Prediction of eukaryotic protein subcellular location using a novel feature extraction method and support vector machine

The rapidly increasing number of sequences entering into the genome databank has created the need for fully automated methods to analyze them. Knowing the cellular location of a protein is a key step towards understanding its function. The development in statistical prediction of protein attributes generally consists of two cores: one is to construct a training dataset and the other is to formulate a predictive algorithm. The latter can be further separated into two subcores: one is how to give a mathematical expression to effectively represent a protein and the other is how to find a powerful algorithm to accurately perform the prediction. To predict the subcellular location of eukaryotic protein, a systematic prediction approach comprised of a novel feature extraction method, an idea of combining this feature extraction method with support vector machine (SVM) algorithm, and 'one-versus-rest' and 'all-versus-all' strategies have been proposed in this paper. Consequently, the total predictive accuracies reach 95.5% for four locations. Compared with existing methods, this new approach provides better predictive performance. For example, it is 13.5%, 5.1% higher than Yuan's and Hua's methods respectively. These results demonstrate the applicability of this new method and concept and possible improvement of prediction for the protein subcellular location. It is anticipated that the current approach may also have a series of impacts on the prediction of other protein features.