Single-cell phenotype-associated subpopulation identification via transfer foundation model and statistical ensemble...

Single-cell RNA sequencing (scRNA-seq) enables the characterization of cell types, states, and lineages within heterogeneous tissues, thereby providing unprecedented opportunities to dissect cellular heterogeneity. However, single-cell data alone cannot directly establish cell-phenotype relationships, which pose major challenges in linking cellular heterogeneity to complex traits and disease outcomes.

Here, we introduce scPASI, which integrates single-cell and bulk-level information to uncover phenotype-associated cell subpopulations. scPASI combines a pre-trained foundation model (PFM) with a residual variational autoencoder (Res-VAE) to extract feature embeddings of cells and samples. Cell clusters are calculated using the Leiden algorithm, after which phenotype associations are inferred based on regression coefficients derived from LASSO and sparse group LASSO (SGL) models. This design enables scPASI to stratify cells into four subpopulations with different levels of phenotype association: strongly positive (SP), weakly positive (WP), strongly negative (SN), and weakly negative (WN) groups. Furthermore, scPASI characterizes phenotype-relevant genes within subpopulations and provides insights into the relationship between cellular heterogeneity and bulk phenotypes. Extensive evaluations across diverse datasets show that scPASI outperforms existing methods and generalizes well across multiple phenotype settings, including tumor status, genetic mutations, and clinical prognosis. Biological analyses demonstrate that signature genes derived from the identified subpopulations can distinguish tumor cells, genetic alterations, and survival outcomes.

By bridging single-cell transcriptomics with phenotype information, scPASI can uncover biologically meaningful cell-phenotype associations underlying tumor biology, enabling the identification of disease-relevant subpopulations and providing a framework for potential therapeutic targeting.