Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PFP/MM: A Hybrid Approach Combining a Universal Neural Network Potential with Classical Force Fields for Large-Scale Reactive Simulations

Yu Miyazaki, Atsuhiro Tomita, Akihide Hayashi, So Takemoto, Mizuki Takemoto, Hodaka Mori

Abstract

Universal machine-learning interatomic potentials (uMLIPs) enable reactive molecular simulations with near-DFT accuracy, yet applying them efficiently to large, realistic condensed-phase systems remains computationally demanding. Here we present PFP/MM, a hybrid approach that combines a uMLIP, PreFerred Potential (PFP), with molecular mechanics (MM) to enable both large-scale and long-time simulations that are challenging for uMLIP-only calculations. Using an alanine dipeptide in explicit water, we achieve multi-ns/day enhanced sampling and obtain a Ramachandran plot consistent with established basins. For an intramolecular nucleophilic addition reaction in a polar solvent environment, we reproduce the expected solvent-induced stabilization in the free-energy profile. We further apply the approach to a cytochrome P450 Compound I hydroxylation reaction and obtain a free-energy landscape consistent with the accepted reaction mechanism. These results demonstrate that uMLIP-based reactive simulations can be applied to diverse condensed-phase processes in large, realistic environments.

PFP/MM: A Hybrid Approach Combining a Universal Neural Network Potential with Classical Force Fields for Large-Scale Reactive Simulations

Abstract

Universal machine-learning interatomic potentials (uMLIPs) enable reactive molecular simulations with near-DFT accuracy, yet applying them efficiently to large, realistic condensed-phase systems remains computationally demanding. Here we present PFP/MM, a hybrid approach that combines a uMLIP, PreFerred Potential (PFP), with molecular mechanics (MM) to enable both large-scale and long-time simulations that are challenging for uMLIP-only calculations. Using an alanine dipeptide in explicit water, we achieve multi-ns/day enhanced sampling and obtain a Ramachandran plot consistent with established basins. For an intramolecular nucleophilic addition reaction in a polar solvent environment, we reproduce the expected solvent-induced stabilization in the free-energy profile. We further apply the approach to a cytochrome P450 Compound I hydroxylation reaction and obtain a free-energy landscape consistent with the accepted reaction mechanism. These results demonstrate that uMLIP-based reactive simulations can be applied to diverse condensed-phase processes in large, realistic environments.
Paper Structure (7 sections, 2 equations, 4 figures)

This paper contains 7 sections, 2 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Method
  3. Results and Discussion
  4. Benchmark of Alanine Dipeptide in Water
  5. Intramolecular Nucleophilic Addition of an Amine to a Carbonyl: Solvent Effects
  6. P450-Cpd-I-Mediated Hydroxylation Reaction
  7. Conclusion

Figures (4)

  • Figure 1: Speed benchmark in alanine dipeptide in water system. In PFP/MM, the 22 atoms of the alanine dipeptide molecule are treated as the PFP region, and all remaining water molecules are treated as the MM region. In PFP-only, all atoms are treated within PFP. Error bars represent standard errors over 5 runs.
  • Figure 2: (a) Schematic overview of major conformational basins of the alanine dipeptide in water. Inset shows the definition of the backbone dihedral angles $\phi$ and $\psi$. Adapted from the original work of Feig Feig2008-of. (b--d) Free energy surfaces from one-day simulations: (b) PFP-only, (c) PFP/MM with V100, and (d) PFP/MM with MN-Core 2.
  • Figure 3: (a) Skeletal formulas of the open and cyclized NCO molecules. (b) Snapshots of the PFP region in the open and cyclized conformations. (c) Free-energy profiles of NCO in water (using FIRES) and in vacuum. Regions around the lines indicates standard error estimated by block averaging of 5 blocks.
  • Figure 4: (a) Structure of the P450-Cpd-I enzyme-ligand model used in the PFP/MM simulation. The PFP region is shown in color, while the MM region is shown in gray. (b) Hydroxylation reaction mechanism of P450-Cpd-I proposed in ref. Wang2021-sb. (c) Collective variables used in the metadynamics and umbrella sampling simulations. (d) Free energy surface obtained from umbrella sampling and the minimum free energy path. (e) Free energy profile along the reaction coordinate.