🔬 physics

Machine learning kinetics from molecular dynamics data

This review surveys modern self-supervised machine learning approaches for estimating the committor and other kinetic statistics from molecular dynamics data, unifying various methods under a common operator framework to overcome timescale limitations and offering guidance for practical applications and future research directions.

Original authors: Jonathan Weare, Aaron R. Dinner

Published 2026-09-17

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Jonathan Weare, Aaron R. Dinner

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Molecules in a liquid are never still. They jostle, collide, and constantly rearrange themselves, driven by the thermal energy of their environment. For scientists trying to understand how life works or how new materials form, the most critical moments are not these constant, small movements, but the rare, dramatic shifts where a molecule changes its shape or breaks apart to become something new. These events, like a protein folding into its functional form or a drug binding to a target, happen on time scales that are often millions of times longer than the tiny steps of a computer simulation can follow. It is a fundamental problem: to see the event, one must wait longer than a computer can reasonably calculate.

To bridge this gap, researchers rely on a statistical concept known as the committor. Imagine a molecule sitting in a valley between two hills. One hill represents the starting shape, and the other represents the finished shape. The committor is simply the probability that if you nudged the molecule from its current spot, it would roll over the first hill to finish, rather than rolling back to the start. If this probability is zero, the molecule is safely in the starting valley. If it is one, it is safely in the finished valley. If the probability is exactly one-half, the molecule is perched right on the ridge, the precise moment of decision where the path forward or backward is equally likely. Knowing this probability for every possible position of the molecule allows scientists to map the entire journey of a reaction, identifying the exact path the molecule takes and how fast it moves, without needing to wait for the event to happen naturally in a simulation.

For decades, calculating this probability required a brute-force approach. Scientists would take a snapshot of a molecule, launch dozens of new simulations from that exact spot, and count how many reached the finish line before the start. This method is incredibly expensive because it requires running simulations all the way to the end for every single point tested, and the most interesting points—the ones on the ridge—require the most simulations to get an accurate answer. A new review by Jonathan Weare and Aaron R. Dinner outlines a modern shift away from this brute force. Instead of launching endless simulations to count outcomes, they describe a suite of machine learning methods that learn the rules of the journey directly from the data of short, incomplete simulation snippets.

The core of this new approach is a change in how the computer is asked to learn. Rather than being told "this path leads to the finish, that one leads to the start," the machine is given a set of short movie clips of molecules moving and is asked to find a mathematical function that satisfies the laws of motion. Specifically, the method looks for a function where the average change in probability over a tiny step of time is zero, unless the molecule has already reached a destination. This is a self-supervised learning strategy. The computer does not need a teacher to label the end of every path; it only needs to ensure that its predictions are consistent with the physics of the system. By using neural networks—flexible mathematical structures capable of learning complex patterns—the researchers can represent the probability of reaching the finish line as a smooth surface over the entire landscape of molecular shapes.

The paper details several ways to achieve this. One method treats the problem like a puzzle where the computer tries to minimize the error in a physical equation. Another approach, which the authors highlight as particularly powerful, involves a technique called "stopping." In a standard simulation, a molecule might wander from the start, cross the finish line, and then wander back. This creates noise in the data. The new methods stop the simulation the instant the molecule hits either the start or the finish. This simple rule allows the computer to learn the probability of reaching the finish much more efficiently, even from very short simulation runs. The authors show that by using these stopped trajectories, the machine can learn the correct probability map without needing to know the detailed forces acting on every atom, a significant advantage when studying complex systems where those forces are hard to calculate.

The power of these methods is demonstrated through real-world examples. In one study, the researchers analyzed how a small protein called Trp-cage folds. Previous simulations had captured only a handful of folding events, making it difficult to see the details of the transition. Using the new machine learning techniques on a dataset of many short, carefully placed simulations, the team reconstructed the full probability map. They found that the protein does not follow a single, simple path but explores a wide region of shapes before committing to the final form. In another example involving insulin, a hormone that must separate into two parts to function, the method revealed four distinct pathways for the molecules to pull apart. Traditional theories had predicted a single, dominant path, but the data showed a much more complex reality with multiple routes and intermediate states that were previously hidden.

