Towards molecular design with desired property profiles and 3D conformer generation using Deep Generative Models

Abstract

As COVID-19 has made sure to show, drug discovery is a very relevant but also very challenging process: the goal is to find a target compound with some set of desired properties (de novo design) among the 1060 − 10100 theoretically possible drug-like molecules. Machine learning methods have proven to be effective to address this problem, allowing for efficient exploration of the vast chemical space. Whilst string based generative models have been thoroughly studied, there has not been yet a great focus on taking into account either the molecular graph or its 3D structure. That is why we propose two generative models: on the one hand, a graph-based deep generative model for targeted design using reinforcement learning; and, on the other hand, a deep auto-regressive generative model of small-molecules conformations. In this work we show the ability of the reinforcement learning framework to fine-tune the graph-based model towards generation of molecules with different desired sets of properties, even when few molecules have the goal attributes initially. We also present some early results in the search of a transferable molecular conformations generator that is able to learn from sample conformations and known rules of physics. These results suggest the potential of including structural information in molecular deep generative models. And, specially, motivate future work towards developing a reinforcement learning-based 3D model for de novo design.