MHCcluster, a method for functional clustering of MHC molecules
- First Online:
- Cite this article as:
- Thomsen, M., Lundegaard, C., Buus, S. et al. Immunogenetics (2013) 65: 655. doi:10.1007/s00251-013-0714-9
The identification of peptides binding to major histocompatibility complexes (MHC) is a critical step in the understanding of T cell immune responses. The human MHC genomic region (HLA) is extremely polymorphic comprising several thousand alleles, many encoding a distinct molecule. The potentially unique specificities remain experimentally uncharacterized for the vast majority of HLA molecules. Likewise, for nonhuman species, only a minor fraction of the known MHC molecules have been characterized. Here, we describe a tool, MHCcluster, to functionally cluster MHC molecules based on their predicted binding specificity. The method has a flexible web interface that allows the user to include any MHC of interest in the analysis. The output consists of a static heat map and graphical tree-based visualizations of the functional relationship between MHC variants and a dynamic TreeViewer interface where both the functional relationship and the individual binding specificities of MHC molecules are visualized. We demonstrate that conventional sequence-based clustering will fail to identify the functional relationship between molecules, when applied to MHC system, and only through the use of the predicted binding specificity can a correct clustering be found. Clustering of prevalent HLA-A and HLA-B alleles using MHCcluster confirms the presence of 12 major specificity groups (supertypes) some however with highly divergent specificities. Importantly, some HLA molecules are shown not to fit any supertype classification. Also, we use MHCcluster to show that chimpanzee MHC class I molecules have a reduced functional diversity compared to that of HLA class I molecules. MHCcluster is available at www.cbs.dtu.dk/services/MHCcluster-2.0.
KeywordsMHC HLA Binding motif Functional clustering MHC specificity Supertypes
Major histocompatibility complex (MHC) molecules play a central role in generating specific T cell-mediated immune responses. T cells scrutinize small peptide fragments, also called epitopes, presented in a complex with MHCs on the surface of most cells in the host. Cytotoxic T cells kill cells that present peptides of foreign or abnormal origin in a complex with MHC class I molecules. T helper cells, on the other hand, orchestrate the immune response by stimulating other immune cells and are stimulated by antigen-presenting cells that display peptides in complex with MHC class II molecules. The binding of peptides to MHC molecules is hence a prerequisite for T cell immunogenicity. Identifying which peptides will be presented in complex with a given MHC molecule is therefore of pivotal importance for our understanding of cellular immunity.
In general, MHC molecules are highly specific, binding only a minor fraction of the set of possible peptides (Yewdell and Bennink 1999; Rao et al. 2009). Moreover, the genomic region encoding MHC molecules is extremely polymorphic comprising several thousand alleles, many encoding a distinct molecule, making the peptide-binding repertoire of each individual unique. The most recent IMGT/human leukocyte antigen (HLA) database (Robinson and Marsh 2007) contains close to 5,000 HLA (the human version of MHC) class I protein sequences. This immense polymorphism of MHC molecules makes it a very costly endeavor to experimentally characterize the binding specificity of each molecule. Despite the significant experimental progress in high-throughput screening technologies (Harndahl et al. 2009, 2011), less that 80 HLA class I molecules have to this day been characterized with peptide binding data, allowing an accurate characterization of their binding motif (data taken from the IEDB; Vita et al. 2010). For nonhuman species including life-stock animals, the situation is even worse. Here, only a minor fraction of the known MHC molecules have been functionally characterized.
Due to the high selectivity of the MHC molecules, major efforts have been dedicated to characterize their binding specificity and several in silico methods have been developed allowing prediction of the binding affinity of peptides to MHC molecules (reviewed in Lundegaard et al. (2010) and Nielsen et al. (2010b)). These state-of-the-art methods make it possible to predict not only the MHC binding repertoire of any MHC molecule of interest (Hoof et al. 2009; Nielsen et al. 2010a; Karosiene et al. 2011), but also to characterize the subtle difference in the MHC specificities imposed by the allelic difference (Erup Larsen et al. 2011).
Not all MHC molecules are equally different in term of function, and several approaches have been described aiming to perform clustering of MHC molecules based on different measures of (functional) similarity (Sette and Sidney 1999; Doytchinova et al. 2004; Lund et al. 2004; Hertz and Yanover 2007). In 1999, Sette and Sidney proposed the HLA class I supertype concept, proposing that allelic variants within a supertype would share a large functional overlap, and nine such supertypes could cover the HLA class I functional space (Sette and Sidney 1999). Using data of known HLA class I ligands, Lund et al. refined this in 2004, and suggested the presence of three additional functional clusters (Lund et al. 2004). A limiting factor for the HLA clustering approach suggested by Lund et al. is the need for known ligands for the MHC molecules interest. In the original NetMHCpan publication (Nielsen et al. 2007), we therefore suggested the use of correlations between predicted binding affinities to perform functional clustering of HLA molecules and demonstrated that this approach could accurately reproduce the earlier proposed 12 HLA supertypes (and similar results have been shown for MHC class II; Nielsen et al. 2008). The functional clustering proposed by NetMHCpan demonstrated that many HLA molecules are characterized by specificities that are poorly characterized by the common 12 supertypes. This underlines an important shortcoming of the supertype concept.
