A Priori Sampling of Transition States with Guided Diffusion

NeurIPS 2025 Workshop FPI Submission28 Authors

Track: Main Track

Keywords: Transition State Search, Score-based Generative Models, Inference-time Control

TL;DR: We introduce ASTRA, a method that uses a guided score-based generative model to directly find conformational and chemical transition state guesses without any prior knowledge about the transition state region.

Abstract:

Transition states, the first-order saddle points on high-dimensional potential energy surfaces, govern the kinetics and thus the mechanisms of chemical reactions and conformational changes.
Thus, identifying them is a central challenge in molecular simulations for physical and life sciences.
Existing methods for locating them often require knowledge about the transition, such as a good initial guess of the transition pathway or reaction coordinates.
We introduce ASTRA, **A** Priori **S**ampling of **TRA**nsition States with Guided Diffusion, a workflow that reframes this search problem as a direct generative task.
ASTRA utilizes a score-based diffusion model trained exclusively on configurations from known metastable states, requiring no prior data from the transition region or prior information about the reaction coordinate.
During inference, ASTRA guides the generative process to draw samples from the isodensity surface that separates the basis of stable states with a principled composition of conditional scores.
This process is coupled with a Score-Aligned Ascent mechanism that maximizes the energy along the score-based reaction coordinate approximation, effectively collapsing the sampling onto the transition state ensemble.
We validate our approach on a series of benchmarks, from 2D potentials to the high-dimensional conformational changes of alanine dipeptide, the folding of the chignolin protein, and a chemical reaction.
Our results demonstrate that ASTRA not only locates transition states with high precision but also discovers competing reaction pathways in mechanistic studies.

Submission Number: 28