Understanding the effectiveness of enzyme pre-reaction state by a quantum-based machine learning model

Prediction of enzymatic stereochemistry is a significant challenge in computational chemistry because of targeting very small energy gaps in highly complicated macromolecular systems. Here, we report a scenario of four substrates (2-pentanone, 2-hexanone, 2-heptanone, and 2-octanone) within four enzyme variants (wild type, W116A, F285A, and W286A) of a medium-chain dehydrogenase from Candida parapsilopsis. The relative stabilities of pro-R and pro-S pre-reaction states are calculated by umbrella sampling, approximately consistent with the observed stereoselectivity in experiment. Besides, a LASSO-SVM machine-learning model is constructed with structural information of 704 pairs of quantum-mechanistic/molecular-mechanistic transition states (TSs) and pre-reaction states (PRSs), achieving the explanatory power of 99.6% for the calculated barriers. Intriguingly, the explanatory power with the PRS-alone structural information reaches 90.7%, but it decreases to 55.4% with the TS-alone structural information. Thus, the outcomes support that the enzymatic stereoselectivity is substantially determined by the frontier-molecular-orbital-related pre-organization of the enzyme-substrate reacting complexes. [Display omitted] •Machine-learning models with transition states and PRSs•Machine-learning, umbrella sampling, and frontier molecular orbital validate PRS theory•Rationalization of the overthrow of Prelog’s rule in CpRCR-catalyzed ketone reductions Luo et al. present a quantum-based machine learning model to validate the effectiveness of pre-reaction state theory for a zinc-dependent medium-chain dehydrogenase CpRCR. Specifically, the anti-Prelog rule of asymmetric reductions is rationalized by the pre-reaction state and frontier molecular orbital analysis.