Machine-Learned Molecular Surface and Its Application to Implicit Solvent Simulations.

Area of Science:

• Computational chemistry
• Biomolecular modeling
• Machine learning applications

Background:

• Implicit solvent models are crucial for biomolecular simulations.
• The solvent-excluded surface (SES) is a key component of these models.
• Classical SES algorithms face limitations in parallelization and derivative calculation.

Purpose of the Study:

• To develop a machine learning strategy for computing the SES.
• To create a level set formulation for the SES.
• To overcome limitations of traditional SES computation methods.

Main Methods:

• A three-step machine learning training process was employed.
• The developed model was integrated into the Amber/PBSA program.
• Performance was evaluated using molecular surfaces and reaction field energy calculations.

Main Results:

• The machine-learned SES achieved over 95% agreement with classical methods.
• The new SES demonstrated high stability and overlap with traditional SES.
• Computational efficiency was improved by approximately 2.5 times on CPU platforms.
• Reaction field energies showed high consistency with a 1% average deviation.

Conclusions:

• The machine-learned SES offers a viable, efficient alternative to classical methods.
• Its level set formulation is suitable for applications requiring surface derivatives or parallel computing.
• Potential for significant performance gains on GPU platforms is anticipated.