Autodifferentiable Geometric Restraints for Enhanced Sampling Simulations with Classical and Machine Learned Force Fields
Gustavo R. Pérez-Lemus, Cintia A. Menendez, Yinan Xu, Pablo F. Zubieta Rico, Yezhi Jin, Juan J. de Pablo · 2025
Read Autodifferentiable Geometric Restraints for Enhanced Sampling Simulations with Classical and Machine Learned Force Fields by Gustavo R. Pérez-Lemus, Cintia A. Menendez, Yinan Xu, Pablo F. Zubieta Rico, Yezhi Jin, Juan J. de Pablo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-13575.txt.
Introduction
Gustavo Pérez-Lemus, Cintia Menendez, and co-authors bring shape based holds that can train by slope maths into boost sample sims, in a 2025 paper. Holds shrink the space a sim must roam, which cuts its cost a lot. Rich shape holds were hard to pair with boost ways that use nets or wave splits.
Slope ready holds let boost loops steer by exact pulls, not rough ones. Net based and wave based boost ways both take the new holds. Test beds span stuff design tasks where roam cost bites most. Wide shape holds thus join the boost kit for hard sims. That join may cut run time on shape rich tasks.
Catalog record
| CREATOR | Gustavo R. Pérez-Lemus, Cintia A. Menendez, Yinan Xu, Pablo F. Zubieta Rico, Yezhi Jin, Juan J. de Pablo |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-13575
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Autodifferentiable Geometric Restraints for Enhanced Sampling Simulations with Classical and Machine Learned Force Fields Gustavo R. Pérez-Lemus, Cintia A. Menendez, Yinan Xu, Pablo F. Zubieta Rico, Yezhi Jin, Juan J. de Pablo Abstract The use of external restraints is ubiquitous in advanced molecular simulation techniques. In general, restraints serve to reduce the configurational space that is available for sampling, thereby reducing the computational demands associated with a given simulations. Examples include the use of positional restraints in docking simulations or positional restraints in studies of catalysis. Past work has sought to couple complex restraining potentials with enhanced sampling methods, including Metadynamics or Extended Adaptive Biasing Force approaches. Here, we introduce the use of more general geometric potentials coupled with enhanced sampling methods that incorporate neural networks or spectral decomposition to achieve more efficient sampling in the context of advanced materials design. 1 Introduction Simulation of the free-energy landscapes that govern complex molecular transformations is a central element of statistical and quantum mechanics. This is the case for biological processes such as protein-ligand binding [ 1 ] , chemical processes such as heterogeneous catalysis [ 2 ] , and physical processes such as reorientation of liquid crystals at interfaces [ 3 ] . These systems are commonly modeled using molecular models and molecular dynamics simulations, where atoms within the system interact through potential energies calculated from, for instance, density functional theory ( dft ) [ 4 ] , classical models [ 5 ] , or, recently, machine-learned force fields ( mlff ) [ 6 ] . However, accurate calculation of free energies via molecular dynamics often requires the use of enhanced sampling methods. This requirement is due to the significant energy barriers that separate various states in many processes of interest, which, in many cases, cannot be overcome merely by the thermal motion of the particles. Enhanced sampling methods are essential for molecular dynamics simulations of complex systems as they facilitate the exploration of extensive portions of the configuration space and enable the calculation of free-energy profiles. Given the large family of enhanced sampling methods [ 7 ] , in this work, we focus on those that build a bias potential to later calculate the free energy using a collective variable in the system. Examples of these methods include Metadynamics [ 8 ] and its new upgrade opes [ 9 ] , Umbrella Sampling [ 10 ] , Adaptive Biasing Force ( abf ) [ 11 ] , and the recently-proposed Extended Adaptive Biasing Force (e abf ) [ 12 ] . These methods have been successfully applied to a wide variety of biological processes [ 13 ] . A key application is the accurate determination of the binding free energies of potential drugs against target proteins. In this context, the enhanced sampling method is used to calculate the potential of mean force ( pmf ) to remove the ligand from its binding site. On the other hand, to address potential sampling challenges in pmf calculations, it is essential to use external restraints that limit the configurational space and promote convergence in the pmf calculation. In addition, it is necessary to analyze the effect of the restraining potential in the calculation and adjust for its influence to accurately report binding energies, which are critical for differentiating between drug candidates. For these purposes, protocols such as bfee [ 14 , 15 ] have integrated e abf with a set of positional and orientational restraints to calculate the pmf along a defined spatial path of a ligand. Funnel Metadynamics is another powerful enhanced sampling method that has been used to study a wide variety of systems [ 16 ] . This method applies a funnel-shaped spatial restraint on a grid to control the movement of a ligand as it is detached from the binding pose following a determined path. However, methods such as abf have not yet incorporated external restraints. Recent advancements in the abf family of methods, such as the incorporation of neural networks or spectral decomposition, have shown to markedly enhance the speed of convergence in free energy calculations. The integration of spatial restraints in these enhanced methods can lead to even greater efficiency and accuracy in their application. Such integration of spatial restraints to the abf family enhanced sampling is especially relevant given the recent advancements in Machine Learned Force Fields ( mlff s). mlff s are surging as a new paradigm for molecular dynamics simulations [ 17 ] . These mlff s rely on training a neural network from a dataset of atomic configurations labeled with energies and forces. From that dataset, the neural network can learn the potential energy surface of the system based on local descriptors that can be invariant [ 18 ] or equivariant [ 19 ] . The growing interest in mlff is driven by their ability to maintain dft -level accuracy in energy and force predictions at a fraction of the computational cost. Moreover, the expansive capability of these models makes them well-suited for complex scenarios such as catalysis, which often involve bond formation and breaking – tasks that are challenging for traditional classical force fields that rely on fixed system topologies [ 20 ] . For this latter application, mlff has been successfully applied in heterogenous catalysis [ 21 ] . These modeled systems share features with biological systems like protein-ligand binding, where there is an active binding site and a complex detaching path that the reactants can follow. In such systems, employing geometrical restraints can be particularly beneficial, accelerating the calculation of free energies when combined with enhanced sampling techniques and computational tools like just-in-time compilation and automatic differentiation ( ad ). In this work we are proposing the incorporation of a general geometric restraints in the family of abf methods implemented in the p y sages enhanced sampling library that facilitated the application of the restraints using ad and gpu support [ 22 ] . We show how this methods works in classical systems and in mlff applications. Additionally, we incorporate Funnel Metadynamics into p y sages with the advantage of grid-free restraint potentials that reduces computation cost and ad facilitates the incorporation of more complex funnel shape potentials for general application. We hope this work serves the scientific community on applications of free energy methods where the use of restraints are imperative for an accurate calculation of free energies. 2 Results and Discussion In order to incorporate restraint potentials into the abf family of methods, like abf , funn [ 23 ] , cff [ 24 ] , and Spectral abf [ 25 ] , we need to show how we can recover the contribution of the restraint potential on the free energy calculation. For this goal, we use the Darve et al. equation [ 26 ] : \nabla_{\xi}A=-\left\langle\frac{d}{dt}(w\cdot p)\right\rangle_{\xi} (1) where \nabla_{\xi}A is the gradient of the free energy with respect a collective variable \xi , p is the momenta of the particles of the system and w=M_{\xi}J_{\xi}M^{-1} is the mass matrix of the jacobian transformation of \xi . Then, Darve et al. show that the time derivative of the w\cdot p matrix is related to the original abf equation: \left\langle\frac{d}{dt}(w\cdot p)\right\rangle_{\xi}=\left\langle-\nabla U\cdot w+\nabla\cdot w\right\rangle_{\xi} (2) where U is