- Research article
- Open Access
Derivation of an amino acid similarity matrix for peptide:MHC binding and its application as a Bayesian prior
© Kim et al; licensee BioMed Central Ltd. 2009
- Received: 13 May 2009
- Accepted: 30 November 2009
- Published: 30 November 2009
Experts in peptide:MHC binding studies are often able to estimate the impact of a single residue substitution based on a heuristic understanding of amino acid similarity in an experimental context. Our aim is to quantify this measure of similarity to improve peptide:MHC binding prediction methods. This should help compensate for holes and bias in the sequence space coverage of existing peptide binding datasets.
Here, a novel amino acid similarity matrix (PMBEC) is directly derived from the binding affinity data of combinatorial peptide mixtures. Like BLOSUM62, this matrix captures well-known physicochemical properties of amino acid residues. However, PMBEC differs markedly from existing matrices in cases where residue substitution involves a reversal of electrostatic charge. To demonstrate its usefulness, we have developed a new peptide:MHC class I binding prediction method, using the matrix as a Bayesian prior. We show that the new method can compensate for missing information on specific residues in the training data. We also carried out a large-scale benchmark, and its results indicate that prediction performance of the new method is comparable to that of the best neural network based approaches for peptide:MHC class I binding.
A novel amino acid similarity matrix has been derived for peptide:MHC binding interactions. One prominent feature of the matrix is that it disfavors substitution of residues with opposite charges. Given that the matrix was derived from experimentally determined peptide:MHC binding affinity measurements, this feature is likely shared by all peptide:protein interactions. In addition, we have demonstrated the usefulness of the matrix as a Bayesian prior in an improved scoring-matrix based peptide:MHC class I prediction method. A software implementation of the method is available at: http://www.mhc-pathway.net/smmpmbec.
- Major Histocompatibility Complex
- Major Histocompatibility Complex Class
- Prediction Performance
- Major Histocompatibility Complex Molecule
- Amino Acid Similarity
Amino acid similarity matrices define a quantitative measure of likeness between each of the 20 canonical amino acids. They are utilized throughout computational biology in areas such as phylogenetics, protein structure modeling, and prediction of protein ligand interactions. Depending on the application, different measures of similarity are appropriate. For example, the commonly used PAM and BLOSUM matrices [1, 2] have been built based on the frequencies of amino acid substitutions observed in aligned protein sequences. This measure, routinely used in programs such as BLAST , represents both evolutionary and functional similarity between amino acids.
Our group has been interested in amino acid similarity in the context of peptides binding to proteins. Given binding data for several peptide ligands, the challenge is to predict the affinity of any peptide of arbitrary sequence. Our specific interest is in peptide binding to proteins involved in antigen processing and presentation, such as the TAP transporter [4, 5] and MHC molecules. In recent large-scale benchmark studies, the best performing prediction method for peptide:MHC class I binding is the NetMHC artificial neural network, outperforming linear methods such as SMM [6, 7]. NetMHC is trained using a BLOSUM matrix based encoding of peptide sequences [8–13]. This provides the neural network with information on amino acid similarity, and allows it to predict the impact of residues on binding that are not represented in the training set.
In this study, we tested the hypothesis that amino acid similarity in the context of peptide binding to MHC molecules is distinct from previously defined metrics. We further examined if this similarity measure can be used to improve peptide:MHC binding predictions, and if incorporating it into the SMM approach can close the gap in prediction quality to NetMHC.
