Prediction of protein residue contacts with a PDB-derived likelihood matrix

Proteins with similar folds often display common patterns ofresidue variability. A widely discussed question is how thesepatterns can be identified and deconvoluted to predict proteinstructure. In this respect, correlated mutation analysis (CMA)has shown considerable promise. CMA compares multiple membersof a protein family and detects residues that remain constantor mutate in tandem. Often this behavior points to structuralor functional interdependence between residues. CMA has beenused to predict pairs of amino acids that are distant in theprimary sequence but likely to form close contacts in the nativethree-dimensional structure. Until now these methods have usedevolutionary or biophysical models to score the fit betweenresidues. We wished to test whether empirical methods, derivedfrom known protein structures, would provide useful predictivepower for CMA. We analyzed 672 known protein structures, derivedcontact likelihood scores for all possible amino acid pairs,and used these scores to predict contacts. We then tested themethod on 118 different protein families for which structureshave been solved to atomic resolution. The mean performancewas almost seven times better than random prediction. Used inconcert with secondary structure prediction, the new CMA methodcould supply restraints for predicting still undetermined structures.