PURPOSE: Assessment of the accuracy of experimental and theoretical methods of pKa determination for acids and bases as separate classes. METHODS: Four literature pKa datasets were checked for errors and pKa values assigned unambiguously to a single acidic and/or basic ionisation centre. A new chemically diverse and drug-like dataset was compiled from high-throughput UV-vis spectrophotometry pKa data. Measured pKa values were compared with data obtained by alternative methods and predictions by the Epik, Chemaxon and ACD pKa DB software packages. RESULTS: The pKa values of bases were considerably less accurately predicted than those of acids, in particular for structurally complex bases. Several new chemical motifs were identified for which pKa values were particularly poorly predicted. Comparison of pKa values obtained by UV-vis spectrophotometry and different literature sources revealed that low aqueous solubility and chromophore strength can affect the accuracy of experimental pKa determination for certain bases but not acids. CONCLUSIONS: The pKa prediction tools Epik, Chemaxon and ACD pKa DB provide significantly less accurate predictions for bases compared to acids. Certain chemical features are underrepresented in currently available pKa data sets and as a result poorly predicted. Acids and bases need to be considered as separate classes during pKa predictor development and validation.