Machine learning builds full-QM precision protein force fields in seconds.

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning in science

Background:

• Full-quantum mechanics (QM) calculations offer high precision but are computationally expensive for large systems like proteins.
• There is a growing need for efficient computational methods to study large biomolecular systems at a QM level.

Purpose of the Study:

• To develop a computationally efficient and highly accurate method for calculating energies and atomic forces of proteins.
• To leverage machine learning to accelerate QM-level calculations for large biomolecular systems.

Main Methods:

• Designed a neural network-based two-body molecular fractionation with conjugate caps (NN-TMFCC) approach.
• Developed residue-based fragment models for neural network potential energy surfaces.
• Validated the method on 15 representative proteins, comparing results against full-QM calculations.

Main Results:

• Achieved high precision with energy root-mean-squared errors (RMSEs) < 1.0 kcal/mol and force RMSEs < 1.3 kcal/mol/Å.
• NN-TMFCC calculated protein energies and forces with full-QM precision in seconds (10-100 s), thousands of times faster than full-QM.
• Computational complexity is independent of system size, enabling massive acceleration for very large proteins.

Conclusions:

• The NN-TMFCC method provides a highly precise and efficient approach for QM-level energy and force calculations in proteins.
• This method holds significant potential for accelerating structure predictions and molecular dynamics simulations of large biomolecular systems.
• Enables large-scale, high-accuracy computational studies previously limited by computational cost.