Single nucleotide polymorphisms (SNPs) are the most common form of genetic variation in humans. The number of SNPs identified in the human genome is growing rapidly, but attaining experimental knowledge about the possible disease association of variants is laborious and time-consuming. Several computational methods have been developed for the classification of SNPs according to their predicted pathogenicity. In this study, we have evaluated the performance of nine widely used pathogenicity prediction methods available on the Internet. The evaluated methods were MutPred, nsSNPAnalyzer, Panther, PhD-SNP, PolyPhen, PolyPhen2, SIFT, SNAP, and SNPs&GO. The methods were tested with a set of over 40,000 pathogenic and neutral variants. We also assessed whether the type of original or substituting amino acid residue, the structural class of the protein, or the structural environment of the amino acid substitution, had an effect on the prediction performance. The performances of the programs ranged from poor (MCC 0.19) to reasonably good (MCC 0.65), and the results from the programs correlated poorly. The overall best performing methods in this study were SNPs&GO and MutPred, with accuracies reaching 0.82 and 0.81, respectively. Hum Mutat 32:1–11, 2011. © 2011 Wiley-Liss, Inc.