Reinforcement Learning using a Molecular Fragment Based Approach for Reaction Discovery

Ajnabiul Hoque, Mihir Jitendra Surve, Shivaram Kalyanakrishnan, Raghavan B Sunoj

Keywords: reinforcement learning, transfer learning, reaction discovery, deep generative model

TL;DR: A multi-pronged deep learning approach using a fragment based method is applied to chemical reaction discovery

Abstract: Deep learning methods have recently been applied to both predictive and generative tasks in the molecular space. While molecular generation and prediction of an associated property are now reasonably common, studies on reaction outcome due to the generated molecules remain less explored. Chemical reactions present a complex scenario as they involve multiple molecules and the breaking/forming of bonds. In reaction discovery, one aims to maximise yield and/or selectivity, which depends on a multitude of factors, including partner reactants and reaction conditions. We propose a multi-pronged approach that combines policy gradient reinforcement learning with a recurrent neural network-based deep generative model to identify prospective new reactants, whose yield/selectivity is estimated by a pre-trained regressor. Using SMILES (simplified molecular-input line-entry system) as the raw representation, our approach involves attaching a user-defined core fragment to the generated molecules for reaction-specific learning. On three distinct reaction types (alcohol deoxyflourination, imine-thiol coupling, asymmetric hydrogenation of imines and alkenes), we obtain notable improvements in yield and enantioselectivity. The generated molecules are diverse, while remaining synthetically accessible.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )