Immune epitope database analysis resource (IEDB-AR)

T-cell epitope prediction

T-cell epitopes can be classified as either class I or class II epitopes. Class I epitopes are typically peptides presented by MHC class I (MHC-I) molecule and recognized by cytotoxic T lymphocytes (CTLs) that can kill infected cells. Class II epitopes are generally peptides that are presented by MHC class II (MHC-II) molecules and recognized by helper T lymphocytes (HTLs). Presentation of peptides-MHC-I complexes to T lymphocytes is a multistep process. In the IEDB-AR there are both tools that can predict the individual steps—such as cleavage of the polypeptide chain by the proteasome, transport of peptides into the endoplasmatic reticulum (ER) by the transporter associated with processing (TAP) and binding to MHC molecules—as well as methods that integrate these steps into one prediction.

All T-cell epitope prediction methods implemented in the IEDB-AR take protein sequences encoded by their single letter symbols as an input. In addition, the user can specify the MHC molecule for which they want to make predictions, as different MHC molecules have different binding specificities. Each input protein sequence is split into all possible peptides meeting the length preference of the selected MHC molecule. The output contains this list of peptides, each with a predicted score that reflects its likelihood of being a T-cell epitope. Depending on the chosen tool, this score reflects the disposition of a peptide to bind MHC, be processed from its parent protein or both.

MHC class I binding predictions

Three MHC class I peptide binding prediction methods are provided through the IEDB-AR: artificial neural network (ANN), stabilized matrix method (SMM) and average relative binding (ARB). Two of the methods, SMM and ARB, model binding specificity of an MHC molecule using position-specific scoring matrices (PSSM) (2–4). ANN, on the other hand, uses neural networks with diverse sequence encoding schemes to model-binding specificity (5).

The three prediction methods have been previously trained and evaluated (6) using 5-fold cross-validation. Since then, a new set of peptide-binding data for MHC class I molecules has become available, motivating the retraining and testing of all three MHC-I peptide prediction methods. For all MHC alleles with additional data available, new algorithms were trained that should exhibit improved prediction quality compared to the previous version, as the prediction accuracy is largely dependent on the amount of training data available. Also, the number of different MHC molecules for which predictions can be made was enhanced by eight new additions, including alleles from humans, mice, chimpanzees, macaques and gorillas. To accommodate historic comparisons, the IEDB-AR website now hosts two versions of the prediction methods, the previously published one and the updated implementations.

The availability of the new binding dataset gives us an opportunity to re-evaluate the 5-fold cross-validation used in ref. (6) to measure performance of these methods. Specifically, we made binding predictions for all peptides in the newly available data (called the ‘Independent Data Set’ from here on). The ability of each method to discriminate high affinity binders (IC50 < 500 nM) from those with lower affinity was measured using the Area under Receiver-Operating-Characteristic curves (AUC) values (7). Figure 1 shows AUC values published in ref. (6) compared to those attained with the Independent Data Set. This clearly indicates that the two performance evaluation methods give similar results overall. Importantly, there is no evidence for the previous cross-validation overestimating the performance of the prediction methods: out of 127 cases total, the independent dataset had the higher AUC in 64 cases, while the cross-validated results were higher in the remaining 63 cases. Confirming our earlier analysis (6), SMM and ANN perform better than ARB using a different performance evaluation measure. While both ARB and SMM use PSSM to model-binding specificity, they arrive at these matrices in a different manner. For each element in a PSSM, ARB bases the matrix entry on the average affinity of peptides sharing the corresponding residue. In contrast, all elements of PSSM for SMM are calculated simultaneously using a least squares fitting approach. This apparently leads to a better representation of binding specificity.

Open in a separate window

Figure 1.

Performances of prediction methods measured using either cross-validation or independent data set. Each data point represents an area under a receiver-operating-characteristic curve (AUC) for an MHC molecule (and its preferred peptide length). AUCs generated using cross-validation were taken from ref. (6), while those using independent data set are presented here for the first time. For independent data set, prediction methods trained on the old data used in ref. (6) were tested on the new, and thus independent, peptide-binding data that recently became available. In the case of cross-validation, however, an AUC value is an average of AUCs generated by training a prediction method on 4/5 of a data set and testing on the remaining one. Each data set for an MHC molecule had at least 50 binding affinity measurements.

MHC class I processing predictions

Additional events apart from peptide–MHC binding determine if a peptide is likely to be a T-cell epitope. These include proteasomal cleavage of the source protein and transport of peptides into the ER from the cytosol by the TAP transporter. The IEDB-AR provides an implementation of such predictions described in ref. (8), which can be combined with the updated MHC class I binding predictions described above. In the new release, alternative approaches were implemented: NetChop (9) uses novel sequence encoding methods to predict proteasomal cleavage, and NetCTL (10,11) combines these with predictions of MHC binding and TAP transport. Both tools were implemented with a common user interface, where a user can select to use either NetChop or NetCTL methods.

