomo, A. S. E., Stegle, O. & Marioni, J. C. Computational principles and challenges in single-cell data integration. Nat. Biotechnol. 39, 1202–1215 (2021).

    Article  CAS  PubMed  Google Scholar 

  33. Fang, J. et al. Clustering deviation index (CDI): a robust and accurate internal measure for evaluating scRNA-seq data clustering. Genome Biol. 23, 269 (2022).

  34. Duò, A., Robinson, M. D. & Soneson, C. A systematic performance evaluation of clustering methods for single-cell RNA-seq data. F1000Res. 7, 1441 (2018).

  35. Street, K. et al. Slingshot: cell lineage and pseudotime inference for single-cell transcriptomics. BMC Genomics 19, 477 (2018).

  36. Ji, Z. & Ji, H. TSCAN: pseudo-time reconstruction and evaluation in single-cell RNA-seq analysis. Nucleic Acids Res. 44, e117 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

  37. Stasinopoulos, D. M. & Rigby, R. A. Generalized additive models for location scale and shape (GAMLSS) in R. J. Stat. Softw. 23, 1–46 (2008).

  38. Zhang, Y., Parmigiani, G. & Johnson, W. E. ComBat-seq: batch effect adjustment for RNA-seq count data. NAR Genom. Bioinform. 2, lqaa078 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  39. Wood, S. N. Generalized Additive Models: An Introduction with R (Chapman and Hall/CRC, 2006).

  40. Kammann, E. E. & Wand, M. P. Geoadditive models. J. R. Stat. Soc. C 52, 1–18 (2003).

  41. Czado, C. Analyzing Dependent Data with Vine Copulas (Springer, 2019).

  42. Lun, A. T. L., McCarthy, D. J. & Marioni, J. C. A step-by-step workflow for low-level analysis of single-cell RNA-seq data with Bioconductor. F1000Res. 5, 2122 (2016).

  43. Stuart, T., Srivastava, A., Madad, S., Lareau, C. A. & Satija, R. Single-cell chromatin state analysis with Signac. Nat. Methods 18, 1333–1341 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  44. Hao, Y. et al. Integrated analysis of multimodal single-cell data. Cell 184, 3573–3587 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  45. Zhu, J., Sun, S. & Zhou, X. SPARK-X: non-parametric modeling enables scalable and robust detection of spatial expression patterns for large spatial transcriptomic studies. Genome Biol. 22, 184 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  46. Li, B. et al. Benchmarking spatial and single-cell transcriptomics integration methods for transcript distribution prediction and cell type deconvolution. Nat. Methods 19, 662–670 (2022).

    Article  CAS  PubMed  Google Scholar 

  47. Lütge, A. et al. CellMixS: quantifying and visualizing batch effects in single-cell RNA-seq data. Life Sci. Alliance 4, e202001004 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

  48. Newman, A. M. et al. Robust enumeration of cell subsets from tissue expression profiles. Nat. Methods 12, 453–457 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  49. Zeng, D. et al. IOBR: multi-omics immuno-oncology biological research to decode tumor microenvironment and signatures. Front. Immunol. 12, 687975 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  50. Biancalani, T. et al. Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram. Nat. Methods 18, 1352–1362 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

  51. Moriel, N. et al. Novosparc: flexible spatial reconstruction of single-cell gene expression with optimal transport. Nat. Protoc. 16, 4177–4200 (2021).

    Article  CAS  PubMed  Google Scholar 

  52. Song, D., Wang, Q. & Li, J. J. scDesign3 generates realistic in silico data for multimodal single-cell and spatial omics. Zenodo https://doi.org/10.5281/zenodo.7110761 (2022).

Download references

Acknowledgements

We appreciate the comments and feedback from the members of the Junction of Statistics and Biology at UCLA (http://jsb.ucla.edu). This work was supported by the following grants: National Science Foundation grants no. DBI-1846216 and no. DMS-2113754, NIH/NIGMS grants no. R01GM120507 and no. R35GM140888, Johnson & Johnson WiSTEM2D Award, the Sloan Research Fellowship, the UCLA David Geffen School of Medicine W. M. Keck Foundation Junior Faculty Award and the Chan-Zuckerberg Initiative Single-Cell Biology Data Insights Grant (to J.J.L.). J.J.L. was a fellow at the Radcliffe Institute for Advanced Study at Harvard University in 2022–2023 while she was writing this paper.

