Interpretable Factors in scRNA-seq Data with Disentangled Generative Models

Single-cell RNA sequencing (scRNA-seq) experiments measure transcriptional profiles that encode diverse sources of variation, both biological and technical, and complex coordinations among them. PCA is a popular method for interpretable dimension reduction. However, it assumes a linear mapping between data and latent components and this may not be warranted in complex data. Single cell variational inference (scVI) offers a nonlinear method for latent space mapping, but the latent factors are not in general interpretable. In light of disentangled representations that learn independent data generative factors of single cell data in an unsupervised way, we propose factor variational inference (factorVI). FactorVI learns the disentangled factors among biologically relevant latent variables directly by penalizing correlations between them. We evaluate the factorVI through clustering in publicly available datasets and we observe high accuracy. We also propose biological interpretation of the latent factors.