The identification of MHC class II restricted peptide epitopes is an important goal in immunological research. A number of computational tools have been developed for this purpose, but there is a lack of large-scale systematic evaluation of their performance. Herein, we used a comprehensive dataset consisting of more than 10,000 previously unpublished MHC-peptide binding affinities, 29 peptide/MHC crystal structures, and 664 peptides experimentally tested for CD4+ T cell responses to systematically evaluate the performances of publicly available MHC class II binding prediction tools. While in selected instances the best tools were associated with AUC values up to 0.86, in general, class II predictions did not perform as well as historically noted for class I predictions. It appears that the ability of MHC class II molecules to bind variable length peptides, which requires the correct assignment of peptide binding cores, is a critical factor limiting the performance of existing prediction tools. To improve performance, we implemented a consensus prediction approach that combines methods with top performances. We show that this consensus approach achieved best overall performance. Finally, we make the large datasets used publicly available as a benchmark to facilitate further development of MHC class II binding peptide prediction methods.
A critical step in developing immune response against pathogens is the recognition of antigenic peptides presented by MHC class II molecules. Since experiments for MHC class II binding peptide identification are expensive and time consuming, computational tools have been developed as fast alternatives but with inferior performance. Here, we carried out a large-scale systematic evaluation of existing prediction tools with the aim of establishing a benchmark for performance comparison and to identify directions that can further improve prediction performance. We provide an unbiased ranking of the performance of publicly available MHC class II prediction tools and demonstrate that the MHC class II prediction tools did not perform as well as the MHC class I tools. In addition, we show that the size of training data and the correct identification of the binding core are the two factors limiting the performance of existing tools. Finally, we make available to the immunology community a large dataset to facilitate the evaluation and development of MHC class II binding prediction tools.
Citation: Wang P, Sidney J, Dow C, Mothé B, Sette A, Peters B (2008) A Systematic Assessment of MHC Class II Peptide Binding Predictions and Evaluation of a Consensus Approach. PLoS Comput Biol 4(4): e1000048. https://doi.org/10.1371/journal.pcbi.1000048
Editor: Gary Stormo, Washington University, United States of America
Received: November 8, 2007; Accepted: February 29, 2008; Published: April 4, 2008
Copyright: © 2008 Wang et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This work was supported by the National Institutes of Health contract HHSN26620040006C and PSWRCE grant U54 AI065359 to BP. Kirin number 956.
Competing interests: The authors have declared that no competing interests exist.
The activation of CD4+ helper T cells is essential for the development of adaptive immunity against pathogens –. A critical step in CD4+ T cell activation is the recognition of epitopes presented by MHC class II molecules . MHC class II molecules are heterodimers expressed on the surface of professional antigen presenting cells that bind peptide fragments derived from protein antigens . X-ray crystallographic studies demonstrated that the MHC class II epitope binding site consists of a groove and several pockets provided by a β-sheet and two α-helices ,. Unlike class I, the class II binding groove is open at both ends. As a result, peptides binding to class II molecules tend to be of variable length, but typically between 13 and 25 residues.
A hallmark of the MHC class II binding peptide groove is that there are four major pockets. These pockets accommodate side-chains of residues 1, 4, 6, and 9 of a 9-mer core region of the binding peptide. This core region interaction largely determines binding affinity and specificity . In addition, peptide residues immediately flanking the core region have been indicated to make contact with the MHC molecule outside of the binding groove, and to contribute to MHC-peptide interaction .
MHC class II molecules are highly polymorphic, and this polymorphism largely corresponds with differences along the peptide binding groove. However, the binding motifs derived for MHC class II molecules are highly degenerate, and many promiscuous peptides have been identified that can bind multiple MHC class II molecules . Promiscuous peptides are a prime target for vaccine and immunotherapy and computational tools have been developed to facilitate systematic scanning for promiscuous peptides .
Computational prediction of MHC class II epitopes is of important theoretical and practical value, as experimental identification is costly and time consuming ,. The basis of a successful computational prediction is a sufficiently large set of high quality training data. There are several databases hosting MHC epitope related data such as SYFPEITHI , MHCBN , Antijen , FIMM , HLA Ligand  and our own project, the Immune Epitope Database (IEDB) ,. Information from those databases is, for the most part, extracted from the literature. These databases typically combine data from different sources and different experimental approaches, which can complicate the generation of consistent training and evaluation datasets.
The establishment of numerous MHC class II epitope databases has facilitated the development of a large number of algorithms aimed at predicting peptide binding to MHC molecules. Early works focused on finding peptide patterns and deriving motifs for MHC molecules –. With the accumulation of epitope data, more sophisticated algorithms were developed. Several methods have derived scoring matrices that evaluate the contribution to binding of different residues in a peptide based on quantitative binding data (ARB , SMM-align ). Others base similar scoring matrices on multiple peptide alignments (RANKPEP ,) or domain expert knowledge (SYFPEITHI method ). By combining the similarities of key residues forming the pockets of the binding groove with quantitative matrices derived from experiments, the TEPITOPE  algorithm can predict binding to MHC alleles for which no binding affinities were determined. Other machine learning algorithms that have been applied include hidden Markov models , evolutionary algorithms  and linear programming . The MHC class II binding prediction problem has also been modeled with a distance function in a recently developed method PepDist . In addition to the previously listed models that are directly interpretable, “black box” approaches, such as support vector machines  and artificial neural networks –, have also been applied to MHC class II binding prediction with success.
