Generative adversarial networks simulate gene expression and predict perturbations in single cells
Arsham GhahramaniiD, Fiona M. WattiD, Nicholas M. LuscombeiD
Abstract
Recent advances have enabled gene expression profiling of single cells at lower cost. As more data is produced there is an increasing need to integrate diverse datasets and better analyse underutilised data to gain biological insights. However, analysis of single cell RNA-seq data is challenging due to biological and technical noise which not only varies between laboratories but also between batches. Here for the first time, we apply a new generative deep learning approach called Generative Adversarial Networks (GAN) to biological data. We apply GANs to epidermal, neural and hematopoietic scRNA-seq data spanning different labs and experimental protocols. We show that it is possible to integrate diverse scRNA-seq datasets and in doing so, our generative model is able to simulate realistic scRNA-seq data that covers the full diversity of cell types. In contrast to many machine-learning approaches, we are able to interpret internal parameters in a biologically meaningful manner. Using our generative model we are able to obtain a universal representation of epidermal differentiation and use this to predict the effect of cell state perturbations on gene expression at high time-resolution. We show that our trained neural networks identify biological state-determining genes and through analysis of these networks we can obtain inferred gene regulatory relationships. Finally, we use internal GAN learned features to perform dimensionality reduction. In combination these attributes provide a powerful framework to progress the analysis of scRNA-seq data beyond exploratory analysis of cell clusters and towards integration of multiple datasets regardless of origin.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 100
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 2 (30 Jul 2018) · CC BY 4.0
A Wasserstein GAN trained on three batch-uncorrected mouse epidermal scRNA-seq datasets simulates realistic cells, predicts differentiation via latent arithmetic, and infers gene networks recovering Gata6 targets.
Single-cell RNA-seq noise varies across labs and batches, so data from different protocols are hard to integrate, and differential expression or marker analyses rarely reveal functional gene regulatory relationships. One jointly trained generator and discriminator integrates uncorrected datasets and its internals also yield perturbation simulation, state-determining genes, gene association networks and dimensionality reduction, analyses otherwise done with separate tools.
Key results
ShownA single Wasserstein GAN with gradient penalty (WGAN-GP), trained without batch correction, generates cells overlapping every real cluster in pooled mouse epidermal data, including Joost's eleven in vivo cell types, and does likewise when retrained on mouse neural and human hematopoietic data.
1,763 mouse epidermal training and 500 unseen cells (6605 genes); diversity peaks between 12,500 and 15,000 steps; generated-vs-unseen Pearson correlations centre near 0.4 early, broadening after 10,000 steps; 500 closely related generated cells average 0.71 correlation. Output also spans clusters of 2,343 mouse neural, 7,108 human CD34+ and 13,034 pooled human blood cells. · 2.1 Generative adversarial networks integrate…; 2.2 Generative modeling of large and sparse…
SuggestedAdding a basal-to-differentiated latent vector to unseen basal cells and interpolating simulates per-cell mouse epidermal differentiation, recovering marker trends, non-monotonic profiles, and cell-variable or late-onset expression (e.g., Uxs1) missed by Monocle's averaged pseudotime trajectory.
Vector from 30 unseen mouse basal/differentiated pairs; 1000 interpolated timepoints for 30 cells; eight profile types, early/late genes enriched for stem-cell/differentiation gene ontology (q-value<0.01); Monocle misorders early Itgb1-high cells and calls Uxs1 invariant; 223 predicted transient (Group IV) genes include PTPN1, PTPN13, MAF, also transient in human keratinocytes. · 2.3 Simulating cellular perturbations using…; Latent space mapping and interpolation
ShownCorrelations of the mouse epidermal generator's final-layer weights define sparse gene association networks that overlap GRNBoost2 (SCENIC) and co-expression predictions; Gata6 local network genes are upregulated in GATA6+ cells, and Gata6 ChIP-seq targets receive high association ranks.
Top-50 target sets for 512 transcription factors overlap significantly across all three methods; GAN and SCENIC usually tie each factor strongly to fewer than 20 genes. Gata6 network genes show higher GATA6+/GATA6- log fold-changes than all other genes (Kolmogorov-Smirnov, p = 1.3e-4); for Gata6 ChIP-seq targets both beat random genes, GAN with higher median score than SCENIC. · 2.5 GAN-derived gene association networks…; Local gene association networks; Figure legends
ShownSweeping each gene's expression in held-out cells and reading the discriminator's critic output ranks genes by maximum sensitivity gradient; top-ranked genes are enriched for known epidermal regulators regardless of abundance, and Lars2 emerges as a candidate pan-epidermal state gene.
In held-out mouse epidermal cells, each of 6605 genes was swept from observed minimum to maximum in 100 linear steps; Actb shows little change whereas Ivl, Krt10, Krt14, Cav2, Brd4 and Sox9 respond strongly, Krt10 with two sensitive ranges; the top 100 genes show strong enrichment for known epidermal regulators, with Krt6a, Krt10 and Krt14 in the top 10. · 2.4 Discriminator network identifies…; Discriminator network sensitivity analysis; Figure legends
Suggestedt-SNE on the discriminator's 200 hidden-layer features (GANt-SNE) separates cells by differentiation and interfollicular epidermis (IFE) versus hair follicle origin, whereas leading PCA components track dataset/laboratory batch and subtracting the first two before t-SNE loses biological signal.