The review also addresses a major hurdle in these calculations: the "memory" of the system. When scientists simplify a complex molecule by tracking only a few key distances or angles, they lose information about the rest of the molecule. This loss means the simplified model remembers its past, violating the assumption that the future depends only on the present. The authors explain how newer methods can account for this memory, either by looking at a history of previous positions or by mathematically correcting for the missing information. This allows the machine learning models to remain accurate even when the description of the molecule is simplified, which is essential for studying large, complex systems.

A critical insight from the paper concerns how the data is collected. The authors argue that the most efficient way to learn is not to sample molecules randomly, as they would appear in nature, but to focus the sampling on the difficult transition region—the ridge where the probability is one-half. Random sampling wastes most of the computer's time on molecules that are safely in the start or finish valleys, where the answer is already known. By using the machine learning model to guide the sampling, researchers can concentrate their efforts exactly where the uncertainty is highest. This adaptive strategy allows them to build accurate models with far less computational power than before.

The paper concludes by connecting these chemical methods to broader fields like reinforcement learning, a branch of artificial intelligence used to teach computers to play games or control robots. The mathematical tools used to predict molecular transitions are essentially the same as those used to teach an agent how to navigate a maze. This cross-pollination suggests that advances in artificial intelligence can directly improve our ability to understand the physical world, and vice versa. The authors emphasize that while the mathematical framework is general and can be applied to many types of data, the true value lies in applying these tools to the diverse and difficult problems of chemistry and biology. By moving away from brute-force counting and toward learning the underlying rules from short, smartly chosen data, scientists can now map the hidden landscapes of molecular change with a clarity that was previously out of reach.

Technical Summary: Machine Learning Kinetics from Molecular Dynamics Data

Problem Statement
Molecular dynamics (MD) simulations are essential for linking molecular structures to function, yet they are limited by a timescale gap. While MD can resolve microscopic dynamics, it typically accesses timescales of $\sim 10^{-5}$ to $10^{-4}$ seconds, whereas many relevant chemical and biological phenomena (e.g., protein folding, catalysis) occur on timescales of $10^{-3}$ to $10^4$ seconds. Directly observing rare events in unbiased simulations is computationally prohibitive.

The central challenge is characterizing the kinetics of these transitions, specifically identifying the committor $q(x)$ . The committor is defined as the probability that a system in microscopic state $x$ reaches a product state ( $B$ ) before a reactant state ( $A$ ). It serves as the optimal reaction coordinate, defines transition states ( $q=0.5$ ) without prior mechanistic knowledge, and is the foundation for Transition Path Theory (TPT) and rate calculations. Traditional methods for estimating $q(x)$ , such as "shooting" (launching many trajectories from $x$ ), are computationally expensive and suffer from high statistical variance, particularly near the transition state where $q \approx 0.5$ . Furthermore, many existing methods rely on labeled data or require explicit knowledge of the dynamical generator (e.g., the drift and diffusion terms of a Langevin equation), which is often unavailable when analyzing reduced sets of variables or black-box simulators.

Methodology
The paper surveys and unifies modern approaches for estimating the committor and related kinetic statistics using machine learning, with a specific emphasis on self-supervised methods. These methods learn solutions to the defining dynamical equations of the committor without relying on labeled shooting data.

The authors establish a common operator viewpoint connecting several frameworks:

Partial Differential Equation (PDE) Approximation: When the generator $L$ of the dynamics is known (e.g., overdamped Langevin dynamics), $q(x)$ satisfies the backward Kolmogorov equation $Lq(x) = 0$ with boundary conditions $q=0$ in $A$ and $q=1$ in $B$ . This can be solved using Physics-Informed Neural Networks (PINNs) by minimizing the PDE residual. However, this requires second-order derivatives and explicit knowledge of the generator, which is often impractical for reduced variables.
Variational Principles: For reversible dynamics, the committor minimizes the Dirichlet form. Variational methods (e.g., Variational Committor Network, VCN) minimize a loss function based on the equilibrium flux. However, these typically require sampling from the stationary distribution $\pi(x)$ , which under-samples the transition region, and often fail to account for finite lag-time errors unless modified.
Trajectory-Based Self-Supervised Learning: The core contribution of the review focuses on methods that estimate statistics directly from trajectory data without an explicit generator.
- Markov State Models (MSMs): Discretize space and solve linear systems for transition probabilities. They suffer from discretization error and finite-lag bias.
- Dynamical Galerkin Approximation (DGA): Generalizes MSMs by expanding the committor in a basis of functions (e.g., pairwise distances, polynomials) and solving equations for the stopped transition operator $S_\tau$ . $S_\tau$ stops trajectories upon reaching $A$ or $B$ , ensuring exact statistics for any finite lag time $\tau$ .
- Neural Network Approaches:
  - Fixed-Point Iteration: Solves the fixed-point equation $q(x) = S_\tau q(x)$ by iteratively updating a neural network. To avoid the "two-trajectory trick" (which requires multiple trajectories per initial condition), this method uses a single trajectory per step to generate noisy targets, minimizing a loss (e.g., binary cross-entropy) to update parameters.
  - Exact Variational Committor Network (EVCN): A symmetrized variational approach that handles finite-lag errors and includes contributions from trajectories crossing between $A$ and $B$ , providing accurate rates and committors for any $\tau$ .
  - Subspace Acceleration: Accelerates convergence by learning the dominant eigenspace of the operator, addressing the slow convergence caused by eigenvalues near 1 in metastable systems.

