Bioinformatics Advance Access originally published online on March 1, 2006
Bioinformatics 2006 22(10):1232-1238; doi:10.1093/bioinformatics/btl071
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Prediction of HLA-DQ3.2ß Ligands: evidence of multiple registers in class II binding peptides
1 Department of Biochemistry, Yong Loo Lin School of Medicine, National University of Singapore 8 Medical Drive, Singapore 117597, Singapore
2 Institute for Infocomm Research 21 Heng Mui Keng Terrace, Singapore 119613, Singapore
3 Division of Biomedical Sciences, John Hopkins Medicine in Singapore 41 Science Park Road, Lobby C, The Gemini, Singapore 117610, Singapore
4 Department of Pharmacology and Molecular Sciences, John Hopkins University School of Medicine Baltimore, MD, USA
5 Australian Centre for Plant Functional Genomics, School of Land and Food Sciences, and Institute for Molecular Bioscience, The University of Queensland Brisbane 4072, Australia
6 Department of Chemistry and Biomolecular Sciences and Biotechnology Research Institute, Macquarie University NSW 2109, Australia
*To whom correspondence should be addressed.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMETHODSDISCUSSIONREFERENCES |
|---|
Motivation: While processing of MHC class II antigens for presentation to helper T-cells is essential for normal immune response, it is also implicated in the pathogenesis of autoimmune disorders and hypersensitivity reactions. Sequence-based computational techniques for predicting HLA-DQ binding peptides have encountered limited success, with few prediction techniques developed using three-dimensional models.
Methods: We describe a structure-based prediction model for modeling peptide-DQ3.2ß complexes. We have developed a rapid and accurate protocol for docking candidate peptides into the DQ3.2ß receptor and a scoring function to discriminate binders from the background. The scoring function was rigorously trained, tested and validated using experimentally verified DQ3.2ß binding and non-binding peptides obtained from biochemical and functional studies.
Results: Our model predicts DQ3.2ß binding peptides with high accuracy [area under the receiver operating characteristic (ROC) curve AROC > 0.90], compared with experimental data. We investigated the binding patterns of DQ3.2ß peptides and illustrate that several registers exist within a candidate binding peptide. Further analysis reveals that peptides with multiple registers occur predominantly for high-affinity binders.
Contact: shoba{at}els.mq.edu.au
Supplementary information: Supplementary data is available at Bioinformatics online.
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSDISCUSSIONREFERENCES |
|---|
Major histocompatibility complex (MHC) class II molecules play a critical role in immune responses. They bind short antigenic peptide fragments and present them on the surface of antigen-presenting cells for recognition by the CD4+ helper T-cells. T-cell recognition of the peptide-MHC (pMHC) complex initiates a cascade of immunological events necessary for initiation and regulation of immune responses. These events are necessary for normal immune responses but may also be involved in the pathogenesis of autoimmune disorders (Klein et al., 2000; Flynn et al., 2004) and hypersensitivity reactions (Neeno et al., 1996; Krco et al., 2000).
The HLA-DQ allele, DQ3.2ß (DQA1*0301/DQB1*0302), commonly known as DQ8, is present in 
20% of the human population (Middleton et al., 2003, http://www.allelefrequencies.net). DQ3.2ß is of particular interest in the study of allergenicity and autoimmunity because of its association to house dust mite allergy (Neeno et al., 1996; Krco et al., 2000) and several human autoimmune disorders including celiac disease (CD) (Sollid and Thorsby, 1993), insulin-dependent diabetes mellitus (IDDM) (Nepom and Kwok, 1998; Erlich et al., 1993), IDDM-associated periodontal disease (Faustman et al., 1991) and autoimmune polyendocrine syndrome type II (APS-II) (Robles et al., 2002). Some 70% of IDDM patients (Kwok et al., 1989) have DQ3.2ß. Improved understanding of peptide binding to this molecule is important for elucidating the role of DQ3.2ß in both autoimmunity and allergies. Peptide-binding studies are invaluable for designing vaccines and immunotherapies for controlling allergic or autoimmune responses.
Computational methods for the identification of peptides that bind to HLA-DR molecules are relatively advanced (Brusic et al., 2004), while methods for prediction of peptide binding to HLA-DQ molecules have encountered limited success because of the paucity of peptides as training data for sequence-based techniques. Computational strategies for DQ3.2ß binding peptides using sequence motifs (Godkin et al., 1997, 1998; Rammensee et al., 1999; Moustakas et al., 2000) have been used with varying degrees of success (Harfouch-Hammoud et al., 1999), but an effective model for large-scale screening is still currently lacking. Up to now, few prediction techniques for HLA-DQ molecules have been developed using three-dimensional (3D) models as the dual issues of docking and scoring must be addressed (Ranganathan et al., 2005).
