DriverOmicsNet: An Integrated Graph Convolutional Network for Multi-Omics Exploration of Cancer Driver Genes

Abstract

AbstractBackgroundCancer is a complex and heterogeneous group of diseases driven by genetic mutations and molecular changes. Identifying and characterizing cancer driver genes (CDgs) is crucial for understanding cancer biology and guiding precision oncology. Integrating multi-omics data can reveal the intricate molecular interactions underlying cancer progression and treatment responses.MethodsWe developed a graph convolutional network (GCN) framework, DriverOmicsNet, that integrates multi-omics data using STRING protein-protein interaction (PPI) networks and correlation-based weighted correlation network analysis (WGCNA). We applied this framework to 15 cancer types, analyzing 5555 tumor samples to predict cancer-related features such as homologous recombination deficiency (HRD), cancer stemness, immune clusters, tumor stage, and survival outcomes.FindingsDriverOmicsNet demonstrated superior predictive accuracy and model performance metrics across all target labels when compared with GCN models based on STRING network alone. Gene expression emerged as the most significant feature, reflecting the dynamic and functional state of cancer cells. The combined use of STRING PPI and WGCNA networks enhanced the identification of key driver genes and their interactions.InterpretationOur study highlights the effectiveness of using GCNs to integrate multi-omics data for precision oncology. The integration of STRING PPI and WGCNA networks provides a comprehensive framework that improves predictive power and facilitates the understanding of cancer biology, paving the way for more tailored treatments.