Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments
Molecular dynamics (MD) simulations allow insights into complex processes, but accurate MD simulations require costly quantum-mechanical calculations. For larger systems, efficient but less reliable empirical force fields are used. Machine-learned force fields (MLFFs) offer similar accuracy as ab initio methods at orders-of-magnitude speedup, but struggle to model long-range interactions in large molecules. This work proposes a general approach to constructing accurate MLFFs for large-scale molecular simulations (GEMS) by training on “bottom-up” and “top-down” molecular fragments, from which the relevant interactions can be learned. GEMS allows nanosecond-scale MD simulations of >25,000 atoms at essentially ab initio quality, correctly predicts dynamical oscillations between different helical motifs in polyalanine, and yields good agreement with terahertz vibrational spectroscopy for large-scale protein-water fluctuations in solvated crambin. Our analyses indicate that simulations at ab initio accuracy might be necessary to understand dynamic biomolecular processes.
