Committor Guided Estimates of Molecular Transition Rates

Andrew Mitchell, Grant M. Rotskoff

TL;DR

This paper introduces a machine-learning framework to estimate molecular transition rates by learning the "committor function"—the gold standard for describing reaction pathways. By combining a novel logarithmic loss function with an adaptive "swarm" sampling strategy, the authors can accurately predict rare transition events in high-dimensional molecular systems without needing a pre-defined reaction coordinate.

What Problem Does This Paper Address?

In molecular biology and chemistry, many critical processes—like protein folding or drug binding—are "rare events." While they happen quickly once they start, the waiting time between events is orders of magnitude longer than standard molecular dynamics (MD) simulations can reach.

To bridge this gap, researchers use "enhanced sampling," which often requires a reaction coordinate (RC)—a simplified, low-dimensional map of the transition. However, picking a good RC is notoriously difficult; a poor choice leads to inaccurate rates. The committor function is the "ideal" reaction coordinate because it represents the exact probability that a specific configuration will transition to the final state before returning to the start. Unfortunately, calculating this function in high-dimensional space has historically been computationally impossible.

Methods

The authors propose a "physics-informed" machine learning approach to solve this:

  1. Neural Network Ansatz: They represent the committor function $q(x)$ as a deep neural network.
  2. Logarithmic Committor Loss: This is a key innovation. Standard Mean Squared Error (MSE) struggles because committor values near the starting state are nearly zero. MSE essentially ignores these regions. By using a logarithmic discrepancy, the objective function "zooms in" on these tiny values, which are crucial for accurate rate estimation.
  3. Adaptive Swarm Sampling: Instead of a long, continuous simulation, they launch "swarms" of short trajectories. The algorithm identifies the endpoints with the highest committor values and uses them as starting points for the next swarm, effectively "guiding" the simulation across the energy barrier.
  4. The Hill Relation: They use a classical statistical mechanics formula (the Hill relation) to convert the learned committor and the flux out of the starting basin directly into a transition rate.

Importantly, this method works for underdamped Langevin dynamics, which is the standard for realistic molecular simulations, whereas previous similar ML methods were often limited to simpler "overdamped" (purely diffusive) models.

Datasets

The authors evaluated their method on three distinct systems:

  • 2D Toy Potential: A synthetic "two-channel" potential used to rigorously compare the logarithmic loss against standard MSE.
  • Alanine Dipeptide: The "Hello World" of molecular kinetics. They tested this under both overdamped and underdamped conditions to ensure the method's robustness.
  • Aib9 (9-residue peptide): A more complex synthetic peptide that features intermediate metastable states. This served as a benchmark for how the method handles "bumpy" energy landscapes where the transition isn't a single jump.

Key Insights and Ideas

  • The "Jensen Gap" is a Feature, Not a Bug: The authors acknowledge that the logarithmic loss introduces a small statistical bias (due to Jensen's inequality). However, they prove mathematically that this bias is controllable and that the trade-off—vastly better accuracy near the basins—is what makes the rate estimation possible.
  • Reaction Coordinate Agnostic: Unlike traditional methods (like Metadynamics), this approach doesn't require the user to guess which atoms or angles are important. The neural network discovers the high-dimensional transition pathway on its own.
  • Parallel Efficiency: Because the method relies on "swarms" of short trajectories, it is "embarrassingly parallel," making it a perfect fit for modern GPU clusters.

Key Results

  • Efficiency: For Alanine Dipeptide, the method produced accurate rate estimates using only 2 ns of total sampling time, even though the actual mean first passage time (the time you'd have to wait in a standard simulation) is roughly 38 ns.
  • Accuracy: In the 2D potential tests, the logarithmic objective was "unambiguously superior" to standard MSE, which failed to capture the correct rates.
  • Complexity Handling: For the Aib9 system, the method successfully identified a non-obvious reaction pathway where the peptide collapses during isomerization, providing physical insights that a simple 1D coordinate would miss.

Limitations and Future Work

  • Metastability Bottlenecks: While the method handles intermediates, the computational cost increases significantly if the system gets "stuck" in many middle states. The authors suggest that combining this with Markov State Models (MSMs) could be a powerful future direction.
  • Scaling to Large Proteins: The current work used simple Multi-Layer Perceptrons (MLPs). To handle massive proteins with thousands of atoms, the authors note that more sophisticated architectures, like Graph Neural Networks (GNNs) or equivariant transformers, will likely be necessary.
  • Lag Time Sensitivity: The accuracy depends on choosing a "lag time" $\tau$ that is long enough for the system to be Markovian but short enough to be computationally efficient. Finding an automated way to tune this parameter remains an open challenge.

Summary updated

Abstract

The probability that a configuration of a physical system reacts, or transitions from one metastable state to another, is quantified by the committor function. This function contains richly detailed mechanistic information about transition pathways, but a full parametrization of the committor requires the construction of a high-dimensional function, a generically challenging task. Recent efforts to leverage neural networks as a means to solve high-dimensional partial differential equations, often called "physics-informed" machine learning, have brought the committor into computational reach. Here, we build on the semigroup approach to learning the committor and assess its utility for predicting dynamical quantities such as transition rates. We show that a careful reframing of the objective function and improved adaptive sampling strategies provide highly accurate representations of the committor. Furthermore, by directly applying the Hill relation, we show that these committors provide accurate transition rates for molecular systems.
BibTeX
@article{mitchell2024committor,
  author = {Andrew Mitchell and Grant M. Rotskoff},
  title = {{Committor Guided Estimates of Molecular Transition Rates}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2024},
  doi = {10.1021/acs.jctc.4c00997},
}

Loading related content...

Figure