Multiscale machine learning molecular mechanics for mechanism and stereoselectivity of Diels-Alderase catalysis

Enzymes catalyze complex chemical transformations with remarkable efficiency and selectivity, yet their atomistic mechanisms remain challenging to capture because conventional simulations trade accuracy for efficiency. Here we introduce a reactive machine learning/molecular mechanics (ML/MM) framework that bridges quantum chemistry with long-timescale sampling, enabling direct exploration of enzymatic transition states and free-energy landscapes. Coupled with metadynamics, this approach achieves nanosecond sampling of bond-forming reactions and quantitatively predicts activation barriers, mutational effects, and stereoselectivity. Applied to Diels-Alderases, the framework not only reproduces experimental activity and endo/exo preferences with sub-kcal mol^-1 accuracy but also uncovers how pathway dynamics and local electrostatics preorganize substrates for selective outcomes. By uniting reactivity, conformational dynamics, and predictive power, this work establishes reactive ML/MM as a broadly applicable strategy for mechanistic enzymology and a foundation for the rational design of new biocatalysts.