We have recently developed a new technique for rapid and accurate docking of flexible peptide ligands to class I alleles and the core recognition residues of peptide ligands to class II alleles (Tong et al., 2004; Ranganathan and Tong, 2005). The starting point involves the use of a probe or ‘base fragment’ to sample different regions of the receptor binding site, followed by loop closure and refinement of the core recognition residues or binding registers (Li et al., 2000; Seamons et al., 2003). The technique was benchmarked against 40 non-redundant peptide/MHC (pMHC) complexes (29 class I and 11 class II) and we successfully modeled 33 peptides with a C
RMSD < 1.00 Å. We used this approach to successfully discriminate between alleles implicated in the autoimmune disorder, pemphigus vulgaris from non-disease implicated and protective alleles (Tong et al., 2006). However, despite the accuracy of our docking experiments, these results are qualitative rather than quantitative, as energy-based scoring was not considered. The earlier model, therefore, cannot be used for effective discrimination of peptide-binding affinities (strong, moderate and weak binders and non-binders). We have now extended our docking protocol by developing a complementary scoring function to effectively identify DQ3.2ß epitopes, with the correct binding register. We also investigated the binding patterns of DQ3.2ß peptides and show that recognition of MHC class II peptides may occur at several registers within the candidate peptide.
|
METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSDISCUSSIONREFERENCES |
|---|
DQ3.2ß binding and non-binding peptide sequences were extracted from the literature and docked into the experimental crystal structure of DQ3.2ß using a four-step protocol (Tong et al., 2004). A customized free energy scoring function was developed to improve the predictive performance of the model.
Data
Crystallographic data
The coordinates of DQ3.2ß was extracted from the crystal structure of DQ3.2ß–insulin B9-23 complex, with Protein Databank (PDB) code 1JK8
[PDB]
(Lee et al., 2001). The structure was relaxed by conjugate gradient minimization, using the Internal Coordinate Mechanics (ICM) 3.0 package (Abagyan et al., 1994).
Experimental binding data
Two sets of data are used in this study: (1) peptides with experimental IC50 values from biochemical studies and (2) peptides with experimental T-cell proliferation values from functional studies.
Dataset I comprises 127 peptides (Supplementary Materials: Table S1) with experimentally determined IC50 values (70 high-affinity, 13 medium-affinity and 23 low-affinity binders and 21 non-binders) derived from biochemical studies (Godkin et al., 1998; Sidney et al., 2002; Suri et al., 2005). Peptides are classified based on their experimental IC50 values (high-affinity binders: IC50 
500 nM, medium-affinity binders: 500 nM < IC50 1500 nM, low-affinity binders: 1500 < IC50 5000 nM and non-binders: 5000 < IC50). In this dataset, 87 binding peptides had experimentally determined binding registers.
Dataset II consists of 12 Dermatophagoides pternoyssinus (Der p) peptides with experimental T-cell proliferation values from functional studies (Krco et al., 2000; Neeno et al., 1996), with seven peptides eliciting DQ3.2ß restricted T-cell proliferation.
Model
Peptide docking
In this study, an overlapping sliding window of size nine is applied to each peptide to generate all combinations of nonameric core regions to be modeled into the binding groove of DQ3.2ß. Docking was performed with the Empirical Conformational Energy Program for Peptides 3 force field parameters (ECEPP/3) (Abagyan et al., 1994) and MMFF partial charges (Abagyan and Totrov, 1999) on a 4-CPU SGI Origin 3200 workstation using an extension of the protocol (Tong et al., 2004), illustrated in Figure 1: (1) pseudo-Brownian rigid body docking of peptide fragments to the ends of the binding groove, (2) central loop closure by satisfaction of spatial constraints, (3) refinement of the backbone and side-chain atoms of the core recognition residues and receptor contact regions and (4) extension of flanking peptide residues by satisfaction of spatial constraints. The first three steps of the docking procedure for generating the binding register were described earlier (Tong et al., 2004). The conformations of the flanking peptide residues are subsequently generated by satisfying the spatial constraints in the allowed subspace for backbone dihedrals (Sali and Blundell, 1993), defined by the conformation of the bound core nonameric peptide docked into the binding groove. In brief, this is performed in three stages: (1) distance and dihedral angle restraints on the entire peptide sequence are derived from its alignment with the nonamer sequence in the binding groove; (2) the restraints on spatial features of the flanking residues are derived by extrapolation from the known 3D structure of flanking residues (PDB code 1JK8
[PDB]
) in the alignment, expressed as probability density functions; and (3) the spatial restraints on the flanking residues are then satisfied by optimization of the molecular probability density function using a variable target function technique that applies the conjugate gradients algorithm to positions of all non-hydrogen atoms.
