M³Surv: Fusing Multi-slide and Multi-omics for Memory-augmented robust Survival prediction.

Area of Science:

• Computational biology
• Biomedical informatics
• Cancer research

Background:

• Personalized oncology relies on multimodal survival prediction.
• Current methods often use only FFPE slides and single omics, neglecting FF slides and multi-omics data.
• Missing data due to clinical constraints limits conventional fusion models.

Purpose of the Study:

• To develop a framework, M³Surv, for integrating multi-pathology slides (FF and FFPE) with multi-omics data.
• To address the challenge of missing modalities in survival prediction models.
• To enhance the accuracy and applicability of cancer survival prediction.

Main Methods:

• Utilized a divide-and-conquer hypergraph learning for multi-slide fusion, capturing intra- and inter-slide relationships.
• Integrated multi-omics data (proteomics, transcriptomics) with pathology features using interactive cross-attention.
• Introduced a prototype-based memory bank to impute missing modalities during inference.

Main Results:

• M³Surv achieved an average 2.2% improvement in C-Index across five TCGA cancer datasets and an in-house dataset.
• Demonstrated superior performance compared to state-of-the-art survival prediction methods.
• Showcased strong stability and robustness in scenarios with missing data modalities.

Conclusions:

• M³Surv effectively integrates diverse data types for multimodal survival prediction.
• The framework's ability to handle missing data enhances its clinical applicability.
• M³Surv offers a promising approach for data-incomplete cancer survival prediction.