Computational prediction of protein--protein interactions

Protein–protein interactions are vital for cellular function. Computational methods that predict the high-resolution structures of protein–protein complexes offer functional insights and guide rational engineering efforts to identify potential therapeutic targets, or modify protein binding affinities and specificities. With the scope of structural biological simulations expanding rapidly, there is a need for the protein complex prediction methods to adapt to the increasing complexity. So I have developed tools to expand their capacity beyond idealized protein–protein docking. In this thesis, I detail my work focusing on developing new methods to account for the effects of a critical environmental factor, pH, on protein complexes, and further advancements to address other protein interaction challenges. First, I developed Rosetta-pH, a fast and efficient method to calculate the pKa values of protein residues that commonly exhibit variable protonation states. I studied the effects of incorporating increasing levels of protein conformational flexibility on the pKa calculations, and tested the method’s efficacy in capturing large pKa shifts in Staphylococcal nuclease point mutants. Second, I utilized the knowledge of residue pKa values to develop pHDock, the first protein–protein docking method that can sample side-chain protonation states on-the-fly during the docking simulation. pHDock generates more accurate models for the protein complexes compared to the conventional docking method (RosettaDock) with fixed protonation states. I also demonstrated that pHDock can be further expanded to include binding affinity calculations by using it to predict a large pH-dependent binding affinity change in the Fc–FcRn complex. Third, I expanded RosettaDock to address diverse challenges in CAPRI (Critical Assessment of PRediction of Interactions), a world-wide blind protein interaction prediction challenge. Specifically, I developed methods to predict structures of water molecules at protein–protein interfaces, dock flexible sugar–protein complexes, discriminate natural and designed protein interfaces, and predict effects of mutations on binding affinities of protein complexes. Fourth, I used a cross-docking study to develop a structural basis to discriminate protein binders from non-binders by identifying native antibody–antigen interaction pairs among cognate and non-cognate complexes. I demonstrated a decrease in the prediction accuracy when using bound and unbound coordinates, and homology models, respectively, for the antibodies, potentially related to a drop in the interface hydrogen bond formation due to backbone inaccuracies. In summary, I have developed methods to incorporate pH effects on protein–protein complexes and expanded existing docking methods to target binding predictions and interactions involving non-protein biomolecules. The diversity of biomolecular problems requires computational methods to be versatile; my contributions expand their capabilities to encompass more biologically realistic docking problems