Developing SVM Classifier for GPI-(Like)-Anchored Proteins Prediction

Glycosylphosphatidylinositol (GPI) anchoring is a means for COOH-terminus post-translational modification of proteins. Although experimental result [1] shows importance of hydrophobicity of COOHterminus of GPI-anchored proteins, Fankhauser et al. reported that accuracy of the GPI-(like)anchored protein prediction by solely using hydrophobicity scale is not better as being expected [2]. Could we improve the accuracy of identifying GPI-(like)-anchored proteins only using hydrophobicity property? For this purpose, we trained a binary classifier by using Support Vector Machine (SVM) [3] and optimizied its input parameters utilizing particle swarm optimization (PSO) algorithm [4]. Not merely does the classifier show high accuracy of 96% under 5-fold cross validation test, but also the area under Receiver Operating Characteristic curve (ROC) so called AUC as a good summary of the performance of the classifier reaches to 0.97. In virtue of being based on SVM, computational efficiency and remarkable generalization ability of our classifier will be helpful for protein annotation in whole genomic-wide.