CoChat

Biologically informed variational autoencoders allow predictive modeling of genetic and drug-induced perturbations

Open access
Doncevic, Daria, Herrmann, Carl · 2023 · Bioinformatics

Published version of a preprint

The DOI matches its Crossref record; 38 of the 38 references it cites were found in a registry or an index.

Abstract

Motivation: Variational autoencoders (VAEs) have rapidly increased in popularity in biological applications and have already successfully been used on many omic datasets. Their latent space provides a low-dimensional representation of input data, and VAEs have been applied, e.g. for clustering of single-cell transcriptomic data. However, due to their non-linear nature, the patterns that VAEs learn in the latent space remain obscure. Hence, the lower-dimensional data embedding cannot directly be related to input features. Results: To shed light on the inner workings of VAE and enable direct interpretability of the model through its structure, we designed a novel VAE, OntoVAE (Ontology guided VAE) that can incorporate any ontology in its latent space and decoder part and, thus, provide pathway or phenotype activities for the ontology terms. In this work, we demonstrate that OntoVAE can be applied in the context of predictive modeling and show its ability to predict the effects of genetic or drug-induced perturbations using different ontologies and both, bulk and single-cell transcriptomic datasets. Finally, we provide a flexible framework, which can be easily adapted to any ontology and dataset. Availability and implementation: OntoVAE is available as a python package under https://github.com/hdsu-bioquant/onto-vae.

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 · 10.1093/bioinformatics/btad387

