Can bespoke MM small-molecule force fields fitted to MLFFs match ML/MM results while increasing speed?

Determine whether bespoke molecular mechanics force fields for small molecules that are parametrized by fitting to machine learning force field (MLFF) energies and forces can reproduce the accuracy and outcomes achieved by mechanical embedding ML/MM schemes for protein–ligand binding affinity simulations while providing improved computational speed.

Background

The paper discusses hybrid ML/MM approaches in which the intramolecular energetics of a ligand are described by a machine learning force field while the rest of the system is modeled with a classical molecular mechanics force field. Mechanical embedding (subtractive) schemes have shown improved binding affinity predictions in several studies and, in a promising demonstration, were used to simulate large solvated protein–ligand complexes with computational efficiency roughly one order of magnitude slower than pure MM.

Given the substantial speed gap between MLFF and MM simulations, the authors raise the question of whether an alternative strategy—fitting bespoke small-molecule MM force fields directly to MLFF predictions (e.g., via automated torsion parametrization tools)—could deliver the same accuracy as ML/MM mechanical embedding approaches while improving speed. This remains explicitly unresolved in the paper.