Combinatorial peptide library binding affinity data
A library of combinatorial peptide mixtures was used to measure the binding affinity contribution of each residue in a 9-mer peptide to an MHC molecule. The library contains mixtures of 9-mer peptides all sharing the same residue type at one position, while the remaining positions are allowed to sample all residue types. For instance, 'XAXXXXXXX' represents a mixture of peptides with an Alanine at position P2, and any one of the possible residues at the remaining positions. A total of 180 mixtures covering 20 residue types in all positions of a 9-mer peptide were synthesized, and tested for binding to 24 MHC class I molecules listed in the methods section. Thus, a total of 180 × 24 = 4320 binding affinities in terms of IC50 values were measured (Additional file 1: Dataset S1). We then transformed these values to approximate a relative binding energy contribution of an amino acid aa at peptide position pos for a given MHC molecule:
Building the peptide:MHC binding energy covariance (PMBEC) matrix
Comparing the PMBEC matrix to previously established measures of amino acid similarity
Numerous amino acid similarity matrices have been published to date. Some of more prominent matrices such as BLOSUM series have been, in fact, used in peptide:MHC binding predictions to represent peptide sequences, despite fundamental differences in context of their use . Apparent success of their use indicates that it is worthwhile to investigate how similar the PMBEC matrix is to these matrices.
Ten most similar matrices to PMBEC out of all complete matrices from the AA Index Database.
Johnson & Overington 
Koshi & Goldstein 
Overington et al. 
Benner et al. 
Vogt et al. 
Gonnet et al. 
Another observation one can make about the matrices in the table is that the PMBEC matrix falls outside the cluster of matrices that most resembles it. Specifically, in contrast to a correlation coefficient of 0.64 between PMBEC and BLOSUM50, seven out of ten matrices in the table have more than 0.93 with respect to BLOSUM50. This further adds to evidence that the PMBEC matrix is novel.
In the following sections, a novel peptide:MHC binding prediction method SMMPMBEC is introduced, and its benefits are illustrated. SMMPMBEC improves upon the original SMM approach  by using PMBEC as a Bayesian prior. The details of how a Bayesian prior used in the SMM approach is described in the Methods section.
SMMPMBEC can compensate for the absence of important subset of binding data
Peptide binding affinity measurements available for MHC class I molecules can vary from less than a hundred to thousands for each allele. Furthermore, peptide sequence space coverage in the binding data can be limited even for alleles with large data sets, due to biases when selecting peptides for binding measurements. Here, we examined how well SMMPMBEC addresses the problem of missing data. We then compared its performance to that of SMMBLOSUM, which uses BLOSUM62 as a prior.
We then derived a subset from the original binding data for HLA A*3101 such that all peptides with an Alanine at P1 were excluded. When SMM was trained on this subset, its scoring matrix gave a value of zero at P1 for residue Alanine since SMM had no information on it available from the training data. When training SMMPMBEC on this subset, its scoring matrix entry for Alanine at P1 gave a non-zero value. This step was repeated for the remaining 19 residues, and their corresponding scoring matrix entry values are shown as gray bars in Figure 4. In the figure, three out of four residues with the most favorable binding energy contribution - Arginine, Lysine, Histidine (R, K, and H) - are shared between SMM and SMMPMBEC. Similarly, three out of four residues with the least favorable binding energy contribution - Glutamic acid, Aspartic acid, and Proline (E, D, and P) - are shared between the two methods. The binding energy profiles of SMM and SMMPMBEC had a Pearson's correlation value of 0.92. Thus, these observations indicate that SMMPMBEC can infer binding energies of missing residues from those present in the binding data.
When we repeated this exercise for SMMBLOSUM, resulting in a binding energy profile shown in blue in Figure 4, we saw that SMMBLOSUM can also infer binding energies of missing residues (r = 0.69). However, approximated binding energies for Lysine (K) and Histidine (H) are sharply different between SMMPMBEC and SMMBLOSUM. The obvious benefit of using SMMPMBEC for HLA A*3101 binding data set is by no means true for all other data sets. However, this example highlights what can be achieved with a properly tuned amino acid similarity matrix.
SMMPMBEC consistently outperforms SMM over a wide range of training data set sizes
Application of SMMPMBEC in peptide:MHC-I binding predictions
Prediction performances as measured by AUC values of ROC curves.
When looking at these measured prediction performances in Table 2, some precautions are in order. It has been observed earlier that accuracy of a predictor tends to increase with more training data . Likewise, reliability of a predictor's accuracy also increases with more training data, because larger peptide sequence space is sampled. Thus, apparently high prediction accuracy of a predictor for Patr B*0101, for instance, should not be taken at its face value, because the predictor was trained on relatively small training data (132 data points).
