Calculation of solvation force in molecular dynamics simulation by deep-learning method.

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Electrostatic calculations are crucial for understanding biomolecular thermodynamics and kinetics in solution.
• Traditional Poisson-Boltzmann equation solvers are computationally intensive and time-consuming.
• Accelerating these calculations is vital for large-scale molecular simulations.

Purpose of the Study:

• To develop a deep neural network (DNN) for predicting atomic solvation free energies and forces.
• To accelerate electrostatic calculations in molecular simulations.
• To enable efficient analysis of long molecular dynamics trajectories.

Main Methods:

• A DNN was trained using internal molecular coordinates as input.
• Solvation free energies and transformed atomic forces from the Poisson-Boltzmann equation served as training labels.
• GPU acceleration was employed for both training and prediction.
• The DNN was integrated into molecular dynamics simulations with enhanced sampling.

Main Results:

• The DNN accurately predicts decomposed solvation free energies and forces for atoms.
• The method demonstrates reasonable accuracy for small molecules with sufficient training data.
• GPU acceleration significantly speeds up the prediction process.
• Free energy landscapes from DNN-accelerated simulations closely matched those from explicit solvent simulations.

Conclusions:

• The developed DNN offers a computationally efficient alternative to traditional Poisson-Boltzmann solvers.
• This approach is highly suitable for processing extensive molecular dynamics simulation data.
• The method enhances the feasibility of advanced sampling techniques in molecular simulations.