Development and assessment of a ChemInformatics model for accurate pKa prediction in aqueous medium

The accurate prediction of the acid dissociation constants (pKa) of organic and drug molecules is known to be a challenging problem in computational quantum chemistry. Specifically, density functional theory-based predictions suffer from a high dependence on the nature of the functional group as well as the underlying exchange–correlation functional. Additionally, the introduction of explicit solvent molecules is known to be important for accurate prediction of the pKa values for many functional groups in water, making it a particularly challenging problem. The inclusion of only implicit solvation effects, though highly efficient, is often inadequate for the prediction of pKas. In this paper, we have considered a data set of 303 molecules containing 13 different functional groups to assess the predictability of DFT for the calculation of pKas. Using just implicit solvation models with DFT, each functional group shows a linear correlation with experiment, though with different slopes for different functional groups. Using simple linear regression-based corrections for systematic errors of different functional groups, we show that DFT including implicit solvation can be used to make reliable predictions of pKas with a mean absolute deviation of only 0.397 pKa units. For a test set of 100 larger and more complex drug molecules, the performance of our model is very good, though with a slightly larger mean absolute deviation of 0.629 pKa units. More importantly, our pKa protocol is general and applicable to any underlying density functional, making it an effective computational tool for pKa predictions.