OBJECTIVES: De novo drug development is an extremely time-consuming process with a high risk of failure and tremendous resource requirements. For these reasons, a drug-repurposing strategy has gained significant attention as a cost-effective alternative strategy. Drug repurposing can rapidly repurpose FDA-approved drugs to other indications than the approved indications. Until now, drug repurposing is serendipitous, such as repurposing Sildenafil for erectile dysfunction, and sometimes requires extensive manual reviews of related literature, clinical trials, and relevant clinical data. With an ever-growing amount of literature and information, an automatic method is much needed.

METHODS: We developed KnowledgeSphere, a framework for processing and integrating diverse knowledge bases. It includes four modules: i) NLP module: building natural language processing (NLP) pipelines to extract biomedical knowledge from literature, ii) integration module: incorporating data from various sources, iii) knowledge module: building deep learning-based models and scoring systems, and iv) evaluation module: validating results.

RESULTS: First, we built NLP pipelines to extract biomedical entities and relations from 35 million PubMed abstracts. Second, we integrated the data from manually curated biomedical resources into our literature-based knowledge graph. As a result, our knowledge graph consists of 20 thousand entities (drugs, diseases, genes, etc.) and 10 million relations ("inhibits", "treats", "stimulates", etc.). Third, we trained deep learning-based knowledge graph embedding models and then predicted the "treats" relations for each drug-disease pair. Finally, in our evaluation module, we excluded all relations involving 15 successful drug repurposing cases collected from review articles during knowledge graph embedding. After applying link prediction for all 15 successful pairs of drugs and their new indications, we found that all are ranked top 0.5% across all diseases.

CONCLUSIONS: In conclusion, KnowledgeSphere, a framework leveraging deep learning-based knowledge graph embedding models, enables the representation of complex and interconnected knowledge from various sources and potentially opens the way for drug repurposing at scale.