Journal of Chemical Theory and Computation, Journal Year: 2025, Volume and Issue: unknown

Published: April 4, 2025

In sequential multiscale molecular dynamics simulations, which advantageously combine the increased sampling and at coarse-grained resolution with higher accuracy of atomistic is altered over time. While coarse-graining straightforward once mapping between defined, reintroducing details still a nontrivial process called backmapping. Here, we present ART-SM, fragment-based backmapping framework that learns from simulation data to seamlessly switch resolution. ART-SM requires minimal user input goes beyond state-of-the-art approaches by selecting multiple conformations per fragment via machine learning simultaneously reflect structure Boltzmann distribution. Additionally, introduce novel refinement step connect individual fragments optimizing specific bonds, angles, dihedral angles in process. We demonstrate our algorithm accurately restores bond length, angle, angle distributions for various small linear molecules Martini beads resulting high-resolution structures are representative conformations. Moreover, reconstruction TIP3P water model fast robust, can be applied larger as well. To illustrate efficiency local autoregressive approach simulated large realistic system containing surfactants TAPB SDS solution using Martini3 force field. The self-assembled micelles shapes were backmapped after training on only short simulations single water-solvated or molecule. Together, these results indicate potential method extended more complex such lipids, proteins, macromolecules, materials future.

Language: Английский