Benchmark of Approximate Quantum Chemical and Machine Learning Potentials for Biochemical Proton Transfer Reactions

Area of Science:

• Computational Chemistry
• Biophysical Chemistry

Background:

• Proton transfer reactions are fundamental to biological processes like enzymatic catalysis and energy conversion.
• Accurate simulation of these reactions requires reliable quantum chemical methods.
• Existing methods like Density Functional Theory (DFT) have limitations, particularly for nitrogen-containing groups.

Purpose of the Study:

• To benchmark various approximate quantum chemical methods and machine learning potentials for simulating proton transfer reactions.
• To evaluate the performance of these methods against high-level reference data for biochemical systems.
• To assess their applicability in both isolated and microsolvated environments using QM/MM.

Main Methods:

• Benchmarking of semiempirical molecular orbital, tight-binding DFT, and machine learning (ML) potentials.
• Comparison against high-level MP2 reference data for relative energies, geometries, and dipole moments.
• Simulation of microsolvated reactions using a hybrid Quantum Mechanics/Molecular Mechanics (QM/MM) approach.

Main Results:

• Traditional DFT methods show deviations for proton transfers involving nitrogen.
• Approximate models like RM1, PM6, PM7, DFTB2-NH, DFTB3, and GFN2-xTB offer reasonable accuracy with varying performance.
• ML-corrected (Δ-learning) model PM6-ML significantly improves accuracy across properties and chemical groups.
• Standalone ML potentials demonstrated poor performance for most reactions.

Conclusions:

• The ML-corrected PM6-ML model shows superior performance for proton transfer simulations in complex biochemical environments.
• Approximate methods have varying degrees of accuracy, necessitating careful selection for specific applications.
• This study provides guidance for choosing appropriate computational methods for simulating proton transfer reactions.