|
Empirical free energy functions
The scoring function presented in the study is based on the free energy potential in ICM3.0 package (Abagyan and Totrov, 1999). The binding free energy is computed as the difference between the energy of the solvated complex and the sum of the energy of the solvated receptor and that of the peptide ligand. The reference state chosen for the peptide is the fully relaxed conformation of the free peptide in water (Schapira et al., 1999). In all binding energy calculations, the protein and the ligand are separated after docking and their relaxed energies computed, following energy minimization in water. The binding free energy function (
Gbind) is expressed as
 | (1) |
GH is the hydrophobic energy computed as the product of solvent accessible surface area (determined by rolling a sphere of 1.4 Å radius along the surface of the molecule) by the surface tension. GS refers to the entropic contribution from the protein side-chains computed from the maximal burial entropies for each type of amino acid and their relative accessibilities. GEL denotes the electrostatic term composed of coulombic interactions between receptor and ligand and the desolvation of partial charges transferred from an aqueous medium to a protein core environment, and it is determined by the numeric solution of the Poisson equation using an implementation of the boundary element algorithm (Zauhar and Morgan, 1985; Bharadwaj et al., 1995; Schapira et al., 1999). An additional constant term C (or K; Rognan et al., 1999) accounts for entropy change in the system due to the decrease of free molecular concentration and the loss of rotational/translational degrees of freedom upon binding (Schapira et al., 1999). In theory, C represents physical parameters which are independent of the dataset used and there are great variations in its value among various research groups (reviewed in Janin, 1995). The coefficients (, ß, 
) assigned to each energy term were optimized in this study, to obtain the best separation of binders and non-binders in the peptide-DQ3.2ß model. This partitioning scheme has been successfully adopted as a framework in many earlier studies (Krystek et al., 1993; Weng et al., 1996; Novotny et al., 1997; Froloff et al., 1997; Schapira et al., 1999) and consists of the most significant potentials contributing to protein–protein, protein–ligand and protein–peptide interactions.
Optimization of the scoring function
Reported IC50 values, representing the concentration of ligand required to saturate half of the available binding sites of the protein (Bock and Gough, 2002), were assumed to be similar to equilibrium dissociation constants Kd as the concentration of the ligand in the unbound state is much lower than the equilibrium dissociation constant Kd of the ligand in the binding assay, so that Gbind 
–RT ln (IC50) (Rognan et al., 1999). Gbind is usually reported in units of pKd [–log10(Kd)], where 1 pKd = –1.364 kcal/mol (Wang et al., 2002) or –5.708 kJ/mol at 298.15 K. To improve the discriminative power of the scoring function, the coefficients of the different energy terms were recalibrated using standard least-square multivariate regression analyses of the training set (Wang et al., 2002). This step was followed by a 10-fold cross-validation (Bock and Gough, 2002) to assess the quality of the scoring function. In k-fold cross-validation, k random, (approximately) equal-sized, disjoint partitions of the sample data are constructed, and a given model is trained on (k – 1) partitions and tested on the excluded partition. The results are averaged after k such experiments, and the observed error rate may be taken as an estimate of the error rate expected upon generalization to new data. The predictive power of the models was assessed by the cross-validation coefficient q2 and the standard error of prediction spress. The robustness of the predictive model was further evaluated using evolutionary regression analysis (Wang et al., 2002), with different subsets representing 5-, 4-, 3- and 2-fold cross-validation.
Training, testing and validation
Peptide data obtained from biochemical studies with experimental IC50 values were divided into training and test datasets. Training of the DQ3.2ß prediction model was performed by sampling (1) the bound conformations of binding peptides with experimentally determined registers that can be recognized by MHC and (2) the best conformations of non-binding peptides without any preferred register in the binding groove. The training set comprised 56 binding conformations with known registers and 30 non-binding conformations generated from three non-binding peptides (from Dataset I) without any binding registers. Two external sets of test data were used: (1) Test set 1: 68 peptides (the rest of Dataset I) with experimental IC50 values (16 high-affinity binders, 13 medium-affinity binders, 21 low-affinity binders and 18 non-binders) from biochemical studies and (2) Test set 2: all peptides from Dataset II, with known T-cell proliferation values.
