Classification methods allow for the development of structure-activity relationship models when the target property is categorical rather than continuous. We describe a classification method which fits descriptor splines to activities, with descriptors selected using a genetic algorithm. This method, which we identify as SFGA, is compared to the well-established techniques of recursive partitioning (RP) and soft independent modeling by class analogy (SIMCA) using five series of compounds: cyclooxygenase-2 (COX-2) inhibitors, benzodiazepine receptor (BZR) ligands, estrogen receptor (ER) ligands, dihydrofolate reductase (DHFR) inhibitors, and monoamine oxidase (MAO) inhibitors. Only 1-D and 2-D descriptors were used. Approximately 40% of compounds in each series were assigned to a test set, "cherry-picked" from the complete set such that they lie outside the training set as much as possible. SFGA produced models that were more predictive for all but the DHFR set, for which SIMCA was most predictive. RP gave the least predictive models for all but the MAO set. A similar trend was observed when using training and test sets to which compounds were randomly assigned and when gradually eliminating compounds from the (designed) training set. The stability of models was examined for the random and reduced sets, where stability means that classification statistics and the selected descriptors are similar for models derived from different sets. Here, SIMCA produced the most stable models, followed by SFGA and RP. We show that a consensus approach that combines all three methods outperforms the single best model for all data sets.