Design and implementation of a machine learning model based on knowledge graphs for drug repurposing

Abstract

Drug discovery is a costly and time-consuming procedure, so it is not suitable to provide a new treatment for a new spreading disease. In addition, less than 6% of all rare diseases have an approved treatment, due to the high cost of R&D [1].

Drug repurposing aims to find new therapeutic applications for existing drugs on the pharmaceutical market, leading to significant time and cost reduction. How-ever, identifying new diseases for which existing drugs can be effective is a com-plex task. The goal of this project is to propose a machine learning approach based on knowledge graphs for the repositioning of existing drugs. To this end, knowl-edge graph completion techniques such as link prediction are studied, which can be categorized in embedding-based, rule-based, and reinforcement-learning-based approaches. For the drug repurposing problem, these methods provide as output a ranked list of diseases for which each drug can be used as a treatment option. Embedding-based approaches are the most widely used for drug repurposing, but do not provide explainability as rule-based and reinforcement-learning-based ap-proaches do. Thus, it is important to compare explainable and non-explainable models to see if the former can perform better. If so, explainable models could re-place non-explainable ones, and their reasoning can be provided to doctors and the pharmaceutical industry.

The experiments were carried out using the biomedical knowledge graph He-tionet. To evaluate the performance of the different models on this dataset, two similar metrics based on ranked lists were applied: Mean Reciprocal Rank (MRR) and Hits@N with three different values for N (1,3 and 10). The former gives a global perception of whether the positive predictions (which were on the dataset and are used to evaluate the performance) are on the top of the ranked lists or not while the latter determines the proportion of positive predictions with respect to the top N predictions. As a result, rule-based approaches have been found to be more effective for the drug repurposing problem.

Finally, the repositioning of Epirubicin proposed by the rule-based approaches, which provide the highest metrics, demonstrates the power of these models for drug repurposing. The top 10 predictions are discussed, especially the novel link predictions discovered in the knowledge graph, as they are diseases for which Epir-rubicin could be an effective treatment.