Ligand-based discovery of coronavirus main protease inhibitors using MACAW molecular embeddings

Abstract

Ligand-based drug design methods are thought to require large experimental datasets to become useful for virtual screening. In this work, we propose a computational strategy to design novel inhibitors of coronavirus main protease, M^pro. The pipeline integrates publicly available screening and binding affinity data in a two-stage machine-learning model using the recent MACAW embeddings. Once trained, the model can be deployed to rapidly screen large libraries of molecules in silico. Several hundred thousand compounds were virtually screened and 10 of them were selected for experimental testing. From these 10 compounds, 8 showed a clear inhibitory effect on recombinant M^pro, with half-maximal inhibitory concentration values (IC₅₀) in the range 0.18-18.82 μM. Cellular assays were also conducted to evaluate cytotoxic, haemolytic, and antiviral properties. A promising lead compound against coronavirus M^pro was identified with dose-dependent inhibition of virus infectivity and minimal toxicity on human MRC-5 cells.

Keywords: Coronavirus; cheminformatics; drug discovery; ligand-based drug design; machine learning.

Figure 1.

MACAW embeddings can help identify molecules able to bind to SARS-CoV-2 M^pro. (a) Precision-recall curve of a kNN hit classifier trained on MACAW embeddings applied to a test set of molecules. (b) A SVR regressor was also trained to predict pK_i values for promising molecules. (c) We computationally screened a custom library of 408 935 lead-like molecules and prioritised 105 that both the classifier and the regressor considered promising (orange region). See Jupyter Notebook 1 for details.

Figure 2.

Selected compounds inhibit the activity of SARS-CoV-2 M^pro. The hydrolytic activity of SARS-CoV-2 M^pro was measured in the presence of increasing concentrations of different test compounds. (a) Ebselen. (b) Compound 1. (c) Compound 2. (d) Compound 3. (e) Compound 4. (f) Compound 5. (g) Compound 6. (h) Compound 7. (i) Compound 8. (j) Compound 9. (k) Compound 10. (l) IC₅₀ values for Ebselen, Compound 1, Compound 7, Compound 8, and Compound 9. The dose-response curves and IC₅₀ values were determined by nonlinear regression. All data are shown as mean ± SEM, n = 3 biological replicates.

Figure 3.

Spectra of compounds 2 and 7. UV-vis absorption spectra of compound 2 (a) and compound 7 (b). Fluorescence emission spectra of compound 2 (c) and compound 7 (d) at an excitation wavelength of 325 nm.

Figure 4.

Addition of detergent does not affect the inhibition of SARS-CoV-2 M^pro by selected compounds. (a) The hydrolytic activity of SARS-CoV-2 M^pro was measured in the presence of increasing concentrations of ebselen. (b–e) IC₅₀ values were determined in the presence and absence of 0.1% Triton X-100. (b) Compound 1. (c) Compound 7. (d) Compound 8. (e) Compound 9. (f) Autofluorescence of selected compounds at different concentrations. All data are shown as mean ± SEM, n = 3 biological replicates.

Figure 5.

(a) Evaluation of cytotoxicity of compounds 1, 7, 8, and 9 on MRC-5 cells at 72 h post-treatment by MTT test, n = 3. (b) Effects of compounds 1 and 7 on the haemolysis of red blood cells compared to a positive control (T+, 10% Triton X-100) and a negative control (Untreated), n = 3.

Figure 6.

Evaluation of antiviral activity of compounds 1 and 7. Effect of (a) compound 1 and (b) compound 7 on the cytopathogenic effect by coronavirus hCoV-229 at different viral loads, n = 3. (c) Tissue culture infectious dose (50%) (TCID₅₀), defined as the dilution of the virus required to infect 50% of the cell culture, in the presence of different concentrations of compound 7, n = 3.

Figure 7.

Docking results of N3, ebselen, and compound 7 against the catalytic site of M^pro. (a) Highest scoring pose (yellow) and the crystallographic N3 pose (green). (b) Highest scoring pose of ebselen. (c) Highest scoring pose of compound 7. (d–f) 2D interaction maps for the highest scoring poses of N3, ebselen and compound 7 with M^pro, respectively.

Figure 8.

Docking results of ebselen and compound 7 in the region between domains II and III of M^pro. The best pose of ebselen (a) and compound 7 (d) in that site. 3D visualisation of the key residues involved in the interaction between M^pro and ebselen (b) or compound 7 (e). 2D interaction maps for the best poses of ebselen (c) and compound 7(f) on M^pro.