Comparative Analysis of Multi-Omics Integration Using Graph Neural Networks for Cancer Classification.

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Integrating multiple omics data offers advanced understanding of cancer.
• Graph neural networks (GNNs) show promise in modeling complex biological relationships for cancer classification.
• Challenges persist in high-dimensional multi-omics data integration and effective graph construction.

Purpose of the Study:

• To evaluate GNN architectures (GCN, GAT, GTN) for multi-omics data integration in cancer classification.
• To develop dimensionality reduction and feature selection methods for multi-omics data.
• To compare different graph structure construction strategies.

Main Methods:

• Applied graph-convolutional networks (GCN), graph-attention networks (GAT), and graph-transformer networks (GTN) for multi-omics integration.
• Utilized differential gene expression and LASSO regression for feature selection, creating LASSO-MOGCN, LASSO-MOGAT, and LASSO-MOGTN models.
• Constructed graph structures using sample correlation matrices and protein-protein interaction networks.

Main Results:

• Multi-omics integration models outperformed single-omics models.
• LASSO-MOGAT achieved the highest accuracy (95.9%) in cancer classification.
• Correlation-based graph structures were more effective than protein-protein interaction networks for identifying shared cancer signatures.

Conclusions:

• GNNs are effective for multi-omics data integration in cancer research.
• LASSO-MOGAT demonstrates superior performance for cancer classification.
• Sample correlation matrices provide a robust foundation for graph construction in multi-omics analysis.