Multi-omic data integration model based on inductive graph learning and cascaded co-attention for cancer classification

Abstract A comprehensive analysis of data from multiple omics is essential for devising precise therapeutic strategies for complex diseases. The effective integration and analysis of multi-omic data can boost the precise classification of disease subtypes. However, existing methods for multi-omic data integration learning suffer from the problem of ignoring the importance of inter-omics features and the latent features within and across different omics data, as well as the fact that different types of omics data can present unique features in the high-level feature space. To address this issue, we propose a multi-omic data integration learning model, called IGLCCA4omic, based on inductive graph learning and cascaded co-attention mechanism to learn the features of multi-omic data better and improve cancer classification in this work. IGLCCA4omic, on the one hand, combines cosine-based patient similarity networks with GraphSAGE to efficiently capture intrinsic associations of latent features within each type of omics data. On the other hand, IGLCCA4omic develops a novel attention mechanism called cascaded co-attention to fuse features learned from multi-omic data and focus on their important feature information, comprehensively capturing and integrating the representative or important features and complementary information across multi-omics to improve cancer classification. We conducted extensive experiments on four benchmark datasets comprising miRNA expression, mRNA expression, and DNA methylation data to demonstrate the effectiveness and superiority of IGLCCA4omic. The experimental results show that IGLCCA4omic outperforms state-of-the-art methods, effectively improving cancer classification.