The comparison of eight tools applicable to ligand-binding site prediction is presented. The methods examined cover three types of approaches: the geometrical (CASTp, PASS, Pocket-Finder), the physicochemical (Q-SiteFinder, FOD) and the knowledge-based (ConSurf, SuMo, WebFEATURE). The accuracy of predictions was measured in reference to the catalytic residues documented in the Catalytic Site Atlas. The test was performed on a set comprising selected chains of hydrolases. The results were analysed with regard to size, polarity, secondary structure, accessible solvent area of predicted sites as well as parameters commonly used in machine learning (F-measure, MCC). The relative accuracies of predictions are presented in the ROC space, allowing determination of the optimal methods by means of the ROC convex hull. Additionally the minimum expected cost analysis was performed. Both advantages and disadvantages of the eight methods are presented. Characterization of protein chains in respect to the level of difficulty in the active site prediction is introduced. The main reasons for failures are discussed. Overall, the best performance offers SuMo followed by FOD, while Pocket-Finder is the best method among the geometrical approaches.Electronic supplementary materialThe online version of this article (doi:10.1007/s10822-010-9402-0) contains supplementary material, which is available to authorized users.