Here, we describe a freely available web server, MHCcluster, implementing the functional clustering procedure described above to functionally cluster MHC molecules based on their predicted binding specificity. The method can be applied for both MHC classes I and II molecules for any MHC molecule with a known protein sequence covered by the NetMHCpan and NetMHCIIpan prediction methods (that is any MHC class I molecule and any HLA-DR class II molecule). The method has a flexible web interface that allows the user to include any MHC of interest in the analysis. The output from MHCcluster consists of a static heat map and graphical tree-based visualizations of the functional relationship between MHC variants and a dynamic TreeViewer interface where both the functional relationship and the individual binding specificities of MHC molecules are visualized.
We illustrate the power of the MHCcluster method in three distinct settings. First, we compare conventional sequence-based clustering to the functional clustering of MHCcluster and demonstrate situations where a sequence-based clustering, in contrast to MHCcluster, fails to identify the correct functional relationship between alleles. Next, we apply MHCcluster to the HLA-A and HLA.B. system investigating to what extent the common 12 HLA supertypes give an accurate representation of the functional diversity. Lastly, we use the method to confirm earlier findings (van Deutekom et al. 2011) demonstrating that chimpanzee MHC class I molecules have a reduced functional diversity compared to that of HLA class I molecules.
Materials and methods
The MHCcluster server allows the user to select a set of MHC alleles of interest including the option of uploading a set of full-length MHC I protein sequences and the server returns an unrooted tree and a heat map visualizing the functional similarities between the MHC molecules. The vehicles underlying the MHCcluster server are the NetMHCpan (version 2.7) and NetMHCIIpan (version 2.1) prediction methods. For each selected MHC allele, the MHCcluster method predicts its binding to a set of predefined natural peptides. Next, the similarity between any two MHC molecules is estimated from the correlation between the predictions of the union of the top 10 % strongest binding peptides for each allele (the threshold value can be altered by the user). This similarity is 1 if the two molecules have a perfect binding specificity overlap and −1 if the two molecules share no specificity overlap. Given this similarity, a distance between two molecules is defined as 1–similarity. The distance matrix is converted to an UPGMA (unweighted pair group method with arithmetic mean distance) tree. To estimate the significance of the MHC distance tree, a large set of distance trees is generated using the bootstrap method and a final tree is summarized in the form of a “greedy” consensus tree with corresponding branch bootstrap values.
As part of the new MHCcluster, a sequence logo for each allele is generated using the Seq2Logo service (Thomsen and Nielsen 2012). The logos are created from the top 1 % strongest binding peptides. For MHCII alleles, the logo is constructed from the predicted 9mer binding cores. The sequences used in the logos are clustered using the Hobohm 1 algorithm (Hobohm et al. 1992) using a similarity threshold of 63 % to remove redundancy, and pseudo counts are applied with a weight on prior of 200 (Altschul et al. 1997).
Prevalent HLA molecules
Prevalent HLA-A, B, and C molecules were identified for the European population from the dbMHC (NCBI Resource Coordinators 2013) using an allele frequency threshold of 0.5 %. The set of alleles defined as “HLA Prevalent and Characterized” consists of the HLA molecules characterized with more than 50 peptide binding data points and more than 0.5 % worldwide prevalence (as defined by the Allele Frequency Net database (Middleton et al. 2003), for populations characterized with more than 500 fully typed samples).
The MHCcluster server
The allele liston the right side of the tree lists all the alleles on the tree and presents the user with several functions. Firstly, it allows the user to locate the allele on the tree by showing the motif next to the node on the tree when the mouse cursor hovers over the allele name. Secondly, it allows the user to permanently add the motifs of selected alleles by selecting them on the list. These can be removed again by double clicking on either the motif or the allele, and thirdly, it shows the user which alleles have their motifs shown on the tree. The estimated accuracy value of the predicted sequence motif, a number between 0 and 1, is also provided next to the allele names.
The comparison bar below the tree shows the motifs of the selected alleles side by side to make it fast and easy for the user to compare the motifs of the alleles. The alleles can be selected in two ways either by selecting alleles on the allele list or by right clicking on the alleles on the tree and choosing a slot in the pop-up menu.
