scCont turns single-cell time courses into readable gene modules: nearest-neighbor pairing plus Shapley attribution splits EMT into interpretable transition units

Time-course single-cell RNA sequencing records how cells change identity over days, but these transitions are continuous and nonlinear, and existing analyses either use linear factorization to cut cells into discrete subpopulations or use neural networks whose embeddings are hard to read in gene terms. scCont puts both jobs in one pipeline: it preserves the continuous topology of the transition and translates each latent dimension into a statistically defined gene set. Earlier interpretable routes are represented by non-negative matrix factorization, which separates cell types well but struggles with highly plastic transition states, and by the linearly decoded variational autoencoder LDVAE, which constrains the decoder to linear mappings to keep gene loadings readable. Contrastive learning in single-cell analysis has mostly built positive pairs from augmentations such as random gene dropout and Gaussian noise, and has not been optimized for state transitions in time-course data. On a SERGIO-simulated bifurcating dataset at k=10, background genes make up 7.0% of the 100 highest-attributed genes per latent, about ten-fold below the 73.3% chance level; in the same setting, a multilayer perceptron probe predicts ground-truth pseudotime from the embedding at a cross-validated R² of 0.84, against 0.60 for a 32-component PCA of the same matrix. Program-gene recovery reaches an AUPRC of 0.26, and the authors report LDVAE is slightly better on this simulated set but the two remain comparable, with the result changing across simulation conditions.

In practice, scCont first finds neighboring cells in normalized expression space with a k-nearest-neighbor graph and treats them as positive pairs, so contrastive learning pulls topologically adjacent cells together in a low-dimensional space and keeps the transition continuous without time labels. After training, it projects latent-unit activations back into the gene-expression PCA space and clusters them spatially, merging units that share spatiotemporal patterns into modular transition units, then scores each gene with Shapley attribution to find the genes driving each latent dimension. Attribution uses the DeepSHAP explainer from SHAP with a reference distribution of 100 cells to compute each gene's mean absolute Shapley value, selects significant genes at a threshold of z > 2.57 and p < 0.01, and runs GO biological-process enrichment with g:Profiler at an adjusted p < 0.05. The encoder is a four-layer multilayer perceptron mapping 3000 input genes to a 32-dimensional latent space, and a two-layer projector outputs 16 dimensions for the InfoNCE loss with a temperature of 0.20. On the epithelial-to-mesenchymal transition data, the authors swept k from 3 to 10 on MCF10A cells treated with TGFB1 and sequenced daily for 8 days, choosing k=3 from the participation ratio of the latent covariance, the mean Jaccard overlap of enriched GO terms between latents, and the fraction of positive pairs from the same timepoint, and used it for all EMT analyses. Spatial clustering on that dataset yields three modules, enriched respectively for anchoring junctions and cell motility, for background processes such as tissue development, and for metabolic processes and response to hormone.

The demonstrated scope is four SERGIO simulated datasets plus MCF10A treated with TGFB1 and the 12 EMT time-course datasets from Cook et al., covering the A549, MCF7, DU145 and OVCA420 cell lines with EGF, TGFB1 and TNF inducers; the downstream cross-dataset module alignment focuses on A549. The authors propose connecting trajectory-inference methods such as Monocle3 to scCont's modular latent space as a next direction.

The authors note that the method's success depends on the initial k-nearest-neighbor graph accurately reflecting true biological progression; compared with PCA and NMF, scCont performs better with GPU acceleration and scales less efficiently with dataset size; SHAP's computational overhead grows linearly with the number of cells and becomes a bottleneck at atlas scale. On simulated data LDVAE is slightly better than scCont at program recovery, and that result changes with simulation conditions.