Phytochemical Drug Discovery for COVID-19 Using High-resolution Computational Docking and Machine Learning Assisted Binder Prediction

The COVID-19 pandemic has resulted in millions of deaths around the world. Although multiple safe and effective vaccines and some pharmaceuticals have been approved for use, the problem is still unsolved for individuals with underlying medical conditions and those living in underserved areas that lack vaccines and/or an adequate medical infrastructure. This is especially challenging as new variants of SARS-CoV-2 emerge. One possible approach to solving this problem lies in using naturally abundant phytochemicals generally regarded as safe that bind to and disrupt SARS-CoV-2. When used in conjunction with a polypharmacological approach, targeting multiple essential viral proteins can lead to stronger functional inhibition and provides a safeguard against escape mutations. Although finding the proper phytochemicals to accomplish a specific therapeutic task is challenging and costly, in-silico screening methods have made this a more tractable problem by expediting the initial lead compound discovery phase. Recent studies have gained mechanistic insights of drug interactions through computational docking against select SARS-CoV-2 proteins, yet several viral proteins remain unexplored as druggable targets. Here we investigate a wide range of drug products against a comprehensive array of SARS CoV-2 proteins using a high-resolution docking workflow. Our initial lead compound discovery phase consisted of a structure-based virtual screening (SBVS) wherein 10 types of structural and non-structural SARS-CoV-2 proteins were computationally docked against a panel of anti-viral phytochemicals from the USDA Phytochemical and Ethnobotanical Databases. In the second phase of the study, we employed ligand-based virtual screening (LBVS) by extracting chemical features of 34 lead compounds from the SBVS using unsupervised clustering based on common motifs. Features among dominant ligand clusters were then used to prioritize subsets of additional phytochemical databases for drug discovery. Among the 53 newly identified phytochemicals generated via LBVS, high-resolution docking predicted that 28 elicit strong binding interactions with SARS-CoV-2 proteins. Thus, the inclusion of LBVS resulted in a 4-fold increase in the rate of lead discovery. Finally, drug-likeness of all lead compounds and phytochemical sourcing was evaluated. As a result, this three-phase workflow gave rise to 18 flavone, alkaloid, and anthraquinone phytochemicals with the greatest potential for therapeutic utility. Among these phytochemicals, multiple lead compounds with favorable drug-likeness can be derived from individual plants (e.g. Camptotheca acuminata and Mahonia japonica). Collectively, this study demonstrates the exciting potential of plant-based drug development for COVID-19 prevention and treatment using a polypharmacological approach. These findings further support the advantage of incorporating machine learning elements into a virtual screening workflow.