MMGCN: Multi-modal multi-view graph convolutional networks for cancer prognosis prediction

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Accurate cancer prognosis is crucial for personalized medicine and treatment strategy formulation.
• Integrating multi-modal data (genetic, clinical) with deep learning models can improve prediction accuracy.
• Existing methods often fail to fully leverage patient similarities or capture diverse information effectively.

Purpose of the Study:

• To propose a novel framework, multi-modal multi-view graph convolutional networks (MMGCN), for enhanced cancer prognosis prediction.
• To address limitations in existing multi-modal deep learning methods regarding patient similarity and information integration.
• To improve the accuracy and effectiveness of personalized cancer treatment strategies.

Main Methods:

• Utilized similarity network fusion (SNF) to merge patient similarity networks from gene expression, copy number alteration, and clinical data.
• Developed multi-view graph convolutional networks (GCNs) with a view-level attention mechanism to capture diverse patient similarities.
• Incorporated an edge homophily prediction module to mitigate the impact of heterophilic edges on GCN performance.

Main Results:

• MMGCN demonstrated superior performance compared to state-of-the-art methods on four public cancer datasets (METABRIC, TCGA-BRCA, TCGA-LGG, TCGA-LUSC).
• Achieved high area under the receiver operating characteristic curve (AUROC) scores, including 0.827 ± 0.005 for METABRIC, 0.805 ± 0.014 for TCGA-BRCA, 0.925 ± 0.007 for TCGA-LGG, and 0.746 ± 0.013 for TCGA-LUSC.
• The framework effectively integrates multi-modal patient information for robust prognosis prediction.

Conclusions:

• The proposed MMGCN framework significantly enhances multi-modal cancer prognosis prediction by deeply exploring patient similarities from various perspectives.
• The study highlights the potential of multi-view graph convolutional networks in leveraging complex biological data for improved clinical outcomes.
• The source code for MMGCN is publicly available, facilitating further research and application in cancer prognosis.