Author information

Authors and Affiliations

  1. Bioinformatics Interdepartmental Ph.D. Program, University of California, Los Angeles, CA, USA

    Dongyuan Song & Jingyi Jessica Li

  2. Department of Statistics, University of California, Los Angeles, CA, USA

    Qingyang Wang, Guanao Yan, Tianyang Liu, Tianyi Sun & Jingyi Jessica Li

  3. Department of Human Genetics, University of California, Los Angeles, CA, USA

    Jingyi Jessica Li

  4. Department of Computational Medicine, University of California, Los Angeles, CA, USA

    Jingyi Jessica Li

  5. Department of Biostatistics, University of California, Los Angeles, CA, USA

    Jingyi Jessica Li

  6. Radcliffe Institute for Advanced Study, Harvard University, Cambridge, MA, USA

    Jingyi Jessica Li

Authors
  1. Dongyuan Song
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Qingyang Wang
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Guanao Yan
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Tianyang Liu
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Tianyi Sun
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Jingyi Jessica Li
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

D.S. and J.J.L. conceived of the study. D.S., Q.W. and J.J.L. wrote the paper. D.S. and Q.W. developed the scDesign3 R package. D.S. and Q.W. performed data analysis with assistance from G.Y. and T.L. D.S. and T.S. discussed the scDesign3 method design at the beginning of the study.

Corresponding author

Correspondence to Jingyi Jessica Li.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature Biotechnology thanks Kin Fai Au and Jean Yang for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Extended Data Fig. 1 Benchmarking scDesign3 against four existing scRNA-seq simulators (scGAN, muscat, SPARSim, and ZINB-WaVE) for generating scRNA-seq data from a single trajectory (mouse pancreatic endocrinogenesis; dataset PANCREAS in Supplementary Table 2).

a, Distributions of eight summary statistics in the test data and the synthetic data generated by scDesign3 and the four simulators. Each number on top of a violin plot (the distribution of a summary statistic in a synthetic dataset) is the Kolmogorov-Smirnov (KS) distance between the synthetic data distribution (indicated by that violin plot) and the test data distribution. A smaller number indicates better agreement between the synthetic data and the test data in terms of that summary statistic’s distribution. b, Heatmaps of the gene-gene correlation matrices (showing top 100 highly expressed genes) in the test data and the synthetic data generated by scDesign3 and the four simulators. Pearson’s correlation coefficient r measures the similarity between two correlation matrices, one from the test data and the other from the synthetic data. c, PCA visualization (top two PCs) of the test data and the synthetic data generated by scDesign3 and the four simulators. Colors label cells’ pseudotime values; note that only the synthetic data generated by scDesign3 contain the pseudotime truths. An mLISI value close to 2 means that the synthetic data resemble the real data well in the low-dimensional space. d, UMAP visualization of the real data and the synthetic data generated by scDesign3 and the four simulators.

Extended Data Fig. 2 Benchmarking scDesign3 against four existing scRNA-seq simulators (scGAN, muscat, SPARSim, and ZINB-WaVE) for generating scRNA-seq data from bifurcating trajectories (myeloid progenitors in mouse bone marrow; dataset MARROW in Supplementary Table 2).

a, Distributions of eight summary statistics in the test data and the synthetic data generated by scDesign3 and the four simulators. Each number on top of a violin plot (the distribution of a summary statistic in a synthetic dataset) is the Kolmogorov-Smirnov (KS) distance between the synthetic data distribution (indicated by that violin plot) and the test data distribution. A smaller number indicates better agreement between the synthetic data and the test data in terms of that summary statistic’s distribution. b, Heatmaps of the gene-gene correlation matrices (showing top 100 highly expressed genes) in the test data and the synthetic data generated by scDesign3 and the four simulators. Pearson’s correlation coefficient r measures the similarity between two correlation matrices, one from the test data and the other from the synthetic data. c, PCA visualization (top two PCs) of the test data and the synthetic data generated by scDesign3 and the four simulators. Colors label cells’ pseudotime values in two trajectories; note that only the synthetic data generated by scDesign3 contain the pseudotime truths. An mLISI value close to 2 means that the synthetic data resemble the real data well in the low-dimensional space. d, UMAP visualization of the real data and the synthetic data generated by scDesign3 and the four simulators.

Extended Data Fig. 3 scDesign3 simulated realistic gene expression patterns in cancer spatial transcriptomics data (datasets OVARIAN and ACINAR in Supplementary Table 2.

Human ovarian cancer (a) and human prostate cancer, acinar cell carcinoma (b). The tissue samples were measured with both H&E (hematoxylin and eosin stain, left) and spatial transcriptomics (right, three cancer-related genes). Large Pearson correlation coefficients (r) represent similar spatial patterns in synthetic data and real (test) data.