The predictive performance of our model was assessed using sensitivity (SE), specificity (SP) and receiver operating characteristic (ROC) analysis as described previously (Brusic et al., 2002). SE = TP/(TP + FN) and SP = TN/(TN + FP), indicate percentages of correctly predicted binders and non-binders, respectively. true positives (TP) stand for experimental binders with at least one predicted binding register and true negatives (TN) for experimental non-binders with no predicted binding register. False negatives (FN) denote experimental binders predicted as non-binders and false positives (FP) represent experimental non-binders predicted as binders. The accuracy of our predictions was assessed by ROC analysis where the ROC curve is generated by plotting SE as a function of (1 – SP) for various classification thresholds. The area under the ROC curve (AROC) provides a measure of overall prediction accuracy, AROC < 70% for poor, AROC > 80% for good and AROC > 90% for excellent predictions (Brusic et al., 2002). We consider values of SP 
80% useful in practice and assessed SE for three values of SP (80, 90 and 95%).
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSDISCUSSIONREFERENCES |
|---|
The accuracy of the DQ3.2ß prediction model was evaluated using (1) peptides with experimental IC50 values obtained from biochemical studies with experimental IC50 values and (2) peptides with T-cell proliferation values obtained from functional studies.
Three threshold binding energy values were used to evaluate the accuracy of the DQ3.2ß prediction model on Test set 1—LMH (low-, medium- and high-affinity binders; AROC = 0.88); MH (medium- and high-affinity binders; AROC = 0.93) and H (high-affinity binders only; AROC=0.93). The results indicate that, overall, 3D models are suitable for discriminating class II binding ligands from the background with good accuracy (AROC > 0.80). The accuracy of our model relies on the scoring function derived from the training dataset of experimentally determined binders with known binding registers and non-binders with no binding registers. A scoring function based on the default ICM coefficients ( = ß = = 1; C = 0) resulted in poor correlation (r2 = 0.43, s = 2.91 kJ/mol) to experimental data when tested with the novel peptide-DQ3.2ß system. The discriminative power of our model improved significantly with better correlation (r2 = 0.89, s = 4.77 kJ/mol) after recalibration of the scoring function by fitting to the training data, using multiple linear regression. The optimal scoring function, after 10-fold cross-validation (q2 = 0.85, spress = 2.20 kJ/mol) is
 | (2) |
The training set of 86 complexes in the current study is too large for the leave-one-out cross-validation done by Rognan et al. (1999) on training datasets of 5 and 37 MHC-peptide complexes. At the same time, it is smaller than the training set of 200 complexes used by Wang et al. (2002) or the 2617 protein–ligand complexes studies by Bock and Gough (2002) for extensive cross-validation analyses. The higher standard error in the training set (s = 4.77 kJ/mol = 1.13 kcal/mol = 0.84 pKd) than the standard error after 10-fold cross-validation (spress = 2.20 kJ/mol = 0.52 kcal/mol = 0.39 pKd) is attributable to the noise in binding energy values in the complete training set spanning three orders of magnitude (Supplementary Materials: Table S1), compared with the subsets in 10-fold cross-validation, with a subset size (N) of 78 or 91% of the training set. These values are lower than error values (1.47–1.62 pKd or 8.36–9.22 kJ/mol) reported by Wang et al. (2002), for a training set of 200 protein–ligand complexes, after several rounds of evolutionary regression analysis. Using a similar but limited evolutionary regression analysis approach, the robustness of our predictive model has been estimated for 5-fold (N = 69, q2 = 0.89, spress = 2.47 kJ/mol = 0.43 pKd), 4-fold (N = 65, q2 = 0.86, spress = 2.70 kJ/mol=0.47 pKd), 3-fold (N = 57, q2 = 0.87, spress = 2.50 kJ/mol = 0.44 pKd) and 2-fold (N = 43, q2 = 0.83 spress = 3.29 kJ/mol = 0.58 pKd) cross-validation. The results indicate that despite a very slight increase in the error value for the 2-fold cross-validation, the cross-validation coefficient q2 and the standard error of prediction spress are stable, with mean values of q2 = 0.86 and spress = 2.63 kJ/mol = 0.46 pKd, and respective SD values of 0.02 and 0.41 kJ/mol = 0.07 pKd. This iterative regression procedure thus validates the internal consistency of the scoring function in the current model, rendering it suitable for predictions on the test datasets.
