Quantum-accurate enhanced sampling and mini-protein folding through replica exchange for multiscale neural network potentials

Neural network potentials (NNPs) have emerged as efficient alternatives to computationally demanding quantum-mechanical (QM) methods. AMPv3-EE-BMS25 (short AMP-BMS) is an equivariant NNP designed for multiscale ML/MM simulations that captures directional and long-range interactions. In classical molecular dynamics (MD) simulations, enhanced-sampling techniques are applied to cross large energy barriers and improve convergence. Among those, replica exchange with solute tempering (REST2) is an efficient method that selectively scales solute interactions while keeping the bath temperature fixed, thereby reducing the number of replicas required. Here, we present a REST2-like algorithm compatible with multiscale ML/MM simulations. This algorithm is combined with AMP-BMS by exploiting the separability of its Hamiltonian into solute-solute, solute-solvent, and solvent-solvent contributions. We validate the REST2-AMP approach on several peptidic systems in explicit solvent and demonstrate folding of the mini-proteins chignolin and CLN025 from extended conformations. Thus, our REST2 approach for ML/MM enables access to biologically relevant timescales while maintaining near-DFT accuracy.