Pair potentials for protein folding: Choice of reference states and sensitivity of predicted native states to variations in the interaction schemes

We examine the similarities and differences between two widely used knowled ge-based potentials, which are expressed as contact matrices (consisting of 210 elements) that gives a scale for interaction energies between the natu rally occurring amino acid residues. These are the Miyazawa-Jernigan contac t interaction matrix M and the potential matrix S derived by Skolnick J et al., 1997, Protein Sci 6:676-688. Although the correlation between the two matrices is good, there is a relatively large dispersion between the elemen ts. We show that when Thr is chosen as a reference solvent within the Miyaz awa and Jernigan scheme, the dispersion between the ill and S matrices is r educed. The resulting interaction matrix B gives hydrophobicities that are in very good agreement with experiment. The small dispersion between the S and B matrices, which arises due to differing reference states, is shown to have dramatic effect on the predicted native states of lattice models of p roteins. These findings and other arguments are used to suggest that for re liable predictions of protein structures, pairwise additive potentials are not sufficient. We also establish that optimized protein sequences can tole rate relatively large random errors in the pair potentials. We conjecture t hat three body interaction may be needed to predict the folds of proteins i n a reliable manner.