Development of a deep learning generative neural network for computer-aided design of potential SARS-Cov-2 inhibitors

Two generative deep learning models have been developed for the computer-aided design of potential inhibitors of the SARS-CoV-2 main protease (M$^{\mathrm{Pro}}$), an enzyme critically important for the virus replication and transcription, and, therefore, presenting a promising target for the design of effective antiviral drugs. To solve this problem, we formed a training library of small molecules containing structural elements capable of providing specific and effective interactions of potential ligands with the SARS-CoV-2 M$^{\mathrm{Pro}}$ catalytic site. The architecture of generative models was developed and implemented to generate new high-affinity ligands of this functionally important SARS-CoV-2 protein. The neural network was trained and tested on the compounds from the training library, and the results of training and operation in two different generation modes were evaluated. The use of generative models in conjunction with the molecular docking demonstrated their great potential for filling the unexplored regions of the chemical space with novel molecules with pre-defined properties, which is confirmed by the obtained results according to which out of 4805 compounds generated by the neural network only one compound was present in the original data set.