Regression Modeling for the Prediction of Hydrogen Atom Transfer Barriers in Cytochrome P450 from Semi-empirically Derived Descriptors

Abstract

The calculation of hydrogen atom transfer (HAT) barriers within the cytochrome P450 catalytic cycle is of central importance for the prediction of metabolites formed from medicinally relevant compounds. We report the accurate estimation of hydrogen atom transfer barriers using inexpensive descriptors computed for a panel of 21 compounds. By a simple univariate linear regression between barriers previously computed using density functional theory (DFT) and newly computed “frozen radical” bond dissociation energies using the GFN2-xTB method, a mean absolute error of 1 kcal mol⁻¹ is achieved. Other affordable levels of theory are studied to assess differences in performance and computational cost. Multiple linear regression incorporating additional descriptors using GFN2-xTB is shown to predict HAT barriers with mean absolute errors of 0.82 kcal mol⁻¹. With computing times in milliseconds on modest computing hardware, this systematic approach is accessible and extensible to large scale screening workflows.