CIM-DTA: Conditional Interaction Modulation for Cold-Start Drug-Target Affinity Prediction

Drug-target affinity prediction under cold-start set tings suffers from distribution shifts between training and testing data. This leads to poor generalization for unseen drugs or targets. To address this issue, we propose a conditional inter action modulation framework. The model explicitly captures pair-specific interaction contexts. It integrates molecular graph representations with contextual protein embeddings from ESM-2. A conditional response vector is constructed to guide interaction modeling. This vector modulates the network through hierarchical interaction modulation, including attention bias adjustment, feature activation gating, and feature distribution recalibration. In addition, a condition-aware hypernetwork head enables adaptive affinity prediction. Experiments on the Davis and BindingDB Kd datasets are conducted under warm, cold-drug, and cold target settings. CIM-DTA achieves competitive performance on the Davis and BindingDB-Kd datasets, particularly under cold start settings. It also maintains stable performance under strict cold-start scenarios. Ablation studies and case analyses further verify its effectiveness in modeling pair-specific interaction con texts.