Multi-omics data integration via novel interpretable k-hop graph attention network for signaling network inference

With the advent of sequencing technology, large-scale multi-omics data have been generated to understand the diversity and heterogeneity of genetic targets and associated complex signaling pathways at multiple levels in diseases, which are critical targets to guide the development of personalized precision medicine. However, it remains a challenging task to computationally mine a few essential targets and pathways from a large number of variables characterized by the multi-level multi-omics data. In this study, we proposed a novel interpretable k-hop graph attention network model, k-hop GAT, to integrate the multi-omics data to infer the essential targets and related signaling networks. We evaluated the proposed model using the multi-omics data, i.e., genetic mutation, copy number variation, methylation, gene expression data, of 332 cancer lines; and the experimentally identified essential targets. The validation and comparison results indicated that the proposed model outperformed the GAT and graph convolutional network (GCN) models.