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

Single-cell RNA sequencing data has become a valuable tool for studying the complex regulatory processes involved in continuous cell-state transitions, such as the epithelial-to-mesenchymal transition. However, the intricate dynamics of these gene programs remain challenging to unravel using current methods. Deep neural network models can map these complex topologies, but they lack the gene-level interpretability necessary for experimental practitioners.

To address this issue, researchers have introduced scCont, a novel unsupervised contrastive learning framework designed to extract interpretable functional transition units from single-cell RNA-seq data.

Unlike traditional approaches that rely on explicit temporal labels or clustering priors, scCont employs an unsupervised k-nearest neighbors method to project local transcriptomic neighborhoods into a low-dimensional embedding space. During post-training analysis, scCont utilizes Shapley network attribution to determine relationships between genes and latent dimensions, effectively mapping these connections to functional gene programs.

To validate the effectiveness of their method, the researchers compared scCont against baseline methods such as LDVAE across 13 diverse epithelial-to-mesenchymal transition timecourse datasets, as well as four simulated datasets.

The results demonstrated that scCont captures biologically interpretable dynamics more effectively than the baseline methods, providing a valuable tool for bridging the gap between deep learning representation and translational biological utility. By exposing interpretable functional dynamics and effectively mapping context-dependent functional groups, scCont offers a promising solution for understanding the complex gene programs underlying cellular state transitions. The software is now available as open-source code on GitHub (https://github.com/nramani611/scCont).