Session 8 : Knowledge Graph Embedding Models for Drug Repurposing

Approximately 7,000 rare diseases impact 300-400 million people worldwide, with 30 million in the United States alone. Discovering treatments relies on unraveling their molecular basis, a costly and labor-intensive endeavor. Artificial intelligence and deep learning aim to expedite this process by integrating biomedical databases like UniProtKB, OMIM, and BioGrid. However, current machine learning techniques face limitations in modeling biological entities, integrating diverse data sources, and providing explanations. Knowledge Graph (KG) interconnects heterogeneous biomedical entities for more comprehensive analysis using Knowledge Graph embedding (KGE). KGE models compute low dimensional vector representation for the knowledge graph entities and relations to build downstream predictive models such as link prediction, node prediction.The primary benefits of KGs are 1) easy integration of data from multiple sources, 3) explainable decision making while using existing machine learning techniques. In this research we, 1) curate a KG to include external biomedical information about rare diseases, 2) develop a machine learning pipeline to predict the repurpose-able drugs for rare diseases. Experiment-wise, we have run 27 KGE models on a biomedical knowledge graph with 111,171 entities, 16 relation types, and 2,132,796 edges prepared using BioKG. We identified 6 models that performed with a hit@10 at least 0.50. We then applied these models to find repurposable drugs for Mesothelioma, a rare disease. This work provides extended insights on the performance as well as challenges in terms of variability of the model's outcome to retrieve meaningful drug repurposing outcomes.