Predicting hydrogen atom transfer energy barriers using Gaussian process regression.

Area of Science:

• Computational chemistry
• Materials science
• Biochemistry

Background:

• Predicting reaction barriers is crucial for catalyst design and simulating reactions in complex materials.
• Current methods often require extensive density functional theory (DFT) calculations, limiting efficiency.

Purpose of the Study:

• To introduce Gaussian process regression (GPR) as a data-efficient method for predicting reaction barriers.
• To evaluate GPR's performance for hydrogen atom transfer reactions in proteins.

Main Methods:

• Utilized Gaussian process regression (GPR) with SOAP descriptors and a marginalized graph kernel.
• Compared GPR performance against a graph neural network-based model.
• Focused on datasets with hundreds to thousands of DFT barrier calculations.

Main Results:

• Achieved a mean absolute error of 3.23 kcal mol⁻¹ for hydrogen atom transfer barriers in proteins.
• Demonstrated robust estimation of reaction barriers within proteins' complex chemical and conformational spaces.
• GPR models showed comparable predictive power to graph neural networks, outperforming them in low-data scenarios.

Conclusions:

• Gaussian process regression (GPR) offers a valuable, data-efficient approach for approximating chemical reactivity.
• GPR is suitable for modeling reactions in complex and variable environments, especially when DFT data is limited.
• This method enhances the efficiency of catalyst design and reaction simulations in materials and biological systems.