Unsupervised Deep Disentangled Representation of Single-Cell Omics with DRVI

Single-cell genomics allows for the unbiased exploration of cellular heterogeneity.Representation learning methods summarize high-dimensional single-cell data into a manageable latent space in a typically nonlinear fashion, allowing cross-sample integration or generative modeling.However, these methods often produce entangled representations, limiting interpretability and downstream analyses.Existing disentanglement methods instead either require supervised information or impose sparsity and linearity, which may not capture the complexity of biological data. We, therefore, introduce Disentangled Representation Variational Inference (DRVI), an unsupervised deep generative model that learns nonlinear, disentangled representations of single-cell omics. This is achieved by combining recently introduced additive decoders with nonlinear pooling, for which we theoretically prove disentanglement under reasonable assumptions. We validate DRVI's disentanglement capabilities across diverse relevant biological problems, from development to perturbational studies and cell atlases, decomposing, for example, the Human Lung Cell Atlas into meaningful, interpretable latent dimensions. Moreover, we demonstrate that if applied to batch integration, DRVI's integration quality does not suffer from the disentanglement constraints and instead is on par with entangled integration methods.With its disentangled latent space, DRVI is inherently interpretable and facilitates the identification of rare cell types, provides novel insights into cellular heterogeneity beyond traditional cell types, and highlights developmental stages.