Towards accurate machine learning predictions of properties of the P–O bond cleaving in ATP upon enzymatic hydrolysis

Molecular dynamic simulations using QM/MM potentials are performed for the enzyme–substrate complex of adenosine triphosphate (ATP) with the motor protein myosin. Machine learning methods are applied to a dataset consisting of the geometry parameters of the active site in the enzyme–substrate complex to predict the Laplacian of electron density at the bond critical point of the P_G–O_3B bond being broken in ATP. Using a gradient boosting machine learning model, a mean absolute error of 0.01 a.u. and an R² score of 0.99 are achieved, and it is found that the P_G–O_3B bond length is the most important feature, contributing 2/3, while other geometry features contribute 1/3.