After everything is drawn, a few user interaction events are added to the tree elements. The nodes and labels receive a hover event, which shows a motif for the corresponding MHC allele next to the cursor, and they also receive a right-click event, which activates a menu where the user can add the corresponding MHC allele motif to a slot in the comparison bar. In addition, the labels and the permanent motifs (motifs from the selected alleles in the allele list) can now be dragged to any location in the box by left clicking, holding and dragging the element with the cursor. When the user left clicks on a motif or a label, the motif/label is brought to the front of the screen. This feature makes it possible for the user to arrange the overlapping images and labels as preferred. Finally, a copy of TreeViewer, including the tree file and corresponding sequence logos, can be downloaded as a zip file allowing the user to work and generate figures locally.
One thing that is important to note in this figure is the very close distance between the HLA-A*68:01 and HLA-A*68:02 molecules. On the sequence level, these two molecules share close to 99 % similarity differing by only five amino acid substitutions. However, when we look at the binding specificity as represented by the logos, it is apparent that these two molecules are very different in terms of function. The HLA-A*68:01 molecule has an A3 supertype specificity with a preference for basic amino acids at the C terminal whereas HLA-A*68:02 has a mixed A2/A26 specificity matching A26 at the N terminal and A2 at the C terminal.
Discussion and conclusion
Functional clustering of MHC molecules is a highly challenging task due to the vast polymorphism of the MHC genomic region and the very delicate relationship between subtle amino acid substitutions and dramatic variations in binding specificity. Here, we have illustrated how conventional sequence-based methods due to this subtle relationship in many cases will fail to produce a correct clustering and functional annotation for MHC molecules.
Given this observation, we argue that clustering and functional annotation for MHC molecules must be made based on information reflecting the peptide binding preference for each molecule and propose the MHCcluster method as an effective visual tool to compare functional similarities between large sets of MHC molecules.
The MHCcluster method estimates the functional relationship between two molecules from the overlap in prediction binding specificity, and returns a heat map and graphical tree-based visualizations of the functional relationship between MHC variants. Besides these conventional representations of the functional map of the MHC molecules of interest, the MHCcluster method provides a dynamic TreeViewer interface where both the functional relationship and the individual binding specificities of MHC molecules are visualized (the later in terms of sequence logos). This TreeViewer is a unique feature of the MHCcluster server that allows in a highly intuitive manner for functional interpretations of the MHC map proposed by the MHCcluster method. Earlier methods have been proposed for functional clustering of MHC molecules (Sette and Sidney 1999; Doytchinova et al. 2004; Lund et al. 2004; Hertz and Yanover 2007), and for the browsing of predicted binding motifs of MHC molecules (Rapin et al. 2008; Rapin et al. 2010). But to the best of our knowledge, no method has combined these two approaches allowing for the direct functional mapping of MHC molecules in terms of both clustering and visualization of binding motifs.
Using the MHCcluster method, we confirm the existence of the 12 HLA supertypes earlier proposed to characterize the specificity space of HLA-A and HLA-B molecules. However, the analysis also clearly revealed that not all HLA molecules fit equally well into a supertype classification scheme, and that some supertype “clusters” consist of molecules with highly divergent specificities. Finally, moving to nonhuman primates, we compare the MHC class I specificity space of human and chimpanzee using the MHCcluster method and demonstrate that the Patr A locus has significantly reduced functional diversity compared to the human HLA-A locus manifested by the almost complete loss of HLA-A2 and HLA-A26 supertype specificities.
In this work, we have focused on demonstrating the use of the MHCcluster method to analyze functional diversities of MHC class I molecules. The method is equally well suited for making functional analysis and clustering for MHC class II molecules, and the server does include an option to analyze MHC class II molecules. However, as no pan-specific prediction algorithm currently exists to allow for the prediction of peptide binding to any MHC class II molecule, the analysis is limited to the HLA-DR loci molecules covered by the NetMHCIIpan method (Nielsen et al. 2008, 2010a).
In conclusion, we have demonstrated that the MHCcluster method can be used as an effective visual tool to compare functional similarities between MHC molecules. The method is highly flexible and allows the user to analyze any MHC variant of interest. MHCcluster is available at www.cbs.dtu.dk/services/MHCcluster-2.0.
Even though we here have limited the applications of the MHCcluster method to the comparison of functional similarities between large sets of MHC molecule, many other types of important questions that can be addressed by the method. Some could be as a guide to help researchers interpret immunological phenotypic similarities between patients using information about HLA types (i.e., understand for instance why patients with no overlap in HLA types can share an overlap in T cell epitopes), as a guide to see where a specific allele (maybe present at a high frequency in a particular cohort) fits in to the specificity space covered by the common MHCs.
MN is researcher at the Argentinean National Research Council (CONICET). This work was supported by NIH grant HHSN272200900045C.