Compared to the highly selective peptide binding to MHC molecules, proteasomal cleavage and TAP transport have little influence on epitope generation. This makes predictions made by these tools less informative, but can be helpful to explain why some peptides that bind MHC molecules well are not naturally presented.

MHC class II binding predictions

Contrary to the closed binding groove of MHC class I molecules, the MHC class II groove that binds peptides is open at both ends and can thus accommodate peptides of variable length (12). This flexibility in binding makes MHC class II binding prediction substantially more difficult as compared to MHC class I binding prediction (6,13,14), leading to an overall lower prediction accuracy.

The MHC class II binding prediction service hosted at the IEDB-AR provides four prediction methods: ARB, SMM_align, Sturniolo's method (which is also the basis of TEPITOPE), and a consensus approach. ARB was the only method implemented in the last release, and is based on ARB matrices (2). SMM-align (15) is a matrix-based method with extensions incorporating flanking residues outside of binding grooves. Both methods can predict the IC50 values of peptides. The Sturniolo method implemented at IEDB-AR is a direct implementation of the matrices published in Sturniolo's original article (16), which provides scores on a nominal scale rather than IC50 values.

The consensus method is developed to combine the results of those three methods. We first scanned a random set of Swiss-Prot proteins and generated scores for 2 000 000 random peptides. The set of scores were then used as a reference to rank new predictions. The consensus approach then uses the median rank of the three methods as the final prediction score.

In order to independently evaluate the performance of the four methods, we generated an independent dataset by combining MHC class II binding data from several recent publications (17–20). The average AUCs for the four methods tested on the independent dataset are as follows: ARB-0.75, SMM-align-0.74, Sturniolo-0.72 and consensus-0.77. In addition to this new dataset, we have previously evaluated the performance of those methods with ∼20 000 peptide–MHC binding affinities generated by the Sette group (manuscript submitted for publication). The four methods implemented at the IEDB-AR are the top performing ones with average AUC values ranging from 0.71 to 0.76, which are consistent with the new independent evaluation.

B-cell epitope prediction

B-cell epitopes (or, antibody epitopes) can be defined as regions on molecules that are specifically bound by B-cell receptors (antibodies). They can include native protein structures, linear peptides, carbohydrates and other biological macromolecules. The majority of short linear peptides have the potential to be B-cell epitopes (21,22). This makes their prediction much harder than for T-cell epitopes, which require binding to the highly selective MHC molecules. The IEDB-AR includes seven methods for B-cell epitope prediction; five of them are previously published approaches for which no web implementation had been available. Two recently developed tools, BepiPred and DiscoTope, were implemented in the new release of IEDB-AR. BepiPred combines a PSSM with a propensity scale method to predict linear epitopes (23). DiscoTope is a method for discontinuous epitope prediction that uses protein 3D structural data as input (24). It is based on amino acid statistics, spatial information and surface accessibility for a set of discontinuous epitopes determined by X-ray crystallography of antibody/antigen protein complexes. It exhibits better performance for predicting residues of discontinuous epitopes than methods based solely on sequence information (24). Both BepiPred and DiscoTope were implemented on the IEDB-AR website with new features that include structural visualization of predicted discontinuous epitopes on protein 3D structures (Figure 2), based on the Jmol package (Jmol: an open-source Java viewer for chemical structures in 3D. http://www.jmol.org/).

Open in a separate window

Figure 2.

Screenshot of the ‘3D View’ page for the DiscoTope tool on IEDB-AR. The input for this prediction is the PDB structure file of AMA1 from Plasmodium falciparum (28) (PDB ID 1Z40). On the left is the 3D display of this structure, with all predicted epitope residues for chain A highlighted in yellow. The table on the right lists details of these predicted residues. Clicking on the ‘CPK’ button in the table highlights that specific residue in the 3D display, as seen on the top.

Epitope analysis

The IEDB-AR provides five analytical tools for: (i) calculating the epitope population coverage; (ii) assessing the degree of conservancy of an epitope; (iii) visualization and analysis of 3D structures of molecules containing epitopes; (iv) clustering of epitope sequences and (v) mapping of epitopes onto 3D protein structures. The last two tools have not been published before.

Clustering of epitope sequences

In evaluating groups of peptide epitopes for vaccine applications, it is important to remove redundant sequences or sequences with high similarity. The clustering tool was designed for this purpose. It groups epitopes into clusters based on sequence identity. A cluster is defined as a group of sequences, which have a sequence similarity greater than the minimum sequence identity threshold specified by a user.

Analysis of conserved epitopes

In an epitope-based vaccine setting, the use of conserved epitopes would be expected to provide broader protection across multiple strains, or even species, than epitopes derived from highly variable genome regions (25). The conservancy tool was designed to analyze the variability or conservation of epitopes. This tool calculates the degree of conservancy of a set of peptide/epitope sequences within a given set of protein sequences. The degree of conservation is defined as the fraction of protein sequences containing the peptide sequence at a given identity level.

The IEDB is funded by NIH contract HHSN26620040006C. Funding to pay the Open Access publication charges for this article was provided by NIH.

Conflict of interest statement. None declared.