DiM2-DRP: Dual-Input Multi-View Dynamic Contrastive Learning Framework for Drug Response Prediction

Accurately modeling cellular states is crucial for pharmacogenomics, yet existing deep learning methods often rely on shallow representations that fail to capture complex drug-cell interactions. These limitations necessitate a multi-view architecture that synergistically leverages multiple biological modalities to explicitly capture drug-cell interactions. Here, we propose DiM2-DRP, a dual-input multi-view framework that integrates heterogeneous and complementary views from cancer cells and small-molecule compounds. The framework develops a network-informed cell encoder that integrates bi-directional network diffusion with stability-guided pathway clustering to derive compact, biologically consistent representations. A novel bi-directional attention-based contrastive fusion module integrates these cellular representations with drug features extracted by a structure-informed drug encoder, dynamically adjusting cross-modal alignment according to interaction strength. By emphasizing biologically relevant drug-cell dependencies, this architecture alleviates the shallow integration limitations of conventional methods. Furthermore, an auxiliary classification module imposes lineage-specific and target-pathway constraints, ensuring the learned representations are both mechanistically grounded and robust for drug response prediction. Experiments show that DiM2-DRP achieves higher predictive accuracy, improved drug-blind and cell-blind generalization, and enhanced biological interpretability compared with existing methods. The pathway signatures recovered by the model remain consistent in patient tumours and are associated with multiple clinical outcomes, supporting the translational relevance of the learned representations. By leveraging a multimodal architecture integrated with dynamic contrastive learning, DiM2-DRP refines the characterization of drug sensitivity, offering a robust framework to support precision oncology strategies.