scSurv: a deep generative model for single-cell survival analysis

Area of Science:

• Computational biology
• Cancer research
• Genomics

Background:

• Single-cell omics reveals tumor cell heterogeneity.
• Current methods lack single-cell resolution for linking heterogeneity to patient survival.
• Understanding cellular contributions to outcomes is crucial for personalized medicine.

Purpose of the Study:

• Introduce scSurv, a novel computational framework.
• Quantify individual cellular contributions to clinical outcomes at single-cell resolution.
• Integrate Cox proportional hazards models with deep generative models of single-cell transcriptomes.

Main Methods:

• Developed scSurv, combining Cox proportional hazards model and deep generative model.
• Applied scSurv to simulated and real single-cell omics datasets.
• Validated accuracy and identified prognostic cells and genes.

Main Results:

• scSurv accurately estimates cellular contributions to patient survival.
• Identified cells and genes associated with favorable or adverse prognoses.
• Reproduced known prognostic macrophage classifications in melanoma and enabled hazard mapping in renal cell carcinoma via spatial transcriptomics.

Conclusions:

• scSurv provides a novel framework for analyzing single-cell data in relation to clinical outcomes.
• The method advances understanding of tumor heterogeneity's impact on survival.
• Demonstrated applicability across various cancers and infectious diseases, highlighting its versatility.