Searching for potential inhibitors of SARS-COV-2 main protease using supervised learning and perturbation calculations

Abstract

Inhibiting the biological activity of SARS-CoV-2 Mpro can prevent viral replication. In this context, a hybrid approach using knowledge- and physics-based methods was proposed to characterize potential inhibitors for SARS-CoV-2 Mpro. Initially, supervised machine learning (ML) models were trained to predict a ligand-binding affinity of ca. 2 million compounds with the correlation on a test set of $R=0.748\pm 0.044$ . Atomistic simulations were then used to refine the outcome of the ML model. Using LIE/FEP calculations, nine compounds from the top 100 ML inhibitors were suggested to bind well to the protease with the domination of van der Waals interactions. Furthermore, the binding affinity of these compounds is also higher than that of nirmatrelvir, which was recently approved by the US FDA to treat COVID-19. In addition, the ligands altered the catalytic triad Cys145 - His41 - Asp187, possibly disturbing the biological activity of SARS-CoV-2.

Keywords: Docking, Simulation; FEP; FEP, Free Energy Perturbation; LIE; LIE, Linear Interaction Energy; ML, Machine Learning; Machine learning; Mpro, SARS-CoV-2 Mpro; SARS-CoV-2 Mpro; SL, Supervised Learning; Supervised learning.

Fig. 1

Computational strategy. (A) Computational approach utilized to search promising inhibitors for SARS-CoV-2 Mpro by using hybrid approach involving supervised machine learning and atomistic simulations. (B) Ligand-binding pose was preliminarily predicted via AutoDock Vina. (C) Configuration of catalytic triad Cys145 - His41 - Asp187. (D) + (E) Initial conformations of SARS-CoV-2 Mpro + inhibitor and individual inhibitors in solution.

Fig. 2

Predicted binding free energy versus experiment for 120 test compounds. Prediction was made using XGBoost model.

Fig. 3

Probability of NBC and HB contacts between SARS-CoV-2 Mpro individual residues and top-lead compounds. Green rectangles denote residues that formed more than 6% HB and 46% NBC to ligands.

Fig. 4

Collective-variable FEL of SARS-CoV-2 Mpro in present and absent ligands. Distances $\mathrm{d}({\mathrm{N}\mathrm{\delta }}_{\mathrm{H}\mathrm{i}\mathrm{s}41}-{\mathrm{O}\mathrm{\delta }}_{\mathrm{A}\mathrm{s}\mathrm{p}187})$ and ($\mathrm{d}({\mathrm{S}\mathrm{\gamma }}_{\mathrm{C}\mathrm{y}\mathrm{s}145}-{\mathrm{N}\mathrm{\epsilon }}_{\mathrm{H}\mathrm{i}\mathrm{s}41})$, which are associated with catalytic triad Cys145 - His41 - Asp187, were utilized as reaction coordinates.

Fig. 5

Interaction diagram between SARS-CoV-2 Mpro + CHEMBL3945443. Complexed structure was obtained by calculating clustering of all equilibrium snapshots of complex with cutoff of 0.2 nm.