Using Fuzzy Support Vector Machine Network to Predict Low Homology Protein Structural Classes

Abstract

Prediction of protein structural classes for low homology proteins is a challenging research task in bioinformatics. A dual-layer fuzzy support vector machine (FSVM) network approach is proposed to predict protein structural classes. A protein sample can be represented by nine representation feature vectors: pair couple amino acid (210-D) and eight pseudo amino acid composition vectoers (PseAAC). Eight physicochemical properties of amino acids extracted from AAIndex databank are used to calculate low frequencies of power spectrum density of sequence-order correlation in protein sequence. In the first layer of FSVM network, nine FSVM classifiers are established, which are trained by different protein feature vectors, respectively. The outputs of the first layer are reclassified by FSVM classifier in 2nd layer of the network. The performance of proposed method is validated by low homology (average 25%) dataset covering 1673 proteins. The promising results indicate that the new method may become a useful tool for predicting not only the structural classification of proteins but also their other attributes.