A Novel Numerical Feature Extraction Method for Protein Subcellular Localization

Although there are many subcellular localization methods currently, the prediction accuracy of them is not good enough. The main reason may be that when converting sequence information into numerical information, these methods lose important information. In order to get more global and local information in the protein sequences, we propose a novel numerical feature extraction method for representation of protein sequences which includes three parts of features that are amino acid composition, compression tripeptide composition and local frequency domain values. Then, we use support vector machine and the nearest neighbor algorithm to predict subcellular localization of two benchmark data sets and compare the prediction results and evaluation index values with other methods. Comparison results prove that our method can effectively extract information of protein sequence and improves the prediction accuracy of subcellular localization.