Leveraging Structural Flexibility to Predict Protein Functio

Proteins are essentially versatile and flexible molecules and understanding protein function plays a fundamental role in understanding biological systems. Protein structure comparisons are widely used for revealing protein function. However,with rigidity or partial rigidity assumption, most existing comparison methods do not consider conformational flexibility in protein structures. To address this issue, this thesis seeks to develop algorithms for flexible structure comparisons to predict one specific aspect of protein function, binding specificity. Given conformational samples as flexibility representation, we focus on two predictive problems related to specificity: aggregate prediction and individual prediction..;For aggregate prediction, we have designed FAVA (Flexible Aggregate Volumetric Analysis). FAVA is the first conformationally general method to compare proteins with identical folds but different specificities. FAVA is able to correctly categorize members of protein superfamilies and to identify influential amino acids that cause different specificities. A second method PEAP (Point-based Ensemble for Aggregate Prediction) employs ensemble clustering techniques from many base clustering to predict binding specificity. This method incorporates structural motions of functional substructures and is capable of mitigating prediction errors.;For individual prediction, the first method is an atomic point representation for representing flexibilities in the binding cavity. This representation is able to predict binding specificity on each protein conformation with high accuracy, and it is the first to analyze maps of binding cavity conformations that describe proteins with different specificities. Our second method introduces a volumetric lattice representation. This representation localizes solvent-accessible shape of the binding cavity by computing cavity volume in each user-defined space. It proves to be more informative than point-based representations. Last but not least, we discuss a structure-independent representation. This representation builds a lattice model on protein electrostatic isopotentials. This is the first known method to predict binding specificity explicitly from the perspective of electrostatic fields.;The methods presented in this thesis incorporate the variety of protein conformations into the analysis of protein ligand binding, and provide more views on flexible structure comparisons and structure-based function annotation of molecular design.