The sensitivity of our prediction model was determined on Test set 1 for three decision thresholds (Table 1) that define levels of specificities suitable for practical applications (Brusic et al., 2002). SP = 0.80 offers high-sensitivity predictions, whereas SP = 0.95 results in very few false positives but fewer true positives. The prediction results for our model were in accordance with expected binding patterns of DQ3.2ß peptides and provided a sensitivity of 90% (SP = 0.80). The sensitivity values decrease with higher levels of specificity (SP = 0.90, SE = 0.84 or SP = 0.95, SE = 0.81), while still correctly predicting more than half of the high-affinity binders in the worst case scenario (high-binders alone, SP = 0.95, SE = 0.63). The efficacy of our model in detecting binding registers was then evaluated with experimentally determined registers. Our external test data comprised 23 peptides from Test set 1, with known binding energy for each register (Suri et al., 2005). At a threshold of –30.82 kJ/mol (SP = 0.80, SE = 0.75), our model accurately detected 87% (20/23) of the experimentally determined binding registers. We also correctly predicted the only experimentally determined register (4–12) for Der p 2 1–20 (Krco et al., 2000), from Test set 2. Next, the predictive performance of the optimized model was tested on the functional dataset of 12 peptides (Test set 2) with experimental T-cell proliferation values using the decision thresholds defined above. The top five predictions (Der p 2 61–80, 51–70, 110–129, 101–120 and 91–110) are experimental positives (Table 2) with binding energy values of –34.52 kJ/mol or less (predicted high-binders for SE = 0.63, SP = 0.95). This is in agreement with existing studies that high-affinity binders have a greater chance of stimulating T-cell proliferation (Deng et al., 1997; Keogh et al., 2001) and this knowledge is crucial for peptide vaccine design. Peptide Der p 2 31–50, ranked #6, is a predicted high-affinity binder at a threshold of –33.59 kJ/mol (SE = 0.63, SP = 0.95). It is possible that Der p 2 31–50 is either a high-affinity binder that failed to stimulate T-cell proliferation (Deng et al., 1997; Keogh et al., 2001) or is a false positive in the prediction. At this cut-off, correct predictions number 5/7 (71%), with one false positive (14%) and two false negatives (28%). Peptide Der p 2 41–60, ranked #12 in our prediction, is possibly an outlier, as it failed to stimulate detectable T-cell response in study of Neeno et al. (1996), despite a similar reported T-cell stimulatory propensity as Der p 2 91–110; deletion experiments confirm the criticality of only residues 55–70 in the region 41–70 (Table 2 in Krco et al., 2000). For T-cell proliferation predictions, the current model is suitable to screen for high-affinity binders at SP = 0.95.
|
|
The specificity/sensitivity results are consistent with the results obtained from ROC analysis. Our results indicate that we have developed a model that can make accurate predictions for peptide binding to DQ3.2ß that has been validated using experimentally verified binding and non-binding peptides obtained from both biochemical and functional studies.
Detection of epitopes that do not conform to binding motifs
Consensus peptide-binding motifs for identifying potential immunodominant epitopes within autoantigenic proteins have been developed for many HLA class II molecules. However, earlier studies (Harfouch-Hammoud et al., 1999) reveal that these motifs do not correlate with binding to a specific allele. In Test set 1, 63 out of 68 binding peptide sequences have amino acid residues that do not conform to available DQ3.2ß binding motifs (Godkin et al., 1998; Rammensee et al., 1999) considering all relevant positions (P1, P4, P6, P7, P9). Table 3 lists 17 LMH predictions from this dataset. A-gliadin 49–63 (#10), MHC Ia 46–63 (#14) and VP16 (#15) are classified negatives using existing DQ3.2ß binding motifs. However, using our scoring function these T-cell epitopes are easily identified, with A-gliadin 49–63 as a high-affinity binder and the MHC Ia 46–63 and VP16 as medium-affinity binders. This reaffirms our earlier observation that binding motifs may be inadequate for defining T-cell epitopes and many other factors, including the physicochemical composition of the peptide, (affecting the overall stability of the pMHC complex) have to be considered in prediction systems for HLA-binding peptides.
|
Detection of multiple registers in experimental binders
Our results support the existence of multiple registers with different nonameric core regions within a candidate binding peptide that serve as recognition sites for MHC class II molecules (Fig. 2). In particular, the results indicate that several binding registers (with different nonameric core recognition regions) exist within an MHC class II binding peptide, facilitating binding to DQ3.2ß in several different conformations. Of the binding peptides in Test set 1 58% exhibit two or more registers that can be docked to DQ3.2ß with favorable binding energy values. Multiple registers occur predominantly in medium- and high-affinity binders, suggesting that recognition using flexible fitting may play a critical role in binding to MHC class II alleles as well as in T-cell recognition and this knowledge should be taken into consideration in vaccine design. For example, two conformations of the high-affinity binding peptide Pf ABRA 487–506 showed
G values less than the decision threshold of –33.59 kJ/mol (SP = 0.95, SE = 0.63), with the 496–504 register (shown in Table 3) being the preferred binding mode.
