Neural-network potentials get solute tempering: REST2-AMP/MM folds chignolin and CLN025 from extended chains without preset reaction coordinates
Synopsis
REST2 enhanced sampling had been used with classical force fields and QM/MM, but not with multiscale ML/MM simulations whose solute is described by a neural-network potential. Exploiting the AMP-BMS Hamiltonian's split into solute–solute, solute–solvent and solvent–solvent terms, the approach scales the first two energies directly and exchanges replicas while holding the solvent at physical temperature. In explicit solvent, chignolin and CLN025 started from extended conformations reach Cα-RMSD plateaus of about 0.2–0.3 nm and about 0.2 nm relative to their NMR structures, with no reaction coordinate specified in advance.
Interpretation
Multiscale ML/MM simulations can now use REST2-like solute tempering, with the solute described by a neural-network potential and the environment by a classical force field, so no reaction coordinate has to be chosen for each folding run. REST2-like schemes had appeared only in classical force-field and QM/MM simulations, and reaction-coordinate-based enhanced sampling already coupled to ML/MM required knowing in advance which quantity to bias. For alanine dipeptide, a REST2 setup with eight replicas and scaling factors distributed geometrically from 1 to 0.125 sampled the right quadrant of the Ramachandran plot within 1 ns, whereas a 160 ns conventional AMP/MM trajectory reached that region only after about 150 ns.
The method directly shrinks the solute internal energy and the solute–solvent interaction energy computed by the neural-network potential, each by its own per-replica factor, while the solvent stays at its physical temperature. Classical REST2 achieves this scaling indirectly by rescaling solute charges, Lennard-Jones parameters and torsional force constants, but AMP predicts all multipoles internally and has no torsional force constant to rescale, so the energy terms themselves are scaled. All AMP/MM simulations used the global AMP-BMS model without further fine-tuning, with systems solvated in TIP4P-FB water and exchange attempted every 100 integration steps; replica counts and scaling ranges per system are listed in the parameter table.
In folding runs started from extended conformations, CLN025's Cα-RMSD reaches a plateau near 0.2 nm after about 30 ns, and chignolin settles to about 0.2–0.3 nm after roughly 20–25 ns and stays there. These folding trajectories specify no reaction coordinate, and the folded state need not be known beforehand, unlike approaches that depend on predefined collective variables. Both systems used 12 replicas with scaling factors linearly distributed from 1 to 0.5 and 50 ns simulations, and most NOE distance differences from the last 50% of the folded trajectories fall within ±0.1 nm.
Perspective
The demonstrated scope is peptide systems covered by the BMS25 training set: alanine dipeptide, a cyclic pentapeptide, polyproline, Trp-cage, chignolin and CLN025, all simulated in explicit TIP4P-FB water with the unfine-tuned AMP-BMS model.
The authors state that the folding simulations show recovery of experimentally characterized folded conformations from extended starting structures but do not establish equilibrium folding populations or folding–unfolding reversibility; the plateaus seen for polyproline and Trp-cage indicate stability of the monitored structural observables over the simulated timescale, not convergence of the full thermodynamic ensemble. Cis/trans isomerization has too high a barrier to be observed even with the enhanced sampling in this setup.