Virtual screening of Mao-B inhibitors using generative deep learning and target focused statistical potential

Abstract

Generative deep learning is actively used to derive models for de novo design of new drug candidates because it can explore the broad chemical space of drug-like molecules efficiently. The aim of this study was to compare several generative deep learning methods such as ORGANIC, INVENT, DrugEx, and ReLease, which were based on the recurrent neural network (RNN) and enhanced with the reinforcement learning (RL). The diversity and uniqueness of generated structures after the application of current generative methods were investigated. The general generation model was trained using about one million number of small drug-like molecules from CheMBL database in SMILES string format, and then the model was refined to be biased to a single protein target. Similar to the generative adversarial network (GAN), our model was consisted of generator and discriminator networks; the generator was trained to generate chemically valid representation of small molecules and the discriminator based upon random forest method was trained to predict the binding activity of new compounds toward the target protein after their conversion into the 2D fingerprints. After the generation of new molecules, their binding affinity was further estimated using a statistical potential developed for the specific target protein for cross checking and filtering of candidate molecules toward next experimental research. We chose Monoamine oxidase-B (MOA-B) enzyme, which is related to the Alzheimer's disease, for the development of new inhibitors, and the result of virtual screening using the generative deep learning will be shown here with the implications of this new approach.