MoGraphDRP: Multi-omics and graph fusion with bilinear attention for predicting drug sensitivity

Abstract

Accurate prediction of drug response in cancer cells is a fundamental step toward achieving precision medicine and designing personalized therapies. In this study, a multi-branch deep learning framework is proposed that integrates multi-omics cellular data including gene expression, mutation, methylation, and biological pathways with structural features of drugs (molecular graphs and various chemical fingerprints) to enable drug response prediction. The graph structure of the drug is modeled using a three-layer Graph Convolutional Network (GCN), and chemical fingerprints are compressed using MLP networks. These multiple representations of drugs are integrated and then combined with cellular features in a Multi-head Bilinear Attention module to model the complex interactions between cells and drugs. In the final stage, an ensemble model based on XGBoost is used to refine the outputs. The MoGraphDRP model demonstrates significantly higher accuracy in drug response prediction compared to existing state-of-the-art methods. Experimental results show that the MoGraphDRP model outperforms advanced methods such as BANDRP, DeepCDR, and DeepTTA, achieving PCC = 0.9689, RMSE = 0.6622, and R² = 0.9388. This model not only accurately reconstructs missing IC50 values but also effectively distinguishes between sensitive and resistant drugs in unknown combinations. The MoGraphDRP framework can serve as a powerful, interpretable, and reliable tool for analyzing drug response and designing preclinical treatments.