scCont: Interpretable Contrastive Learning for Functional Characterization of Cellular State Transitions
- 1Boston University
Abstract
Timecourse single-cell RNA (scRNA) datasets have become indispensable biological tools for studying and uncovering the underlying regulatory processes driving continuous cell-state transitions such as the epithelial-to-mesenchymal transition. However, these highly fluid, non-linear gene program dynamics remain difficult to disentangle. While deep neural network models can effectively map complex topologies, they are biological black boxes, lacking the gene-level interpretability required by experimental practitioners. Here, we introduce scCont, a fully unsupervised contrastive learning framework for extracting interpretable functional transition units from scRNA-seq data. Rather than relying on explicit temporal labels or clustering priors, scCont uses an unsupervised k-nearest neighbors approach to project local transcriptomic neighborhoods into a low-dimensional embedding. During post-training analysis, scCont employs Shapley network attribution to determine gene-latent relationships, mapping latent dimensions directly into functional gene programs. We quantitatively and qualitatively validate this representation by demonstrating that scCont captures biologically interpretable dynamics more effectively than baseline methods such as LDVAE. Evaluated across 13 diverse EMT timecourse datasets and four simulated datasets, scCont exposes interpretable functional dynamics and effectively maps context-dependent functional groups, helping to bridge the gap between deep learning representation and translational biological utility. scCont is an open source software available on GitHub (https://github.com/nramani611/scCont).
Related
- Predicting miRNA–Disease Associations Based on Contrastive Learning of Linear and Nonlinear Features
- Interpretable spherical geometry of single-cell state transitions from dominant principal components
- How effective are contrastive learning-based approaches for activity-cliff prediction?
- Revealing pan-cancer clonal niches of T cells from single-cell RNA sequencing using contrastive learning
- Pocket contrastive learning and conformation fusion for robust structure-conditioned 3D molecular generation under receptor conformational variation.
BibTeX
@misc{kewalramani2026sccont,
title = {{scCont: Interpretable Contrastive Learning for Functional Characterization of Cellular State Transitions}},
author = {N. Kewalramani and S. Wuchty and M. Crovella and I. Paul and A. Emili},
howpublished = {bioRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.05.764599},
url = {https://www.biorxiv.org/content/10.64898/2026.11.05.764599v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - scCont: Interpretable Contrastive Learning for Functional Characterization of Cellular State Transitions
AU - N. Kewalramani
AU - S. Wuchty
AU - M. Crovella
AU - I. Paul
AU - A. Emili
PY - 2026
DA - 2026/11/05
PB - bioRxiv
DO - 10.64898/2026.11.05.764599
UR - https://www.biorxiv.org/content/10.64898/2026.11.05.764599v1
AB - Timecourse single-cell RNA (scRNA) datasets have become indispensable biological tools for studying and uncovering the underlying regulatory processes driving continuous cell-state transitions such as the epithelial-to-mesenchymal transition. However, these highly fluid, non-linear gene program dynamics remain difficult to disentangle. While deep neural network models can effectively map complex topologies, they are biological black boxes, lacking the gene-level interpretability required by experimental practitioners. Here, we introduce scCont, a fully unsupervised contrastive learning framework for extracting interpretable functional transition units from scRNA-seq data. Rather than relying on explicit temporal labels or clustering priors, scCont uses an unsupervised k-nearest neighbors approach to project local transcriptomic neighborhoods into a low-dimensional embedding. During post-training analysis, scCont employs Shapley network attribution to determine gene-latent relationships, mapping latent dimensions directly into functional gene programs. We quantitatively and qualitatively validate this representation by demonstrating that scCont captures biologically interpretable dynamics more effectively than baseline methods such as LDVAE. Evaluated across 13 diverse EMT timecourse datasets and four simulated datasets, scCont exposes interpretable functional dynamics and effectively maps context-dependent functional groups, helping to bridge the gap between deep learning representation and translational biological utility. scCont is an open source software available on GitHub (https://github.com/nramani611/scCont).
ER -