Pharmacophore-Aware Dual-View Learning with Bidirectional Cross-Attention for Drug-Drug Interaction Prediction

Accurate drug-drug interaction (DDI) prediction is critical for ensuring patient safety and guiding clinical decision-making. Existing methods often rely on single-view molecular representations, limiting their ability to capture the complex structural and spatial properties of drugs. In this study, we propose a novel pharmacophore-aware dual-view learning framework (PharmaDual) that integrates both 2D and 3D representations of pharmacophores for enhanced DDI prediction. Specifically, we first extract pharmacophore fragments as key substructures, and independently encode their 2D and 3D spatial information using specialized graph-and geometry-based encoders. To effectively combine the complementary views, we introduce a bidirectional cross-attention fusion module that dynamically aligns and integrates 2D and 3D pharmacophore representations. Extensive experiments on benchmark DDI datasets demonstrate that our method consistently outperforms existing approaches, highlighting the benefit of dual-view modeling and cross-attentive fusion in capturing nuanced pharmacophore-level interactions. The code is available at https://github.com/ZWX1289/PharmaDual.