Extended Data Fig. 4 scDesign3 simulated 10x Visium spatial transcriptomics data (sagital mouse brain slices; dataset VISIUM in Supplementary Table 2).

a, Distributions of eight summary statistics in the test data and the synthetic data generated by scDesign3 using cell type labels (scDesign3-ideal) and spatial locations (scDesign3-spatial), respectively. Each number on top of a violin plot (the distribution of a summary statistic in a synthetic dataset) is the Kolmogorov-Smirnov (KS) distance between the synthetic data distribution (indicated by that violin plot) and the test data distribution. A smaller number indicates better agreement between the synthetic data and the test data in terms of that summary statistic’s distribution. b, Heatmaps of the gene-gene correlation matrices (showing top 100 highly expressed genes) in the test data and the synthetic data generated by scDesign3-ideal and scDesign3-spatial. Pearson’s correlation coefficient r measures the similarity between two correlation matrices, one from the test data and the other from the synthetic data. c, PCA visualization (top two PCs) of the real data and the synthetic data generated by scDesign3-ideal and scDesign3-spatial. Cell types (clusters) are labeled by colors. Since the scDesgin3-spatial dataset was based on spatial locations only, it did not contain cell types. An mLISI value close to 2 means that the synthetic data resemble the real data well in the low-dimensional space. d, UMAP visualization of the real data and the synthetic data generated by scDesign3-ideal and scDesign3-spatial. In summary, scDesign3 realistically simulated 10x Visium data based on spatial locations without needing cell type annotations.

Extended Data Fig. 5 scDesign3 mimicked spatial transcriptomics data so that prediction algorithms had similar prediction performance when trained on real data or scDesign3 synthetic data.

In detail, we first split each of four spatial transcriptomics datasets (VISIUM, SLIDE, OVARIAN, and ACINAR in Supplementary Table 2) into two datasets (training and testing) by randomly splitting the spatial locations into two halves. Second, we used each of the four training datasets to fit scDesign3 and generate the corresponding synthetic dataset. Third, on each pair of training dataset and synthetic dataset (among a total of four pairs), we trained each of three prediction algorithms (gbm: gradient boosting machine; randomForest: random forest; svmRadial: support vector machine with the radial kernel) to predict each gene’s expression at a spatial location (input: spatial location; output: the gene’s log(count+1) expression level at the location), obtaining a pair of prediction models for each gene. Fourth, we applied each pair of prediction models to the corresponding testing dataset and calculated each model’s root-mean-squared error (RMSE) for predicting the corresponding gene, obtaining a pair of RMSEs. As a result, in each panel, we plotted the RMSEs for each prediction algorithm (row) and dataset (column), with each dot in the panel representing a gene. We found all genes’ RMSEs highly similar, indicating that scDesign3’s synthetic data well mimicked real data.

Extended Data Fig. 6 The effect of K on scDesign3’s simulation of spatial transcriptomics data (dataset ACINAR in Supplementary Table 2).

The rows represent three cancer-related genes; column 1 represents real test data; columns 2–8 represent scDesign3’s synthetic data generated using varying K, the input basis number. A large Pearson correlation coefficient (r) represents similar spatial patterns in synthetic and test data. The effective degrees of freedom (edf) represents the wiggliness of the fitted surface. With a larger K, scDesign3 can fit more complex patterns. The overfitting issue is accounted for by the automatic smoothness estimation39: when K is sufficiently large, edf (model complexity) and r (model goodness-of-fit) both become stable.

Extended Data Fig. 7 scDesign3 simulated spot-resolution spatial transcriptomics data for benchmarking cell-type deconvolution algorithms (datasets MOB-SP and MOB-SC in Supplementary Table 2).

a, scDesign3’s synthetic spot-resolution data well mimicked real data (top row), showing similar expression patterns for four cell-type marker genes (columns). scDesign3 used three steps to generate the spot-resolution data. Step 1: every gene’s estimated mean expression level at each spot (as a smooth function of spot location) by scDesign3. Step 2: every gene’s predicted expression level at each spot from CIBERSORT’s estimated cell-type proportions at the spot (considered as the ‘true proportions’) and the gene’s cell-type-specific expression levels (from the reference scRNA-seq data). Step 3: every gene’s simulated expression level at each spot by scDesign3 (from the true cell-type proportions at the spot and scDesign3’s synthetic scRNA-seq data). b, Using scDesign3 synthetic data, we benchmarked three spatial cell-type deconvolution algorithms (CARD6, RCTD7, and SPOTlight8). For each of the four cell types (columns), we used two metrics-Pearson correlation (r) and root-mean-square error (RMSE)-to compare the proportions estimated by each deconvolution algorithm (rows 2-4) to the true proportions (top row). Large r values represent similar spatial patterns of proportions, while small RMSE values represent similar values of proportions. Although all three algorithms well captured the spatial patterns of each cell type’s proportions (evidenced by large r values), CARD and RCTD outperformed SPOTlight by estimating cell-type proportions more accurately (evidenced by smaller RMSE values).

