Dimensionality reduction in patch-signature based protein structure matching

Searching bio-chemical structures is becoming an important application domain of information retrieval. This paper introduces a protein structure matching problem and formulates it as an information retrieval problem. We first present a novel vector representation for protein structures, in which a protein structural region, formed by the vectors within the region, is defined as a patch and indexed by its patch signature. For a k-sized patch, its patch signature consists of 7k − 10 inter-atom distances which uniquely determine the patch’s spatial structure. A patch matching function is then defined. As structures for proteins are large and complex, it is computationally expensive to identify possible matching patches for a given protein against a large protein database. We propose to apply dimensionality reduction to the patch signatures and show how the two problems are adapted to fit each other. The Locality Preservation Projection (LPP) and Singular Value Decomposition (SVD) are chosen and tested for this purpose. Experimental results show that the dimensionality reduction improves the searching speed while maintaining acceptable precision and recall. From a more general point of view, this paper demonstrates that information retrieval techniques can play a crucial role in solving this biologically critical but computationally expensive problem.