Towards Automated Reaction Kinetics with Message Passing Neural Networks

Abstract

Predictive chemistry holds great promise to accelerate scientific discovery and innovation. An approach towards predictive chemistry involves decomposing systems into kinetic mechanisms consisting of elementary reactions and quantitatively describing each of those reactions. Incredibly, the immense progress in computational methods and compute power now allows the calculation of thermodynamic and kinetic parameters at an accuracy necessary for predictive chemistry. Unfortunately, real systems can consist of tens of thousands of elementary reactions, so it is infeasible to calculate these parameters using traditional, labor-intensive computational methods.

This thesis focuses on computing kinetic parameters by both automating and accelerating the computational pipelines used to generate them, relying on modern machine learning frameworks— specifically, message passing neural networks—to facilitate these calculations.

Noting that in the framework of automated kinetic parameter calculation, transition state search is a key bottleneck, this thesis first devises a method to generate transition state geometries with deep learning. The new method achieves improvements in both accuracy and speed compared to existing alternatives. This thesis next investigates a fundamental limitation of message passing neural networks to capture tetrahedral chirality and proposes several fixes to address this limitation. While generating a single transition state structure is an important goal, accurate calculation of kinetic parameters often requires investigating multiple conformations. Hence, this thesis builds a generative framework to predict multiple low-energy conformations directly from the molecular graph. The method is demonstrated for stable species conformer generation and outperforms existing baselines. Integrating all the developed models together, this thesis next develops an end-to-end pipeline to generate transition state conformers directly from the atom-mapped reaction SMILES. While most of presented work investigates reactions in the gas phase, reactions in condensed phase require additional solvation corrections. Therefore, this thesis constructs a large dataset of solution free energies across a range of solvents. It then develops a model to predict relevant conformations of the solute for any given solute-solvent pair.

The tools developed in this thesis will become an integral part of modern computational chemistry pipelines. Undoubtedly, the future of automated predictive chemistry will heavily rely on these and similar deep learning models for fast and accurate parameter estimation.