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1. OntoVAE—a novel VAE architecture that can incorporate biological ontologies and provide pathway activities in its latent space and decoder. (a) Schematic overview of OntoVAE. A non-linear encoder is coupled to a masked, multi-layer linear decoder reflecting a biological ontology. The latent space incorporates the root terms of the ontology, each layer in the decoder represents one depth of the ontology, with “depth” referring to the longest possible path between node and a root node and “level” referring to the shortest possible path between a node and a root node. (b–e) OntoVAE with GO-decoder has been trained on all tissue samples from GTEx. (b) Scatter plots showing example pathway activations for digestive system process and glutamate receptor signaling pathway (top), and aortic valve morphogenesis and axon ensheathment (bottom). (c and d) Naive Bayes classifier with 10-fold cross-validation was trained for each GO term for each tissue (1-vs-all). Median AUC for each tissue for 10 example pathways is displayed in (c), for tissues with AUC higher than 0.5, density over the pathway activity is displayed in (d), with the vertical line corresponding to a pathway activity of zero. (e) Subnetwork of the GO graph, nodes are colored by the median AUC for blood (left) and spleen (right). act., activation; diff., differentiation; prolif., proliferation; deg., degranulation; i.i.i.r., involved in immune response. (f) Top GO term network for liver. Top GO terms were identified using pairwise Wilcoxon testing at each term and then further summarized into networks based on their semantic similarities.
Figure 1. OntoVAE—a novel VAE architecture that can incorporate biological ontologies and provide pathway activities in its latent space and decoder. (a) Schematic overview of OntoVAE. A non-linear encoder is coupled to a masked, multi-layer linear decoder reflecting a biological ontology. The latent space incorporates the root terms of the ontology, each layer in the decoder represents one depth of the ontology, with “depth” referring to the longest possible path between node and a root node and “level” referring to the shortest possible path between a node and a root node. (b–e) OntoVAE with GO-decoder has been trained on all tissue samples from GTEx. (b) Scatter plots showing example pathway activations for digestive system process and glutamate receptor signaling pathway (top), and aortic valve morphogenesis and axon ensheathment (bottom). (c and d) Naive Bayes classifier with 10-fold cross-validation was trained for each GO term for each tissue (1-vs-all). Median AUC for each tissue for 10 example pathways is displayed in (c), for tissues with AUC higher than 0.5, density over the pathway activity is displayed in (d), with the vertical line corresponding to a pathway activity of zero. (e) Subnetwork of the GO graph, nodes are colored by the median AUC for blood (left) and spleen (right). act., activation; diff., differentiation; prolif., proliferation; deg., degranulation; i.i.i.r., involved in immune response. (f) Top GO term network for liver. Top GO terms were identified using pairwise Wilcoxon testing at each term and then further summarized into networks based on their semantic similarities.CC BY · PubMed Central
Figure 2. OntoVAE can predict phenotypic outcome of a gene knockout (here: DMD). OntoVAE has been trained with a GO-decoder on all GTEx tissue samples, knockouts have been performed in muscle samples only. (a) Schematic drawing of how the model can be used for in silico investigation of gene knockouts. Input value for a gene (here: DMD) can be set to zero before running samples through the trained model and obtaining their activations at each node/term. Paired Wilcoxon tests can be performed for all the terms in latent space and decoder (pre-knockout versus post-knockout) to identify the most affected terms (term-level analysis). Paired Wilcoxon tests can also be performed for all the genes in the reconstruction layer to identify the most affected genes, and these can then be further grouped into terms using gene set ORA (gene-level analysis). (b) Knockout of DMD in GTEx muscle samples. Barplot displays the 10 most affected terms from the term-level analysis ranked by significance. (c) Knockout of DMD in GTEx muscle samples. Barplot displays the 10 most affected terms from the gene-level analysis ranked by significance. (b and c) The dotted line represents the significance threshold.
Figure 2. OntoVAE can predict phenotypic outcome of a gene knockout (here: DMD). OntoVAE has been trained with a GO-decoder on all GTEx tissue samples, knockouts have been performed in muscle samples only. (a) Schematic drawing of how the model can be used for in silico investigation of gene knockouts. Input value for a gene (here: DMD) can be set to zero before running samples through the trained model and obtaining their activations at each node/term. Paired Wilcoxon tests can be performed for all the terms in latent space and decoder (pre-knockout versus post-knockout) to identify the most affected terms (term-level analysis). Paired Wilcoxon tests can also be performed for all the genes in the reconstruction layer to identify the most affected genes, and these can then be further grouped into terms using gene set ORA (gene-level analysis). (b) Knockout of DMD in GTEx muscle samples. Barplot displays the 10 most affected terms from the term-level analysis ranked by significance. (c) Knockout of DMD in GTEx muscle samples. Barplot displays the 10 most affected terms from the gene-level analysis ranked by significance. (b and c) The dotted line represents the significance threshold.CC BY · PubMed Central
Figure 3. OntoVAE can predict differential gene expression in disease. OntoVAE has been trained with HPO-decoder on all tissue samples from GTEx, knockouts were performed in muscle samples only. Knockout was performed for all genes one-by-one by setting their input value to zero, and one-tailed paired Wilcoxon tests at the LGMD node in the decoder were computed for each gene, comparing pre- and post-knockout. All genes were then ranked according to the P-value of the paired Wilcoxon test, and GSEA was performed with the genes that were upregulated in patients versus control (LGMD_up) and the genes that were downregulated in patients versus control (LGMD_dn) in the LGMD dataset. (a) GSEA results using ranking of genes that significantly downregulated the LGMD node in the decoder. (b) GSEA results using ranking of genes that significantly upregulated the LGMD node in the decoder. (c) Leading edges of both GSEAs were fused (predicted genes). Venn diagram in the top right corner shows overlap between the predicted genes and LGMD_dn. Venn diagram in the bottom left corner shows overlap between predicted genes and genes directly annotated to the term LGMD in HPO. Hockeystick plot displays the ranking of the predicted genes, the 10 genes directly annotated to LGMD in HPO are labeled. (d and e) For each of the 10 genes, a new model was trained where the link of the gene to LGMD had been removed. Plots show the ranking of the 10 genes before (blue circle) and after (orange circle) removing the link to LGMD. (d) Genes that were found in the leading edge of (a), dotted line indicates leading edge cutoff. € Genes that were found in the leading edge of (b), red dotted line indicates leading edge cutoff.
