HMVPS-DRP: Hybrid multi-view representation learning and pathway-state-aware graph learning for interpretable cancer drug response prediction with multi-omics data

Cancer drug response prediction is an important task in precision oncology and can support personalized treatment and candidate anticancer drug screening. Although many deep learning methods have been developed, existing approaches still do not fully exploit complementary information from different molecular views of drugs. Meanwhile, multi-omics data are often processed by independent encoders or simple fusion strategies, which limits the characterization of cell line molecular states. In addition, jointly modeling pathway structural priors and cell-line-specific pathway activity remains challenging. To address these limitations, we propose HMVPS-DRP, an interpretable multi-omics framework that integrates hybrid multi-view representation learning with pathway-state-aware graph learning. HMVPS-DRP first jointly learns molecular graph topology, SMILES sequence semantics, and substructure information from ECFP fingerprints to obtain complementary and comprehensive drug representations. Meanwhile, pathway connections based on gene-set overlap and cell-line-specific pathway activity states are integrated into pathway-cell line heterogeneous graphs, allowing the model to combine pathway structural information with cell-line-specific functional states. We further design a condition aware aggregation module that uses multi-omics data to regulate pathway information propagation and attention based feature aggregation, thereby enhancing cell line representation learning. Comprehensive experiments on the GDSC and PRISM datasets demonstrate the predictive performance of HMVPS-DRP. On the primary GDSC dataset, HMVPS-DRP achieves an RMSE of 0.9410, a PCC of 0.9364, and an SCC of 0.9165. Pathway and gene attribution analyses provide molecular-level interpretations of model predictions, while a melanoma case study highlights the potential of HMVPS-DRP to predict drug responses and support candidate anticancer drug screening.