Enhancing Gaussian process regression-accelerated QM/MM free energy simulations using atomic environment descriptors.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning in Chemistry

Background:

• Accurate free energy simulations are crucial for understanding chemical reactions in complex systems.
• Achieving ab initio QM/MM accuracy with affordable semiempirical QM/MM methods for efficient sampling is a significant challenge.

Purpose of the Study:

• To extend a Δ-machine learning approach using Gaussian process regression (GPR) for enhanced QM/MM simulations.
• To incorporate atomic environment descriptors and MM-solvent contributions into GPR models for improved accuracy and efficiency.

Main Methods:

• Utilized atom-centered symmetry functions for atomic environment descriptors.
• Employed a system-specific sum kernel for molecular similarity inference.
• Trained energy-only GPR and GPR with derivative observation (GPRwDO) schemes.
• Integrated models into CHARMM simulations via a GPflow/pyCHARMM interface.

Main Results:

• Reduced AM1/MM potential energy errors from ~13.1 to 1.4 (energy-only GPR) and 2.2 (GPRwDO) kcal/mol.
• Decreased force errors from ~14.6 to 4.4 and 2.1 (kcal/mol)/Å.
• Achieved ~100-fold acceleration with AM1-GPR(wDO)/MM reaching target accuracy.

Conclusions:

• The developed GPR-based QM/MM methods significantly improve energetics and force description accuracy.
• Models provide excellent agreement with high-level benchmarks for free energy barriers and reaction energies.
• This approach enables efficient and accurate free energy simulations for complex chemical systems.