Predicting Epitopes for MHC Molecules

Major Histocompatibility Complex (MHC) plays a key role in immune response by presenting antigenic peptides, which are recognizable to T-cells. Identifying MHCbinding peptides is crucial to understand pathogenesis and develop corresponding vaccines. Direct identification of MHC-binding peptides by biological assays is laborious and expensive, because of the huge size (20) of potential combinations. Current computational methods are also not satisfactory: many of them failed to capture the features of nonconserved motifs, MHC molecule polymorphism, and non-specific binding to low-affinity peptides. Large datasets from National Institute of Health (NIH) public data repository and our own research data, with machine learning methods, provide a great opportunity for addressing this problem. Thus, we aim to develop an efficient and accurate method to identify and predict MHC-binding peptides and to further differentiate various MHC subtypes by their peptide-binding specificities. By selecting ~3000 peptide motifs as features, we built classifiers with Naive Bayes and SVM-based approaches. These classifiers achieved accuracy up to 99% on four most frequent human MHC subtypes— HLA-A01, HLA-A02, HLA-B27, and HLAB08—with 10 fold cross-validation. We applied these classifiers onto experimental data—a pool of potential binders. It has been found that up to 98% peptides are classified as binders and the classifier can be used to determine the specific subtypes. We also applied another approach, iterative statistical search in feature optimization. This iterative approach showed very promising results in our preliminary test. By applying this approach in feature generation, we achieved about 85% accuracy with less than 30 features. Further exploration of this approach will optimize feature space and improve algorithm efficiency. Introduction The immune system acts as a physical barrier against pathogen infections by activating B cells and T cells. B cells and T cells are special types of white blood cells that can recognize “nonself” cells, including pathogen-infected cells, and trigger immune response. In particular, cytotoxic T cell receptors can bind to major histocompatibility complex (MHC) that are antigen-specific receptors on the “nonself” cells and alert immune system to kill these infected cells (Smith-Garvin et al., 2009). Among three classes (I, II, and III) of MHC family, class I is by far the most well characterized subgroup, which we focused on in this study. Class I MHC proteins have a special structure with four antiparallel β-strands in the center region and two α-helices on one side. The two αhelices form a groove that contains six aminoacid binding pockets and can only accommodate short peptides of 8 to 11 amino acids—short fragments of antigens, in other words epitopes(Figure 1; Mester et al., 2011). Identifying and predicting MHC-binding epitopes are essential to understand the cause of diseases and develop corresponding vaccines (Lundegaard et al., 2007). Due to the size of the potential binding peptides (20 = 512 billion) for each MHC molecule (Liao and Arthur, 2011), empirical approach by biology experiments is laborious