A Machine Learning Force Field for Bio-Macromolecular Modeling Based on Quantum Chemistry-Calculated Interaction...

Area of Science:

• Computational chemistry
• Biomolecular modeling
• Machine learning

Background:

• Accurate force fields are crucial for molecular dynamics simulations of biomacromolecules.
• Determining appropriate quantum chemistry methods and continuous energy functions are key challenges in force field development.
• Previous work utilized Symmetry-Adapted Perturbation Theory (SAPT0) for interaction energies, creating the SOFG-31 datasets.

Purpose of the Study:

• To determine the optimal Symmetry-Adapted Perturbation Theory (SAPT) level of theory for calculating interaction energies.
• To benchmark SAPT interaction energies against coupled cluster with single, double, and perturbative triple excitations/complete basis set (CCSD(T)/CBS) calculations.
• To develop a general-purpose force field for biomolecular dynamics simulations using machine learning.

Main Methods:

• Recalculated intermolecular interaction energies using the advanced SAPT2 level of theory with extended basis sets.
• Employed the CLIFF (Continuous, Low-dimensional, Interactive, Force-Field) scheme, a machine learning technique, for force field construction.
• Utilized the SOFG-31 and SOFG-31-heterodimer datasets for training and testing the machine learning model.

Main Results:

• The SAPT2 level of theory, with appropriate basis sets, provides interaction energies consistent with CCSD(T)/CBS benchmarks.
• The CLIFF scheme successfully reproduced diverse dimeric interaction energy patterns using a small training dataset.
• Errors in SAPT energy components and total SAPT energy were significantly below the target chemical accuracy of ~1 kcal/mol.

Conclusions:

• SAPT2 is a suitable level of theory for accurate calculation of noncovalent interaction energies in biomolecular systems.
• The CLIFF scheme is effective for developing accurate and efficient force fields from quantum chemical data.
• This approach balances chemical accuracy and computational efficiency for biomolecular simulations.