RLASON-CDR: a reinforcement learning-driven adaptive synergistic optimization network for cancer drug response prediction.

Cancer drug response (CDR) prediction is crucial for advancing precision medicine. Advanced computational predictors extracted CDR patterns from multimodal features of cell lines and drugs under the guidance of known CDRs. However, most existing methods struggle to extract biologically meaningful and generalizable CDR representations due to insufficient semantic alignment among the multimodal features. In addition, semantic inconsistencies between multimodal and topological features further hinder the predictive accuracy of CDRs. To address the challenges, a novel Reinforcement Learning-driven Adaptive Synergistic Optimization Network-CDR (RLASON-CDR) is put forward for CDR prediction. RLASON-CDR first constructs multimodal CDR representations aligned within and across drugs and cell lines. Next, it captures high-order topological representations of CDRs from the cell line-drug response network. Finally, a reinforcement learning-based network is proposed to adaptively explore potential CDR patterns by synergistically optimizing these representations instead of semantic fusion. Extensive evaluations demonstrate that RLASON-CDR consistently outperforms existing methods and is robust for predicting unknown CDRs. Gradient attribution analysis further reveals that RLASON-CDR identifies key modality contributions and uncovers biologically response patterns. Furthermore, biological significance analysis indicates that RLASON-CDR can effectively reveal mechanisms of drug response and provide valuable guidance for precision therapy in clinical applications.