Dual-Channel Structure-Aware Transformer Framework for Drug-Drug Interaction Prediction

Accurate prediction of drug-drug interactions (DDIs) is crucial for ensuring drug safety and supporting pharmaceutical development. Existing methods still face challenges in jointly modeling molecular structural information and biomedical semantic contexts. To address this issue, we propose DSAT-DDI, a Dual-channel Structure-Aware Transformer framework for DDI prediction. The core design of DSAT-DDI lies in incorporating structure-aware attention into both molecular and knowledge graph representation learning, enabling the model to capture local chemical environments and semantic neighborhoods while preserving global dependency modeling. In addition, a bidirectional cross-attention fusion mechanism is introduced to dynamically integrate molecular and biomedical representations and generate pair-aware drug embeddings. Experimental results on the DrugBank and KEGG datasets show that DSAT-DDI consistently outperforms representative molecular-based, knowledge graph-based, and multimodal baselines across multiple evaluation metrics. These results demonstrate the effectiveness of structure-aware attention and bidirectional cross-attention fusion for robust DDI prediction.