Drug target interaction prediction via gated attention and information bottleneck

Accurate drug–target interaction (DTI) prediction is essential for drug repositioning and accelerating drug discovery. Deep learning methods have made remarkable progress over traditional biological experiments, yet existing models often fail to capture local node topologies and multi-view semantic dependencies simultaneously. Moreover, most methods rely on basic loss functions that cannot filter redundant noise, hindering compact and discriminative node representations. In this work, we propose the DTI prediction framework that integrates a hierarchical gated multi-head attention (HGMA) mechanism with an information bottleneck (IB) strategy. HGMA adopts a two-layer architecture: the first layer performs weighted aggregation over semantic meta-paths, and the second layer fuses attention heads via an adaptive gating mechanism, enhancing drug and target representations. The IB module compresses inputs by removing task-irrelevant redundancy while preserving predictive information, improving discriminability and generalization. Extensive experiments show that our model consistently outperforms state-of-the-art methods in both accuracy and robustness.