The paper also addresses non-Markovian effects (memory) arising from reduced variables. It discusses the Mori-Zwanzig formalism for deriving memory kernels from data and delay embedding techniques to augment the state space, showing that memory corrections can significantly reduce the required lag time.

Key Contributions

Unified Framework: The authors provide a cohesive theoretical framework connecting PDEs, variational principles, MSMs, DGA, and neural networks through the lens of operator equations (specifically the stopped transition operator $S_\tau$ ).
Self-Supervised Efficiency: The review highlights that self-supervised methods, which learn from the satisfaction of operator equations rather than labeled outcomes, are more efficient than shooting. They allow for arbitrary sampling distributions $\mu(x)$ , enabling the concentration of computational effort in the transition region ( $q \approx 0.5$ ) where it is most needed.
Handling Finite Lag and Memory: The paper details how to construct exact estimators for finite lag times using stopped operators and how to correct for memory effects using data-driven kernels or state augmentation, avoiding the statistical inefficiency of extending lag times until memory decays.
Sampling Strategies: The authors analyze the statistical error bounds, demonstrating that sampling from the stationary distribution $\pi(x)$ is inefficient for committor estimation. They advocate for adaptive sampling strategies that target the transition region, often guided by preliminary committor estimates.

Results and Applications
The paper presents empirical evidence from several case studies where these methods were applied:

Trp-cage Folding: DGA with pairwise-distance basis functions provided committor estimates robust to lag time, revealing transition states consistent with previous shooting methods.
Insulin Dimer Dissociation: Using DGA with an indicator basis set on 45 features, the authors identified four distinct dissociation pathways and quantified fluxes, revealing a 100 ms transition rate that was missed by Kramers' theory due to its inability to resolve the multi-pathway nature of the barrier.
Voltage Sensing: DGA applied to voltage-sensing domains revealed a five-state microscopic mechanism with intermediates hidden from electrophysiology measurements, explaining discrepancies between fluorescence and electrophysiology data.
Hydrogen on Catalytic Surfaces: DGA was used to study hydrogen association/dissociation on rhodium, capturing non-monotonic rate dependencies on coverage and deviations from Transition State Theory (TST) at high coverages.

In all cases, the machine learning approaches (particularly DGA and fixed-point iteration with NNs) yielded accurate committors and mechanistic insights from data sets that were insufficient for direct observation of rare events.

Significance and Claims
The paper claims that these self-supervised machine learning methods represent a significant advancement in computational kinetics. By moving away from the requirement for explicit dynamical models or labeled shooting data, these methods can be applied to reduced variable sets, irreversible dynamics, and experimental data. The authors argue that the combination of operator-based learning, flexible sampling distributions, and neural network expressivity allows for the efficient estimation of kinetic statistics in complex, high-dimensional systems where traditional methods fail.

The review concludes by identifying future directions, including the integration of these methods with reinforcement learning (viewing the committor as a value function), the use of equivariant graph neural networks to eliminate manual feature engineering, and the application of generative models for transition path sampling. The authors emphasize that realizing the full potential of these methods requires systematic comparisons of sampling strategies and further theoretical development regarding stability and convergence.

Technical Summary: Machine Learning Kinetics from Molecular Dynamics Data

More like this