Multi-Omics Cancer Subtype Classification with Shared-Specific Feature Learning and Graph Attention Network

Advances in sequencing technologies have enabled the generation of vast multi-omics data, offering valuable insights for cancer diagnosis and prognosis. However, the heterogeneity of multi-omics data poses challenges for effective integration. Existing methods often overlook the decoupling and joint use of shared and specific information (SSI), limiting performance and interoperability. Moreover, some methods often ignore the latent structural relationships among patient samples, which are critical for accurate subtype classification. To address these challenges, we propose MCSGAT, a novel framework that explicitly extracts and utilizes SSI from multi-omics data. We employ multilayer autoencoders to capture omics-shared and omics-specific representations and introduce an omics-view classifier to effectively decouple them, ensuring the independence and effectiveness of the information. To further exploit structural relationships, we incorporate the multi-head graph attention network to construct separate classifiers for shared and specific information. Extensive experiments on four public multi-omics datasets validate that MCSGAT consistently outperforms the state-of-the-art methods.