Abstract
Reliable prediction of drug-target interaction (DTI) is essential for accelerating drug discovery, yet remains hindered by data imbalance, limited interpretability, and neglect of protein dynamics. Here, we present GHCDTI, a heterogeneous graph neural framework designed to overcome these challenges through three synergistic innovations. First, cross-view contrastive learning with adaptive positive sampling improves generalization under extreme class imbalance (positive/negative ratio<1:100). Second, heterogeneous data fusion integrates molecular graphs, protein structure graphs, and bioactivity data via cross-graph attention, enabling interpretable residue-level insights. Third, multi-scale wavelet feature extraction captures both conserved and dynamic structural features by decomposing protein conformations into frequency components. GHCDTI achieves state-of-the-art performance on benchmark datasets (AUC: 0.966 ± 0.016; AUPR: 0.888 ± 0.018) and processes 1,512 proteins and 708 drugs in under two minutes, highlighting its potential for scalable virtual screening and drug repositioning. These results demonstrate GHCDTI's ability to effectively identify novel drug-target pairs, providing a practical tool for accelerating drug discovery and improving biomedical knowledge integration.