Explaining Drug-Discovery Hypotheses Using Knowledge-Graph Patterns

Abstract

Drug discovery is an important process in which biomedical experts identify new potential treatments for diseases. Currently, this requires much time and manual effort, so automating any part of the process would be a significant improvement. Thus, we propose to identify drug candidates via explainable automated fact-checking. That is, given a hypothesized drug-disease treatment relationship, we aim to generate explanations for the hypothesis as a means for determining whether or not the drug could be a potential treatment for the disease. Our goal in this paper is to develop such an approach that is well-suited for the biomedical domain.Direct application of existing fact-checking tools faces several challenges since most are not developed for use within the biomedical domain; both the explanation formats and the evaluation metrics are ill-suited for this domain application. We propose explanations in the form of knowledge graph patterns, which directly relate to existing structures used by biomedical experts, as well as evaluation metrics which rely solely on existing evidence present in knowledge graphs and make no domain-specific assumptions. We report experimental results, which suggest that, for the drug discovery task and potentially others, our metrics are accurate, and our explanations are understandable and reasonable to domain experts, as well as useful.

Mining Literature-Based Knowledge Graph for Predicting Combination Therapeutics: A COVID-19 Use Case

“…Gao et al [16] constructed a knowledge graph based on associations and presented a computational approach to drug repurposing through lower-dimensional representation of entities and relations in the knowledge graph; they demonstrated the method for the case of Alzheimer's disease. Schartz et al [17] proposed a new fact-checking mechanism to explaining drug discovery hypotheses using knowledge graph patterns; while interesting, this was not a computational work.…”