 |
Center for Nonlinear Studies |
In this work we propose a novel AI-based computational method for de-novo design of molecular compounds with desired properties. General workflow of the proposed method integrates two deep neural networks – generative and predictive – that are initially trained separately but then trained jointly to generate novel chemical structures with desired properties. The generative model is a recurrent neural network with stack-augmented memory, which contains millions of trainable parameters. Firstly, the generative model is pre-trained on a vast dataset of chemical compounds to produce chemically feasible molecules without any property optimization at this point. This stage of the algorithm is highly computationally intensive and efficient implementation requires GPU usage. Afterwards, the pre-trained model is then used as a starting point for further optimization with reinforcement learning techniques. In parallel, the predictive model is trained to estimate the desired property given the molecular representation. Secondly the generative model is fine-tuned by the predictive model with reinforcement learning algorithm to produce molecules with optimized property values. SMILES representation of molecules is used both for model training and new molecule design escaping the conventional cheminformatics routine of chemical descriptor calculation. In this proof-of-concept study, we have employed this integrative strategy to design chemical libraries biased toward compounds with either maximal, minimal, or specific ranges of physical properties, such as melting point and hydrophobicity, as well as to develop novel putative inhibitors of JAK2. This new approach can find general use for generating targeted chemical libraries optimized for a single desired property or multiple properties.
Host: Ben Nebgen