Despite the large number of available prediction methods, computational prediction of MHC class II epitopes remains a challenging problem. It has been suggested that the prediction performance of class II algorithms is systematically inferior to that of MHC class I epitope prediction methods . To assess the current state of the MHC class II binding predictions, we have here sought to establish a systematic and quantitative benchmark similar to our previous effort for MHC class I molecules . We present a large dataset of unpublished MHC class II-peptide binding affinities that were experimentally determined under uniform conditions. We then proceed to evaluate a set of nine publicly available MHC class II prediction methods using this dataset and systematically compared their performance. Finally, we analyzed the ability of current methods to identify the binding cores of peptides and to predict T-cell responses from peptide sequences.
Overview of MHC Class II Epitope Affinity Dataset and MHC Class II Binding Prediction Methods
We assembled a dataset of peptide binding affinities for various MHC class II molecules experimentally measured in our group (see Materials and Methods for details). Table 1 gives an overview of the dataset, encompassing a total of 10,017 experimentally determined peptide MHC II binding affinities. These data span a total of 16 human and mouse MHC class II types. The number of unique MHC-peptide affinities measured per type varies greatly, from 3,882 for HLA DRB1*0101, to only 39 for H-2-IEd. Compared to datasets publicly available on the IEDB and other MHC class II epitope databases, our new dataset expands the number of measured peptide-MHC class II interactions significantly for a large number of MHC class II molecules. For example, the number of peptides with known IC50 values for HLA DRB1*0101 was more than tripled with the addition of our new dataset.
The MHC class II binding prediction tools evaluated in this study are listed in Table 2. We included as many prediction methods as possible provided that they (1) can perform predictions for MHC class II types in our dataset; (2) were publicly available; and (3) did not specifically disallow the use of automated prediction retrieval scripts. A total of nine methods matched these criteria. A more detailed description of tested methods is provided in the Materials and Methods section.
Performance Evaluation of Publicly Available Prediction Tools
The binding predictions for peptides in our affinity dataset were extracted from the MHC class II binding prediction tools with custom scripts (see Materials and Methods for details). From the experimental data, peptides were classified into binders (IC50<1000 nM) and nonbinders (IC50≥1000 nM) based on measured affinities. The performance of the prediction methods was then measured by ROC curves (see Materials and Methods for details). Since the new dataset was never published before, it is equivalent to a blind test. An important exception is the ARB method. Since it was developed at IEDB and was constantly updating with new data, its performance was instead evaluated via 10-fold cross validation.
Table 3 shows the prediction performance of the various methods in terms of area under ROC curve (AUC). The ROC curves for tested methods were also plotted in Figure 1 using HLA DRB1*0101 as an example. SVMHC was not evaluated separately, since it implements the same TEPITOPE matrices utilized by PROPRED. When overall performance is assessed by averaging across all available MHC class II molecules SMM-align and PROPRED are associated with the best AUC value (0.73). The ARB method has the third best performance with an average of AUC of 0.71. When performance on individual MHC class II molecule is examined, the ARB, PROPRED or SMM-align perform best for all but the H-2 IEd molecule, for which RANKPEP gives the best result.
Prediction results for eight methods for HLA DRB1*0101 are shown in the ROC curve. The curves were generated by plotting the true positive rate (y-axis) against the false positive rate (x-axis). The AUC values for corresponding ROC curves were shown in parentheses.
Since we restrict our testing to publicly available tools, it is important to point out that the methods were trained on different datasets (Table 2). Some databases such as MHCPEP only include positive binding data and the lack of nonbinders would be expected to negatively impact some methods that require negative training data. Two of the top performing methods (SMM-align and ARB) utilize the IEDB dataset, confirming that the size of the training set maybe an important factor contributing to better performance. PROPRED is among the most accurate MHC class II binding prediction methods, despite being based on the TEPITOPE method developed over eight years ago. The good predictive power of TEPITOPE demonstrates the validity of its approach, based on pocket information derived from MHC class II structures and quantitative peptide binding profiles.
The cutoff of 1000 nM to classify peptides into binders and non-binders was chosen following an expert immunologist's recommendation for an immunologically relevant threshold, but it is still somewhat arbitrary. To further our analysis in a systematic fashion, we varied the cutoff from 50 nM to 5000 nM. The changes in cutoffs enable us to evaluate performances of binding prediction to identify peptides with different affinities. A cutoff of 50 nM focuses on identifying strong binders, while a cutoff of 5000 nM will identify all including very weak binders. The results of the evaluation using different cutoffs are shown in Dataset S1. For MHC molecules with large number of binding affinities (such as HLA DRB1*0101), varying the cutoff has little impact on the AUC values. For datasets with smaller number of binding affinities (such as H-2-IEd), the change in AUC values is more significant. Despite the variations in AUC values introduced by different cutoffs, the relative performance of different methods remains largely the same, suggesting our conclusion of different methods' performance is not strongly dependent on the cutoff used to decide binders.
Existing Tools Lack Consistency in Identifying the 9-Mer Core Interacting with the Binding Groove of the MHC Class II Molecule
A key difference between MHC class I and class II molecules is that the binding groove of class II molecules is open at both ends ,. As a result, the length of peptides binding class II molecules can vary considerably, and typically range between 13 and 25 amino acids long. Thus, a requisite for all MHC class II binding prediction approaches is the capacity to identify within longer sequences the correct 9-mer core residues that mediate the binding interaction . All methods included in our study explicitly predict cores when they predict MHC class II binding peptides. They either predict binders as 9-mer peptides, or clearly state in prediction the location of the predicted 9-mer core.
