TRACE-DDI: A Hybrid Framework of Transformer–GAT Context Encoder and Pathway-Anchored Knowledge Graphs for DDI Prediction

Accurate prediction of drug-drug interactions (DDIs) is crucial for preventing adverse drug reactions and improving drug development efficiency. However, most existing DDI prediction models treat chemical structure and biological knowledge separately, failing to capture multiscale biological context. We present TRACE-DDI, a hybrid framework that integrates 3 complementary modalities within a single end-to-end architecture: (a) the Simplified Molecular Input Line Entry System (SMILES)-based chemical sequences, (b) molecular graph topology, and (c) pathway-anchored biological knowledge graphs. TRACE-DDI employs a Transformer-Graph Attention Network (GAT) cross-modal context encoder that performs multilayer graph-based feature propagation across heterogeneous biological representations, enabling both fine-grained chemical understanding and global pathway-level context aggregation. Experimental results on the benchmark dataset demonstrate that TRACE-DDI consistently outperforms state-of-the-art baselines across all major evaluation metrics. Beyond predictive accuracy, TRACE-DDI provides biologically interpretable insights: pathway-level subgraph visualization and centrality analysis prioritize shared pathway nodes that occupy prominent topological positions in merged drug subgraphs, thereby providing hypothesis-generating biological context for predicted DDIs. Together, TRACE-DDI establishes a unified and interpretable framework for biologically grounded DDI prediction and offers new directions for integrating actionable biological context into artificial intelligence-driven pharmacology.