Explainable Drug Repurposing in Context via Deep Reinforcement Learning

Biomedical knowledge graphs encode domain knowledge as biomedical entities and relationships between them. Graph traversal algorithms can make use of these rich sources for the discovery of novel research hypotheses, e.g. the repurposing of a known drug. Traversed paths can serve to explain the underlying causal mechanisms. Most of these models, however, are trained to optimise for accuracy w.r.t. known gold standard drug-disease pairs, rather than for the explanatory mechanisms supporting such predictions. In this work, we aim to improve the retrieval of these explanatory mechanisms by improving path quality. We build on a reinforcement learning-based multi-hop reasoning approach for drug repurposing. First, we define a metric for path quality based on coherence with context entities. To calculate coherence, we learn a set of phenotype annotations with rule mining. Second, we use both the metric and the annotations to formulate a novel reward function. We assess the impact of contextual knowledge in a quantitative and qualitative evaluation, measuring: (i) the effect training with context has on the quality of reasoning paths, and (ii) the effect of using context for explainability purposes, measured in terms of plausibility, novelty, and relevancy. Results indicate that learning with contextual knowledge significantly increases path coherence, without affecting the interpretability for the domain experts.