We next analyzed whether the various class II prediction tools can accurately identify the 9-mer cores of a binding peptide. We extracted MHC-peptide complex structures from the Research Collaboratory for Structural Bioinformatics (RCSB) Protein Data Bank (PDB). A total of 29 structures associated with 14 different MHC class II molecules were identified (Table 4). For each method we compared the predicted cores with the true cores extracted from crystal structures. The results are shown in Table 5. The PROPRED method based on TEPITOPE was associated with the best performance, with all the predicted cores matching the cores determined by PDB structures. This is in good agreement with the fact that TEPITOPE is directly based on experimental assays. SYFPEITHI is the second best method with an accuracy of 0.9. However, this result should be interpreted with caution since seven of the nine correctly predicted peptides are documented in the SYFPEITHI database. Apart from PROPRED and SYFPEITHI, the method most effective in predicting binding affinity (SMM-align) is also the method with highest accuracy in predicting 9-mers cores, with an accuracy of 0.625. RANKPEP and SVRMHC come next with accuracies about 0.55. The remaining three methods had limited success (21–25%), although they still perform above random prediction (the probability to randomly guess the right core for a 15-mer peptide is 1 out of 7 or 0.143). Overall, these data suggest that correctly aligned cores contribute to the superior performance of PROPRED and SMM-align, and that there is substantial room to improve the quality of the core predictions of other methods.
Predicting T Cell Activation from Peptide Sequences with Existing MHC Class II Epitope Prediction Tools
The ultimate goal of MHC binding peptide prediction is to identify epitopes that activate T cells. Recognition of a peptide bound to an MHC molecule by a T cell receptors is the critical step in this activation, and binding of peptide to the MHC molecule is obviously a necessary requirement . In a separate study carried out in our lab, a set of 664 peptides overlapping the LCMV proteome were tested for their abilities to promote H-2 IAb specific IFN gamma production from CD4+ T cells in splenocytes from previously LCMV infected mice (manuscript in preparation). These peptides provided an ideal test set to evaluate MHC class II binding predictions as a tool to identify peptides that trigger an immune response.
For each of the 664 peptides, we obtained H-2 IAb binding predictions from the five methods in our study that cover H-2 IAb following exactly the same procedures as predictions of simple binding. We then evaluated the methods' performance in predicting which peptides triggered an immune response. The ROC curves quantifying the performance of each method are shown in Figure 2. The Consensus method is the best performing methods with AUC of 0.89±0.05. ARB is the second best performing method with an AUC of 0.85±0.05. SMM-align and RANKPEP have similar performance with AUC about 0.76±0.08 and 0.78±0.07, respectively. MHCPRED and MHC2PRED do not perform as well, with AUC values of about 0.67±0.12 and 0.36±0.1 (standard deviations calculated by bootstrapping with replacement). Except MHCPRED, every other method's performance in this evaluation compared favourably to that in predicting peptide binding. Overall, the ranking of prediction performances is well in concert with that for predicting peptide binding, specifically when taking into account the high standard deviations of AUC values. These large standard deviations are due to the limited number of positive datapoints in the set utilized.
ROC curves are generated from the predictions made by five MHC class II peptide binding prediction methods on the LCMV CD4+ T cell activation data. The AUC value for each method is shown in parentheses.
To further analyze the performance of the T cell activation prediction, we classified peptides into predicted binders and non-binders. Since different methods produce scores on different scales, we adopt a rank based classification in that we classify the top 10% highest scoring peptides as binders. We then calculated sensitivity and positive predictive value (PPV) for each method (Table 6). These two measures were chosen since we are primarily interested in identifying T cell activating peptides while minimizing the number of false positive predictions. In our system, sensitivity is the percentage of peptides activating T cells predicted to be binders and PPV is the percentage of predicted binders that actually activate T cells. The data in Table 6 show that results of these two measures are largely consistent with the AUC results. Methods with high AUCs tend to have high PPV and sensitivity. Only the consensus method has sensitivity above 50%, indicating that 5 out of 6 methods missed more than half of the T cell activating peptides when top 10% ranked peptides are classified as binders. In addition, the consensus method also had the highest PPV value of 9.4%, making it again the best prediction method by this measure. The overall low PPV values are expected, as many peptides that are capable of binding MHC are not recognized by T cells following a natural infection, due to other factors such as peptide processing and the available T cell repertoire.
Improving MHC Class II Binding Prediction with a Consensus Approach
Our evaluation of prediction performance suggests that in all cases there is clearly room for improvement, and that no single method is dominantly better than all others. Motivated by the success of a consensus prediction approach to map MHC class I epitopes in vaccinia virus , we implemented the same approach for MHC class II binding predictions. This consensus approach is based on calculating the median rank of the top three predictive methods for each MHC class II molecule (see materials and methods for details).
The consensus prediction performance is shown in the last column of Table 3 for the 14 MHC alleles for which three or more predictions were available. For ten of these fourteen alleles, the consensus method gives similar or higher performance than the best individual method. For each of the remaining four datasets, a single prediction performs better (i.e., ARB for DRB1*1302, SMM-align for DRB1*1101 and DRB1*1501, and PROPRED for DRB1*0405). In terms of overall performance across all molecules in our dataset, the consensus method outperforms all individual MHC class II prediction methods.
The MHC-peptide affinity, MHC-peptide structure and T cell activation datasets are available as supplemental material at http://mhcbindingpredictions.immuneepitope.org/MHCII. These data are presented in this paper for immediate access by the immunology and bioinformatics community. We are currently in the process of depositing these data into the IEDB, making them available through the epitope informatics framework of the IEDB.
In this study we have presented a comprehensive dataset for the systematic evaluation of MHC class II peptide binding prediction methods. This dataset consists of three components. The first component is a large set of 10,017 quantitative peptide-binding affinities for 16 MHC class II types that significantly expands the amount of publicly available data. These data were generated under identical experimental conditions and comprise affinities for binders as well as non-binders. The second component is a set of non-redundant structures of MHC class II molecules complexed with peptide ligands compiled from the PDB. This set of structures provided a “gold standard” for evaluating the ability of prediction methods to locate the 9-mer core of epitopes. The last component is a set of 664 peptides that has been tested experimentally to determine their ability to stimulate CD4+ T cells from widely utilized C57BL/6 (H-2b) strain of laboratory mice. Together, these datasets serve as a benchmark set to facilitate the development and testing of algorithms for predicting peptide binding to MHC as well as T-cell responses.
