Disentangling cellular heterogeneity into interpretable biological factors through structured latent representations

Abstract

Deep generative models have become central to single-cell omics analysis, but their latent spaces remain difficult to interpret biologically. Linear factor models offer dimension-wise interpretability, but often lack the nonlinear flexibility, scalability, and integration quality required for large multi-batch atlases. We bridge this gap with Disentangled Representation Variational Inference (DRVI), an unsupervised deep generative model that learns dimension-wise interpretable representations for single-cell omics without supervised priors on cell types or biological processes. DRVI achieves this through additive decoder subnetworks combined with log-sum-exp pooling, enabling disentangled nonlinear gene programs. Across atlases, perturbation screens, and developmental datasets, DRVI separates cell identity, signaling pathways, stress responses, developmental trajectories, and perturbation effects into distinct interpretable factors. These factors identify rare migratory dendritic cells and recover coherent combinatorial perturbation programs in CRISPR screens. Systematic benchmarks show that this interpretability does not reduce integration performance. DRVI recovers biological factors more accurately while maintaining competitive integration quality. Altogether, DRVI provides a practical route to nonlinear single-cell modeling with factor-level interpretability.