MoESurv: A Zero-Sample and Transferable Survival Prediction Framework for Rare Cancers Using Mixture of Experts

Area of Science:

• Computational biology
• Machine learning in oncology
• Genomics and personalized medicine

Background:

• Accurate cancer survival prediction is vital for personalized treatment but hindered by limited data in rare cancers.
• Deep learning models typically require large datasets, posing a challenge for rare cancer types.
• Existing methods face a clinical bottleneck due to data scarcity in rare cancer research.

Purpose of the Study:

• To develop a zero-sample survival prediction framework for rare cancers.
• To overcome data limitations in rare cancer survival analysis using deep learning.
• To enhance personalized cancer treatment strategies through improved prognostic accuracy.

Main Methods:

• Proposed MoESurv, a zero-sample survival prediction framework utilizing a mixture-of-experts architecture.
• Integrated shared, cancer-specific, and routing experts within an autoencoder to disentangle survival features.
• Leveraged pan-cancer data to extract generalizable prognostic patterns.

Main Results:

• MoESurv achieved state-of-the-art performance on seven rare TCGA cancer types, improving the average C-index by 4% over baselines.
• External validation across diverse cohorts, including Chinese glioma and pan-cancer PCAWG, confirmed MoESurv's robust and generalizable predictive performance.
• The model effectively stratified patient risk groups and identified potential survival-associated genes, demonstrating clinical utility and biological interpretability.

Conclusions:

• MoESurv offers a powerful solution for survival prediction in rare cancers, addressing critical data limitations.
• The framework demonstrates significant potential for advancing personalized cancer treatment through accurate prognostic modeling.
• MoESurv's ability to provide biological insights enhances its value for both clinical application and future research.