Figure 3. OntoVAE can predict differential gene expression in disease. OntoVAE has been trained with HPO-decoder on all tissue samples from GTEx, knockouts were performed in muscle samples only. Knockout was performed for all genes one-by-one by setting their input value to zero, and one-tailed paired Wilcoxon tests at the LGMD node in the decoder were computed for each gene, comparing pre- and post-knockout. All genes were then ranked according to the P-value of the paired Wilcoxon test, and GSEA was performed with the genes that were upregulated in patients versus control (LGMD_up) and the genes that were downregulated in patients versus control (LGMD_dn) in the LGMD dataset. (a) GSEA results using ranking of genes that significantly downregulated the LGMD node in the decoder. (b) GSEA results using ranking of genes that significantly upregulated the LGMD node in the decoder. (c) Leading edges of both GSEAs were fused (predicted genes). Venn diagram in the top right corner shows overlap between the predicted genes and LGMD_dn. Venn diagram in the bottom left corner shows overlap between predicted genes and genes directly annotated to the term LGMD in HPO. Hockeystick plot displays the ranking of the predicted genes, the 10 genes directly annotated to LGMD in HPO are labeled. (d and e) For each of the 10 genes, a new model was trained where the link of the gene to LGMD had been removed. Plots show the ranking of the 10 genes before (blue circle) and after (orange circle) removing the link to LGMD. (d) Genes that were found in the leading edge of (a), dotted line indicates leading edge cutoff. € Genes that were found in the leading edge of (b), red dotted line indicates leading edge cutoff.CC BY · PubMed Central
Figure 4. OntoVAE can predict IFN treatment response. (a) UMAP plot shows the PBMC dataset colored by cell type (right panel) and treatment (left panel). (b–e) Prediction of IFN treatment response by in silico stimulation of genes. All in silico stimulated genes were ranked according to the P-value of the paired Wilcoxon test, and GSEA (b) was performed with the genes that were upregulated in IFN-ß treated CD4T cells versus control CD4T cells in the original dataset (CD4T_IFN-ß_stim_up). Based on GSEA, genes from the leading edge (termed “predicted genes”) were selected. (c and d) Projection of in silico stimulated cells (green) on the UMAP computed using GO pathway activities for control CD4T cells (purple) and IFN-stimulated CD4T cells (red). UMAP and density plots are shown for the 717 leading edge genes (c) and the 717 bottom GSEA genes used as control (d), with stimulation strength increasing from top to bottom. Venn diagram € displaying the overlap between the predicted genes (green circle) and CD4T_IFN-ß_stim_up (red circle). € Hockey stick plot displaying the predicted genes sorted by significance, the 44 genes overlapping with CD4T_IFN-ß_stim_up are labeled. Darker label color indicates higher significance of this gene in CD4T_IFN-ß_stim_up, a black triangle next to the label indicates that the gene is directly annotated to the term type I IFN signaling pathway or to one of its descendant terms. (f and g) Further investigation of predicted genes that do not overlap with CD4T_IFN-ß_stim_up [green circle in Venn diagram € without intersection]. Venn diagrams are showing the overlap of these genes (orange circles) with the genes that are upregulated upon IFN-ß treatment in the other cell types of the PBMC dataset (gray circles) (f). Dotplot is showing results from a gene set ORA using IFN-related terms from MsigDBs C7 immunologic signature gene sets (g).
Figure 4. OntoVAE can predict IFN treatment response. (a) UMAP plot shows the PBMC dataset colored by cell type (right panel) and treatment (left panel). (b–e) Prediction of IFN treatment response by in silico stimulation of genes. All in silico stimulated genes were ranked according to the P-value of the paired Wilcoxon test, and GSEA (b) was performed with the genes that were upregulated in IFN-ß treated CD4T cells versus control CD4T cells in the original dataset (CD4T_IFN-ß_stim_up). Based on GSEA, genes from the leading edge (termed “predicted genes”) were selected. (c and d) Projection of in silico stimulated cells (green) on the UMAP computed using GO pathway activities for control CD4T cells (purple) and IFN-stimulated CD4T cells (red). UMAP and density plots are shown for the 717 leading edge genes (c) and the 717 bottom GSEA genes used as control (d), with stimulation strength increasing from top to bottom. Venn diagram € displaying the overlap between the predicted genes (green circle) and CD4T_IFN-ß_stim_up (red circle). € Hockey stick plot displaying the predicted genes sorted by significance, the 44 genes overlapping with CD4T_IFN-ß_stim_up are labeled. Darker label color indicates higher significance of this gene in CD4T_IFN-ß_stim_up, a black triangle next to the label indicates that the gene is directly annotated to the term type I IFN signaling pathway or to one of its descendant terms. (f and g) Further investigation of predicted genes that do not overlap with CD4T_IFN-ß_stim_up [green circle in Venn diagram € without intersection]. Venn diagrams are showing the overlap of these genes (orange circles) with the genes that are upregulated upon IFN-ß treatment in the other cell types of the PBMC dataset (gray circles) (f). Dotplot is showing results from a gene set ORA using IFN-related terms from MsigDBs C7 immunologic signature gene sets (g).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 38 references

Cited by

Metadata

Authors
Doncevic, D., & Herrmann, C.
Affiliations
Heidelberg University (DE)
Corresponding author
  • Doncevic, Daria
  • Herrmann, Carl
Venue
Bioinformatics
Published
2023-06-01
Volume
39
Issue
6
ISSN
  • 1367-4803 (ISSN-L)
  • 1367-4811
OpenAlex
W4380869622
Cited by
51
References
35
Citations by year
2026: 12 · 2025: 30 · 2024: 5 · 2023: 4
FWCI
4.54
Type
journal article
Language
en
Open access
Locations
  • Bioinformatics (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
  • IPMB, University Heidelberg
  • Deutsche Forschungsgemeinschaft
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Bioinformatics and Genomic Networks
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Health Sciences › Medicine › Health Informatics › Biomedical Text Mining and Ontologies
Keywords
variational autoencoder, interpretability, pathway activity, single-cell transcriptomics, bulk transcriptomics, predictive modeling
MeSH terms
Cluster Analysis, Gene Expression Profiling, Transcriptome

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

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.