GONet: A Molecular Interaction-Guided Graph Learning Framework for Multi-Omics Cancer Classification.

GONet, a hierarchical GNN integrating multi-omics with molecular interaction networks, achieves an average AUROC of 0.9262 on mRNA and miRNA datasets across 10 cancer types.

• Why it matters: Existing multi-omics graph methods often use shallow learning or incomplete fusion, limiting their ability to capture complementary information and complex dependencies, reducing performance and generalizability in cancer classification.
• What they did: GONet builds a graph where nodes are omics features and edges are molecular interactions, learning global representations with an autoencoder, local structure with Deep Graph Infomax (DGI), and fusing heterogeneous omics via gated attention.
• The result: On 13 benchmark datasets, it improves AUROC by 13.77% versus representative GNN methods and 6.35% versus non-GNN methods, and on an independent CPTAC cohort it improves by 25.22% and 13.91%, with proteomics adding 5.56% for cancer stage classification.