Structured multi-view relational learning for drug response prediction

Drug response prediction (DRP) requires models that jointly characterize molecular drug structure and heterogeneous cellular states. Although recent deep learning approaches have integrated multi-omics profiles with drug representations, many treat omics views as parallel feature sources or merge them before relational modeling, which can weaken omics-view-specific cell-line relation structures. Here, we propose MulDR, a structured multi-view relational learning framework that preserves omics-view-specific cell-line relation structures before drug-conditioned cell-drug interaction. MulDR constructs separate relation structures for pathway activity, gene expression, mutation, and DNA methylation, learns view-specific relational cell representations, and uses complementary motif-level, geometry-aware molecular graph, and Morgan fingerprint drug representations to condition their fusion. With this design, cell-line similarity remains view-dependent until the interaction stage, rather than being prematurely compressed into a single feature or relation space. Extensive experiments show that MulDR consistently outperforms representative and recent baselines under the same evaluation protocol, reducing root mean square error by 6.07% relative to the strongest baseline. Cold-start evaluations on unseen drugs and cell lines further show that MulDR generalizes to previously unseen entities, while ablation studies show that removing either view-specific relational modeling or drug-conditioned fusion consistently compromises predictive performance. Together, these results suggest that preserving omics-specific cell-line relations before drug-conditioned interaction modeling provides an effective strategy for robust DRP.