Using Extraction and Evaluation of Explanations for Drug-Repurposing on Knowledge Graphs

Knowledge-graph (KG) databases represent information about real-world entities (via nodes) and relationships between them (via edges) in the form of semantic networks. We consider the problem of reliably inferring novel relationships between KG entities, with applications in drug repurposing, the area of biomedical research that focuses on identifying new uses for existing drugs. Inferring novel drug-disease relationships can be viewed in drug repurposing as generation of testable hypotheses, which can then be validated in preclinical or clinical studies. Thus, reliable KG generation of noteworthy drug - treats - disease hypotheses would be of value in those cases in which biomedical information is available through large-scale KGs. We introduce a scalable domain- and task-agnostic approach for extracting from KGs explainable hypotheses and for providing data-supported evaluations of the explanations for their potential noteworthiness. Our explanations take the form of knowledge-graph patterns that are evaluated using our proposed metrics; in the drug-repurposing domain, they can be interpreted as drug - treats - disease hypotheses that can be further analyzed by biomedical experts. Through broad-ranging experiments, we provide evidence for our extracted explanations being understandable and reasonable to domain experts, and for the proposed metrics being accurate. These findings suggest that the approach can be a viable tool in drug repurposing to extract and evaluate explanations for potential drug-treats-disease hypotheses.