Comparison of SVM-Based Methods for Remote Homology Detection

Abstract

Remote homology detection for protein sequences is one of the important and well-studied problems in Bioinformatics. Many algorithms have been developed for this purpose. The Smith-Waterman (SW) dynamic programming algorithm was developed in early 1980’s [8], and is still used widely today. In 1990’s, many methods were developed based on profiles [1] and hidden Markov models [2, 4]. In 2000’s, methods using SVMs (support vector machines) were developed such as the SVM-Fisher method [3]. Recently, Liao and Noble proposed the SVM-pairwise method [5], which uses a vector of pairwise similarities with all proteins in the training set. Quite recently, we proposed a new SVM based method (SVM-SW), which uses the SW algorithm as a kernel function [7]. Though we do not yet succeed to prove that the SW score is always a valid kernel, SVM-SW worked successfully in all cases we tested. In this poster abstract, we briefly show the results of comparison of algorithms for remote homology detection using the SCOP database [6].