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

Multi-omics integration holds considerable promise for advancing disease understanding and improving the performance of biomedical classification tasks. However, existing methods often struggle to comprehensively capture the complex relationships within and between different omics modalities, thereby limiting both interpretability and predictive power. In this study, we propose MOGANet, a novel deep learning framework designed to integrate multi-omics data through per-view Graph Convolutional Networks (GCNs) for extracting structured features and a Hierarchical Attention-based Fusion Mechanism (HAFM) for capturing hierarchical importance and effectively fusing multi-omics representations. This design enables MOGANet to learn informative and biologically meaningful representations while facilitating the identification of interpretable biomarkers. Comprehensive evaluations on three cancer datasets from The Cancer Genome Atlas (TCGA), including Lower Grade Glioma (LGG), Kidney Pan-Cancer (KIPAN), and Breast Invasive Carcinoma (BRCA), demonstrate that MOGANet consistently outperforms traditional and state-of-the-art methods in classification tasks. Furthermore, functional enrichment analysis confirms the biological relevance of the identified biomarkers. Collectively, these results highlight the effectiveness and interpretability of MOGANet as a robust framework for multi-omics integration in cancer diagnosis and biomarker identification.