The biological activities of proteins depend on specific molecular recognition and binding. Computational methods for predicting binding modes can facilitate the discovery and design of ligands and yield information on the factors governing complementarity. The DOCK suite of programs has been applied to several systems; here, the degree of orientational sampling required to reproduce and identify known binding modes, with and without rigid-body energy minimization, is investigated for four complexes. There is a tradeoff between sampling and minimization. The known binding modes can be identified with intensive sampling alone (10,000 to 20,000 orientations generated per system) or with moderate sampling combined with minimization. Optimization improves energies significantly, particularly when steric clashes are present, and brings many orientations closer to the experimentally observed position. Whether or not minimization is performed, however, sampling must be sufficient to find at least one structure in the vicinity of the presumed true binding mode. Hybrid approaches combining docking and minimization are promising and will become more viable with the use of faster algorithms and the judicious selection of fewer orientations for minimization.