scCont: Interpretable Contrastive Learning for Functional Characterization of Cellular State Transitions

scCont learns interpretable functional transition units from timecourse scRNA-seq by projecting local transcriptomic neighborhoods into a low-dimensional embedding and attributing genes to latent programs.

• Why it matters: Continuous cell-state transitions like epithelial-to-mesenchymal transition involve fluid, non-linear gene dynamics that are hard to disentangle, and existing deep models lack gene-level interpretability.
• What they did: scCont is fully unsupervised: it uses an unsupervised k-nearest neighbors projection and then applies Shapley network attribution to link latent dimensions to functional gene programs.
• The result: Across 13 EMT timecourse datasets and four simulated datasets, scCont captures biologically interpretable dynamics more effectively than LDVAE, enabling context-dependent functional group mapping.

AI summary of the abstract · may contain errors · not medical advice · read the paper.