|
It is possible that the open binding groove of DQ3.2ß (and other class II alleles) accommodates peptides with differing pocket specificities and can recognize multiple regions within a single candidate peptide. While not all binding registers may elicit T-cell response, the existence of multiple registers within a candidate peptide (especially for high-affinity binders) can facilitate binding to a particular allele, enhancing T-cell recognition, with the highest binding affinity register acting as the primary recognition region.
Peptide vaccine development is advancing rapidly with recent successes in malaria (Lopez et al., 2001) and anti-tumor vaccines (Knutson et al., 2001). A key research area is to identify allele-specific candidate T-cell epitopes suitable for designing vaccines and immunotherapies to control allergic or autoimmune responses. The task of identifying candidate class II binding ligands is a challenging process due to the open binding groove that can potentially accommodate multiple binding registers (Li et al., 2000; Seamons et al., 2003) and has wholly occupied the energies of researchers. An additive method (Doytchinova and Flower, 2003) for binding to DRB1*0401, a polynomial derived scoring matrix for DRB1*0401 of Southwood et al. (1998), an iterative stepwise discriminant analysis meta-algorithm (Mallios, 2001) and a genetic algorithm (Brusic et al., 1998) are excellent approaches. However, the nonameric core regions used for training predictive models were often pre-selected based on existing binding motifs, usually extracted from multiple sequence alignment and not experimentally validated. Such methodologies exclude the prediction of other binding registers within a candidate class II binding ligand capable of eliciting a strong T-cell response. Moreover, the possibility of the existence of multiple binding registers, particularly for high-affinity binders, suggests that all possible nonameric core regions within a candidate binding ligand must be carefully examined. For training computational models, the utilization of experimentally validated binding registers is preferred.
Recently, Sinha et al. (A. A. Sinha, personal communication) discovered that DRB1*0402-specific binding motifs are insufficient for the design of pemphigus vulgaris epitopes, because of the presence of register shifts as well as polymorphisms in the binding register. With increasing evidence suggesting the inadequacy of binding motifs in defining class II T-cell epitopes, the current approach of predictive model building and virtual screening for vaccine candidates is independent of sequence motifs and takes into account the presence of multiple registers within class II ligands. In this study, we have illustrated that it is possible to efficiently discriminate between categories of binders from non-binders and predict the binding register of class II ligands with good accuracy. Our docking methodology, combined with a sensitive scoring function, provides a set of sensitive and specific computational tools to facilitate systematic screening of peptides for immunotherapeutic applications.
|
Acknowledgments |
|---|
This work was partly funded by the National Institute of Allergy and Infectious Diseases, National Institute of Health, USA (Grant #5 U19 AI56541 & Contract #HHSN266200400085C).
Conflict of Interest: none declared.
|
FOOTNOTES |
|---|
Associate Editor: Dmitrij Frishman
Received on October 17, 2005; revised on February 22, 2006; accepted on February 23, 2006
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSDISCUSSIONREFERENCES |
|---|
Abagyan, R., et al. (1994) ICM—a new method for protein modeling and design: Applications to docking and structure prediction from the distorted native conformation. J. Comput. Chem, . 15, 488–506[CrossRef][ISI].
Abagyan, R.A. and Totrov, M. (1999) Ab initio folding of peptides by the optimal-bias Monte Carlo minimization procedure. J. Comput. Phys, . 151, 402–421[CrossRef].
Bharadwaj, A., et al. (1995) The fast multipole boundary element method for molecular electrostatics: an optimal approach for large system. J. Comput. Chem, . 16, 898–910[CrossRef].
Bock, J.R. and Gough, D.A. (2002) A new method to estimate ligand–receptor energetics. Mol. Cell. Proteomics, 1, 904–910
Brusic, V., et al. (1998) Prediction of MHC class II-binding peptides using an evolutionary algorithm and artificial neural network. Bioinformatics, 14, 121–130
Brusic, V., et al. (2002) Prediction of promiscuous peptides that bind HLA class I molecules. Immunol. Cell Biol, . 80, 280–285[CrossRef][Medline].
Brusic, V., et al. (2004) Computational methods for prediction of T-cell epitopes—a framework for modelling, testing, and applications. Methods, 34, 436–443[CrossRef][ISI][Medline].
