Leveraging single-cell foundation models for accurate survival outcome prediction

Foundation models trained on large-scale single-cell transcriptomes can capture rich molecular representations of cellular states, yet their potential for cancer survival prediction from bulk RNA-seq data remains largely unexplored.

We applied the single-cell foundation model scFoundation to derive patient-level embeddings across 25 cancer types from TCGA and systematically evaluated their prognostic value under both cancer-specific and pan-cancer settings. To leverage complementary information, we developed an Embedding-Gene-Survival Prediction (EGSP) model that integrates foundation model embeddings with gene expression and clinical variables. EGSP achieved a mean concordance index (C-index) of 0.724 across cancers and exceeded 0.8 in seven cancer types, consistently outperforming single-modality models and existing multi-omics survival approaches. Comparative analyses showed that embeddings derived from pretrained scFoundation weights exhibited lower redundancy with gene expression while retaining complementary prognostic signals relative to pan-cancer fine-tuned embeddings. Explainable AI analyses further revealed that prognostic embeddings capture interpretable biological programs related to tumor differentiation, immune activity, and tumor-intrinsic growth, enabling transparent survival prediction at both cohort and patient levels. Overall, single-cell foundation model embeddings provide biologically meaningful and partially non-redundant survival signals that substantially improve bulk RNA-seq-based prognostic modeling.

https://github.com/weiliu123/EGSP.