Introduction: The emergence of COVID-19 has resulted in a widespread pandemic with a critical need for therapeutic options and comprehensive identification of risk factors governing disease severity. Given the lengthy process required to develop drugs for clinical use, a strong emphasis has been placed on repurposing FDA approved drugs for COVID-19 treatment. With the large literature base of COVID-19 publications, there is a need for tools to quickly draw meaningful conclusions. Here we utilize unsupervised text mining and custom-built knowledge graph simulation tools to identify and rank promising therapeutics and predictive risks.

Materials and Methods: SemNet is a relationship ranking tool which operates on a version of SemMedDB, a repository of semantic predications between Unified Medical Language System (UMLS) concepts scraped from 30 million+ articles on PubMed. These form a knowledge graph where UMLS concepts are nodes and predications are edges [1]. Simulations run in SemNet identify source nodes and rank them with respect to user specified target nodes using an unsupervised machine learning ranking algorithm. Hetesim is a similarity metric commonly used for identifying well-established literature relationships. To identify therapeutic targets and relevant drugs, simulations were guided by broad coronavirus related key terms, specific therapeutic targets, and drug classes. Key findings in simulation results were verified by a combination of expert review, subgraph or node neighborhood visualization, and entity-relation prediction APIs developed in our lab.

Results and Discussion: Results include ranked lists of UMLS biomedical concepts and their rankings. Up to 17 drug names were compiled based on the top 40% of highly ranked nodes for Hetesim scores from four different drug classes targeting endosomes, hemagglutin, neuraminidase, palpain, peptide hydrolases, and RNA-directed RNA polymerase. Besides the known antiretroviral drugs, other promising drugs include HIV protease inhibitors, neuraminidase inhibitors, folic acid antagonists, nucleoside analogs, proton pump inhibitors, H1 antagonists, and nucleoside reverse transcriptase inhibitors. Specifically, zalcitabine (reverse transcriptase inhibitor to treat HIV) and valganciclovir (antiviral to prevent cytomegalovirus following organ transplant) showed strong connections that inhibit mechanisms of hypothesized COVID-19 replication. Additionally, 10 risk factor simulations were performed with >1,500 concepts relevant to high risk conditions, including age, diabetes, immunosuppressive disorders, chronic lung disease, and cardiovascular disease, were identified and ranked by literature prevalence. Literature review and cross-verification of results was conducted to confirm up to 22 potential risk factors while providing insight into their clinical presentation and mechanism of interaction with the SARS-CoV-2 virus.

Conclusion: Literature based discovery (LBD) tools have potential to provide meaningful insight for fast initial screening of repurposed drugs for novel diseases, such as COVID-19. Multiple different mechanisms of action were identified to positively correlate with COVID-19 treatment and illustrate promising drug lists for laboratory screening. Next, we will use the knowledge graphs to pair high-ranking mechanism of action (drug classes) with high risk comorbidities to identify which sets of drugs are best for specific COVID-19 patient populations.