A computer program for de novo molecular design was used to explore the diversity of molecules complementary to the binding sites of enzymes. The program, GrowMol(1) (preliminary results presented at the XIIth International Symposium on Medicinal Chemistry, Basel, Switzerland, 1992), generates molecules with steric and chemical complementarity to the three-dimensional structure of a host binding site. The molecules are created in the host binding site one atom or functional group at a time. At each step the position and type of atom to be added is randomly selected from a set of possible values which are consistent with internal bond lengths and bond angle requirements as well as the spatial and chemical properties of the binding site. The selection is achieved using Boltzmann statistics to bias acceptance toward atoms which can form favorable interactions with the binding site. Rings are generated by connecting closely positioned nonbonded atoms. This process ensures that a highly diverse set of molecules is generated unbiased by knowledge of previous guest molecules. When applied to several enzyme binding sites, the program produced structures which were identical to or closely resembled known inhibitors of these enzymes. In addition, the program rapidly produced tens of thousands of distinct molecules which display a large variety of structural motifs. A detailed analysis of the potential diversity of thiol inhibitors of thermolysin was carried out. Twenty-two thousand structures were generated in the thermolysin binding site. Twelve thousand of these were unique. After energy minimization in the active site and rejection of structures with a high conformational strain energy, four thousand unique structures were found which had an estimated potency comparable to those of known inhibitors. To investigate the range of diversity of these structures, they were clustered into distinct families. A representative example from each cluster was selected, resulting in a set of approximately three hundred structures which were examined visually. The resulting set included structures which were substrate mimics, non-peptidic structures which satisfied the hydrogen-bonding requirements of the binding site in novel ways, a large variety of different groups filling the hydrophobic binding pockets, and various macrocycles and fused-ring structures which spanned adjacent binding pockets. Many of these structures differ considerably from each other and have little in common with known inhibitors. To identify structures most likely to bind well to the enzyme, a scoring algorithm based on complementarity was developed. The estimated potency determined with this algorithm correlated well with the experimentally determined values of known inhibitors. The results of this study clearly demonstrate that molecular structures generated by computer programs incorporating algorithms for de novo design can reveal a wealth of opportunities for the design of novel enzyme inhibitors.