MoESurv: a zero-sample and transferable survival prediction framework for rare cancers using mixture of experts.

MoESurv achieves a 4% higher C-index than existing models in predicting survival for seven rare cancer types using a zero-sample, transfer learning approach.

• Why it matters: Accurate survival prediction for rare cancers is hindered by limited data, impeding personalized treatment and clinical decision-making.
• What they did: The framework employs a mixture-of-experts architecture within an autoencoder, integrating shared, cancer-specific, and routing experts trained on pan-cancer data, with validation across diverse cohorts.
• The result: MoESurv demonstrates superior robustness and generalizability, effectively stratifies patient risk, and identifies survival-related genes, supporting its clinical utility and biological interpretability.