Efficient and Accurate pKa Prediction Enabled by Pre-Trained Machine-Learned Interatomic Potentials

Quickly and accurately predicting the pKa of small molecules is an important unsolved challenge in computational chemistry: while approaches based on electronic structure theory have shown great promise, the utility of these methods is limited by the considerable expense of the requisite computations. In this study, we investigate AIMNet2, a machine-learned interatomic potential, as a low-cost surrogate for electronic structure theory in pKa prediction. The accuracy of the AIMNet2-based pKa prediction workflow is evaluated over a wide range of compounds and functional groups, and potential sources of error are discussed.