Evaluation of predictive machine learning models for drug repurposing against delta variant of SARS-CoV-2 spike protein

Abstract

Drug repurposing is a major approach used by researchers to tackle the COVID-19 pandemic which has been worsened by the current surge of delta variant in many countries. Though drugs like Remdesivir and Hydroxychloroquine have been repurposed, studies prove these drugs have insignificant effect in treatment. So, in this study, we use the already FDA approved database of 1615 drugs to apply semi-flexible and flexible molecular docking methods to calculate the docking scores and identify the best 20 potential inhibitors for our modelled delta variant spike protein RBD. Then, we calculate 2325 1-D and 2-D molecular descriptors and use machine-learning algorithms like K-Nearest Neighbor, Random Forest, Support Vector Machine and ensemble stacking method to build regression-based prediction models. We identify 15 best descriptors for the dataset all of which were found to be inversely correlated with ligand binding. With only these few descriptors, the models performed excellently with an area under curve (AUC) value of 0.952 in Regression Error Characteristic curve for ensemble stacking. Therefore, we comment that these 15 descriptors are the most important features for the binding of inhibitors to the spike protein and hence these should be studied properly in terms of drug repurposing and drug discovery.