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

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. Developing interpretable models is a major challenge in single cell deep learning. Here we show that the VEGA variational autoencoder model, whose decoder wiring mirrors gene modules, can provide direct interpretability to the latent space further enabling the inference of biological module activity.