Several previous studies have compared the performances of various MHC class II binding prediction methods –. The Borras-Cuesta study  from 2000 only had a limited number of peptides, alleles and methods to compare. The two recent studies were published after we finished our initial comparison. Gowthaman et al  compared six commonly used method with data spanning seven MHC class II alleles. However, their evaluation dataset comprised only 179 peptides, limiting the significance of their results. Rajapakse et al  compared their multi-objective evolutionary algorithms (MOEA) with five other algorithm using two datasets. The first dataset consisted of 1 training and 10 testing datasets on HLA DRB1*0401 assembled from different sources. The second dataset was extracted from the IEDB and comprised more than 5,000 peptides covering 16 MHC class II alleles. We couldn't include MOEA in our comparison since it is not publicly available at the moment. Despite the difference in datasets used in comparison, their conclusion is consistent with ours in that SMM-align, TEPITOPE and ARB are the better performing methods.
We have carried out a comprehensive unbiased evaluation of existing MHC class II epitope prediction algorithms using these datasets. Except binding prediction for ARB, all the other MHC class II prediction algorithms are evaluated in a completely blinded fashion. In our analysis, the better performing methods proved to be those that are based on quantitative matrices extended by method specific features. For example, SMM-align is the only method tested that considers the contribution of residues outside of the binding groove, and TEPITOPE is the only method whose matrices are based on experiments aimed to determine individual amino acid's contribution to binding. Merely using quantitative matrices alone is not sufficient to ensure good performance, since pure position specific scoring matrix based methods such as RANKPEP and SYFPEITHI do not perform as well.
One potential reason for the differential performance of various methods is the likely different number of data points utilized by the various methods in the training stage. In this respect, we anticipate that the datasets described herein, and now made publicly available, could be utilized to retrain several of the methods and further increase their performance.
Despite the large number of existing MHC class II epitope prediction methods, the best performance is generally not as good as that for MHC class I methods. Indeed, it is notable that the majority of methods examined in the present study have also been employed to make predictions for MHC class I peptide binding, and almost invariably their performance is appreciably better in the context of class I . For example, when SMM  was applied to predict epitopes for several MHC class I molecules, it achieved an average AUC of 0.874, which is substantially higher than that for class II (0.783).
In an attempt to identify what limits the performance of MHC class II binding prediction, we tested the ability of prediction methods to identify the 9-mer peptide cores revealed in crystal structures of MHC-peptide complexes. Except for PROPRED and SYFPEITHI, the methods examined performed poorly, suggesting that difficulties in identification the correct binding core contribute to the inferior performance of class II binding prediction. It is noteworthy that the two methods with the best core predictions do not take all positions of a peptide into account when making binding predictions, but rather focus on anchor positions in the peptide. This may explain why especially the ARB method performs much poorer in the core identification rather than the binding predictions: It treats all positions in the peptide identically and relies on automated peptide alignments to derive an overall peptide profile. While this inclusion of weakly interacting positions can be an advantage to predict overall peptide binding, it may lower the accuracy when picking the correct core.
In an attempt to improve upon the prediction performance realized by individual prediction tools, we implemented a consensus approach for class II binding predictions. The consensus approach was found to clearly outperform each individual prediction approach when measured over the entire dataset, and provided the best predictions for 10 out of 14 molecules. This shows that the consensus approach is just as useful for MHC class II peptide binding prediction as its recent successful application for MHC class I molecules . In a smaller study addressing 3 different prediction methods in the context of a single DR type, Mallios previously came to a similar conclusion .
Other types of meta approaches have been successfully applied to MHC binding prediction. For example, Mallios  has used an iterative stepwise discriminant analysis meta-algorithm to successfully classify binders and non-binders for HLA-DR1. Stern and co-workers effectively used a two-dimensional dot plot to combine the prediction results of SYFPEITHI and TEPITOPE . Finally, Trost et al  have reported achieving greater accuracy in MHC class I binding predictions by combing results from multiple prediction tools. Compared to these methods, our median rank approach does not depend on the absolute values of scores and it has exceptional scalability since typical sorting algorithms have running times proportional to n log n where n is the number of cases needed to be sorted. Overall, it is astonishing that the systematic use of consensus predictions comes rather late (see Mallios ,) to the problem of MHC peptide binding since consensus approaches have for quite some time proven their superiority in a number of fields, notably protein structure prediction .
In any case, it is also likely that the remarkable increase in performance obtained by the use of the consensus approach hinges on the fact that it combines information derived from methods trained on large numbers of data points with methods incorporating structural considerations leading to effective core predictions. We are currently working on development of algorithms specifically combining these two different features.
We also tested the ability of MHC class II binding prediction methods to predict a peptide's ability to activate CD4+ T cells. Most of the methods were associated with good performance. This was somewhat surprising since T cell activation is a multi-step process where multiple signals are needed for successful activation –. In addition, a peptide that binds well to MHC molecules is not necessarily a good stimulator for T-cell response as different amino acids are interacting with T cell receptor. It is important to point out that the performance was based on a set of 664 peptides of which only 9 activated CD4+ T cells. The limited number of positive cases makes the ROC curve jagged and the AUC values calculated less robust. Despite the encouraging AUC values achieved by several methods, it is still necessary to test a large number of peptides to identify most of the T cell activating peptides. In addition, all those methods still have high numbers of false positives peptides that are predicted binders but will not activate T cells. Since experimental efforts to test T cell activation are even more time consuming than testing peptide-MHC binding, significant efforts are needed to develop tools that can identify T cell activating peptides with high sensitivity and specificity.
