Functional classification of proteins by pattern discovery and top-down clustering of primary sequences

Given a functionally heterogeneous group of proteins, such as a large superfamily, or an entire database, two important problems in biology are the automated inference of subsets of functionally related proteins and the identification of functional regions and residues. The former is typically performed in an unsupervised bottom-up manner, by clustering based on pairwise sequence similarity. The latter is performed independently, in a supervised top-down manner starting from functional groups that have already been identified by either biological or computational means. Clearly, however, the two processes remain inextricably linked, because functional motifs and residues are related to corresponding functional clusters. This paper introduces a high-performance, unsupervised, top-down clustering technique and the corresponding system that determines functionally related clusters and functional motifs by coupling a pattern discovery algorithm, a statistical framework for the analysis of discovered patterns, and a motif refinement method based on Hidden Markov Models. The high performance comes in two ways. First, the functional motifs are determined by first discovering regular expressions from the database, which can be done relatively fast and easily, and then converting them into statistical models, which offer computational complexity and theoretical soundness. This approach, as opposed to the one where rigorous treatments of functional motifs are attempted from the very beginning, is expected to achieve both efficiency as well as superior performance. Second, the system constructs a binary tree during the top-down clustering process. Since the two child nodes of a parent node in the tree are independent, the construction and manipulation of the two child nodes can be carried out simultaneously, therefore resulting in greater efficiency. Results are reported for the G-Protein Coupled Receptor superfamily. These show that a significant majority of wellknown functional groups and biologically relevant motifs are correctly recovered. They also show that a majority of the important functional residues reported in the literature occur in the inferred functional motifs. This technique has relevant implication in functional clustering and it could be used as a highly predictive aid to mutagenesis experiments.