Computational prediction methods for peptide binding to human leukocyte antigen (HLA) molecules have played an instrumental role in the development of epitope-based vaccines. These methods are based on experimentally verified peptides. However, the available peptide data continue increasing and contain significant biases. In this study, we report the feedback effect of peptide data on a frequently used matrix-based prediction method. We implemented the weighted and unweighted models of this method and evaluated the relative performance of the two models on several benchmark datasets. Improvements on both models were obtained by optimizing the components of a training dataset based on the effect of peptide data on the performance of prediction models. Moreover, the variation of the relative performance of the weighted and unweighted models with the evaluated data indicated that the increased number of binding peptides required the modification of the predictive engine. Our results suggest that prediction methods for HLA-binding peptides should be updated as HLA-peptide-binding knowledge increases.
Keywords: AUC; HLA; HLA-binding peptides; Improvement; MHC; NMN; Prediction; ROC; human leukocyte antigen; major histocompatibility complex; non-motif-containing non-binder; receiver operating characteristic; the area under the ROC curve.
Copyright © 2013 Elsevier B.V. All rights reserved.