UCSF

Successfully incorporating backbone flexibility into the computational modeling and design of proteins and protein interactions is a key challenge that has yet to be fully solved. Most existing techniques for perturbing the backbone make changes that are not localized but propagate to distant parts of the protein, which can cause inefficiencies when working in a restricted region of the protein. The methods that do make localized changes have mostly been based on complex mathematical formulations that can inhibit their widespread application. In this thesis I describe a simple, automated approach, termed "backrub," for sampling the protein backbone. The method is based on a recurring motif of backbone motion previously observed in ultrahigh resolution (1Å) crystal structures, and involves backbone rotations around axes between C-alpha atoms. It is shown to be useful for a variety of applications, including recapitulating the backbone/side chain bias in known instances of the backrub motif, predicting the conformations of point mutants, and modeling the opening and closing of a loop around an enzyme active site. After these initial results, I undertook a large-scale study in retrospective and prospective prediction of the peptide binding specificities of natural and synthetic PDZ domains. Here, backrub backbone flexibility was shown to significantly improve the accuracy of amino acid frequency prediction. The developed method was able to capture a large fraction of the amino acids frequently observed in phage display experiments both with natural PDZ domains and a large dataset of point mutants. Finally, in an effort to broaden the application and use of the PDZ peptide specificity work, I generalized the method to also predict fold stability, using GB1 phage display as a benchmark, and produced a detailed protocol for others to apply to a wide variety of systems.