In protein-structure prediction, simplified energy functions are necessarily used to allow fast sorting over many conformations. As a rule, these functions are derived from residue-residue approximation, which attributes all atomic interactions between residues to a single point within each residue. Physically, the simplified energies should result from averaging of the atomic interactions over various positions and conformations of the interacting amino acid residues, as well as the surrounding solvent molecules. Unfortunately, direct calculation of such mean-force potentials is not possible today both because of methodological difficulties and the lack of reliable atom-based energy functions.