The recent review of the ICM method and its applications to protein folding, homology modeling, docking, domain movements and protein design. 'Low-Energy Alternative Folds' (LEAF) hypothesis is introduced. It is argued that the accuracy of energy calculations in protein structure prediction should not exceed about 1 kcal/mole/residue for most of the protein topologies. Therefore fast and accurate electrostatic solvation term, surface term and the entropy term should be added to the optimized energy function. Omission of any of these terms may lead to an impermissible level of energy error. A number of peptides up to 23 residues, having different experimentally characterized topologies can be predicted ab initio in a detailed, convergent, full atom ICM global optimization of the free energy function with probability biased random steps. Exclusion of the entropic term leads to deformations. The improved accuracy of energy calculations may be necessary to predict larger molecules.