Complete sets of low-resolution conformations are generated for eight small proteins by rotating the C ␣ -C ␣ virtual bonds at selected flexible regions, while the remaining structural elements are assumed to move in rigid blocks. Several filtering criteria are used to reduce the ensemble size and to ensure the sampling of well-constructed conformations. These filters, based on structure and energy constraints deduced from knowledgebased studies, include the excluded volume requirement, the radius of gyration constraint, and the occurrence of sufficiently strong attractive inter-residue potentials to stabilize compact forms. About 8,000 well-constructed decoys or ''probable folds'' (PFs) are constructed for each protein. A correlation between rootmean-square (rms) deviations from X-ray structure and total energies is observed, revealing a decrease in energy as the rms deviation decreases. The conformation with the lowest energy exhibits an rms deviation smaller than 3.0 A ˚, in most of the proteins considered. The results are highly sensitive to the choice of flexible regions. A strong tendency to assume native state rotational angles is revealed for some flexible bonds from the analysis of the distributions of dihedral angles in the PFs, suggesting the formation of foldons near these locally stable regions at early folding pathway.