In this paper, we introduce a method to account for the shape of the potential energy curve in the evaluation of conformational free energies. The method is based on a procedure that generates a set of conformations, each with its own force-field energy, but adds a term to this energy that favors conformations that are close in structure (have a low rmsd) to other conformations. The sum of the force-field energy and rmsd-dependent term is defined here as the ''colony energy'' of a given conformation, because each conformation that is generated is viewed as representing a colony of points. The use of the colony energy tends to select conformations that are located in broad energy basins. The approach is applied to the ab initio prediction of the conformations of all of the loops in a dataset of 135 nonredundant proteins. By using an rmsd from a native criterion based on the superposition of loop stems, the average rmsd of 5-, 6-, 7-, and 8-residue long loops is 0.85, 0.92, 1.23, and 1.45 Å, respectively. force field ͉ energy minimization ͉ protein structure prediction P rotein loops are usually defined as segments of the polypeptide chain that do not contain regular units of secondary structure. Although some loops seem to serve as no more than connectors between secondary structure elements, others have been implicated as determinants of protein stability and folding pathways, while others may play important functional roles. The loop prediction problem involves finding the correct conformation for a given loop under the constraint that both ends are fixed through their connection to the rest of the protein. The problem has taken on considerable importance with the increased application of homology modeling methods in protein structure prediction. Although secondary structure elements can, in many cases, be predicted with considerable accuracy because they are often well conserved, sequence and structural variability are integral properties of many loops where, as in the case of antibodies, specificity differences among family members often reside. This variability makes the problem of loop prediction particularly complicated because, by its very nature, homology methods will often not be applicable. Indeed, loop prediction can to some extent be viewed as a mini ab initio folding problem, because the necessary information will not necessarily be found in databases. As such, the problem also serves as an important test of our understanding of the physical chemical principles that determine protein structure.Two general approaches have been applied to the prediction of loop conformation: database search and ab initio techniques. In the database search method (1-6), a library of segments derived from known protein structures is searched for conformations that fit the topological constraint of the loop stems. The stems correspond to the main-chain atoms that precede and follow the loop, but are not part of the loop itself. Loop candidates found in this way then can be evaluated by different criteria such as sequence...