KG-DFI: A Prediction of Drug-Food Interactions Based on Knowledge Graph Embedding

Drug-food interactions (DFIs) can substantially affect patient safety and therapeutic outcomes. However, most existing computational approaches represent foods by decomposing each food into individual constituent molecules, which captures only limited aspects of foods and aligns poorly with real-world clinical practice, where dietary decisions are made at the level of whole foods rather than isolated components. This study constructed a knowledge graph centered on DFIs by integrating diverse biomedical databases covering food constituents, drug-related knowledge, and compound-gene interactions. Here, we propose a knowledge graph-based drug-food interaction (KG-DFI) prediction model that treats foods as complete entities rather than being decomposed into individual compounds. Notably, a KG-DFI model employs relational graph convolutional networks to learn food-entity representations that capture both structural and relational information from the knowledge graph. In addition, we incorporated a cross-attention mechanism that enables food entities to selectively attend to relevant drug characteristics, effectively capturing complex interaction patterns. In predicting DFIs, the KG-DFI model achieved an area under the receiver operating characteristic curve value of 0.9292 and an F1-score of 0.8640. Additionally, the KG-DFI model predicted the 4 specific interaction types (possible, positive, negative, and harmful) among observed DFIs, achieving an area under the receiver operating characteristic of 0.9163 and a weighted F1-score of 0.8801. This framework enables practical clinical decision support by predicting interactions using representations of food as entities, aligning with how healthcare providers deliver dietary recommendations to patients.