In conclusion, we have presented a set of benchmarks to facilitate the evaluation and development of MHC class II binding predictions. While several good methods are available, these do not reach the performance of those for MHC class I molecules. We have shown that a simple and robust consensus approach can improve the prediction performance for the great majority of the MHC class II molecules tested. Finally, we speculate that novel approaches that capture distinct features of MHC class II peptide interactions could lead to more successful predictions than the current approaches, which are commonly developed as extensions of MHC class I predictions.
Materials and Methods
Peptides utilized for the assessment of MHC binding, antigenicity and immunogenicity were purchased as crude material from Mimotopes (Minneapolis, MN and Clayton, Victoria, Australia), Pepscan Systems B.V. (Leylstad, Netherlands) or A and A Labs (San Diego, CA). Quality control analyses of crude syntheses were performed by mass spectrometry on randomly selected peptides. Peptides selected for additional deconvolution and HLA peptide binding assays were resynthesized by A and A as purified material. Peptides were purified to >95% by reversed-phase HPLC, and the purity assessed by amino acid sequence and/or composition analysis.
Experimental Procedures to Measure MHC Class II Peptide Affinity
Quantitative assays to measure the binding affinities of peptides to purified soluble class II molecules are based on the inhibition of binding of a radiolabeled standard peptide. Binding assays were performed essentially as described previously ,. Briefly, 0.1–1 nM radiolabeled peptide was coincubated for 2 days at room temperature with 1 µM to 1 nM purified MHC in the presence of a cocktail of protease inhibitors. Following a two-day incubation, the amount of MHC bound labelled peptide was determined by capturing MHC/peptide complexes on LB3.1 antibody coated Lumitrac 600 microplates (Greiner Bio-one, Longwood, FL), and measuring bound cpm using the TopCount microscintillation counter (Packard Instrument Co., Meriden, CT). Individual peptides were typically tested in 3 or more independent experiments for its capacity to inhibit the binding of the radiolabeled peptide. The concentration of peptide yielding 50% inhibition of the binding of the radiolabeled peptide was calculated. Under the conditions used, in which [label]<[MHC] and IC50≥[MHC], the measured IC50 values are reasonable approximations of the true Kd values. The binding affinities are expressed in terms of IC50 and are capped at 50,000 nM, reflecting the experimental sensitivity threshold.
Dataset of Binding Affinities Used in the Study
The assembled MHC class II peptide binding affinities are listed in Table 1. The peptide binding affinities for various MHC class II molecules were generated in the context of various projects currently ongoing in our laboratory. Because they have been recently generated, to the best of our knowledge, none of the binding affinities in this dataset has been previously published. This assessment was confirmed by comparing our dataset to publicly available records contained in the IEDB (Table 1) or elsewhere. There are total 10,017 measured affinities in our dataset spanning thirteen human and three mouse MHC class II types. Peptides for 114 proteins from 30 organisms were synthesized and tested. While peptide sizes ranged form 9 to 37 amino acids, the vast majority of the measured affinities are for 15-mers (9,632 out of 10,017). The present dataset is currently in the process of being deposited in the IEDB.
PDB Structures of MHC Class II and Epitope Complexes
Structures of MHC class II were retrieved from the Protein Data Bank with a keyword search (using keyword “MHC class II”). The retrieved structures were then examined to select complexes have epitopes with at least 9 amino acids. In addition, the structures were examined to identify entries with identical MHC and binding peptide sequences. For duplicated structures of the same MHC and epitope, we retained the structure with the highest resolution. The final dataset contains 29 non-redundant structures.
MHC Class II Binding Prediction Tools Evaluated in This Study
The eight MHC class II binding prediction tools evaluated in this study are listed in Table 2. Five of the prediction methods are based on various scoring matrices. The method developed at IEDB utilizes the Average Relative Binding (ARB) matrix . PROPRED  and SVMHC  are web servers based on TEPITOPE's pocket profile . Both SYFPEITHI  and RANKPEP  use position specific matrices. Another matrix based approach, SMM-align , utilizes the stabilized matrix method (SMM ), but introduces a novel step to identify peptide binding cores, which makes it applicable to MHC class II predictions. Two of the methods, SVRMHC  and MHC2PRED (http://www.imtech.res.in/raghava/mhc2pred/index.html), apply support vector machine or support vector regression to predict epitopes. Finally, MHCPRED is a quantitative structure activity relationship (QSAR) regression based method .Three of the nine methods, ARB, MHC2PRED and SMM-align, give predictions in terms of the quantitative affinity of a peptide for a MHC class II molecule. The predictions of the other six methods are given as a score which is not directly translatable into an affinity of peptide-MHC binding.
In terms of the number of MHC class II types covered, the two TEPITOPE based methods (PROPRED and SVMHC) have the broadest coverage with 51 types, 11 of which also appear in our dataset. The next most comprehensive method is RANKPEP which covers 46 types, 16 of which overlap with our dataset. ARB, MHC2PRED and SMM-align make predictions for about 20 MHC class II types and the majority of the types (15 to 16) also appear in our dataset. The three remaining methods (MHCPRED, SVRMHC and SYFPEITHI) have less coverage, as they only predict peptide binding for 5 to 6 MHC class II types in our dataset.
Table 2 also lists the dataset used by each method to train their predictive models. Training on larger sets of data would be expected to yield better performance when tested on independent new data. In this context, the IEDB has HLA-DRB1*0101 binding information for 1390 peptides, AntiJen for 730, and MHCBN for 588. By contrast, SYFPEITHI lists only 42 entries for HLA-DRB1*0101. Thus the ARB and SMM-align methods which use data from the IEDB, had access to the largest training set compared to other methods, while the SYFPEITHI method had access to the smallest dataset.
