A multiclass SVM classifier with teaching learning based feature subset selection for enzyme subclass classification

Abstract

The functional characterization of proteins aims to understand life at macroscopic and microscopic levels, thus having wide and extensive applications in biological and pharmaceutical research. The classical methods used for the functional characterization of proteins rely on the sequence similarity approach i.e. identifying a similar protein whose functions are known. However, the performance of these methods reduces in the event of the sequences exhibiting less similarity. Hence, to tide over this impediment, many computational function prediction methods that do not use the sequence similarity approach were developed. This paper focuses on a Support Vector Machine (SVM) based method which classifies a protein at three levels. At the first level, a protein is identified and earmarked as an enzyme or non-enzyme. In the case of it being an enzyme, its functional class, and subclass are predicted in one step by taking 7 EC classes with 63 of their subclasses. Each protein is initially represented by its 32 physicochemical properties. Then a multiclass SVM (MSVM) classifier, that solves an n class classification problem with log2n binary classifiers, along with a Modified Teaching Learning Based Optimization method (MTLBO) to identify the significant features for classification, is designed to predict the functional class, and subclass of an enzyme in one step. The recall of the proposed MSVM-MTLBO using only 25 features ranges from 92.97% to 98.14% for class label prediction and 60% to 98.25% for subclass label prediction, which is significantly better than the existing methods for class and subclass classification. Moreover, the proposed method predicts the subclass labels with less number of binary classifiers than a standard multiclass SVM.