Conformer ensemble generation is an integral component of computational modeling of catalytic systems. Reaction path tools for investigating catalytic mechanisms have been shown sensitive to reactant, product, and ligand conformers because these methods search along locally minimal reaction paths corresponding to individual conformers. However, due to the computational and intellectual expense of conformer search, lacking physically relevant conformers is often a limiting factor in modelling pipelines. While the number of potential conformers is combinatorically large, individual conformer candidates generated algorithmically can be evaluated computationally to provide immediate feedback. The field of reinforcement learning provides a formal framework for such sequential interactions of an algorithmic agent with a simulated environment. We formally characterize conformer ensemble generation within the context of this framework. The algorithmic agent chooses torsion angles, and the environment evaluates a resulting ensemble of locally optimized, unique conformers using a thermodynamically natural reward metric based on free energy. We demonstrate that standard reinforcement learning agents can learn to generate ensembles of low energy conformers. We also provide our framework for analyzing and evaluating reinforcement learning agents on conformer ensemble generation in a systematic way. We hope this framework can be assistive in our ongoing investigations into a polyolefin catalyst, extended with additional macromolecular environments of interest to the community, and used as a training ground for algorithmic agent development.

TorsionNet: Generating conformer ensembles with reinforcement learning

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.