Comparative evaluation of pK(a) prediction tools on a drug discovery dataset

Due to their impact on pharmacokinetic and pharmacodynamic properties the accurate prediction of dissociation constants is of outmost importance in drug discovery settings. The prediction accuracy, however, is typically assessed on public datasets most likely included in the training sets of the available tools. In this work we therefore tested five pK(a) prediction softwares such as ACD, Epik, Marvin, PharmaAlgorithm and Pallas on novel, never-published compounds. Our dataset consists of 177 pK(a) values of 95 structurally diverse in-house compounds prepared for real-life drug discovery programs. The thorough analysis of prediction accuracy allowed us identifying the best practice and exploring the limitations of the current methods. Mean absolute errors (0.86-1.28) obtained for this set of discovery compounds indicates the potential in the improvement of the available pK(a) prediction approaches. Limitations were further characterized by measuring and evaluating 39 pK(a) values of additional 28 commercially available compounds representing the most challenging chemotypes. We believe that these results would facilitate further developments and hopefully contribute to the necessary improvement of the prediction accuracy.