Chemistry

Objectives The pKa value of a drug is a key factor in its solubility and other ADME properties, including receptor-binding. Consequently, numerous attempts have been made to predict pKa values from structural properties. Most of these attempts have been based on the Hammett constant, and so have involved congeneric series. There have, however, been a few attempts, such as those of Klopman & Fercu (1994) and of Klamt et al (2003), to model pKa values of diverse data-sets. Recently, commercial software has become available for the prediction of pKa values. However, no comparison of the performance of pKa software has been published. In an attempt to rectify this situation, we have tested the predictive ability of ten such software programs, using a large test-set. Methods Our test-set comprised 653 compounds, and included a large number of drugs. There were about 40 tautomeric compounds in the test-set. Many of the compounds had multiple ionisation sites, but we used only the primary measured pKa values, as these were considered to be the most accurate. In addition, some of the software programs calculated only the primary pKa value of a compound. The software programs tested were: ACD/Labs (www.acdlabs.com), ADME Boxes (www.ap-algorithms.com), ADMET Predictor (www.simulationsplus.com), ChemAxon (www.chemaxon.com), CSpKa (www.chemsilico.com), PALLAS (www.compudrug.com), Pipeline Pilot (www.scitegic.com), QikProp (www.schrodinger.com), SPARC (ibmlc2.chem.uga.edu/sparc) and VCCLAB (www.vcclab.org). Two of the software programs (SPARC and VCCLAB) are freely usable on-line, and two programs were already in use in our laboratory. Results from the other software programs were kindly provided by the software companies. Results Some of the programs did not predict pKa values for all of the test-set compounds. The predictive abilities of the ten programs are shown in Table 1. The r value is the coefficient of determination for the correlation between observed and predicted pKa values, and MAE is the mean absolute error of prediction. Conclusions There is wide variation between the predictive abilities of the software programs. The weak performance of the CSpKa software could be due, in part at least, to the fact that this was the only software for which we could be certain that none of our test-set compounds was in the training set used to develop each software program. If a calculated pKa is required, it is recommended that predictions be obtained from three sources, and the mean taken. Klamt, A., et al (2003) J. Phys. Chem. A 107: 9380–9386 Klopman, G., Fercu, D. (1994) J. Comput. Chem. 15: 1041–1050