Protein Sequences Classification Based on String Weighting Scheme

Motivation: We present a new technique to recognize remote protein homologies relies on combining probabilistic modeling and supervised learning in high-dimensional feature spaces. The main novelty of our technique is the method of constructing feature vectors using Hidden Markov Model and the combination of this representation with a classifier capable of learning in very sparse high-dimensional spaces. Each feature vector records the sensitivity of each protein domain to a previously learned set of sub-sequences (strings). Unlike other previous methods, our method takes in consideration the conserved and non-conserved regions within the protein sequences of interest. The system subsequently utilizes Support Vector Machines (SVM) classifiers to learn the boundaries between structural protein classes. Results: Experiments show that this method, which we call the String Weighting Scheme-SVM (SWS-SVM) method, significantly improves on previous methods for the classification of protein domains based on remote homologies. Our method is then compared to five existing homology detection methods. Contact: [email protected], http://www.kp.fsksm.utm.my/Main/research/geno me/zakiall1.html Supplementary Information: The datasets, models, and results files used in this paper are available online: (http://www.kp.fsksm.utm.my/Main/research/geno me/paper1.html).