Expert-Augmented Computational Drug Repurposing Identified Baricitinib as a Treatment for COVID-19

Abstract

The onset of the 2019 Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic necessitated the identification of approved drugs to treat the disease, before the development, approval and widespread administration of suitable vaccines. To identify such a drug, we used a visual analytics workflow where computational tools applied over an AI-enhanced biomedical knowledge graph were combined with human expertise. The workflow comprised rapid augmentation of knowledge graph information from recent literature using machine learning (ML) based extraction, with human-guided iterative queries of the graph. Using this workflow, we identified the rheumatoid arthritis drug baricitinib as both an antiviral and anti-inflammatory therapy. The effectiveness of baricitinib was substantiated by the recent publication of the data from the ACTT-2 randomised Phase 3 trial, followed by emergency approval for use by the FDA, and a report from the CoV-BARRIER trial confirming significant reductions in mortality with baricitinib compared to standard of care. Such methods that iteratively combine computational tools with human expertise hold promise for the identification of treatments for rare and neglected diseases and, beyond drug repurposing, in areas of biological research where relevant data may be lacking or hidden in the mass of available biomedical literature.

Keywords: COVID-19; SARS-CoV-2; drug repurposing; human computer interaction; knowledge discovery and data mining; knowledge graph.

FIGURE 1

Example use-case of graph pattern querying: in search of targets regulating autoantibody production. Question mark symbols represent stages of asking questions of the knowledge graph, which can result in undesirable results or a failed attempt at querying the knowledge graph (represented by red cross symbols), or desirable results and successful attempts at querying the knowledge graph (represented by green check mark symbols). A diverging path represents the user exploring possibilities down both routes, either as a result of a failed attempt, or as a result of there being two equally valuable options to pursue. A converging path represents the user linking the results of two patterns together, in a new pattern.

FIGURE 2

(A) Pivotal proteins (represented by nodes A and B) are loosely defined as proteins that facilitate cross-talk between network modules. This has some overlap with the notion of a node with high betweenness centrality (Freeman 1977), but there is emphasis on the node’s connectivity across network modules. Node A represents a target belonging to the yellow network module and interacts with the highest number of targets in the blue module. Node B represents a target belonging to the green network module and interacts with the highest number of targets in the yellow module. (B) Pathway membership, represented by nodes colored red, can be scattered across different network modules. While the modules in the network may represent distinct GO processes, biological pathways serve multiple such processes and are therefore seldom confined to one module.

FIGURE 3

The full drug repurposing workflow consisted of three stages. The first stage incorporated the evaluation of a customised knowledge graph via graph pattern querying as a means of rapid qualitative evaluation. The second was a mechanistic analysis of viral and host processes, via the creation of an initial network using a gene set representation of COVID19-related processes extracted from the customised knowledge graph. This was iterated three more times to achieve a final network containing a clear hypothesis and therapeutic targets. The third stage resulted in a hypothesis as a partially defined graph pattern, from which targets and drug candidates were retrieved, producing a final set of candidate treatments for COVID-19.

FIGURE 4

Initial network - the 556 imported genes from the graph. The different colored modules reflect clusters of protein interactions which reflect specific pathways and processes. The modules 1 and 2 reflect inflammatory processes, module 3 signaling pathways, module 4 cholesterol metabolism and modules 5 and 6 coagulation cascades. The three processes most strongly associated with each module are listed in Table 6.

FIGURE 5

Third iteration, fourth and final network - 657 genes in 20 network modules identified via Louvain-based community detection using protein-protein interactions. The largest 7 modules are annotated.

FIGURE 6

Graph pattern for finding approved drugs that are selective and effective against the two hypothesised driving mechanisms behind SARS-CoV-2 induced COVID-19.