SADRP: A structurally aware multi-modal framework for drug response prediction

Motivation: Predicting drug sensitivity across diverse cancer samples is a fundamental challenge in personalized therapy and drug discovery. This task requires comprehensive characterization of both drug and cancer cell line features. Current approaches often fail to fully capture the complex structural characteristics and interactions of drug compounds and struggle to integrate multi-omics features derived from the same sample, limiting their feature representation capacity. Consequently, there is a need for modeling frameworks that can jointly model molecular structural features and multi-modal cellular contexts in a unified manner. Results: We present SADRP, a structurally aware multi-modal framework for drug response prediction, designed to model complex molecular structures and multi-omics data. For drug representation, SADRP incorporates Shortest Distance Path (SDP) encoding and edge encoding to capture both topological relationships and bond-type chemical semantics. For cell line profiling, a multi-view interactive encoder integrates gene expression and DNA methylation, capturing cross-omics synergistic complementarity. SADRP consistently outperforms state-of-the-art methods, while also identifying potential therapeutic candidates and biomarkers, providing mechanistic insights and supporting applications in drug discovery and personalized medicine. Author summary: Cancer patients often respond differently to the same anticancer drug, making effective treatment selection a major challenge in precision oncology. Although computational models have been widely developed for drug response prediction, many existing approaches remain limited in their ability to accurately represent complex drug structures and integrate multiple layers of cellular information. To address these challenges, we developed SADRP, a structurally aware deep learning framework that models both the topological relationships and chemical bond semantics of drug molecules while jointly integrating gene expression and DNA methylation profiles from cancer cell lines. By combining these complementary sources of information, SADRP provides more comprehensive representations of both drugs and cellular states. We evaluated SADRP on large-scale benchmark datasets and found that it consistently outperformed several state-of-the-art methods. In addition to improving prediction accuracy, SADRP identified potential therapeutic candidates and biologically relevant biomarkers associated with drug sensitivity. Our results demonstrate the value of integrating molecular structural information with multi-omics data to support drug discovery and precision cancer therapy.