A Novel Approach of Protein Secondary Structure Prediction by SVM Using PSSM Combined by Sequence Features

Abstract

Knowledge of protein secondary structure is a useful step toward prediction of the 3D structure of a particular protein. In this paper, a support vector machine (SVM) based method used for the prediction of secondary structure is introduced in details. Protein sequence data is in a hybrid representation combining the Position-specific Scoring Matrix (PSSM), the Hydrophobicity Sequence Feature (HSF), and the Structural Sequence Feature (SSF). Protein sequences are obtained from CB513 dataset, corresponding PSSM profiles are obtained from PSI-BLAST Program and sequence features are computed based on amino acid scales offered by Expasy website (http://web.expasy.org/protscale/). Basically, PSSM profiles are used as input data to the SVM-PSSM classifier of the secondary structure prediction. Furthermore, to construct more accurate classifiers, more than 40 SFs (sequence features) are examined as accessional input vector to SVM-PSSM classifier for feature selection. The most accurate classifier in this study is constructed using a combination of PSSM and few relevant sequence features. The experimental results show that relevant sequence features extracted from Hydrophobicity index and Structural conformational parameters can improve the SVM-PSSM classifier for the prediction of protein secondary structure elements. Our proposed final SVM-PSSM-SF method achieved an overall accuracy of 78%.