Deciphering the protein folding code: Ab initio prediction of protein structure

A solution to the protein folding problem has been the Holy Grail in structural molecular biology for the past 50 years. Early experiments in the field of protein folding discovered that all of the information coding for the 3-dimensional structure of a protein is contained in its sequence. The work described in this dissertation is largely devoted to the development of a method for the prediction of protein structures from amino acid sequence. The incredible challenge of the protein folding problem can be separated into two major obstacles: reducing the conformational complexity of the polypeptide chain and distinguishing the correct native structure from incorrect misfolded conformations.; To reduce the possible number of local structures of the protein chain, the approach of our laboratory is to combine fragments of known protein structures with similar local sequence and predicted secondary structure to predict the complete tertiary structure. The tertiary structures are evaluated on the basis of similarity to features of known protein structures including but not limited to the formation of hydrophobic cores and the correct geometry of beta-strands in beta-sheets. Combined with the frequency at which structures are generated, the scoring function is able to detect the differences between the native and misfolded structures.; Ab initio simulations provide the true test of a scoring function. There are two potentially fatal outcomes: either the scoring function is too exact, relying on the recognition of the fine details of native structure, or the scoring function is too inexact, not allowing the discrimination of numerous artificial non-native minima from the minimum containing the native structure. Not using any information about the native structure, simulations using our scoring functions on protein sequences of known structure have been successful for proteins of less than 100 residues, including proteins with beta-strands. Tests on blind predictions have shown that the method has not been crafted on a particular set of sequences; structures with less than 7A Calpha rmsd and over 90 residues have been predicted for two of the targets with hopefully more successes to follow as the coordinates of experimentally determined structures are released publicly.