We report a preliminary study of the use of mean-force potentials (MFPs) for predicting protein tertiary structure. For three leucine zipper sequences, we have calculated ensembles of structures spanning all possible backbone conformations consistent with the canonical coiled-coil geometry. MFPs were measured with the program PROSA. The MFP alone was poor at discriminating the native structure from very divergent structures, and the global minimum of the MFP sometimes occurred far from the native structure. We found that adding an internal energy function (a subset of the CHARMM potential that describes only interactions between backbone atoms), the resultant total energy (CHARMM+PROSA) performed much better; in each case, there was a clear positive correlation between total energy and root-mean-square deviation (RMSD) from the experimental structure, and the lowest-energy structures were about 1 A RMSD from the experimental structures. Thus, we conclude that the combined potential is a powerful method for predicting leucine zippers and is very promising for general 3D structure prediction.