Benchmarking pK_a Prediction Algorithms against an Extensive, Public Data Set

Accurate prediction of proton dissociation constants (pK_a) is essential for downstream drug discovery and molecular modeling workflows. While several proprietary pK_a prediction tools have been established as popular choices in the field, open-source solutions provide viable alternatives for large-scale computational workflows. In particular, machine learning approaches have recently emerged as a promising orthogonal route to traditional empirical methods. However, many of these algorithms were benchmarked on small, disjoint data sets, with inconsistencies in pK_a data interpretation, particularly for polyprotic molecules. To address these challenges, we assembled a comprehensive data set of over 90,000 experimental aqueous pK_a values spanning over 31,000 unique molecules from scattered online resources, with each entry annotated for charge state transitions and microspecies distributions. This data set, made accessible through the pKahub online database, represents one of the largest publicly available collections of annotated pK_a data to date. We used this resource to benchmark seven pK_a prediction methods, including three commercial tools (ACD/Labs, Chemaxon, and Epik) and four open-source machine learning models (MolGpKa, pKaSolver, QupKake, and Uni-pK_a).