TK-DDI: Accurate and efficient drug-drug interaction prediction via token encoding

Accurate prediction of drug-drug interactions (DDIs) is paramount for preventing adverse drug events and ensuring patient safety. While existing computational methods show promise, they often struggle to effectively model long-range intramolecular dependencies and identify the most salient substructures for interaction. To address these limitations, we introduce TK-DDI, a novel deep learning framework based on molecular tokenization. TK-DDI first converts drug molecules into sequences of tokens, creating a unified representation that captures both 2D structural and 3D conformational information. A Transformer encoder is then employed to learn the contextual relationships between all token pairs, effectively modeling the influence of distant functional groups. To elucidate the interaction mechanism, TK-DDI incorporates a two-stage attention strategy: an intra-drug attention module to highlight key substructures within each molecule, followed by an inter-drug attention module to fuse the representations of the drug pair. Comprehensive experiments on benchmark datasets demonstrate that TK-DDI robustly outperforms state-of-the-art methods, establishing a new standard for DDI prediction.