Efficient and Automated Analysis of Protein Structures

In recent years, computational complexity in structural bioinformatics attained a new level with the vast increase in the amount of structural data available. The Protein Data Bank (PDB), which is the single worldwide repository for 3-D macromolecular structure data, contains more than 25k structures as of July 2004. However, existing methods for protein structure analysis are unable to cope with this increase in the amount of available data. Therefore, this wealth of data requires computationally efficient methods to be developed for the analysis of large numbers of protein structures and their associated functions. In this dissertation, we present methods for protein structure analysis that can scale well with the amount of protein structure data available. Our work can be described under three main categories: (1) visualization and surface modelling, (2) structure comparison and similarity search, and (3) automated classification. For efficiently visualizing protein structures using a scene-graph based graphics API, we have developed methods to optimize the constructed scene-graph to enable real-time visualization of very large protein complexes. Our method (FPV) achieves up to 8 times interactive speed compared to existing methods. For generation of molecular surfaces we recently developed a method based on a level set formulation that can compute the surface and interior inaccessible cavities very efficiently (1.5 to 3.14 times faster on the average than compared methods). For comparison and similarity search of protein structures we have developed a method that utilizes local shape signatures based on the theory of differential geometry. Our method (CTSS) is up to 30 times faster than CE, a widely used method for structure comparison, while achieving the similar level of accuracy. We have also developed an integrated sequence and structure analysis method (ProGreSS), which enables biologists to perform joint sequence and structure similarity queries while improving on the accuracy and efficiency of existing methods. For an up-to-date view of the protein structure universe with the help of automated classification, we have developed an ensemble classifier based on decision trees rooted in machine learning. We show that higher classification accuracy can be achieved using the joint hypothesis of the ensemble classifier.