MHC Class II Epitope Prediction with External Tools
We identified eight publicly available MHC class II prediction tools through literature search and the IMGT link list at http://imgt.cines.fr/textes/IMGTbloc-notes/. For each tool, we mapped the MHC types for which predictions could be made to the four-digit HLA nomenclature (e.g., HLA-DRB1*0101). If this mapping could not be done exactly, we left that type/tool combination out of the evaluation. For example, HLA-DR4 could refer to HLA-DRB1*0401, DRB1*0402 etc, which do have distinct binding specificities.
For the ARB evaluation, the 10-fold cross validation results stored at IEDB was used to estimate performance since ARB was trained on datasets overlapping with the one used in this study. For the other seven tools in the evaluation, we wrote python script wrappers to automate prediction retrieval. For the SYFPEITHI prediction, we patched each testing peptide with three Glycine residues at both ends before we submitted it for prediction. This was recommended by the creators of SYFPEITHI method to ensure that all potential binders are presented to the prediction algorithm. For all other methods, the original testing peptides were submitted directly for prediction. Peptide sequences were sent to the web servers one at a time and predictions were extracted from the server's response. To assign a single prediction for peptides longer than nine amino acids in the context of tools predicting the affinity of 9-mer core binding regions, we took the highest affinity prediction of all possible 9-mers within the longer peptide as the prediction result.
Consensus Approach to Predict MHC Class II Binding Peptides
For each MHC class II molecules whose binding can be predicted by three or more algorithms, we employed the following approach to generate a consensus prediction. First, we selected the top three methods that give the best performance. For each method, the tested peptides are ranked by their scores with higher ranks for better binders. For each tested peptide, the three ranks from different methods are then taken and the median of the three is calculated. This median rank is taken as the consensus score.
Performance Measure of External Tools
Receiver operating characteristic (ROC) curves  were used to measure the performance of MHC class II binding prediction tools. For binding assays, the peptides were classified into binders (experimental IC50<1000 nM) and nonbinders (experimental IC50≥1000 nM), which was determined as a practical cutoff in a previous study . For CD4+ T cell activation assays, the peptides were classified into T-cell epitopes (experimental SFC count≥100) or non-epitopes (experimental SFC count <100). For a given prediction method and a given cutoff for the predicted scores, the rate of true positive and false positive predictions can be calculated. An ROC curve is generated by varying the cutoff from the highest to the lowest predicted scores, and plotting the true positive rate against the false positive rate at each cutoff. The area under ROC curve is a measure of prediction algorithm performance where 0.5 is random prediction and 1.0 is perfect prediction. The plotting of ROC curve and calculation of AUC are all carried out with the ROCR  package for R .
LCMV Epitope Identification
C57BL/6 (H-2b) mice were purchased from The Jackson Laboratory (Bar Harbor, ME), and infected intraperitoneally with 2×105 PFU of LCMV Armstrong (i.p.). Spleens were harvested eight days post infection, and IFN-γ ELISPOT assays were performed as previously described  using CD4+ T cells isolated with anti-CD4+ magnetic beads (Miltenyi Biotech Inc., Auburn, CA). Experimental values were expressed as the mean net spots per million CD4+ cells ±SD for each peptide pool or individual peptide. For the initial screening of the 83 pools, responses against each pool were considered positive if a) the number of spot forming cells (SFCs) /106 CD4+ T cells exceeded the absolute value of the mean negative control wells (effectors plus APCs without peptide) by two-fold, b) the value exceeded 200 SFCs/106 CD4+ cells and c) these conditions were met in at least two replicate independent experiments. Positive pools were deconvoluted into their eight individual components and tested again, to determine which individual peptides were responsible for the pooled IFN-γ response. Responses against individual peptides were considered positive if they exceeded the threshold of the mean negative control wells (effectors plus APCs without peptide) by at least 2 standard deviations and exceeded a threshold of 200 SFCs/106 CD4+ cells.
Conceived and designed the experiments: PW AS BP. Performed the experiments: PW JS CD BM. Analyzed the data: PW BP. Contributed reagents/materials/analysis tools: JS CD BM. Wrote the paper: PW JS BM AS BP.
- 1. Dranoff G, Jaffee E, Lazenby A, Golumbek P, Levitsky H, et al. (1993) Vaccination with irradiated tumor cells engineered to secrete murine granulocyte-macrophage colony-stimulating factor stimulates potent, specific, and long-lasting anti-tumor immunity. Proc Natl Acad Sci U S A 90: 3539–3543.
- 2. Jenkins MK, Khoruts A, Ingulli E, Mueller DL, McSorley SJ, et al. (2001) In vivo activation of antigen-specific CD4 T cells. Annu Rev Immunol 19: 23–45.
- 3. Germain RN (1994) MHC-dependent antigen processing and peptide presentation: providing ligands for T lymphocyte activation. Cell 76: 287–299.
- 4. Rosenberg ES, Billingsley JM, Caliendo AM, Boswell SL, Sax PE, et al. (1997) Vigorous HIV-1-specific CD4+ T cell responses associated with control of viremia. Science 278: 1447–1450.
- 5. Rudolph MG, Stanfield RL, Wilson IA (2006) How TCRs bind MHCs, peptides, and coreceptors. Annu Rev Immunol 24: 419–466.
- 6. Cresswell P (1994) Assembly, transport, and function of MHC class II molecules. Annu Rev Immunol 12: 259–293.
