DEMONet: A Dual-channel Multi-omics Integration Hypergraph Network for Cancer Gene Identification

Chao Deng, Kai Zheng, Zeyu Zhong, Hongdong Li, Jianxin Wang

Keywords: Cancer Gene Identification; Hypergraph; Multi-omics; Hypergraph Neural Network

Abstract: Identifying cancer driver genes is a critical challenge in cancer genomics. While hypergraph neural networks are powerful tools for identifying cancer genes by modeling higher-order functional relationships, they face a critical limitation in integrating multi-omics features. To address this, we propose DEMONet, a Dual-channel Multi-Omics Integration Hypergraph Network. DEMONet enhances multi-omics integration through three synergistic modules: (1) a tree-based sparse encoder that transforms raw multi-omics features into a robust, structured representation; (2) a biologically-informed node-weighted hypergraph convolutional layer to capture gene importances within functional hyperedges; and (3) a dual-channel architecture to prevent information interference between different hypergraph sources before final fusion. Benchmark results demonstrate that DEMONet outperforms existing state-of-the-art methods, improving AUROC by 1.9\% and AUPRC by 2.3\% over prior methods. Its generalization and robustness are further validated on two independent test sets. Analysis of multiple independent functional genomics data validated the significant biological associations between the DEMONet-predicted top candidate genes and cancer genes. Furthermore, TCGA survival analysis further reveals that 16 novel cancer genes identified by DEMONet are significantly associated with patient outcomes, highlighting the potential of our model to discover actionable targets for cancer research.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 11029