We propose a novel computational strategy based on deep and reinforcementlearning techniques for de-novo design of molecules with desired properties.This strategy integrates two deep neural networks -generative and predictive -that are trained separately but employed jointly to generate novel chemicalstructures with the desired properties. Generative models are trained toproduce chemically feasible SMILES, and predictive models are derived toforecast the desired compound properties. In the first phase of the method,generative and predictive models are separately trained with supervisedlearning algorithms. In the second phase, both models are trained jointly withreinforcement learning approach to bias newly generated chemical structurestowards those with desired physical and biological properties. In thisproof-of-concept study, we have employed this integrative strategy to designchemical libraries biased toward compounds with either maximal, minimal, orspecific range of physical properties, such as melting point andhydrophobicity, as well as to develop novel putative inhibitors of JAK2. Thisnew approach can find a general use for generating targeted chemical librariesoptimized for a single desired property or multiple properties.
展开▼
