Robust generative transition-state models for unseen chemistry

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Reaction Dynamics

Background:

• Predicting transition states (TSs) is crucial for understanding chemical reaction rates and outcomes.
• Current machine learning (ML) models show high accuracy for small organic reactions but lack generalization.
• The ability of generative ML models to predict TSs for novel chemical environments is largely unexplored.

Purpose of the Study:

• To develop and evaluate methods for improving the generalization of generative ML models for transition state prediction.
• To create targeted benchmarks assessing model performance on chemically novel systems.
• To address limitations in predicting TSs for systems with unseen elements and transition metal complexes.

Main Methods:

• Curated extensions of the Transition1x dataset with elemental substitutions and transition metal complexes (TMCs).
• Introduced targeted benchmarks to probe chemical and structural novelty in generative TS prediction.
• Developed a self-supervised pretraining strategy using equilibrium conformers for generative TS models.

Main Results:

• Generative models exhibit fundamental limitations in generalizing to unseen elements.
• Self-supervised pretraining significantly enhances TS prediction for novel systems.
• Reduced median root-mean-square deviation of TS geometries on Transition1x-TMC reactions.
• Improved performance in low-data regimes, reducing fine-tuning data requirements.

Conclusions:

• Self-supervised pretraining is a viable strategy to improve the generalization of generative TS models.
• The proposed method enhances TS prediction accuracy and data efficiency for unseen chemical systems.
• This approach enables reliable TS prediction even with limited fine-tuning data, advancing computational chemistry.