Extended Data Fig. 8 scDesign3 simulated scATAC-seq data (human PBMCs; dataset ATAC in Supplementary Table 2).

a, Distributions of eight summary statistics in the test data and the synthetic data generated by scDesign3 using cell type labels. Each number on top of a violin plot (the distribution of a summary statistic in a synthetic dataset) is the Kolmogorov-Smirnov (KS) distance between the synthetic data distribution (indicated by that violin plot) and the test data distribution. A smaller number indicates better agreement between the synthetic data and the test data in terms of that summary statistic’s distribution. b, Heatmaps of the peak-peak correlation matrices in the test data and the synthetic data generated by scDesign3. Pearson’s correlation coefficient r measures the similarity between two correlation matrices, one from the test data and the other from the synthetic data. c, PCA visualization (top two PCs) of the test data and the synthetic data generated by scDesign3. Cell types are labeled by colors. An mLISI value close to 2 means that the synthetic data resemble the test data well in the low-dimensional space. d, UMAP visualization of the test data and the synthetic data generated by scDesign3.

Extended Data Fig. 9 scDesign3 simulated CITE-seq data (human PBMCs; dataset CITE in Supplementary Table 2).

a, Distributions of eight summary statistics in the test data and the synthetic data generated by scDesign3. The CITE-seq dataset contains simultaneous measurements of each cell’s gene expression and surface protein abundance captured by Antibody-Derived Tags (ADTs). Each number on top of a violin plot (the distribution of a summary statistic in a synthetic dataset) is the Kolmogorov-Smirnov (KS) distance between the synthetic data distribution (indicated by that violin plot) and the test data distribution. A smaller number indicates better agreement between the synthetic data and the test data in terms of that summary statistic’s distribution. b, Heatmaps of the gene and protein correlation matrices (10 proteins with names starting with ‘ADT’ and their corresponding genes) in the test data and the synthetic data generated by scDesign3. Pearson’s correlation coefficient r measures the similarity between two correlation matrices, one from the test data and the other from the synthetic data. scDesign3 preserved the correlations between the RNA and protein expression levels of the 10 surface proteins. c, PCA visualization (top two PCs) of the test data and the synthetic data generated by scDesign3. Cell types are labeled by colors. An mLISI value close to 2 means that the synthetic data resemble the real data well in the low-dimensional space. d, UMAP visualization of the test data and the synthetic data generated by scDesign3.

Extended Data Fig. 10 scDesign3 provides unsupervised measures of the goodness-of-fit of pseudotime, clusters, and inferred spatial locations.

For visual clarity, we plot the relative BIC or AIC (rBIC or rAIC) by re-scaling scDesign3’s marginal BIC or AIC to [0, 1]. a, The scDesign3 rBIC (unsupervised) is negatively correlated with the R2 (supervised). Each R2 was calculated between the set of perturbed or inferred pseudotimes and the set of true pseudotimes in each of the eight datasets (the column names). The P value is from the one-sided test of Spearman’s rank correlation ρ. The true pseudotime is the ground truth used for generating the synthetic data. b, Comparison of the scDesign3 rBIC and the Clustering Deviation Index (CDI) rBIC (rescaled to [0, 1])33. The color scale shows the number of clusters, and the shapes represent clustering algorithms. We found the scDesign3 rBIC (unsupervised) negatively correlated with the ARI (supervised). The P value is from the one-sided test of Spearman’s rank correlation ρ. We also found the scDesign3 rBIC to perform better or similarly to the CDI on six out of the eight datasets (the column names). c, The scDesign3 rAIC (unsupervised) is negatively correlated with the mean cosine similarity (supervised). The mean cosine similarity was calculated between the set of perturbed or inferred locations and the set of true locations in each of the two spatial datasets (the column names). The P value is from the one-sided test of Spearman’s rank correlation ρ. The true locations are the ground truth used for generating the semi-synthetic data. Due to the high complexity of spatial patterns, the scDesign3 rAIC (left) outperformed the scDesign3 rBIC (right) for penalizing the model complexity less.

Supplementary information

Supplementary Information

Supplementary Methods, Figs. 1–5 and Tables 1–5.

Reporting Summary

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Song, D., Wang, Q., Yan, G. et al. scDesign3 generates realistic in silico data for multimodal single-cell and spatial omics. Nat Biotechnol 42, 247–252 (2024). https://doi.org/10.1038/s41587-023-01772-1

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1038/s41587-023-01772-1

This article is cited by

Search

Advanced search

Quick links

Nature Briefing AI and Robotics

Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing: AI and Robotics