Multiview-cooperated graph neural network enables novel multi-omics cancer subtype classification

Cancer presents a significant challenge in the field of public health due to its high incidence, mortality rate, and inherent heterogeneity. Integrating multi-omics biological data offers a comprehensive and intricate understanding of biological processes, disease mechanisms, and cancer subtyping, rendering it an influential tool for scientific research. Nevertheless, current approaches to integrating multi-omics data often fail to consider the scale of data in feature information and overlook the analysis of the individual and shared feature expressions in multi-omics data. The proposed study introduces a Multiview-Cooperated graph neural network (MCgnn), an end-to-end cancer subtype classifier that effectively integrates and analyzes complex multi-omics data. Firstly, MCgnn innovatively constructs a similarity network using Mahalanobis distance and density methods. Then, by employing stacked graph convolution layers, MCgnn effectively captures potential local structural features. Subsequently, MCgnn extracts and fuses complementary information from different omics data through the attention mechanism among views to achieve effective integration of multiple views and attain higher classification performance. Finally, MCgnn performs multi-task learning across omics data using a cross-omics tensor to seamlessly integrate the feature learning component with the classification component. Experiments on four publicly available The Cancer Genome Atlas (TCGA) datasets were conducted to demonstrate that MCgnn outperforms most comparison classification algorithms in effectively addressing the cancer subtype classification problem and exhibits remarkable robustness and generalization capabilities. Additionally, MCgnn was employed to identify pivotal biomarkers in cancer, providing a valuable reference for precision medicine.