Rapid Generation of Transition-State Conformer Ensembles via Constrained Distance Geometry

Stefan P. Schmid, Henrik Seng, Thibault Kläy, Kjell Jorner

TL;DR

This paper introduces racerTS, a new, computationally efficient method for generating transition-state (TS) conformer ensembles. By leveraging constrained distance geometry and the RDKit library, racerTS achieves significant speedups over existing methods like CREST and GOAT, while maintaining comparable or better accuracy and validity, making it ideal for high-throughput computational chemistry and machine learning applications.

What Problem Does This Paper Address?

Accurately modeling chemical reactions, particularly in catalyst design, requires understanding their transition states (TSs). Crucially, this involves not just a single TS structure but an ensemble of possible conformers, as different conformers can significantly impact predicted reaction properties like selectivity. While existing methods like CREST and GOAT can generate these ensembles, they are computationally expensive, creating a bottleneck for large-scale studies, especially for generating data for machine learning models. This paper addresses the need for a faster, yet reliable, TS conformer ensemble generation method.

Methods

The core of racerTS is its adaptation of RDKit's distance geometry algorithm for TS conformer generation. Unlike standard conformer generation, racerTS incorporates constraints on the positions and distances of atoms within the "reaction center" (those involved in bond breaking/forming) to preserve the TS's unique topology. The process involves:

  1. Input: A TS guess (3D structure, charge) and optionally reactant/product SMILES.
  2. Topology Inference: Using RDKit to determine atom and bond types.
  3. Constrained Distance Geometry: Sampling conformers by generating random distances within a defined bounds matrix for the reaction center and its neighbors.
  4. Force-Field Minimization: Initial refinement using the MMFF94 force field.
  5. Pruning: Removing duplicate conformers based on RMSD and filtering out high-energy conformers within a specified energy window.

This approach is designed to be significantly faster than physics-based simulation methods like metadynamics used in CREST.

Datasets

The study evaluated racerTS using a benchmark set of 20 diverse reactions, including organometallic complexes, sourced from existing literature. This dataset was used to compare racerTS against CREST and GOAT across several metrics. No new datasets were introduced for training.

Key Insights and Ideas

  • Distance Geometry for TSs is Viable: The paper demonstrates that adapting distance geometry, a fast method for ground-state conformers, can be effectively applied to TSs by introducing appropriate constraints.
  • Computational Cost is the Major Bottleneck: The significant speedup achieved by racerTS highlights how computational cost has been a major barrier in TS conformer generation, limiting its application in high-throughput screening.
  • Constraints are Key for TS Validity: The careful selection and constraint of reaction center atoms and their neighbors are crucial for generating chemically meaningful TS conformers.
  • Post-Optimization Matters: While racerTS is fast, the study shows that a GFN2-xTB//GFN-FF post-optimization step significantly improves the quality of conformer ranking and accuracy, suggesting a trade-off between raw speed and refined accuracy.

Key Results

  • Speedup: racerTS offers a substantial speedup, achieving up to a 4x improvement over CREST and a 500x improvement over GOAT in terms of wall-time for conformer ensemble generation.
  • Exhaustiveness: racerTS demonstrates comparable conformer space coverage to CREST, though slightly less comprehensive than GOAT.
  • Validity: racerTS exhibits a higher success rate in generating valid TS structures that can be successfully optimized at the DFT level, outperforming both CREST and GOAT. It successfully generated at least one valid TS for all 20 benchmarked reactions, whereas CREST and GOAT had failures.
  • Accuracy: The method provides sufficient accuracy in the low-energy region for computational chemistry applications, with a median error of 0.17 kcal/mol for activation energy calculations.
  • Computational Cost: The racerTS (no xTB) variant is the fastest but sacrifices some accuracy, indicating that the GFN2-xTB//GFN-FF post-optimization is important for high-quality results.

Limitations and Future Work

  • Outliers: While generally accurate, racerTS can produce high-energy outliers in specific cases (e.g., due to unfavorable rotamers or boat-like structures), which might be mitigated by sampling more conformers at an increased computational cost.
  • CPU vs. GPU: The current benchmarking was performed on CPUs. Future work could explore GPU acceleration for racerTS to further reduce computational costs, leveraging recent GPU-accelerated RDKit functions.
  • Dataset Size: The study acknowledges that benchmarking on 20 reactions, while diverse, might lead to variances in results, and a larger dataset could provide more robust conclusions.
  • Semiempirical/Force-Field Ranking: The relatively low top-1 accuracy in identifying the overall lowest-energy conformer is hypothesized to be due to inaccuracies in semiempirical or force-field energy rankings, suggesting that optimizing every conformer at the DFT level or using more advanced methods could improve this.

Summary updated

Abstract

Consideration of transition-state (TS) conformer ensembles is required to accurately model a reaction, and thus plays a key role in computational catalyst design. While CREST and GOAT are established methods for TS conformer ensemble generation, the associated computational cost remains a major bottleneck in computational chemistry pipelines, including for the generation of large machine learning data sets for catalyst design. To this end, we present racer<sup>TS</sup> (<b>RA</b>pid <b>C</b>onformer <b>E</b>nsembles with <b>RDK</b>it for <b>T</b>ransition <b>S</b>tates), a method for efficient TS conformer ensemble generation. In this work, we describe the algorithm behind racer<sup>TS</sup>, which is based on constrained distance geometry. To benchmark the performance of racer<sup>TS</sup> against CREST and GOAT, we created conformer ensembles for transition states of 20 diverse reactions. To assess the utility of each conformer generator in computational chemistry workflows, we optimize selected low-energy and diverse conformers at the DFT level. We use the generated conformer ensembles and the results of this pipeline to assess conformer generators according to the following metrics: computational cost, exhaustiveness, validity, and accuracy in low-energy regions. Considering the generated ensembles, we find that racer<sup>TS</sup> covers the conformer space similarly to CREST and slightly less comprehensively than GOAT, while the validity of the DFT-optimized TSs is better and the accuracy in the low-energy region is sufficient for computational chemistry applications (median error of 0.17 kcal/mol). Remarkably, racer<sup>TS</sup> achieves these results with a significant reduction in required wall-time. Our results demonstrate that racer<sup>TS</sup> is a highly efficient TS conformer ensemble generator, allowing for rapid TS conformer sampling in computational chemistry pipelines. Additionally, racer<sup>TS</sup> paves the way to create meaningful TS data sets to advance machine learning methods for the discovery of novel and sustainable catalysts.
BibTeX
@article{schmid2026rapid,
  author = {Stefan P. Schmid and Henrik Seng and Thibault Kläy and Kjell Jorner},
  title = {{Rapid Generation of Transition-State Conformer Ensembles via Constrained Distance Geometry}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2026},
  doi = {10.1021/acs.jcim.5c02794},
}
Figure