LASSO-MOGAT: A Multi-Omics Graph Attention Framework for Cancer Classification

The application of machine learning methods to analyze changes in geneexpression patterns has recently emerged as a powerful approach in cancerresearch, enhancing our understanding of the molecular mechanisms underpinningcancer development and progression. Combining gene expression data with othertypes of omics data has been reported by numerous works to improve cancerclassification outcomes. Despite these advances, effectively integratinghigh-dimensional multi-omics data and capturing the complex relationshipsacross different biological layers remains challenging. This paper introducesLASSO-MOGAT (LASSO-Multi-Omics Gated ATtention), a novel graph-based deeplearning framework that integrates messenger RNA, microRNA, and DNA methylationdata to classify 31 cancer types. Utilizing differential expression analysiswith LIMMA and LASSO regression for feature selection, and leveraging GraphAttention Networks (GATs) to incorporate protein-protein interaction (PPI)networks, LASSO-MOGAT effectively captures intricate relationships withinmulti-omics data. Experimental validation using five-fold cross-validationdemonstrates the method's precision, reliability, and capacity for providingcomprehensive insights into cancer molecular mechanisms. The computation ofattention coefficients for the edges in the graph by the proposedgraph-attention architecture based on protein-protein interactions provedbeneficial for identifying synergies in multi-omics data for cancerclassification.