NepoIP/MM: Toward Accurate Biomolecular Simulation with a Machine Learning/Molecular Mechanics Model Incorporating...

Area of Science:

• Computational chemistry
• Biophysics
• Materials science

Background:

• Machine learning force fields (MLFFs) offer quantum mechanical accuracy for biomolecular simulations at reduced computational cost.
• Hybrid machine learning/molecular mechanics (ML/MM) models enhance realism by incorporating molecular mechanics (MM) for solvent and long-range interactions.
• Accurate modeling of electrostatic embedding in multiscale simulations necessitates accounting for polarization effects in the ML region.

Purpose of the Study:

• To develop a polarizable machine learning force field, NepoIP, capable of modeling polarization effects induced by an external electrostatic potential.
• To enable more realistic simulations of complex biomolecular systems in solution using hybrid ML/MM approaches with electrostatic embedding.

Main Methods:

• Adaptation of the NequIP architecture into a polarizable machine learning force field named NepoIP.
• Integration of NepoIP with molecular mechanics (MM) for hybrid ML/MM simulations.
• Nanosecond molecular dynamics (MD) simulations of a periodic solvated dipeptide system.

Main Results:

• NepoIP/MM simulations demonstrated stability for solvated dipeptide systems.
• Converged sampling from NepoIP/MM simulations showed excellent agreement with reference quantum mechanics/molecular mechanics (QM/MM) calculations.
• A single NepoIP model exhibited transferability across different MM force fields and diverse environments (water and proteins).

Conclusions:

• NepoIP successfully models polarization effects in ML/MM simulations with electrostatic embedding.
• The developed NepoIP force field provides a foundation for general ML biomolecular force fields suitable for hybrid simulations.
• This work advances the accuracy and applicability of MLFFs in computational biophysics.