Discovering Active Motifs in Sets of Related Protein Sequences and Using Them for Classi cation

We describe a method for discovering active motifs in a set of related protein sequences. The method is an automatic two step process: (1) nd candidate motifs in a small sample of the sequences; (2) test whether these motifs are approximately present in all the sequences. To reduce the running time, we develop two optimization heuristics based on statistical estimation and pattern matching techniques. Experimental results obtained by running these algorithms on generated data and functionally related proteins demonstrate the good performance of the presented method compared with visual method of O'Farrell and Leopold. By combining the discovered motifs with an existing ngerprint technique, we develop a protein classi er. When we apply the classi er to the 698 groups of related proteins in the PROSITE catalog, it gives information that is complementary to the BLOCKS protein classi er of Heniko and Heniko . Thus, using our classi er in conjunction with theirs, one can obtain high con dence classi cations (if BLOCKS and our classi er agree) or suggest a new hypothesis (if the two disagree). Department of Computer and Information Science, New Jersey Institute of Technology, Newark, NJ 07102. Cold Spring Harbor Laboratory, 100 Bungtown Road, Cold Spring Harbor, NY 11724. Courant Institute of Mathematical Sciences, New York University, 251 Mercer Street, New York, NY 10012. Image Processing Section, Laboratory of Mathematical Biology, Division of Cancer Biology and Diagnosis, National Cancer Institute, National Institutes of Health, Frederick, MD 21701. Department of Computer and Information Science, New Jersey Institute of Technology, Newark, NJ 07102.