Prediction of HLA-DRB1*0401 binding peptides using support vector machine.

In recent years, many machine learning methods have been developed to predict HLA binding peptides. However, because only limited types of descriptors characterising the protein features are included in these approaches, these methods have poor prediction accuracy. In this study, we applied support vector machine methods to predict the peptides that bind to the major histocompatibility complexes Class II molecule HLA-DRBl*0401 using six sets of molecular descriptors characterising the primary structures of the peptides. We found that some feature groups provided good prediction accuracies and the overall accuracies were greater than 95% and some feature groups had poor accuracies of only 50%. The performance was improved significantly by additional feature selection and the overall accuracies from each group or combination of descriptors were greater than 90%. Of note, the inclusion of necessary informative and discriminative descriptors improved the prediction accuracies.