DTBNet: Multimodal 3D Geometric Perception with Hierarchical Attention for Robust Drug–Target Interaction and Binding Affinity Prediction

MOTIVATION: Accurate prediction of drug-target interactions (DTIs) and binding affinity (DTA) constitutes a core challenge in accelerating drug discovery. While existing computational methods have made progress, most models struggle to capture critical physicochemical interactions in 3D space due to reliance on 1D sequences or 2D graph structures, and exhibit limited generalization in cold-start scenarios. This study proposes DTBNet, a unified prediction framework integrating multimodal geometric perception and hierarchical attention. The framework constructs heterogeneous embeddings combining local chemical details with global biological information by incorporating protein 3D geometric features and multilevel drug representations. During cross-modal interaction, atom-residue cross-attention first captures pair-specific local associations, followed by group-level multi-head self-attention that integrates complementary chemical-semantic, global structural, and local geometric interaction representations. RESULTS: Experiments on benchmark datasets (Davis, KIBA, PDBBind) demonstrate that compared with the optimal baseline models, DTBNet achieves up to 19.8% reduction in mean squared error and 8.0% improvement in Pearson correlation coefficient, maintaining optimal performance in drug cold-start, target cold-start, and completely novel pairing scenarios, showcasing robust cross-domain adaptability. Independent validation on GPCR targets achieves RMSE 0.783, confirming real-world applicability. Visualization analyses reveal DTBNet precisely identifies interaction sites between key transmembrane helical domain residues and drug aromatic rings, with attention distribution highly consistent with experimentally resolved binding conformations. These results establish DTBNet as a new performance benchmark through 3D geometric perception and dynamic attention mechanisms, providing interpretable molecular insights for structure-based rational drug design and demonstrating broad potential in virtual screening and target discovery. AVAILABILITY: The source code and data are available at https://github.com/wyx2012/DTB_Net. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.