BEACON: predicting side effects and therapeutics outcomes to drugs by Bridging knowlEdge grAph with CONtextual language model.

Biomedical knowledge graphs encode millions of relationships between drugs, proteins, pathways, and diseases, yet translating this structured knowledge into accurate predictions remains challenging. Existing deep learning approaches, including graph neural networks and knowledge graph embeddings, assign fixed representations to entities regardless of biological context, limiting their ability to capture how the same gene or pathway functions differently across scenarios. These methods also lack interpretability and often fail when applied to novel drugs outside their training distribution. Here we present BEACON (Bridging knowlEdge grAph with CONtextual language model), a framework that transforms knowledge graphs into contextual sentence representations processable by language models. BEACON converts biomedical entities into tokens and relationships into syntactic dependencies, creating "sentence trees" that preserve graph structure while enabling contextual processing. A visibility matrix ensures that attention patterns respect the underlying knowledge graph topology, and a perturbation-based evaluation module identifies the specific genes, enzymes, and pathways driving each prediction. We demonstrate BEACON's versatility through two clinically important applications. For drug sensitivity prediction in cancer cell lines, BEACON achieves 0.941 AUROC and Spearman ρ = 0.919, outperforming existing methods (DrugCell, DeepCDR and DeepTTA). For drug-drug interaction (DDI) prediction, BEACON achieves 0.964 AUROC on the TwoSIDES benchmark and 0.84 AUROC on temporally held-out FDA adverse event data (2013-2023), demonstrating robust generalization to newly approved drugs. Applying BEACON to the BTK inhibitor acalabrutinib revealed that predicted interactions are enriched for drugs metabolized by CYP3A enzymes (OR = 3.01, P = 4.3 × 10