MACHINE LEARNING-BASED VIRTUAL SCREENING, MOLECULAR DOCKING, DRUG-LIKENESS, PHARMACOKINETICS AND TOXICITY ANALYSES TO IDENTIFY NEW NATURAL INHIBITORS OF THE GLYCOPROTEIN SPIKE (S1) OF SARS-CoV-2

To identify natural bioactive compounds (NBCs) as potential inhibitors of spike (S1) by means of in silico assays. NBCs with previously proven biological in vitro activity were obtained from the ZINC database and analyzed through virtual screening and molecular docking to identify those with higher affinity to the spike protein. Eight machine learning models were used to validate the results: Principal Component Analysis (PCA), Artificial Neural Network (ANN), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Partial Least Squares-Discriminant Analysis (PLS-DA), Gradient Boosted Tree Discriminant Analysis (XGBoostDA), Soft Independent Modelling of Class Analogies (SIMCA) and Logistic Regression Discriminate Analysis (LREG). Selected NBCs were submitted to drug-likeness prediction using Lipinski's and Veber's rule of five. A prediction of pharmacokinetic parameters and toxicity was also performed (ADMET). Antivirals currently used for COVID-19 (remdesivir and molnupiravir) were used as a comparator. A total of 170,906 compounds were analyzed. Of these, 34 showed greater affinity with the S1 (affinity energy <-7 kcal mol-1). Most of these compounds belonged to the class of coumarins (benzopyrones), presenting a benzene ring fused to a lactone (group of heterosides). The PLS-DA model was able to reproduce the results of the virtual screening and molecular docking (accuracy of 97.0%). Of the 34 compounds, only NBC5 (feselol), NBC14, NBC15 and NBC27 had better results in ADMET predictions. These had similar binding affinity to S1 when compared to remdesivir and molnupirvir. Feselol and three other NBCs were the most promising candidates for treating COVID-19. In vitro and in vivo studies are needed to confirm these findings.