- 7. Stern LJ, Brown JH, Jardetzky TS, Gorga JC, Urban RG, et al. (1994) Crystal structure of the human class II MHC protein HLA-DR1 complexed with an influenza virus peptide. Nature 368: 215–221.
- 8. Zhu Y, Rudensky AY, Corper AL, Teyton L, Wilson IA (2003) Crystal structure of MHC class II I-Ab in complex with a human CLIP peptide: prediction of an I-Ab peptide-binding motif. J Mol Biol 326: 1157–1174.
- 9. Jones EY, Fugger L, Strominger JL, Siebold C (2006) MHC class II proteins and disease: a structural perspective. Nat Rev Immunol 6: 271–282.
- 10. Godkin AJ, Smith KJ, Willis A, Tejada-Simon MV, Zhang J, et al. (2001) Naturally processed HLA class II peptides reveal highly conserved immunogenic flanking region sequence preferences that reflect antigen processing rather than peptide-MHC interactions. J Immunol 166: 6720–6727.
- 11. Consogno G, Manici S, Facchinetti V, Bachi A, Hammer J, et al. (2003) Identification of immunodominant regions among promiscuous HLA-DR-restricted CD4+ T-cell epitopes on the tumor antigen MAGE-3. Blood 101: 1038–1044.
- 12. Zhang GL, Khan AM, Srinivasan KN, August JT, Brusic V (2005) MULTIPRED: a computational system for prediction of promiscuous HLA binding peptides. Nucleic Acids Res 33: W172–W179.
- 13. Sidney J, Southwood S, Oseroff C, Del Guercio M, Grey H, et al. (1998) Measurement of MHC/peptide interactions by gel filtration. Current protocals in immunology. New york: John Wiley & Sons, Inc. pp. 18.13.11–18.13.19.
- 14. Sylvester-Hvid C, Kristensen N, Blicher T, Ferre H, Lauemoller SL, et al. (2002) Establishment of a quantitative ELISA capable of determining peptide - MHC class I interaction. Tissue Antigens 59: 251–258.
- 15. Rammensee H, Bachmann J, Emmerich NP, Bachor OA, Stevanovic S (1999) SYFPEITHI: database for MHC ligands and peptide motifs. Immunogenetics 50: 213–219.
- 16. Bhasin M, Singh H, Raghava GP (2003) MHCBN: a comprehensive database of MHC binding and non-binding peptides. Bioinformatics 19: 665–666.
- 17. Toseland CP, Clayton DJ, McSparron H, Hemsley SL, Blythe MJ, et al. (2005) AntiJen: a quantitative immunology database integrating functional, thermodynamic, kinetic, biophysical, and cellular data. Immunome Res 1: 4.
- 18. Schonbach C, Koh JL, Sheng X, Wong L, Brusic V (2000) FIMM, a database of functional molecular immunology. Nucleic Acids Res 28: 222–224.
- 19. Sathiamurthy M, Hickman HD, Cavett JW, Zahoor A, Prilliman K, et al. (2003) Population of the HLA ligand database. Tissue Antigens 61: 12–19.
- 20. Peters B, Sette A (2007) Integrating epitope data into the emerging web of biomedical knowledge resources. Nat Rev Immunol 7: 485–490.
- 21. Peters B, Sidney J, Bourne P, Bui HH, Buus S, et al. (2005) The immune epitope database and analysis resource: from vision to blueprint. PLoS Biol 3: e91. doi:10.1371/journal.pbio.0030091.
- 22. Sette A, Buus S, Colon S, Smith JA, Miles C, et al. (1987) Structural characteristics of an antigen required for its interaction with Ia and recognition by T cells. Nature 328: 395–399.
- 23. Hammer J, Bono E, Gallazzi F, Belunis C, Nagy Z, et al. (1994) Precise prediction of major histocompatibility complex class II-peptide interaction based on peptide side chain scanning. J Exp Med 180: 2353–2358.
- 24. Marshall KW, Wilson KJ, Liang J, Woods A, Zaller D, et al. (1995) Prediction of peptide affinity to HLA DRB1*0401. J Immunol 154: 5927–5933.
- 25. Bui HH, Sidney J, Peters B, Sathiamurthy M, Sinichi A, et al. (2005) Automated generation and evaluation of specific MHC binding predictive tools: ARB matrix applications. Immunogenetics 57: 304–314.
- 26. Nielsen M, Lundegaard C, Lund O (2007) Prediction of MHC class II binding affinity using SMM-align, a novel stabilization matrix alignment method. BMC Bioinformatics 8: 238.
- 27. Reche PA, Glutting JP, Reinherz EL (2002) Prediction of MHC class I binding peptides using profile motifs. Hum Immunol 63: 701–709.
- 28. Reche PA, Glutting JP, Zhang H, Reinherz EL (2004) Enhancement to the RANKPEP resource for the prediction of peptide binding to MHC molecules using profiles. Immunogenetics 56: 405–419.
- 29. Sturniolo T, Bono E, Ding J, Raddrizzani L, Tuereci O, et al. (1999) Generation of tissue-specific and promiscuous HLA ligand databases using DNA microarrays and virtual HLA class II matrices. Nat Biotechnol 17: 555–561.
- 30. Noguchi H, Kato R, Hanai T, Matsubara Y, Honda H, et al. (2002) Hidden Markov model-based prediction of antigenic peptides that interact with MHC class II molecules. J Biosci Bioeng 94: 264–270.
- 31. Brusic V, Rudy G, Honeyman G, Hammer J, Harrison L (1998) Prediction of MHC class II-binding peptides using an evolutionary algorithm and artificial neural network. Bioinformatics 14: 121–130.
- 32. Murugan N, Dai Y (2005) Prediction of MHC class II binding peptides based on an iterative learning model. Immunome Res 1: 6.
