A Machine Learning-Driven, Probability-Based Approach to Enzyme Catalysis

The workings of enzymes depend crucially on transition state structures, which encode critical chemical information necessary to control their efficiency and selectivity. However, capturing these configurations and describing them on a statistical basis remains a significant challenge due to their transient nature. Here, we leverage a novel enhanced sampling scheme based on a machine-learned committor function to provide a probabilistic characterization of transition states in enzymatic reactions. Applied to the glycolysis reaction of maltopentaose catalyzed by human pancreatic α-amylase, this approach successfully reveals the critical role of water molecules in shaping the catalytic landscape, dictating whether the reaction follows a water-assisted or water-mediated mechanism, and providing atomistic insight on how specific hydrogen bonding interactions within the catalytic pocket can influence the stability of transition states. Our findings highlight the potential of this machine-learning-based enhanced sampling scheme to study rare events in complex biochemical systems, offering a powerful tool for unveiling mechanistic details that are often elusive with traditional simulation approaches and paving the way for accelerating the rational design of novel enzymes through more accurate dynamics-activity correlations targeting the transition state ensemble.