Combined virtual screening and machine learning approach to finding novel SARS-CoV-2 protease inhibitors.

Our aim is to combine machine learning (ML) and molecular modelling to improve virtual screening and drug discovery applications targeting COVID-19. We have developed a genetic algorithm (based on work by Jan Jensen and incorporating a synthesizability filter) capable of searching chemical space surrounding existing antiviral drugs and a deep learning based classification model based on existing public coronavirus binding data (for the SARS-CoV-2 main protease). Here we will combine and extend these tools through a combination of docking and simulation which we can use as inputs to a regression based deep learning model. A key component of our approach will be to use an enhanced version of the out of distribution classification algorithms we created previously to design novel kinase (CDK9) inhibitors to identify molecules which have maximum value in terms of expanding the validity of our model. Enhancing our model from a classification model to one capable of regression in this way should provide greatly enhanced capabilities to identify both existing drugs with potential to treat COVID-19 (virtual screening) as well as the discovery of new active compounds.