- 33. Hertz T, Yanover C (2006) PepDist: a new framework for protein-peptide binding prediction based on learning peptide distance functions. BMC Bioinformatics 7: Suppl 1S3.
- 34. Bhasin M, Raghava GP (2004) SVM based method for predicting HLA-DRB1*0401 binding peptides in an antigen sequence. Bioinformatics 20: 421–423.
- 35. Burden FR, Winkler DA (2005) Predictive Bayesian neural network models of MHC class II peptide binding. J Mol Graph Model 23: 481–489.
- 36. Honeyman MC, Brusic V, Stone NL, Harrison LC (1998) Neural network-based prediction of candidate T-cell epitopes. Nat Biotechnol 16: 966–969.
- 37. Noguchi H, Hanai T, Honda H, Harrison LC, Kobayashi T (2001) Fuzzy neural network-based prediction of the motif for MHC class II binding peptides. J Biosci Bioeng 92: 227–231.
- 38. Peters B, Bui HH, Frankild S, Nielson M, Lundegaard C, et al. (2006) A community resource benchmarking predictions of peptide binding to MHC-I molecules. PLoS Comput Biol 2: e65. doi:10.1371/journal.pcbi.0020065.
- 39. van der Merwe PA, Davis SJ (2003) Molecular interactions mediating T cell antigen recognition. Annu Rev Immunol 21: 659–684.
- 40. Moutaftsi M, Peters B, Pasquetto V, Tscharke DC, Sidney J, et al. (2006) A consensus epitope prediction approach identifies the breadth of murine T(CD8+)-cell responses to vaccinia virus. Nat Biotechnol 24: 817–819.
- 41. Rajapakse M, Schmidt B, Feng L, Brusic V (2007) Predicting peptides binding to MHC class II molecules using multi-objective evolutionary algorithms. BMC Bioinformatics 8: 459.
- 42. Gowthaman U, Agrewala JN (2008) In silico tools for predicting peptides binding to HLA-class II molecules: more confusion than conclusion. J Proteome Res 7: 154–163.
- 43. Borras-Cuesta F, Golvano J, Garcia-Granero M, Sarobe P, Riezu-Boj J, et al. (2000) Specific and general HLA-DR binding motifs: comparison of algorithms. Hum Immunol 61: 266–278.
- 44. Peters B, Sette A (2005) Generating quantitative models describing the sequence specificity of biological processes with the stabilized matrix method. BMC Bioinformatics 6: 132.
- 45. Mallios RR (2003) A consensus strategy for combining HLA-DR binding algorithms. Hum Immunol 64: 852–856.
- 46. Mallios RR (2001) Predicting class II MHC/peptide multi-level binding with an iterative stepwise discriminant analysis meta-algorithm. Bioinformatics 17: 942–948.
- 47. Calvo-Calle JM, Strug I, Nastke MD, Baker SP, Stern LJ (2007) Human CD4(+) T cell epitopes from vaccinia virus induced by vaccination or infection. PLoS Pathog 3: e144. doi:10.1371/journal.ppat.0030144.
- 48. Trost B, Bickis M, Kusalik A (2007) Strength in numbers: achieving greater accuracy in MHC-I binding prediction by combining the results from multiple prediction tools. Immunome Res 3: 5.
- 49. Venclovas C, Zemla A, Fidelis K, Moult J (2003) Assessment of progress over the CASP experiments. Proteins 53: 585–595.
- 50. Samelson LE (2002) Signal transduction mediated by the T cell antigen receptor: the role of adapter proteins. Annu Rev Immunol 20: 371–394.
- 51. Chambers CA, Kuhns MS, Egen JG, Allison JP (2001) CTLA-4-mediated inhibition in regulation of T cell responses: mechanisms and manipulation in tumor immunotherapy. Annu Rev Immunol 19: 565–594.
- 52. Acuto O, Cantrell D (2000) T cell activation and the cytoskeleton. Annu Rev Immunol 18: 165–184.
- 53. Doolan DL, Southwood S, Freilich DA, Sidney J, Graber NL, et al. (2003) Identification of Plasmodium falciparum antigens by antigenic analysis of genomic and proteomic data. Proc Natl Acad Sci U S A 100: 9952–9957.
- 54. Singh H, Raghava GP (2001) ProPred: prediction of HLA-DR binding sites. Bioinformatics 17: 1236–1237.
- 55. Donnes P, Kohlbacher O (2006) SVMHC: a server for prediction of MHC-binding peptides. Nucleic Acids Res 34: W194–W197.
- 56. Wan J, Liu W, Xu Q, Ren Y, Flower DR, et al. (2006) SVRMHC prediction server for MHC-binding peptides. BMC Bioinformatics 7: 463.
- 57. Guan P, Doytchinova IA, Zygouri C, Flower DR (2003) MHCPred: A server for quantitative prediction of peptide-MHC binding. Nucleic Acids Res 31: 3621–3624.
- 58. Swets JA (1988) Measuring the accuracy of diagnostic systems. Science 240: 1285–1293.
- 59. Sidney J, del Guercio MF, Southwood S, Sette A (2002) The HLA molecules DQA1*0501/B1*0201 and DQA1*0301/B1*0302 share an extensive overlap in peptide binding specificity. J Immunol 169: 5098–5108.
- 60. Sing T, Sander O, Beerenwinkel N, Lengauer T (2005) ROCR: visualizing classifier performance in R. Bioinformatics 21: 3940–3941.
- 61. Team RDC (2006) R: A Language and Environment for Statistical Computing. Vienna, Austria.
- 62. Tangri S, Mothe BR, Eisenbraun J, Sidney J, Southwood S, et al. (2005) Rationally engineered therapeutic proteins with reduced immunogenicity. J Immunol 174: 3187–3196.