Generative Autoencoders for Designing Novel Small-Molecule Compounds as Potential SARS-CoV-2 Main Protease Inhibitors

Abstract

AbstractTwo generative autoencoder models for designing novel drug-like compounds able to block the catalytic site of the SARS-CoV-2 main protease (MPro) critical for mediating viral replication and transcription were developed using deep learning methods. To do this, the following steps were performed: (i) architectures of two neural networks were constructed; (ii) a virtual compound library of potential anti-SARS-CoV-2 MPro agents for training two neural networks was formed; (iii) molecular docking of all compounds from this library with MPro was made and calculations of the values of binding free energy were carried out; (iv) two neural networks were trained followed by estimation of the learning outcomes and work of two autoencoders involving several generation modes. Validation of autoencoders and their comparison revealed the best combination of the neural network architecture with the generation mode, which allows one to generate good chemical scaffold for the design of novel antiviral drugs with suitable pharmaceutical properties.KeywordsSARS-CoV-2Main proteaseDeep learningGenerative autoencoderSemi-supervised learningVirtual screeningMolecular dockingBinding free energy calculationsAnti-SARS-CoV-2 drugs