Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 51 of the 51 references it cites were found in a registry or an index.

VEGA is an interpretable generative model for inferring biological network activity in single-cell transcriptomics

Open access

Abstract

Deep learning architectures such as variational autoencoders have revolutionized the analysis of transcriptomics data. However, the latent space of these variational autoencoders offers little to no interpretability. To provide further biological insights, we introduce a novel sparse Variational Autoencoder architecture, VEGA (VAE Enhanced by Gene Annotations), whose decoder wiring mirrors user-provided gene modules, providing direct interpretability to the latent variables. We demonstrate the performance of VEGA in diverse biological contexts using pathways, gene regulatory networks and cell type identities as the gene modules that define its latent space. VEGA successfully recapitulates the mechanism of cellular-specific response to treatments, the status of master regulators as well as jointly revealing the cell type and cellular state identity in developing cells. We envision the approach could serve as an explanatory biological model for development and drug treatment experiments.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41467-021-26017-0

Cites

51 of 51 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Designing a novel VAE architecture with interpretable latent space. a Overview of the VEGA model. Composed of a deep nonlinear encoder (μ, Σ) and a masked linear decoder, VEGA represents single-cell transcriptomics data into a lower-dimensional interpretable latent space z that approximates a set of user-supplied gene modules (GMV). Additionally, VEGA can integrate batch information as another variable s to condition its generative process on batch labels. b UMAP embedding of the latent space of VEGA retains the biological signal of the Kang et al. PBMCs dataset17. c Inferred interferon-alpha/beta signaling pathway activity segregates stimulated cells from the control population. d Bivariate GMV plot showing the ability of the model to recover the tryptophan catabolism activity, an innate (Dendritic cells, FCGR3A+monocytes, CD14+monocytes) immune cell-specific response to the perturbation. e Volcano plot showing differentially active GMVs between stimulated and control innate immune cells. The red dots indicate GMVs with ∣loge(Bayes Factor)∣ > 3 and a mean absolute difference (MD) in the latent space of at least 5. f Comparison of VEGA Bayes Factor with GSEA -log10(FDR). The size of the dots indicates the gene set size. The red, blue, and purple quadrants correspond respectively to significant hits unique to our model, unique to GSEA, and common to both.
Fig. 1 Designing a novel VAE architecture with interpretable latent space. a Overview of the VEGA model. Composed of a deep nonlinear encoder (μ, Σ) and a masked linear decoder, VEGA represents single-cell transcriptomics data into a lower-dimensional interpretable latent space z that approximates a set of user-supplied gene modules (GMV). Additionally, VEGA can integrate batch information as another variable s to condition its generative process on batch labels. b UMAP embedding of the latent space of VEGA retains the biological signal of the Kang et al. PBMCs dataset17. c Inferred interferon-alpha/beta signaling pathway activity segregates stimulated cells from the control population. d Bivariate GMV plot showing the ability of the model to recover the tryptophan catabolism activity, an innate (Dendritic cells, FCGR3A+monocytes, CD14+monocytes) immune cell-specific response to the perturbation. e Volcano plot showing differentially active GMVs between stimulated and control innate immune cells. The red dots indicate GMVs with ∣loge(Bayes Factor)∣ > 3 and a mean absolute difference (MD) in the latent space of at least 5. f Comparison of VEGA Bayes Factor with GSEA -log10(FDR). The size of the dots indicates the gene set size. The red, blue, and purple quadrants correspond respectively to significant hits unique to our model, unique to GSEA, and common to both.CC BY · PubMed Central
Fig. 2 The flexibility in the latent space specification sheds light on the activity of core cellular processes and transcription factors. a tSNE embedding of the latent space of VEGA for the MIX-Seq data21. The color indicates the treatment condition, and the arrow indicates the median shift in coordinates of each cell line between the two conditions. b Inferred G2M checkpoint activity of each cells, showing a decreased activity in the treated condition, as expected from the MoA of Trametinib. c Heatmap with hierarchical clustering showing the average loge(Bayes Factor) of each pathway for each cell line/drug treatment pair (test between DMSO and treatment condition). Each row corresponds to a hallmark gene set and each column to a different cell line/drug pair. The first row of color indicates the drug, and the second row of color indicates the tissue identity (Tissue legend available in SFig. 2). Highlighted cell lines correspond to BRAF-mutant melanoma. Highlighted activities correspond to Trametinib-specific responses. d tSNE embedding of the latent space of the model for the glioblastoma dataset29, colored by cell type or e Inferred activity of the master regulators STAT3 and OLIG2.
Fig. 2 The flexibility in the latent space specification sheds light on the activity of core cellular processes and transcription factors. a tSNE embedding of the latent space of VEGA for the MIX-Seq data21. The color indicates the treatment condition, and the arrow indicates the median shift in coordinates of each cell line between the two conditions. b Inferred G2M checkpoint activity of each cells, showing a decreased activity in the treated condition, as expected from the MoA of Trametinib. c Heatmap with hierarchical clustering showing the average loge(Bayes Factor) of each pathway for each cell line/drug treatment pair (test between DMSO and treatment condition). Each row corresponds to a hallmark gene set and each column to a different cell line/drug pair. The first row of color indicates the drug, and the second row of color indicates the tissue identity (Tissue legend available in SFig. 2). Highlighted cell lines correspond to BRAF-mutant melanoma. Highlighted activities correspond to Trametinib-specific responses. d tSNE embedding of the latent space of the model for the glioblastoma dataset29, colored by cell type or e Inferred activity of the master regulators STAT3 and OLIG2.CC BY · PubMed Central
