Robust generative transition-state models for unseen chemistry.

Self-supervised pretraining enhances generative transition-state models, reducing median RMSD by over 50% on unseen transition metal complexes in Transition1x-TMC reactions.

• Why it matters: Accurate prediction of transition states is crucial for understanding reaction mechanisms, yet current models struggle to generalize beyond small organic molecules, limiting their applicability to diverse chemical systems.
• What they did: The study developed targeted benchmarks with elemental substitutions and transition metal complexes, and implemented a self-supervised pretraining approach based on equilibrium conformers to improve model generalization.
• The result: Pretraining significantly boosts prediction accuracy for unseen systems, decreases the need for extensive fine-tuning data, and enables reliable transition state modeling in low-data scenarios, advancing computational chemistry capabilities.