CellUntangler: Separating distinct biological signals in single-cell data with deep generative models

Single-cell RNA sequencing has provided new insights into both intracellular and intercellular processes. However, multiple processes, such as cell-type programs, differentiation, and the cell cycle, often occur simultaneously within one cell. Existing methods typically target a single process and impose restrictive assumptions, risking the loss of valuable biological information. We introduce CellUntangler, a deep generative model that embeds cells into a latent space composed of multiple subspaces, each tailored with an appropriate geometry to capture a distinct signal. Applied to datasets of cycling-only and mixed cycling/non-cycling cells, CellUntangler disentangles the cell cycle from other processes such as cell type. The framework generalizes to disentangle additional signals, including spatial, tissue dissociation, interferon response, and cell-type identity. By providing flexible embeddings to capture various signals, CellUntangler enables selective enhancement or filtering of signals at the gene-expression level, offering a powerful tool for disentangling complex biological processes in single-cell data.