A systems-based method to repurpose marketed therapeutics for antiviral use: a SARS-CoV-2 case study

Abstract

This study describes two complementary methods that use network-based and sequence similarity tools to identify drug repurposing opportunities predicted to modulate viral proteins. This approach could be rapidly adapted to new and emerging viruses. The first method built and studied a virus-host-physical interaction network; a three-layer multimodal network of drug target proteins, human protein-protein interactions, and viral-host protein-protein interactions. The second method evaluated sequence similarity between viral proteins and other proteins, visualized by constructing a virus-host-similarity interaction network. Methods were validated on the human immunodeficiency virus, hepatitis B, hepatitis C, and human papillomavirus, then deployed on SARS-CoV-2. Comparison of virus-host-physical interaction predictions to known antiviral drugs had AUCs of 0.69, 0.59, 0.78, and 0.67, respectively, reflecting that the scores are predictive of effective drugs. For SARS-CoV-2, 569 candidate drugs were predicted, of which 37 had been included in clinical trials for SARS-CoV-2 (AUC = 0.75, P-value 3.21 × 10^-3). As further validation, top-ranked candidate antiviral drugs were analyzed for binding to protein targets in silico; binding scores generated by BindScope indicated a 70% success rate.

Figure 1.. Complementary methods to identify drug repurposing candidates that directly target viral proteins.

(A) A network-based approach using link prediction to identify drug repurposing opportunities that are at a path length of three from viral proteins on the virus–host–physical interaction network. (B) A sequence similarity approach that identifies drug target proteins with protein sequence homology to viral proteins.

Figure 2.. Method validation in HIV, hepatitis B virus, hepatitis C virus, and HPV.

(A) Receiver operating characteristic curve evaluating predictive power of network-based approach in ranking known drugs in HIV, hepatitis B virus, hepatitis C virus, and HPV. (B) Box and whisker plot of network mean scores generated for all drugs with a non-zero value and for drugs identified by the sequence similarity approach for HIV. The sequence similarity approach predicts drugs with high network mean scores. (C) Virus–host–similarity interaction network representation of drugs predicted by both methods for HIV.

Figure S1.. DPI score distribution of predicted antiviral drugs for each SARS-CoV-2 viral protein.

Viral proteins are color coded based on the protein class: structural (red), non-structural (green), or accessory (yellow). The drug with the highest score in each viral protein class from method 1 is highlighted.

Figure 3.. Method implementation on SARS-CoV-2.

(A) Receiver operating characteristic curve evaluating predictive power of network-based approach in ranking drugs currently under investigation for COVID-19 in clinical trials. (B) Box and whisker plot of network mean scores generated for all drugs with a non-zero value and for drugs identified by the sequence similarity approach. (C) Visualization of predicted antiviral drugs and SARS-CoV-2 Protein Data Bank structures on the virus–host–similarity interaction network.