Learning to Discover Rare Molecular Transition Pathways

This direction is merged into another direction: rare event sampling, collective variable-free sampling. Its last version stays readable and citable here.

Version 1 of 120 papers25 conceptsfrozen 2026-10-04

In the study of molecular systems, the most interesting events—protein folding, chemical reactions, or crystal nucleation—are also the most frustrating to simulate. These are "rare events": the system spends the vast majority of its time vibrating in a stable state (a local energy minimum) and only occasionally hops over a high energy barrier to reach a new configuration.

Standard molecular dynamics (MD) simulations are like watching a ball roll around the bottom of a bowl, waiting for a random gust of wind to blow it over a mountain range. To speed this up, researchers have traditionally used "enhanced sampling" methods that apply a bias force to push the system along a specific path. However, this requires knowing the path in advance.

The Collective Variable Bottleneck

To guide a simulation, you need a map. In computational chemistry, this map is usually defined by Collective Variables (CVs)—a few hand-picked coordinates (like the distance between two atoms) that are thought to capture the essence of the transition.

The problem, as Alberto Baiardi and colleagues argue, is that human intuition is a poor guide for high-dimensional energy landscapes. When we pre-define CVs, we bake our own biases into the simulation. We might miss "kinetic traps"—hidden metastable states that don't align with our chosen descriptors—or fail to see complex, multi-atom mechanisms that are structurally non-obvious Rodriguez et al..

A new research direction is emerging to solve this "chicken-and-egg" problem: how do you find a transition path if you need the path to guide the search? The answer lies in automated, CV-free discovery.

Transitions as Stochastic Optimal Control

One of the most powerful ways to bypass human-defined coordinates is to treat the molecular transition as a Schrödinger Bridge Problem. In this framework, we aren't just pushing the molecule; we are looking for the "cheapest" way (in terms of energy and probability) to get from State A to State B.

Lars Holdijk, Max Welling, and their team developed PIPS (Path Integral Path Sampling) to solve this using Stochastic Optimal Control Holdijk et al.. Instead of a human picking a few distances to monitor, a neural network looks at the full 3D atomic coordinates of the system. The network learns a "bias potential"—a custom-shaped force field—that guides the molecule across the energy barrier. Because the network operates on the raw geometry, it can discover complex mechanisms, like specific hydrogen bond formations, that a human might never think to include in a CV Seong et al..

Blind Exploration and Trajectory Space

What if you don't even know where State B is? This is the "blind exploration" problem. Timothée Devergne and Michele Parrinello have explored methods to drive a system out of its current basin using only raw descriptors like interatomic distances, without a specific target in mind Devergne et al..

A related approach, Metadynamics of Paths (MoP), shifts the focus from the position of the atoms to the trajectory itself. By adding repulsive potentials in "path space," researchers like Davide Mandelli and Barak Hirshberg can force the simulation to find entirely new reaction mechanisms and metastable states that were previously hidden Mandelli & Hirshberg. This turns the simulation into an open-ended explorer rather than a guided tour.

The Committor: The Ultimate Descriptor

If we want to move away from arbitrary CVs, what should we use instead? The gold standard is the committor function—the probability that a system starting at a specific point will reach the final state before returning to the initial one.

Recent work has turned to neural networks to solve for this function iteratively. Pei‐Lin Kang and Enrico Trizio demonstrated that by using the committor’s own gradient as a bias, they can "trap" a simulation right on the transition state—the thin ridge between two energy basins Kang & Trizio. Similarly, Andrew Mitchell uses "swarms" of simulations guided by a learned committor to map out high-dimensional pathways without any pre-defined reaction coordinates.

Generative Models as Pathfinders

The most recent frontier involves repurposing the generative AI models used for images and text. Sanjeev Raja and colleagues have shown that pre-trained diffusion models can be used to find transition pathways in a "zero-shot" manner Raja et al.. By minimizing the Onsager-Machlup action functional—a physics-based cost function—within the learned space of a generative model, they can identify high-probability paths between molecular states without expensive retraining.

By moving from manual descriptors to automated, high-dimensional learning, this research direction is transforming molecular simulation from a process of verifying known mechanisms into a tool for discovering new ones.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_learning_to_discover_rare_molecular_tran_7998dc0b,
  title = {Learning to Discover Rare Molecular Transition Pathways},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-to-discover-rare-molecular-transition-pathways/rdr_7998dc0b5c9d4fc687bdcfdc46b92a51},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_learning_to_discover_rare_molecular_tran_7998dc0b_v1,
  title = {Learning to Discover Rare Molecular Transition Pathways},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-to-discover-rare-molecular-transition-pathways/rdr_7998dc0b5c9d4fc687bdcfdc46b92a51/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}