Drug-Drug Interaction Prediction on a Biomedical Literature Knowledge Graph

Knowledge Graphs provide insights from data extracted in various domains. In this paper, we present an approach discovering probable drug-to-drug interactions, through the generation of a Knowledge Graph from disease-specific literature. The Graph is generated using natural language processing and semantic indexing of biomedical publications and open resources. The semantic paths connecting different drugs in the Graph are extracted and aggregated into feature vectors representing drug pairs. A classifier is trained on known interactions, extracted from a manually curated drug database used as a golden standard, and discovers new possible interacting pairs. We evaluate this approach on two use cases, Alzheimer's Disease and Lung Cancer. Our system is shown to outperform competing graph embedding approaches, while also identifying new drug-drug interactions that are validated retrospectively.