Development and application of a free energy force field for all atom protein folding

Proteins are the workhorses of all cellular life. They constitute the building blocks and the machinery of all cells and typically function in specific three-dimensional conformations into which each protein folds. Currently over one million protein sequences are known, compared to about 40,000 structures deposited in the Protein Data Bank (the world-wide database of protein structures). Reliable theoretical methods for protein structure prediction could help to reduce the gap between sequence and structural databases and elucidate the biological information in structurally unresolved sequences. In this thesis we explore an approach for protein structure prediction and folding that is based on the Anfinsen’s hypothesis that most proteins in their native state are in thermodynamic equilibrium with their environment. We have developed a free energy forcefield (PFF02) that locates the native conformation of many proteins from all structural classes at the global minimum of the free-energy model. We have validated the forcefield against a large decoy set (Rosetta). The average root mean square deviation (RMSD) for the lowest energy structure for the 32 proteins of the decoy set was only 2.14 from the experimental conformation. We have successfully implemented and used stochastic optimization methods, such as the basin hopping technique and evolutionary algorithms for all atom protein structure prediction. The evolutionary algorithm performs exceptionally well on large supercomputational architectures, such as BlueGene and MareNostrum. Using the PFF02 forcefield, we were able to fold 13 proteins (12-56 amino acids), which include helix, sheet and mixed secondary structure. On average the predicted structure of these proteins deviated from their experimental conformation by only 2.89 RMSD.