Deng, Y., et al. (1997) MHC affinity, peptide liberation, T cell repertoire, and immunodominance all contribute to the paucity of MHC class I-restricted peptides recognized by antiviral CTL. J. Immunol, . 158, 1507–1515[Abstract].
Doytchinova, I.A. and Flower, D.R. (2003) Towards the in silico identification of class II restricted T-cell epitopes: a partial least squares iterative self-consistent algorithm for affinity prediction. Bioinformatics, 19, 2263–2270
Erlich, H.A., et al. (1993) HLA class II alleles and susceptibility and resistance to insulin dependent diabetes mellitus in Mexican-American families. Nat. Genet, . 3, 358–364[CrossRef][ISI][Medline].
Faustman, D., et al. (1991) Linkage of faulty major histocompatibility complex class I to autoimmune diabetes. Science, 254, 1756–1761
Flynn, J.C., et al. (2004) Pathogenic human thyroglobulin peptides in HLA-DR3 transgenic mouse model of autoimmune thyroiditis. Cell. Immunol, . 229, 79–85[CrossRef][ISI][Medline].
Froloff, N., et al. (1997) On the calculation of binding free energies using continuum methods: application to MHC class I protein-peptide interactions. Protein Sci, . 6, 1293–1301[Abstract].
Godkin, A., et al. (1997) Use of eluted peptide sequence data to identify the binding characteristics of peptides to the insulin-dependent diabetes susceptibility allele HLA-DQ8 (DQ 3.2). Int. Immunol, . 9, 905–911
Godkin, A.J., et al. (1998) Use of complete eluted peptide sequence data from HLA-DR and –DQ Molecules to Predict T Cell epitopes, and the influence of the nonbinding terminal regions of ligands in epitope selection. J. Immunol, . 161, 850–858
Harfouch-Hammoud, E., et al. (1999) Identification of peptides from autoantigens GAD65 and IA-2 that bind to HLA class II molecules predisposing to or protecting from type 1 diabetes. Diabetes, 48, 1937–1947[Abstract].
Janin, J. (1995) Protein–protein recognition. Prog. Biophys. Mol. Biol, . 64, 145–166[CrossRef][ISI][Medline].
Keogh, E., et al. (2001) Identification of new epitopes from four different tumor-associated antigens: recognition of naturally processed epitopes correlates with HLA-A*0201 binding affinity. J. Immunol, . 167, 787–796
Klein, L., et al. (2000) Shaping of the autoreactive T-cell repertoire by a splice variant of self protein expressed in thymic epithelial cells. Nat. Med, . 6, 56–61[CrossRef][ISI][Medline].
Knutson, K.L., et al. (2001) Immunization with a HER-2/neu helper peptide vaccine generates HER-2/neu CD8 T-cell immunity in cancer patients. J. Clin. Invest, . 107, 477–484[CrossRef][ISI][Medline].
Krco, C.J., et al. (2000) Immune response of HLA-DQ transgenic mice to house dust mite allergen p2: identification of HLA-DQ restricted minimal epitopes and critical residues. Clin. Immunol, . 97, 154–161[CrossRef][ISI][Medline].
Krystek, S., et al. (1993) Affinity and specificity of serine endopeptidase-protein inhibitor interactions. Empirical free energy calculations based on X-ray crystallographic structures. J. Mol. Biol, . 234, 661–679[CrossRef][ISI][Medline].
Kwok, W.W., et al. (1989) Mutational analysis of the HLA-DQ3.2 insulin-dependent diabetes mellitus susceptibility gene. Proc. Natl Acad. Sci. USA, 86, 1027–1030
Lee, K.H., et al. (2001) Structure of a human insulin peptide-HLA-DQ8 complex and susceptibility to type 1 diabetes. Nat. Immunol, . 2, 501–507[CrossRef][ISI][Medline].
Li, Y., et al. (2000) Structural basis for the binding of an immunodominant peptide from myelin basic protein in different registers by two HLA-DR2 proteins. J. Mol. Biol, . 304, 177–188[CrossRef][ISI][Medline].
Lopez, J.A., et al. (2001) A synthetic malaria vaccine elicits a potent CD8 (+) and CD4 (+) T lymphocyte immune response in humans. Implications for vaccination strategies. Eur. J. Immunol, . 31, 1989–1998[CrossRef][ISI][Medline].
Mallios, R.R. (2001) Predicting class II MHC/peptide multi-level binding with an iterative stepwise discriminant analysis meta-algorithm. Bioinformatics, 17, 942–948
Middleton, D., et al. (2003) New allele frequency database. Tissue Antigens, 61, 403–407[CrossRef][ISI][Medline].
