Adverse drug-drug interactions (DDIs) remain a critical public health concern, responsible for approximately 125,000 deaths annually in the United States and accounting for nearly 30% of all adverse drug events. Current computational approaches face two fundamental challenges: the inability to capture distant atomic relationships within molecular structures and poor generalization performance when predicting interactions for newly approved pharmaceuticals. This research introduces an advanced deep learning architecture that addresses these limitations through the integration of global attention mechanisms, knowledge graph embeddings, and self-supervised representation learning specifically designed for pharmaceutical applications. We propose KE-GraphFormer (Knowledge-Enhanced Graph Transformer), a unified framework incorporating five innovative components: (1) a hierarchical transformer encoder utilizing spatial-aware attention for modeling both proximal and distal molecular patterns, (2) a knowledge-guided attention module that integrates drug-related knowledge from biomedical knowledge graphs, (3) pharmaceutical-specific augmentation strategies for contrastive pre-training, (4) a dual-modality fusion module combining structural graphs with sequential SMILES representations, and (5) an adaptive pooling mechanism for multi-resolution feature extraction. Comprehensive experiments across four benchmark datasets reveal substantial performance gains. Our approach achieves 98.73% accuracy on DrugBank (2.28% improvement), 90.12% on TWOSIDES (2.82% improvement), and 87.34% accuracy for previously unseen drugs (9.41% improvement). Statistical validation confirms significance across all metrics (p < 0.001). KE-GraphFormer establishes new performance benchmarks for computational DDI prediction while providing interpretable insights through attention visualization, demonstrating strong potential for clinical decision support systems and pharmaceutical safety assessment.
Amiri et al. (Wed,) studied this question.