scDisent: disentangled representation learning with causal structure for multi-omic single-cell analysis

Abstract

Single-cell multi-omic technologies measure complementary aspects of cellular identity and regulatory state, yet most integration models compress these signals into one entangled latent space. Such representations are useful for clustering but poorly suited to regulator-centered interpretation or perturbation-oriented analysis. We present scDisent (https://github.com/xiguoren/scDisent), a generative framework that separates expression-associated variables (zexpr) from regulation-associated variables (zreg) and links them through a sparse directed mapping. scDisent combines modality-specific encoding, variational disentanglement, total-correlation and orthogonality regularization, and a Gumbel-gated causal module protected by detach-based gradient isolation. Across benchmark datasets with matched modalities, scDisent achieved the strongest clustering performance among the tested methods while exposing regulatory structure that competing integration models do not represent explicitly. The learned causal atlas remained sparse, perturbation analyses recovered biologically coherent lineage-associated programs, and branch-separation analyses showed that benchmark-label information concentrated in zexpr rather than zreg. These results position scDisent as a multi-omic representation model that improves both integration quality and biological interpretability