Fig. 3 Disentangling cellular states and cell types in the early development of cortical organoids. a UMAP embedding of the latent space of our model for the week 2 cortical organoid dataset31. The cell type annotation corresponds to the original paper annotation. b, c, d The inferred activity of each cell type GMVs (as defined by marker genes) correctly identifies the three main subpopulations of cells. e One-vs-rest differential GMV analysis of each cell type population provides a statistical significance for each cell type signature. The significance threshold for positive enrichment was set to loge(BF)> 3. f Identification of dividing and quiescent subpopulations of neural progenitors using pathway and cell-type activity projection. g CELL_CYCLE_MITOTIC pathway activity correctly identifies dividing cells as reported by its correlation with MKI67 gene expression (an external canonical marker of dividing cells).
Fig. 3 Disentangling cellular states and cell types in the early development of cortical organoids. a UMAP embedding of the latent space of our model for the week 2 cortical organoid dataset31. The cell type annotation corresponds to the original paper annotation. b, c, d The inferred activity of each cell type GMVs (as defined by marker genes) correctly identifies the three main subpopulations of cells. e One-vs-rest differential GMV analysis of each cell type population provides a statistical significance for each cell type signature. The significance threshold for positive enrichment was set to loge(BF)> 3. f Identification of dividing and quiescent subpopulations of neural progenitors using pathway and cell-type activity projection. g CELL_CYCLE_MITOTIC pathway activity correctly identifies dividing cells as reported by its correlation with MKI67 gene expression (an external canonical marker of dividing cells).CC BY · PubMed Central
Fig. 4 Generalization of VEGA architecture to out-of-sample data. a Violin plot (n = 10,000 randomly sampled cells per condition) representing the distribution of the interferon-α/β pathway activity in control CD4-T cells, stimulated CD4 T cells unseen at the time of training (out-of-sample), and stimulated CD4-T cells when included in the training procedure (in-sample). Boxes inside the violins represent the median of the distribution bounded by the first and third quartile. Violin limits correspond to data extrema. b Proportion of overlap in the top 50 differentially activated GMVs in the in-sample and out-of-sample settings with stimulated vs control differential procedures for the seven main cell types in the study. Data were presented as mean values ± standard deviation over 100 random sampling. c R2 between the mean expression of real and reconstructed cells in the in-sample and out-of-sample settings for the seven main cell types of the study. Data were presented as mean values ± standard deviation over 100 random samplings. d Violin plot (n = 2000 randomly sampled cells per condition) of distribution of the interferon-α/β pathway activity in control CD4-T cells of study A (Kang et al.17), stimulated CD4-T cells of study A and control CD4-T cells of study B (Zheng et al.32). Boxes inside the violins represent the median of the distribution bounded by the first and third quartile. Violin limits correspond to data extrema. e Proportion of overlap in the top 50 differentially activated GMVs of each study with one-vs-rest differential procedures for the control cells of the seven main PBMC cell types. Data were presented as mean values ± standard deviation over 100 random samplings. f R2 between the mean expression of real and reconstructed cells of study A (Study A), mean expression of real and reconstructed cells of study B (Study B), and mean expression of real cells of study A and real cells of study B (Study A vs Study B). Data were presented as mean values ± standard deviation over 100 random samplings.
Fig. 4 Generalization of VEGA architecture to out-of-sample data. a Violin plot (n = 10,000 randomly sampled cells per condition) representing the distribution of the interferon-α/β pathway activity in control CD4-T cells, stimulated CD4 T cells unseen at the time of training (out-of-sample), and stimulated CD4-T cells when included in the training procedure (in-sample). Boxes inside the violins represent the median of the distribution bounded by the first and third quartile. Violin limits correspond to data extrema. b Proportion of overlap in the top 50 differentially activated GMVs in the in-sample and out-of-sample settings with stimulated vs control differential procedures for the seven main cell types in the study. Data were presented as mean values ± standard deviation over 100 random sampling. c R2 between the mean expression of real and reconstructed cells in the in-sample and out-of-sample settings for the seven main cell types of the study. Data were presented as mean values ± standard deviation over 100 random samplings. d Violin plot (n = 2000 randomly sampled cells per condition) of distribution of the interferon-α/β pathway activity in control CD4-T cells of study A (Kang et al.17), stimulated CD4-T cells of study A and control CD4-T cells of study B (Zheng et al.32). Boxes inside the violins represent the median of the distribution bounded by the first and third quartile. Violin limits correspond to data extrema. e Proportion of overlap in the top 50 differentially activated GMVs of each study with one-vs-rest differential procedures for the control cells of the seven main PBMC cell types. Data were presented as mean values ± standard deviation over 100 random samplings. f R2 between the mean expression of real and reconstructed cells of study A (Study A), mean expression of real and reconstructed cells of study B (Study B), and mean expression of real cells of study A and real cells of study B (Study A vs Study B). Data were presented as mean values ± standard deviation over 100 random samplings.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 51 references

