Despite the simplicity of the model and the energy function, many of the lowest energy structures, using exact measures, are near the native structures (in terms of RMSD). As expected, when using predicted measures, the fraction of good decoys decreases, but in all cases tested, we obtained structures within 6 Å RMSD in a set of low-energy decoys. To the best of our knowledge, this is the first de novo branch and bound algorithm for protein decoy generation that only depends on such one-dimensional predictable measures. Another important advantage of the branch and bound approach is that the algorithm searches through the entire conformational space. Contrary to search heuristics, like Monte Carlo simulation or tabu search, the problem of escaping local minima is indirectly solved by the branch and bound algorithm when good lower bounds can be obtained.