AI maps 1‑in‑10²⁰ molecular events
- University of Amsterdam researchers Rik Breebaart and Peter Bolhuis, with Goethe University Frankfurt collaborators, reported an artificial-intelligence sampling method that maps whole rare-event reaction pathways, including ligand binding, from start to finish. - In Physical Review Letters, the team said its neural-network-guided simulation resolved probabilities as small as 1 in 10^20 and uncovered metastable intermediate states in a host-guest binding system. - The work targets a long-standing bottleneck in molecular simulation: rare transitions that standard dynamics almost never see, but that shape binding and reaction mechanisms. (aps.org)
Chemists at the University of Amsterdam and Goethe University Frankfurt reported an artificial-intelligence method that reconstructs extremely rare molecular events, including binding pathways, across an entire reaction. (aps.org) (uva.nl) These events are the outliers in molecular motion: a ligand finding a binding site, or a protein crossing an energy barrier, after huge numbers of failed tries. Standard molecular dynamics often misses them because it advances in tiny time steps while the decisive transition may be astronomically uncommon. (aps.org) (science.org) The new paper, published April 24 in *Physical Review Letters*, centers on the “committor,” a probability score for whether a molecular configuration will end in products or fall back to reactants. The authors said that quantity is the ideal reaction coordinate, but has been too hard to compute for high-dimensional systems. (aps.org) Breebaart, Gianmarco Lazzeri, Roberto Covino, and Peter Bolhuis built an iterative scheme that alternates between path sampling and a neural network trained on those sampled trajectories. Each round improves the estimate of the reaction coordinate, which in turn improves the next round of sampling. (aps.org) The group said the method predicted probabilities down to 1 in 10^20, which the University of Amsterdam described as one reactive trajectory in a billion times a billion unreactive ones. The paper tested the approach on a benchmark two-dimensional potential and on a host-guest binding and unbinding process in explicit solvent. (uva.nl) (aps.org) In plain terms, the system is trying to map not just the start and finish of a molecular event, but the narrow mountain pass between them. That pass can include “metastable” waypoints, short-lived shapes that persist long enough to redirect how binding or unbinding happens. (uva.nl) The Amsterdam team said those hidden intermediates appeared in its binding simulations and helped explain which ingredients matter at different stages of the process. The university’s write-up said the analysis tracked those changes as a function of the committor, rather than from a single static structure. (uva.nl) That fits a wider shift in computational chemistry away from single snapshots and toward ensembles, pathways, and transient pockets. Recent work in *Science* and *Science Advances* has pushed generative and machine-learning models toward protein motions and cryptic binding sites that conventional structure methods can miss. (science.org 1) (science.org 2) This study does not claim a finished drug-discovery product, and the flagship demonstration in the paper is a host-guest system rather than a full therapeutic target. But it gives chemists a way to ask a harder question with more detail: not only whether binding happens, but which fleeting structures make it possible. (aps.org) (uva.nl) The paper was published as a Physical Review Letters Editors’ Suggestion on April 24. Its core claim is simple: the rare molecular event is still rare, but the path to it no longer has to stay hidden. (aps.org)
