Evaluation of pKa estimation methods on 211 druglike compounds.

The pK(a) values of 211 discovery (druglike) compounds were determined experimentally using capillary electrophoresis coupled with ultraviolet spectroscopy and a novel fitting algorithm. These values were compared to those predicted by five different commercially available pK(a) estimation packages: ACDLabs/pK(a), Marvin (ChemAxon), MoKa (Molecular Discovery), Epik (Schrodinger), and Pipeline Pilot (Accelrys). Even though the topological method MoKa was noticeably faster than ACD, the accuracy of those two methods and Marvin was statistically indistinguishable, with a root-mean-squared error of about 1 pK(a) unit compared to experiment. Pipeline Pilot and EpiK both produced pK(a) estimates in significantly worse agreement with the experiment. Interestingly, on a number of compounds, the predictions due to ACD v12 were in poorer agreement with the experiment than ACD v10. Microscopic and "apparent" pK(a) predictions were also compared using ACD v10. Microscopic pK(a)s gave significantly worse agreement with the experiment than the "apparent" values. In all cases, the errors appeared to be randomly distributed across chemical series.