Research works printed but not linked: 1

  1. Baldi, P. Autoencoders, unsupervised learning, and deep architectures. Proc. Mach. Learn. Res. 27, 37_49 (2012).

Cited by

See all 107

Metadata

Authors
Seninge, L., Anastopoulos, I., Ding, H., & Stuart, J.
Affiliations
University of California, Santa Cruz (US)
Corresponding author
  • Ding, Hongxu
  • Stuart, Joshua
Venue
Nature Communications
Published
2021-09-28
Volume
12
Issue
1
Pages
5684-5684
ISSN
2041-1723 (ISSN-L)
OpenAlex
W3204708585
Cited by
107
References
51
Citations by year
2026: 16 · 2025: 29 · 2024: 27 · 2023: 28 · 2022: 5 · 2021: 2
FWCI
5.02
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • eScholarship (California Digital Library) (submittedVersion)
  • eScholarship (California Digital Library) (submittedVersion, cc-by)
Funders
  • California Institute for Regenerative Medicine (GC1R-06673-C, GC1R-06673)
  • Seagate Technology
  • National Institutes of Health
  • National Institute of Mental Health (R01MH120295)
  • National Cancer Institute (U24CA210990)
  • National Institute of General Medical Sciences (R01GM109031)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene Regulatory Network Analysis
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Bioinformatics and Genomic Networks
Keywords
Vega, interpretable generative models, single-cell transcriptomics, gene modules, latent space interpretability, gene regulatory networks, cell identity, cell states, drug response, master regulators
MeSH terms
Deep Learning, RNA-Seq, Animals, Humans, Models, Genetic, Mice, Gene Regulatory Networks, Single-Cell Analysis, Datasets as Topic

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: VEGA is an interpretable generative model for inferring biological network activity in single-cell transcriptomics. https://app.cochat.ai/papers/10.1038/s41467-021-26017-0
The paper
Seninge, L., Anastopoulos, I., Ding, H., & Stuart, J. (2021). VEGA is an interpretable generative model for inferring biological network activity in single-cell transcriptomics. Nature Communications. https://doi.org/10.1038/s41467-021-26017-0

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.