Subspace Fokker–Planck Transformer: An Ensemble-Based Framework for Accelerated Free-Energy Sampling

Rare transitions between metastable states limit the convergence of molecular free-energy calculations. Here, we introduce the subspace Fokker–Planck Transformer (sFPT), an ensemble-based enhancedsampling method that transports the instantaneous replica distribution in a selected collective-variable (CV) subspace. sFPT estimates the score of the evolving CV distribution using a kernel representation of the replica ensemble, maps the resulting entropy force to atomic coordinates through the CV Jacobian, and incorporates this force into underdamped Langevin dynamics. We show that the self-consistent stationary CV distribution satisfies PCV ∞ (s) ∝ [PCV phys(s)] λ/(1+λ), yielding a controllably flattened free-energy surface, and establish convergence of the ideal mean-field dynamics through a nonincreasing Lyapunov functional. Post-convergence trajectories can be accumulated to improve the statistical precision of freeenergy reconstruction. We evaluate sFPT on alanine dipeptide in vacuum and explicit water, alanine tetrapeptide in a three-dimensional CV space, and a Claisen rearrangement described by an ML/MM potential in explicit water. The reconstructed free-energy landscapes agree with umbrella-sampling, replica-exchange, and well-tempered metadynamics (WT-MetaD) references. For alanine dipeptide in vacuum and explicit water, comparisons at matched walker numbers show that both methods can benefit from parallel sampling. For the Claisen rearrangement, sFPT reaches sustained agreement with the reference basin free-energy difference after approximately 24 ps per replica, whereas the selected 64-walker WT-MetaD calculation retains slow deviations over 400 ps per walker. This advantage remains visible when the integration steps over all replicas are counted. These results establish instantaneous ensemble transport as a practical route to stable free-energy estimates from many short parallel trajectories.

Authors

Institutions

Publication Details

Journal
ChemRxiv
Published
2026-11-18
DOI
https://doi.org/10.26434/chemrxiv.15012433/v1
Primary Topic
Protein Structure and Dynamics
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Subspace Fokker–Planck Transformer: An Ensemble-Based Framework for Accelerated Free-Energy Sampling

Xiaoxuan Yu, Zichen Sun, Yi Qin Gao, Jun Zhang
ChemRxiv
Protein Structure and Dynamics
preprint

Subspace Fokker–Planck Transformer: An Ensemble-Based Framework for Accelerated Free-Energy Sampling

Xiaoxuan Yu, Zichen Sun, Yi Qin Gao, Jun Zhang
preprint en

Abstract

Rare transitions between metastable states limit the convergence of molecular free-energy calculations. Here, we introduce the subspace Fokker–Planck Transformer (sFPT), an ensemble-based enhancedsampling method that transports the instantaneous replica distribution in a selected collective-variable (CV) subspace. sFPT estimates the score of the evolving CV distribution using a kernel representation of the replica ensemble, maps the resulting entropy force to atomic coordinates through the CV Jacobian, and incorporates this force into underdamped Langevin dynamics. We show that the self-consistent stationary CV distribution satisfies PCV ∞ (s) ∝ [PCV phys(s)] λ/(1+λ), yielding a controllably flattened free-energy surface, and establish convergence of the ideal mean-field dynamics through a nonincreasing Lyapunov functional. Post-convergence trajectories can be accumulated to improve the statistical precision of freeenergy reconstruction. We evaluate sFPT on alanine dipeptide in vacuum and explicit water, alanine tetrapeptide in a three-dimensional CV space, and a Claisen rearrangement described by an ML/MM potential in explicit water. The reconstructed free-energy landscapes agree with umbrella-sampling, replica-exchange, and well-tempered metadynamics (WT-MetaD) references. For alanine dipeptide in vacuum and explicit water, comparisons at matched walker numbers show that both methods can benefit from parallel sampling. For the Claisen rearrangement, sFPT reaches sustained agreement with the reference basin free-energy difference after approximately 24 ps per replica, whereas the selected 64-walker WT-MetaD calculation retains slow deviations over 400 ps per walker. This advantage remains visible when the integration steps over all replicas are counted. These results establish instantaneous ensemble transport as a practical route to stable free-energy estimates from many short parallel trajectories.

ChemRxiv
Peking University (CN), Westlake University (CN), Changping Laboratory (CN)
National Natural Science Foundation of China
Protein Structure and Dynamics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon