Machine learned potentials with electrostatic embedding accurately capture Kemp eliminase reactivity

Abstract

Kemp elimination has become a benchmark for de novo enzyme design due to its simplicity and detailed mechanistic understanding but accurate barrier calculations are required to understand the difference in activity between designed Kemp eliminases. QM/MM MD simulations are capable of calculating reaction barriers, but are limited by a cost/accuracy tradeoff in which the most accurate methods are too computationally expensive for extensive screening as would be required in a prospective enzyme design campaign. Recent advances in embedded ML/MM simulations using the electrostatic machine learning embedding (EMLE) method have enabled transferable potentials for ML/MM MD simulations with low computational cost but QM-level accuracy. Here, we trained a MACE MLIP and EMLE embedding model for fast and accurate simulations of the enzymatic, base-catalysed Kemp elimination of 6-nitro benzisoxazole. Applying our EMLE ML/MM scheme to a designed Kemp eliminase and its evolved counterpart captured the {approx}4 kcal/mol difference in barrier between the two observed in experiment. Furthermore, training was done using structures generated from simulations of a single variant and was applied to the second variant, achieving near-experimental barriers, with no further finetuning or computational overhead. Thus, this work establishes EMLE-based ML/MM MD simulations as a potential route for fast and accurate assessment of barriers in computational screening for enzyme design.