Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

BP: Fachverband Biologische Physik

BP 20: Protein Structure and Dynamics

BP 20.1: Vortrag

Mittwoch, 11. März 2026, 15:00–15:15, BAR/0106

Enhanced Conformational Dynamics of Biomolecules with Quantum-Accurate Machine Learning Force Fields — •Naziha Tarannam — University of Luxembourg, Luxembourg

Machine Learning Force Fields (MLFFs) address a long-standing challenge in computational biophysics: lifting quantum accuracy in the treatment of biomolecules to the large scales, above nanometres and nanoseconds, normally available only to classical-like molecular mechanics force fields (MMFFs). Our recently developed SO3LR model was trained on a diverse set of over four million configurations and achieves ab initio-level accuracy while remaining computationally efficient for large systems in explicit solvent. We explore the dynamics of several globular proteins in aqueous systems comprising up to 35,000 atoms using SO3LR, and compare its performance against a set of widely used MMFFs. While SO3LR faithfully reproduces the static structural properties of proteins, it exhibits enhanced exploration of conformational space during molecular dynamics simulations compared to MMFFs, showing qualitatively different properties of sampling and convergence. This stems from its relatively accurate treatment of quantum many-body forces. Notably, SO3LR enables proteins to explore a broader spectrum of dihedral angles (ψ/φ distributions) that are inaccessible to conventional force fields, leading to higher configurational entropy over comparable simulation timescales. These results imply that classical treatments of biomolecules, even if they reach an adequate thermodynamic accuracy, may often do so for the wrong reasons.

Keywords: Molecular Dynamics; Protein; Machine Learning Force Field; Computational Biophysics