On 2263 real mouse epidermal cells, marker overlays show GANt-SNE grouping Krt14-high basal IFE cells, splitting Krt10-high suprabasal from Krt10-low terminally differentiated cells, and separating Rps29-defined IFE from hair follicle cells; PCA and PC removal plus t-SNE fail to resolve these marker-defined states (visual assessment only). · 2.6 Discriminator network + t-SNE allows better…; Dimensionality reduction and clustering; Figure legends
Methods, data and reproducibility
A Wasserstein GAN with gradient penalty (generator: 100-dimensional latent input, 600 hidden units, 6605 log2(TPM+1) gene outputs; discriminator: 200 hidden units) is trained on uncorrected pooled cells, then probed via latent arithmetic, discriminator sensitivity sweeps, generator final-layer weight correlations and t-SNE of discriminator features.
- Pooled mouse epidermal scRNA-seq from three labs: Yang hair follicle subpopulations, Joost dorsal epidermis, Ghahramani in vitro keratinocytes; TPM-normalised, uncorrected · mouse · n = 1,763 training + 500 unseen cells; 6605 genes · GSE90848; GSE67602; GSE99989
- Joost et al. whole dorsal epidermis subset (Fluidigm C1), eleven epidermal cell types; source of basal/differentiated pairs and interfollicular epidermis cells for Monocle · mouse · n = 1422 cells · GSE67602
- Mouse embryonic stem cell to motor neuron differentiation time course over seven days (Sagner et al., neural progenitors, Fluidigm C1) · mouse · n = 2,343 cells; 14,000 genes · E-MTAB-5466 (ArrayExpress)
- FACS-isolated CD34+ human bone marrow hematopoietic cells across myeloid, erythroid and lymphoid lineages (Buenrostro et al., 10x Genomics Chromium) · human · n = 7,108 cells; median 4,400 genes detected per cell · GSE96772
- FACS-sorted CD14+ human blood monocytes (Zheng et al.), droplet-based, pooled with the Buenrostro cells for retraining · human · n = 13,034 cells in the combined set · SRP073767 (Short Read Archive)
- Bulk expression of human keratinocytes during methylcellulose suspension-induced differentiation at 0, 4, 8 and 12 hours (Mishra et al.), used to check predicted dynamics · human
- Bulk expression of GATA6+ versus GATA6- junctional zone and sebaceous duct epidermal cells from a previous study by the same lab (Donati et al.)
- Gata6 ChIP-seq data (Donati; Wamaitha et al.) defining direct Gata6 targets for comparing GAN and SCENIC association ranks
- Methods & tools
- Wasserstein GAN with gradient penalty (WGAN-GP), TensorFlow, Python (numpy, sklearn, networkx), R (data preparation), enrichR (gene ontology enrichment), RMSProp optimizer, t-SNE (perplexity 30), Principal component analysis (PCA), k-means clustering, Monocle with Census mRNA quantification, GRNBoost2 (SCENIC), Kolmogorov-Smirnov test
- Data & accessions
- https://github.com/luslab/scRNAseq-WGAN-GP/tree/master/data
- Software & environment
- TensorFlow in Python (numpy, sklearn, networkx); R for data preparation; enrichR for gene ontology enrichment; GAN implementation shared as a Jupyter Notebook.
- Reproducibility gaps
- No package versions, hardware, runtime or random seeds reported; full GAN parameter list deferred to Supplementary Table 2; validation data (Mishra time course, GATA6+/- bulk profiles, Gata6 ChIP-seq) cited without accessions.
Limitations (5)
- Generated subpopulations overlap real ones but mostly do not span their full within-cluster variation, which the authors partly attribute to the generator's continuous output approximating discrete, dropout-prone count data. · stated by the authors
- No standard method exists for evaluating generative models of expression data, and loss convergence alone cannot signal completed GAN training, so output diversity (median cell-cell Pearson distance) was also tracked to decide when to stop. · stated by the authors
- Simulated differentiation assumes synchronously progressing cells; asynchronous in vitro differentiation from heterogeneous starting cells limits detection of transient genes in the human keratinocyte data and may explain why Group IV splits into three peak-time subgroups. · stated by the authors
- Integration and embedding quality are judged from t-SNE overlap and marker overlays; no quantitative batch-mixing or clustering scores, and no comparison with dedicated scRNA-seq batch-correction methods, are reported. · noted in this summary
- Real cells are mapped to latent space by random sampling until a generated profile reaches 0.7 correlation, an approximate inversion whose effect on predicted trajectories is untested; network validation against independent expression and ChIP-seq data covers only Gata6. · noted in this summary
Who should read:Read this if you work on epidermal cell states or generative deep learning for scRNA-seq and want to see one GAN's generator and discriminator reused for integration, perturbation simulation, gene networks and embedding.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Deep feature extraction of single-cell transcriptomes by generative adversarial network
Mojtaba Bahrami, Malosree Maitra, … Yue Li·Computer applications in the biosciencesPeer-reviewed·Nov 11, 2020
- scGAIN: Single Cell RNA-seq Data Imputation using Generative Adversarial Networks
Mohamed K. Gunady, Jayaram Kancherla, … Soheil Feizi·bioRxiv·Nov 12, 2019
- scIGANs: single-cell RNA-seq imputation using generative adversarial networks
Yungang Xu, Zhigang Zhang, … Xiaobo Zhou·Nucleic Acids ResearchPeer-reviewed·Jun 17, 2020
- Realistic in silico generation and augmentation of single cell RNA-seq data using Generative Adversarial Neural Networks
Matteo Marouf, Pierre Machart, … Stefan Bonn·Nature CommunicationsPeer-reviewed·Jan 9, 2020
- Sampling from Disentangled Representations of Single-Cell Data Using Generative Adversarial Networks
Hengshi Yu, Joshua D. Welch·Genome biologyPeer-reviewed·May 20, 2021
- Optimization and redevelopment of single-cell data analysis workflow based on deep generative models