ABT-DDI: A Graph Transformer Model with Atomic-Bond Structure Awareness for Drug-Drug Interaction Prediction

The widespread use of polypharmacy has significantly increased the risk of drug-drug interactions (DDIs), underscoring the critical need for developing accurate drug-drug interaction events (DDIEs) prediction methods. However, current DDI studies inadequately account for the intrinsic relationships between atoms and bonds in drug molecules, while also overlooking the three-dimensional conformational information on these molecules. To address these limitations, we propose ABT-DDI, an innovative DDI prediction model based on a graph transformer architecture, capable of extracting multimodal information from drug molecules to predict DDI risk levels. ABT-DDI introduces the pioneering systematic modeling of spatial relationships including atom-atom, atom-bond, and bond-bond interactions through a multiscale attention mechanism, which effectively captures atomic and bonding interaction patterns to enhance substructure perception. Furthermore, we introduce two dedicated virtual nodes representing global atom and bond embeddings, which systematically aggregate and propagate overall structural information to refine high-level feature learning. Additionally, the model integrates molecular fingerprint features with 3D spatial distance descriptors to establish a comprehensive molecular representation system. Experimental results demonstrate that our model significantly outperforms existing state-of-the-art methods across multiple metrics on two benchmark data sets, showing important application value in drug development and polypharmacy risk warning systems.