Computing Statistical Profiles of Active Sites in Proteins

Active sites in proteins are three dimensional substructures that cause them to perform their function. The problem of finding substructures in a protein that are “similar” to the active sites of another protein has several important applications in biological sciences such as drug design, genetic engineering, and diagnostic tools for analysis of genetically engineered pathogens. Active sites can be grouped into families whose members are related by similarity of their functions. Since similar sites exhibit variability in their physico-chemical and structural features, statistical profiling methods capture the shared features robustly in the presence of such variations. In this paper, we adapt Profile Hidden Markov Models (PHMMs) that have been successfully used for analyzing biological sequences, to statistically profile active site families. Since PHMMs can only profile one dimensional sequences, we develop a serialization of the three dimensional active sites that captures certain shared physico-chemical and geometric features of the family. PHMM parameters are learnt using these serialized sequences. While traditional PHMM learning algorithms deal with discrete physico-chemical feature only, we expand it to include geometric features drawn from a continuous probability distribution. Experimental results with our PHMM based method for profiling active sites suggest that it is effective in practice.