Numerous chemicals released into the environment can interfere with normal, hormonally regulated biological processes to adversely affect development and/or reproductive function in wildlife and humans. Due to the ability of these chemicals to interfere with the endocrine systems, they have been labeled as endocrine disruptors (EDs). SARs and QSARs are powerful screening tools to detect potential EDs and to prioritize them for more intensive and costly evaluations based on in vitro and in vivo assays. In this context, androgen-receptor binding data (active/inactive) for a large set of about 200 structurally diverse chemicals, described by CODESSA descriptors encoding topological and physicochemical properties, were used for deriving structure-activity models. Different types of artificial neural networks and support vector machines with different kernel functions were tested as statistical tools. The performance of a classical discriminant analysis was also estimated. The comparison exercise was performed on the basis of the same learning and testing sets as well as from the same set of selected descriptors. The modeling performances as well as the technical advantages and limitations of each statistical method have been critically analyzed.