Leveraging single-cell foundation models for accurate survival outcome prediction

Area of Science:

• Computational biology
• Cancer research
• Genomics

Background:

• Foundation models trained on single-cell transcriptomes offer deep molecular insights into cellular states.
• The application of these models for cancer survival prediction using bulk RNA sequencing data is underexplored.

Purpose of the Study:

• To evaluate the prognostic value of single-cell foundation model embeddings for cancer survival prediction.
• To develop a multi-modal model integrating these embeddings with gene expression and clinical data.

Main Methods:

• Applied the scFoundation model to generate patient-level embeddings from TCGA data across 25 cancer types.
• Developed the Embedding-Gene-Survival Prediction (EGSP) model, integrating embeddings, gene expression, and clinical variables.
• Conducted comparative analyses against single-modality and existing multi-omics survival models.

Main Results:

• The EGSP model achieved a mean concordance index (C-index) of 0.724 across cancers, exceeding 0.8 in seven types.
• Embeddings from pretrained scFoundation weights showed lower redundancy with gene expression and retained complementary prognostic signals.
• Explainable AI revealed embeddings capture interpretable biological programs related to differentiation, immune activity, and tumor growth.

Conclusions:

• Single-cell foundation model embeddings provide biologically meaningful and partially non-redundant survival signals.
• These embeddings substantially improve bulk RNA-seq-based cancer prognostic modeling.
• The EGSP model offers transparent survival prediction at both cohort and patient levels.