MOGANet: A Multi-omics Graph Attention Network for Cancer Diagnosis and Biomarker Identification

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Multi-omics data integration is crucial for understanding complex diseases like cancer.
• Current methods often fail to capture intricate inter-omics relationships, limiting predictive performance and biomarker discovery.

Purpose of the Study:

• To introduce MOGANet, a novel deep learning framework for effective multi-omics data integration.
• To enhance biomedical classification tasks and enable the identification of interpretable biomarkers.

Main Methods:

• Utilizes per-view Graph Convolutional Networks (GCNs) for structured feature extraction.
• Employs a Hierarchical Attention-based Fusion Mechanism (HAFM) to capture hierarchical importance and fuse multi-omics data.
• Evaluated on three The Cancer Genome Atlas (TCGA) cancer datasets: LGG, KIPAN, and BRCA.

Main Results:

• MOGANet consistently outperforms existing methods in cancer classification tasks.
• The framework successfully learns biologically meaningful representations from multi-omics data.
• Identified biomarkers are confirmed to be biologically relevant through functional enrichment analysis.

Conclusions:

• MOGANet provides an effective and interpretable framework for multi-omics integration.
• The approach advances cancer diagnosis and biomarker identification.
• Demonstrates the potential of deep learning for complex biological data analysis.