In this presentation, we introduce PFP/MM, a hybrid approach that combines our universal 96-element neural network potential, called PFP, with classical molecular mechanics (MM) force fields. By applying PFP to reactive regions and MM to the rest of the system, PFP/MM overcomes both spatial and temporal scale limitations.This strategy enables simulations involving hundreds of thousands of atoms over nanosecond timescales.
We first demonstrate PFP/MM in the context of organic reactions in solvent. As an illustrative example, we consider the nucleophilic attack of an amine on a carbonyl group, which produces a zwitterionic intermediate that is stabilized by polar solvents. We compare two metadynamics simulations at the nanosecond timescale: one is the purely solute-based system with solely PFP; the other is the system in which solute and a few solvent molecules are treated with PFP and the remaining solvent molecules are treated with MM. These two simulations show remarkable differences in free energy surfaces (FESs), suggesting that the zwitterionic intermediate cannot remain stable without explicit solvent molecules. While solely UNNP or QM/MM approaches require enormous computational resources, the accelerated performance of PFP/MM enables longer simulations and thus facilitates high-accuracy FES calculations.
Additionally, we apply PFP/MM to large metalloproteins, which are challenging targets due to the difficulty of accurately treating the interaction between metal and organic molecules. Such interactions are difficult to handle with the conventional MM force fields or machine learning potentials which are trained only on organic molecules. In contrast, PFP/MM—capable of describing 96 elements—provides the breadth of a QM/MM approach while achieving dramatically faster computation, thereby making large-scale metalloprotein simulations feasible.