Given the three-dimensional structure of a receptor site, there are several methods available for designing ligands to occupy the site; frequently, the three-dimensional structure of interesting receptors is not known, however. The GERM program uses a genetic algorithm to produce atomic-level models of receptor sites, based on a small set of known structure-activity relationships. The evolved models show a high correlation between calculated intermolecular energies and bioactivities; they also give reasonable predictions of bioactivity for compounds which were not included in model generation. Such models may serve as starting points for computational or human ligand design efforts.