MOGAT: A Multi-Omics Integration Framework Using Graph Attention Networks for Cancer Subtype Prediction

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate cancer subtype prediction is vital for personalized medicine.
• Multi-omics data integration offers a comprehensive view of cancer pathophysiology.
• Existing graph convolutional network (GCN) models like MOGONET and SUPREME lack the ability to weigh the importance of neighboring nodes in omics data.

Purpose of the Study:

• To propose MOGAT, a novel multi-omics integration approach using graph attention networks (GAT) for improved cancer subtype prediction.
• To address the limitations of GCNs by incorporating an attention mechanism to prioritize informative neighbors in omics data.
• To evaluate the efficacy of MOGAT in cancer subtype prediction using breast cancer datasets.

Main Methods:

• Developed MOGAT, a multi-omics integration model leveraging graph attention networks (GAT).
• Implemented a multi-head attention mechanism within MOGAT to assign unique attention coefficients to neighboring samples.
• Evaluated MOGAT on TCGA and METABRIC breast cancer datasets.

Main Results:

• MOGAT significantly outperforms MOGONET (32-46% improvement) and SUPREME (2-16% improvement) in cancer subtype prediction.
• The GAT embeddings generated by MOGAT demonstrate superior prognostic capability compared to raw features in distinguishing high-risk from low-risk cancer groups.
• The study is the first to explore GAT for multi-omics integration in cancer subtype prediction.

Conclusions:

• MOGAT, by incorporating graph attention mechanisms, enhances multi-omics integration for more accurate cancer subtype prediction.
• The attention mechanism in MOGAT effectively captures the importance of neighboring nodes, leading to superior performance.
• MOGAT represents a significant advancement in leveraging multi-omics data for precision oncology and improved patient stratification.