Protein Constellations

In this project, I applied several well-known ideas developed in the computer vision field to problems in protein structure matching. Shape indexing and generalized Hough transforms are classical ideas and have been applied to protein structure matching before (eg, [1, 11, 14]). However, I am using a different representation of protein shape and different matching criteria, which yield results that seem promising. In the shape indexing (or geometric hashing) paradigm, a database of proteins is split into small, overlapping segments (“snippets”), and the local geometry of each snippet is described as a vector of real values. In the generalized Hough transform phase, we take a query protein, apply the same transformation to it, and search for snippets with similar local geometry descriptors. Each descriptor that is within a given tolerance is a vote for a match between a snippet from the query protein and a snippet from the database. By accumulating votes, we arrive at a ranking of matching database proteins. In this work, I extend these ideas by adding relevance-weighted matching and by building global alignments from chains of local matches. I also make the observation that the database of snippets is a valuable source of statistical information about protein structures. By searching for snippets that have a large number of neighbours, I produce a library of structural motifs that is nearly as good as other published libraries which are produced by more complex processes.