MCKG-SL: Knowledge graph-based multi-feature cross-aggregation synthetic lethality prediction for KRAS gene.

Abstract

KRAS (Kirsten rat sarcoma viral oncogene homolog) is the most commonly mutated oncogene in human cancer. Targeting synthetic lethal (SL) partners in the setting of oncogenic KRAS is an alternative therapeutic strategy for KRAS-mutant malignancies. However, existing SL prediction algorithms are limited by incomplete understanding of complex biological system interaction networks or ignore some information in the local association of gene pairs. To overcome these challenges, we propose a novel Knowledge Graph-based Synthetic Lethality model named MCKG-SL, which learns the interaction information between genes with multi-feature cross aggregation. First, MCKG-SL extract local association subgraph of gene pairs from the knowledge graph, to focus on the local association information around gene pairs. Then, we utilize Relational Graph Convolutional Network (RGCN) for global relational awareness and Graph Attention Network (GAT) for partial connection concern to learn the gene feature information in the subgraph. Subsequently, we design a multi-feature cross aggregation module to cross-fuse the relational features learned from the local association subgraph with biological features extracted from multi-omics data, enhancing the interactive learning of gene pair features. Across repeated random pair-wise splits, MCKG-SL achieved strong mean performance relative to the evaluated baseline methods. Besides, pathway analysis and SL analysis with MCKG-SL suggest that there is a potential synthetic lethal relationship between KRAS and CDK3 (Cyclin-Dependent Kinase 3), and synthetic lethal between KRAS and TP53 (Tumor Protein P53) play an important role in the Bladder cancer. The code and data are available on https://github.com/Qian0711/MCKG-SL.