React-OT: Optimal Transport for Generating Transition State in Chemical Reactions (2404.13430v2)

Abstract: Transition states (TSs) are transient structures that are key in understanding reaction mechanisms and designing catalysts but challenging to be captured in experiments. Alternatively, many optimization algorithms have been developed to search for TSs computationally. Yet the cost of these algorithms driven by quantum chemistry methods (usually density functional theory) is still high, posing challenges for their applications in building large reaction networks for reaction exploration. Here we developed React-OT, an optimal transport approach for generating unique TS structures from reactants and products. React-OT generates highly accurate TS structures with a median structural root mean square deviation (RMSD) of 0.053{\AA} and median barrier height error of 1.06 kcal/mol requiring only 0.4 second per reaction. The RMSD and barrier height error is further improved by roughly 25\% through pretraining React-OT on a large reaction dataset obtained with a lower level of theory, GFN2-xTB. We envision that the remarkable accuracy and rapid inference of React-OT will be highly useful when integrated with the current high-throughput TS search workflow. This integration will facilitate the exploration of chemical reactions with unknown mechanisms.

• The paper introduces React-OT, an optimal transport model that deterministically generates transition state structures without stochastic sampling.
• It achieves a median RMSD of 0.044 Å and a barrier height error of 0.74 kcal/mol, outperforming previous diffusion-based models.
• Its integration into high-throughput workflows reduces inference times to 0.4 sec per reaction, significantly lowering computational costs.

Implementing Optimal Transport in Transition State Generative Models: A Detailed Analysis of React-OT

Introduction

Transition states (TSs) occupy a crucial role in the paper of chemical reactions, defining the highest energy states along the reaction path. Accurately capturing these TS structures could vastly improve the understanding and management of kinetic pathways in computational chemistry. Despite advances, the direct observation and characterization of TSs remain a challenge in experimental settings, leading to reliance on computational methods like Density Functional Theory (DFT). However, these methods are computationally expensive. Recently, machine-learning models have been proposed as an alternative to predict TSs efficiently. Our main advancement discussed here is React-OT, an optimal transport model for generating transition states deterministically with high accuracy and low computational cost.

React-OT Overview

React-OT is fundamentally different from its predecessors in that it deterministically generates TS structures by employing an optimal transport approach, simulating the sampling process as an ordinary differential equation (ODE). This deterministic nature of React-OT diverges from traditional stochastic diffusion processes used in models like OA-ReactDiff, thus eliminating a need for multiple sampling and subsequent ranking to identify the most probable TS structure.

• Key features and improvements:
  • React-OT achieves a median RMSD (root mean square deviation) of 0.044 Å and a barrier height error of about 0.74 kcal/mol when pretrained on a substantial dataset (RGD1-xTB).
  • The model outperforms previous stochastic models in both the accuracy of the TS structures and the computational efficiency, reducing the typical inference time to about 0.4 seconds per reaction.

React-OT Performance and Optimization

One of the striking adaptations in React-OT is the shift from Gaussian-based random initial guesses to correlated, more chemically relevant ones derived from linear interpolations of reactants and products. This adaptation not only preserves the necessary symmetry properties of the reactions but also simplifies the inference process by eliminating randomness and enhancing stability.

• Findings from direct comparisons:
  • Compared to OA-ReactDiff, React-OT has significantly reduced the mean RMSD from 0.180 Å to 0.103 Å.
  • React-OT considerably decreases barrier height errors and achieves closer estimations towards experimental or high-level theoretical methods.

Future Potential and Practical Implications

React-OT's architecture allows it to integrate seamlessly into larger computational workflows, particularly those involving high-throughput DFT calculations. This integration promises substantial reductions in computational costs while maintaining high accuracy. The model's ability to deliver quick and reliable TS predictions could revolutionize how chemists and material scientists explore reaction mechanisms and design catalysts.

• Integration into larger systems:
  • When incorporated into a hybrid machine learning-DFT workflow, React-OT can reduce computational overheads by up to seven-fold while still achieving results within chemical accuracy.
  • The model's robustness even extends to cases where reactants and products are only approximate, showing minimal degradation in performance when using structures optimized with lower-cost methods like GFN2-xTB.

Conclusions

The development of React-OT marks a significant step forward in the computational exploration of chemical kinetics, especially in the ability to efficiently and accurately predict TS structures. By effectively leveraging an optimal transport framework, React-OT not only improves upon the accuracy and speed of existing machine learning models but also addresses the practical and operational limitations associated with expansive quantum chemical computations. The potential extensions of this model to other domains, including materials science and biological systems, provide a promising avenue for future research and application.