Achieving all-atom molecular dynamics accuracy from the Poisson-Boltzmann method through machine learning.

Area of Science:

• Computational chemistry
• Biophysics
• Molecular modeling

Background:

• All-atom molecular dynamics (MD) simulations provide high accuracy for biomolecular systems but are computationally expensive.
• Implicit models and coarse-graining reduce computational cost but often sacrifice accuracy.
• Poisson-Boltzmann (PB) theory efficiently models long-range electrostatics but neglects crucial short-range interactions.

Purpose of the Study:

• To develop a computational method that combines the accuracy of all-atom MD with the efficiency of implicit models.
• To accurately capture both short-range electrostatic correlations and long-range electrostatic interactions in biomolecular systems.
• To reduce the computational cost of simulating large biomolecular systems while maintaining high accuracy.

Main Methods:

• Introduced a graph neural network (GNN) Δ-learning approach, termed DIS-PB (deep implicit solvation model using the PB potential as a prior).
• Trained the GNN on the difference between all-atom MD and PB calculations.
• Modeled solutes and salt ions explicitly using MD, while water was coarse-grained.

Main Results:

• DIS-PB successfully captures both short-range electrostatic correlations and long-range electrostatic interaction tails.
• Simulations of a DNA molecule in a 1 mol/L salt solution using DIS-PB reproduced structural properties with high fidelity.
• The GNN-corrected PB approach achieved accuracy comparable to all-atom MD but at a significantly reduced computational cost.

Conclusions:

• DIS-PB offers a computationally efficient yet accurate method for modeling biomolecular systems.
• This approach effectively bridges the accuracy gap between all-atom simulations and traditional implicit solvent models.
• The GNN-corrected PB method shows promise for studying complex biomolecular systems, including DNA-protein interactions.