[Paper Review] React-OT: Optimal Transport for Generating Transition State in Chemical Reactions

Chenru Duan, Guan-Horng Liu+7 more|arXiv (Cornell University)|Apr 20, 2024

Radioactive element chemistry and processing4 citations

TL;DR

React-OT is a deterministic optimal transport framework that generates highly accurate transition state (TS) structures for chemical reactions in 0.4 seconds per reaction, achieving a median RMSD of 0.053 Å and barrier height error of 1.06 kcal/mol. Pretraining on GFN2-xTB data improves accuracy by 25%, reducing RMSD to 0.044 Å and error to 0.74 kcal/mol, enabling fast, reliable, and scalable TS prediction without stochastic sampling or ranking models.

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 0.053Å 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.

Motivation & Objective

• To develop a fast, deterministic method for generating accurate transition state (TS) structures in chemical reactions to overcome the high computational cost of DFT-based TS optimization.
• To eliminate the need for stochastic sampling and post-hoc ranking models required by existing generative models like OA-ReactDiff, which hinder scalability in high-throughput workflows.
• To improve structural and energetic accuracy of TS prediction by leveraging optimal transport theory and pretraining on large-scale, low-cost quantum chemistry data.
• To enable efficient construction of large reaction networks by drastically reducing inference time and computational overhead per reaction.
• To provide a scalable, reliable, and symmetry-aware framework for transition state prediction that integrates seamlessly into high-throughput reaction mechanism exploration pipelines.

Proposed method

• React-OT employs optimal transport to map reactant and product structures to a unique, deterministic transition state by minimizing a transport cost between their geometric distributions.
• The model uses an SE(3)-equivariant graph neural network (LEFTNet) as the scoring network to preserve 3D spatial symmetries and ensure geometric consistency in the generated TS.
• It formulates the TS generation as a conditional inpainting problem, where reactants and products are used as conditioning inputs during the denoising process of a diffusion model.
• The method integrates a confidence model fine-tuned from OA-ReactDiff to predict reliability scores for generated TS structures without additional training.
• React-OT is trained via fine-tuning on an existing OA-ReactDiff checkpoint with a new optimal transport objective, using a reduced initial learning rate of 0.0001.
• Pretraining on a large dataset of 760,615 reactions computed with GFN2-xTB (low-cost semi-empirical DFT) enhances generalization and accuracy, particularly in barrier height prediction.

Experimental results

Research questions

• RQ1 Can optimal transport be used to generate a single, deterministic transition state structure from reactants and products with high structural and energetic accuracy?
• RQ2 How does the performance of a deterministic optimal transport-based method compare to stochastic diffusion models like OA-ReactDiff in terms of inference speed and accuracy?
• RQ3 To what extent can pretraining on low-cost quantum chemistry data (GFN2-xTB) improve the accuracy of transition state prediction in a diffusion-based generative model?
• RQ4 Can the integration of optimal transport with equivariant deep learning eliminate the need for post-hoc ranking models in TS generation?
• RQ5 What is the computational efficiency gain of React-OT over existing high-throughput TS search workflows based on DFT or diffusion models?

Key findings

• React-OT generates transition state structures with a median root mean square deviation (RMSD) of 0.053 Å from reference DFT structures, indicating high structural accuracy.
• The median barrier height error is 1.06 kcal/mol, demonstrating strong energetic prediction capability.
• Inference time is only 0.4 seconds per reaction, enabling rapid deployment in high-throughput reaction network generation.
• After pretraining on GFN2-xTB data, the median RMSD improves to 0.044 Å and barrier height error drops to 0.74 kcal/mol, representing a 25% improvement in accuracy.
• React-OT reduces the number of denoising function evaluations from 40,000 (in OA-ReactDiff) to just 50, achieving a near 1000-fold acceleration in inference.
• The method eliminates the need for stochastic sampling and additional ranking models, enabling deterministic, scalable, and reliable TS prediction.