| All-atom models for investigating protein folding and protein-ligand binding are presented. A novel Monte Carlo folding simulation combining full structural realism with coarse-grained energetics is first described. In the first half of this thesis, the Gō potential is used, where the native structure is used to parametrize atom-atom interactions. Ensemble folding data on (1) the protein crambin and two its structural components, the helix and helix hairpin and (2) the experimentally well-characterized protein G are collected. In all cases, folding was cooperative and many kinetic pathways to the native state were observed. A low temperature kinetic trap arising from the incorrect packing of sidechains was also observed. Independently, we determined that there are many different sidechain conformations that are as efficiently packed as the native structure. We attribute this large sidechain entropy as the likely cause of the slow relaxation at low temperatures. With protein G, when folding is monitored using a frequently used reaction coordinate, key experimental observations, such as single exponential kinetics, the burst phase, and mutational data are reproduced. However, more detailed analysis reveals that folding occurs over three distinct, three-state pathways, suggesting that ensemble averaging under this reaction coordinate disguises multiple pathways and intermediates. Interestingly, all pathways eventually converge to a common rate-limiting step, which is the formation of a specific nucleus involving hydrophobic core residues.; In order to move towards general, predictive all-atom models, methods for deriving sequence-based potentials, are next presented. Using a simple method that balances hydrophilic and hydrophobic interactions, a potential capable of folding a three-helix bundle to less than 2 Å Cα dRMS from the native structure is derived. Similar results were obtained for a β-hairpin and helix. In addition, we propose a general, rigorous method for deriving any sequence-based contact potential. It analytically minimizes the Z score in order to establish an energy gap between the native state and competing decoys. Using this method, we show conclusively that residue-based potentials (featuring a maximum of 210 parameters) cannot be used to fold a small β-hairpin. Finally, a self-consistent approach to analyze knowledge-based protein-ligand binding potentials is given. We observed a statistically significant correlation between free energy estimates obtained from this method and experimental binding scores for a diverse set of complexes. |
