Deep learning based computational drug discovery to inhibit the RNA Dependent RNA Polymerase: application to SARS CoV and COVID 19

There is an urgency to find drugs and vaccines for the 2019 coronavirus disease (COVID-19). Therapeutic options include repurposing existing drugs or finding new ones. One approach is to target the RNA-dependent RNA polymerase (RdRp) and block viral RNA synthesis. Currently clinical trials to repurpose remdesivir, a RdRp targeting pro-drug for Ebola, to COVID-19 is under way. More such potential drugs need to be identified to efficiently find best therapeutic options. To address this need, a Long Short Term Memory (LSTM) model from literature was trained to read the SMILES fingerprint of a molecule and predict the IC50 of the molecule when binding to an RdRp. This model was trained using IC50 binding data from the PDB database. 310,000 drug-like compounds from the ZINC database were then screened using the trained LSTM model. Additionally, the 310,000 molecules with their predicted IC50s were used to train a generative Semi-Supervised Variational AutoEncoder (SSVAE) model from literature. Although not trained by actual experimental data (sufficient data are not available), the SSVAE model was used to generate 10 new molecules by sampling from the latent space to demonstrate its utility. These 10 molecules and the 1025 molecules with the lowest predicted IC50s from the LSTM model were docked onto the SARS coronavirus (a virus similar to COVID-19) RdRp using AutoDock Vina. Top four most stable inhibitors from the screened ZINC database compounds had binding energies of less than -33.89 kJ/mol. These binding energies were less than the binding energies of the comparison group consisting of prior drugs remdesivir, favipiravir, and galidesivir. Among the ten new molecules generated by the SSVAE model, the most stable new molecule had binding energy lower than the comparison group of prior drugs. The low binding energies of these molecules indicate they could potentially be good drug candidates for the SARS CoV and COVID-19. These results also show the utility of deep learning-based models in screening existing compound and generating new molecules to find drugs for COVID-19.