MMOSurv: meta-learning for few-shot survival analysis with multi-omics data

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• High-throughput techniques generate vast multi-omics data, enhancing survival prediction accuracy.
• Integrating multi-omics data for few-shot survival prediction, especially for rare cancers, remains a significant challenge.

Purpose of the Study:

• To develop a meta-learning framework (MMOSurv) for accurate multi-omics few-shot survival analysis.
• To enable effective survival prediction from limited samples by leveraging meta-knowledge across related cancer types.

Main Methods:

• MMOSurv employs a deep Cox survival model integrating multiple omics data.
• It learns adaptable parameter initialization from abundant data of relevant cancers.
• Parameters are rapidly adapted for target cancer tasks using few training samples.

Main Results:

• MMOSurv effectively utilizes meta-information from similar omics data across relevant cancers.
• It outperforms single-omics meta-learning methods in few-shot survival prediction.
• MMOSurv demonstrates superior performance compared to multitask learning and pretraining strategies.

Conclusions:

• MMOSurv provides a robust solution for few-shot survival prediction using multi-omics data.
• The framework enhances prediction accuracy by transferring knowledge across cancer types.
• MMOSurv represents a significant advancement in personalized cancer survival analysis.