Application of the LSTM-based deep generative model for de novo design of potential HIV-1 entry inhibitors

Abstract

A Long Short-Term Memory (LSTM) autoencoder model for the design of novel inhibitors of gp120, the HIV-1 envelope glycoprotein critically important for the virus pathogenesis, was repurposed and used to generate a series of compounds potentially active against this therapeutic target. Training and validation of this neural network was carried out using a set of small-molecule compounds collected by a public web-oriented virtual screening platform Pharmit allowing one to search for small molecules based on their structural and chemical similarity to another small molecule. The trained neural network was then evaluated for validity, and the values of binding free energy to the target protein were estimated. As a result, it was shown that the LSTM-based autoencoder model is an effective tool for the design of potent inhibitors against gp120 and may be used for the development of new drugs able to combat other dangerous diseases