Moustakas, A.K., et al. (2000) Structure of celiac disease-associated HLA-DQ8 and non-associated HLA-DQ9 alleles in complex with two disease-specific epitopes. Int. Immunol, . 12, 1157–1166
Neeno, T., et al. (1996) HLA-DQ8 transgenic mice lacking endogenous class II molecules respond to house dust allergens: identification of antigenic epitopes. J. Immunol, . 156, 3191–3195[Abstract].
Nepom, G.T. and Kwok, W.W. (1998) Molecular basis for HLA-DQ associations with IDDM. Diabetes, 47, 1177–1184[Abstract].
Novotny, J., et al. (1997) Empirical free energy calculations: a blind test and further improvements to the method. J. Mol. Biol, . 268, 401–411[CrossRef][ISI][Medline].
Rammensee, H., et al. (1999) SYFPEITHI: database for MHC ligands and peptide motifs. Immunogenetics, 50, 213–219[CrossRef][ISI][Medline].
Ranganathan, S. and Tong, J.C. (2005) A practical guide to structure-based prediction of MHC binding peptides. In Flower, D.R. (Ed.). Methods in Molecular Biology: Immunoinformatics: Predicting Immunogenicity in silico, (in press) Humana Press.
Ranganathan, S., et al. (2005) Structural Immunoinformatics: Understanding MHC-peptide binding. In Schoenbach, C. (Ed.), et al. Advances in Experimental Medicine Series—Immunoinformatics: Opportunities and Challenges of Bridging Immunology with Computer and Information Sciences, (in press) Kluwer Plenum Press.
Robles, D.T., et al. (2002) The genetics of autoimmune polyendocrine syndrome type II. Endocrinol. Metab. Clin. North. Am, . 31, 353–368[Medline].
Rognan, D., et al. (1999) Predicting binding affinities of protein ligands from three-dimensional models: application to peptide binding to class I major histocompatibility proteins. J. Med. Chem, . 42, 4650–4658[CrossRef][ISI][Medline].
Sali, A. and Blundell, T.L. (1993) Comparative protein modelling by satisfaction of spatial restraints. J. Mol. Biol, . 234, 779–815[CrossRef][ISI][Medline].
Schapira, M., et al. (1999) Prediction of the binding energy for small molecules, peptides and proteins. J. Mol. Recognit, . 12, 177–190[CrossRef][ISI][Medline].
Seamons, A., et al. (2003) Competition between two MHC binding registers in a single peptide processed from myelin basic protein influences tolerance and susceptibility to autoimmunity. J. Exp. Med, . 197, 1391–1397
Sidney, J., et al. (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
Sollid, L.M. and Thorsby, E. (1993) HLA susceptibility genes in celiac disease: genetic mapping and role in pathogenesis [Erratum (1994) Gastroenterology 106 1133.]. Gastroenterology, 105, 910–922[ISI][Medline].
Southwood, S., et al. (1998) Several common HLA-DR types share largely overlapping peptide binding repertoires. J. Immunol, . 160, 3363–3373
Suri, A., et al. (2005) Natural peptides selected by diabetogenic DQ8 and murine I-Ag7 molecules show common sequence specificity. J. Clin. Invest, . 115, 2268–2276[CrossRef][ISI][Medline].
Tong, J.C., et al. (2004) Modeling the structure of bound peptide ligands to major histocompatibility complex. Protein Sci, . 13, 2523–2532
Tong, J.C., et al. (2006) Modeling the Bound Conformation of Pemphigus Vulgaris-associated Peptides to MHC Class II DR and DQ Alleles. Immunome Res, . 2, 1[CrossRef][Medline].
Wang, R., et al. (2002) Further development and validation of empirical scoring functions for structure-based binding affinity prediction. J. Comput. Aided Mol. Des, . 16, 11–26[CrossRef][ISI][Medline].
Weng, Z., et al. (1996) Prediction of protein complexes using empirical free energy functions. Protein Sci, . 5, 614–626[Abstract].
Zauhar, R.J. and Morgan, R.S. (1985) A new method for computing the macromolecular electric potential. J. Mol. Biol, . 20, 815–820.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  J. C. Tong, T. W. Tan, and S. Ranganathan Methods and protocols for prediction of immunogenic epitopes Brief Bioinform, March 1, 2007; 8(2): 96 - 108. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. C. Tong, T. W. Tan, and S. Ranganathan In silico grouping of peptide/HLA class I complexes using structural interaction characteristics Bioinformatics, January 15, 2007; 23(2): 177 - 183. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||