Utilizing Machine Learning to Greatly Expand the Range and Accuracy of Bottom-Up Coarse-Grained Models through Virtual Particles - Abstract & Introduction

Aug 6, 2026

Coarse-grained models parameterized from atomistic reference data, or “bottom up” CG models, are widely useful for biomolecules and soft matter, yet building highly accurate, low-resolution representations remains difficult. This work shows how virtual particles—CG sites without atomistic correspondence—can be incorporated as latent variables in relative entropy minimization using variational derivative relative entropy minimization (VDREM). Gradient-descent optimization, assisted by machine learning, tunes virtual-particle interactions. Application to a solvent-free DOPC lipid bilayer demonstrates improved solvent-mediated behavior and higher-order correlations beyond standard REM based only on atom-to-site mapping.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

Abstract

Coarse-grained modeling and bottom-up parametrization

Virtual particles as latent variables in VDREM

Case study: solvent-free DOPC lipid bilayer

Introduction

Limits of atomistic molecular dynamics

Top-down vs bottom-up CG approaches

Many-body potentials of mean force (mbPMF)

Practical approximations and higher-order expressivity

This article is licensed under CC-BY-NC-ND 4.0 [pubs.acs.org/JCTC](pubs.acs.org/JCTC) Article Utilizing Machine Learning to Greatly Expand the Range and Accuracy of Bottom-Up Coarse-Grained Models through Virtual Particles Patrick G. Sahrmann, Timothy D. Loose, Aleksander E. P. Durumeric, and Gregory A. Voth * Cite This: J. Chem. Theory Comput. 2023, 19, 4402−4413 Read Online ACCESS Metrics & More Article Recommendations *sı Supporting Information ABSTRACT: Coarse-grained (CG) models parametrized using atomistic reference data, i.e.,“bottom up” CG models, have proven useful in the study of biomolecules and other soft matter. However, the construction of highly accurate, low resolution CG models of biomolecules remains challenging. We demonstrate in this work how virtual particles, CG sites with no atomistic correspondence, can be incorporated into CG models within the context of relative entropy minimization (REM) as latent variables. The methodology presented, variational derivative relative entropy minimization (VDREM), enables optimization of virtual particle interactions through a gradient descent algorithm aided by machine learning. We apply this methodology to the challenging case of a solvent-free CG model of a 1,2-dioleoyl-sn-glycero-3-phosphocholine (DOPC) lipid bilayer and demonstrate that introduction of virtual particles captures solvent-mediated behavior and higher-order correlations which REM alone cannot capture in a more standard CG model based only on the mapping of collections of atoms to the CG sites. 1. INTRODUCTION Atomistic molecular dynamics (MD) simulations have enabled key insights into biological and material processes,1−3 but modern hardware limits the practicality of MD to the study of millions of atoms on the multi-microsecond scale. This precludes sampling of many biologically relevant phenomena, such as macromolecular assembly.4 Coarse-grained (CG) models aim to extend the spatiotemporal scales of MD by simulating increasing a lower-resolution representation of a system, the efficiency of the simulation.5−8 CG model construction and parametrization often follow either a top-down or bottom-up approach.7,8 In the top-down approach, a CG model is parametrized to directly reproduce macroscopic properties such as thermodynamic data. Popular top-down approaches such as MARTINI have been used to simulate biomolecular structures such as proteins and multicomponent lipid bilayers.9−11 However, these models are not parametrized to reproduce the microscopic correlations and enthalpy−entropy decompositions underpinning these properties. 12 Bottom-up models instead aim to reproduce the microscopic behavior of a reference atomistic model, with the intent of indirectly capturing emergent behavior.8, 13−18 While bottom-up CG models have an explicit correspondence variables implied by the atomistic reference simulation. The exact CG model Hamiltonian whose Boltzmann statistics reproduce the reference distribution is referred to as the CG variable many-body potential of mean force (mbPMF).14, 15,22,23 In the limit of infinite sampling and a perfect basis set to CG methods such CG) and Relative represent the CG interactions, bottom-up as Multiscale Coarse-Graining 13, 15, 16 (MSEntropy Minimization17, 18,24 (REM) are guaranteed to reproduce the mbPMF. However, practical considerations often relegate the CG force-field (basis set) to pairwise nonbonded interactions. Enhancing the expressivity of CG force fields beyond a pairwise basis set through explicit higher-order terms25 and order parameter (e.g., local density) based interactions26−30 enables the capture of certain manybody statistics. However, for biological systems, it is generally not clear which higher-order terms should be included. Machine-learned CG force fields can be utilized to construct general approximations to many-body statistics,31−35 albeit at increased computational cost. Practical implementa

FAQs

What problem does the paper address in bottom-up coarse-grained modeling?

Highly accurate, low-resolution bottom-up CG models are challenging to construct, especially when practical force-field basis sets are limited to low-order interactions.

How do virtual particles improve the CG model in this approach?

Virtual particles are introduced as CG sites without atomistic correspondence, treated as latent variables in relative entropy minimization so their interactions can be optimized to capture missing solvent-mediated and higher-order effects.

What method is used to optimize virtual particle interactions?

The work uses variational derivative relative entropy minimization (VDREM) and performs gradient-descent optimization aided by machine learning.

Aug 6, 2026

Upload by Olivia Brown