machine-learned SES

machine-learned SES approximates the solvent-excluded surface (SES) using machine learning to produce a level set formulation for implicit solvent Poisson-Boltzmann calculations.

Key Features:

• Machine Learning Integration: Employs machine learning models to predict the solvent-solute interface for SES determination.
• Level Set Formulation: Represents the SES as a level set, enabling analytic surface derivatives and numerical operations suited to parallelization.
• Three-step Training Process: Trains models through a three-step procedure that targets agreement with classical SES computations.
• High Agreement with Classical SES: Demonstrates over 95% concordance with geometry-based classical SES used in Poisson-Boltzmann models.
• Stability Across Rotations: Produces SES predictions that are stable under rotational transformations of molecular structures.
• Efficiency Gains: Achieves approximately 2.5× faster performance on tested CPU platforms compared with classical SES routines implemented in Amber/PBSA.
• Parallel/GPU Compatibility: The level set and ML approach offers computational advantages on parallel platforms such as GPUs.

Scientific Applications:

• Implicit Solvent Simulations: Provides an alternative SES definition for implicit solvent models and Poisson-Boltzmann calculations in biomolecular simulations.
• Reaction Field Energy Calculations: Integrates with Amber/PBSA workflows and yields reaction field energies within about 1% of classical SES-based calculations.

Methodology:

Trains machine learning models in a three-step process to approximate the SES as a level set and compares outputs to classical SES computations; the formulation supports parallel execution on GPUs.