With these precautions in mind, as shown in Table 2, SMMPMBEC outperformed SMM in 39 out of 46 cases, and the average performance of SMMPMBEC was higher than SMM (0.894 vs. 0.887 AUC values). The small but consistent improvements are statistically significant with p-value of 0.001 using Student's t-test (paired two tailed; the test assumes that AUC values over data sets follow a normal distribution). The same comparison for SMMPMBEC and NetMHC indicated that their difference of average performances was not statistically significant (p-value = 0.47). Thus, the use of PMBEC as a Bayesian prior improves performance for SMMPMBEC, yielding a performance that is overall comparable to that of NetMHC. Given that NetMHC uses neural networks to represent a model of peptide:MHC binding specificity, it is noteworthy that the use of a simple scoring matrix coupled with a Bayesian prior can match its prediction performance.
A closer look at Table 2 reveals that the most significant performance improvements of SMMPMBEC over SMM tend to come from alleles with smaller data sets (e.g. HLA A*2301, HLA B*4501). This is in agreement with what has been observed earlier that at smaller data size, the benefits of PMBEC as a Bayesian prior compensating for missing binding data are more apparent.
We have also carried out performance evaluations for SMMBLOSUM (Additional file 3: Dataset S3). The difference between averaged prediction performances of SMMBLOSUM and SMMPMBEC was 0.0003, which is not statistically significant (p-value = 0.83). We attribute their similar performances mainly to a feature of SMM approach where prior is relied on less as more binding data become available for training. In fact, when prediction performance differences of SMMPMBEC and SMMBLOSUM are plotted against training data set sizes, a clear pattern emerges where larger training data sizes correlate with smaller performance differences (data not shown). About 80% of data sets are contained within ± 0.01 AUC of zero performance difference.
Of the remaining 20% of the data sets, we see that three MHCs with the highest performance differences favoring PMBEC have binding motifs with strong preference for Glutamic acid at the anchor residue position. (Of these three MHCs, HLA B*4403, followed by B*4002 and H-2 KK, has the largest AUC difference of 0.037.) This latter observation supports one of the main arguments made in the present work that amino acid similarities involving Glutamic acid are the most prominent difference between PMBEC and BLOSUM. This in turn suggests that those MHC's with Glutamic acid as anchor residues is where SMMPMBEC's prediction performances will be superior to those of SMMBLOSUM.
We have derived a novel amino acid similarity matrix (PMBEC) for peptide:MHC class I binding. Rather than relying on sequence alignments, the matrix was derived from experimentally measured binding affinities of combinatorial peptide mixtures. The use of combinatorial peptide mixtures afforded us an unbiased sampling of peptide sequence space. In total, a panel of 24 MHC class I molecules was probed, corresponding to 4320 binding affinity measurements of individual residues in diverse contexts of both peptide ligand and receptor molecule. This approach can, therefore, directly evaluate amino acid similarities in the context of peptide:protein binding.
Once the PMBEC matrix was derived, its comparisons with all complete 135 amino acid similarity matrices taken from the AA Index Database have shown that the matrix is different from previously established amino acid similarity matrices. To determine where PMBEC most differs, detailed comparisons with BLOSUM62 (representative of those matrices most similar to PMBEC) have revealed that PMBEC considers pairs of amino acids with opposite charges to be very dissimilar, while BLOSUM62 considers them a neutral exchange. These differences are most likely due to different molecular consequences of substituting an amino acid in a protein, and in a peptide bound to MHC. Specifically, in protein sequences, most charged residues are on the surface, and a reversal of charge can often be tolerated well, as it preserves hydrophilicity. For peptide ligands, on the other hand, a reversal of charge is likely going to adversely affect their binding affinities.
Because of these peptide:protein specific features, we expect the PMBEC matrix will be useful in modelling peptide similarities in the context of immune recognition. In addition, we also expect that the matrix will be of use to those studying other types of peptide:protein interactions (e.g. SH3, PDZ, and WW protein domains). In hindsight, we are surprised that a novel amino acid similarity matrix such as PMBEC can still be discovered. Our work further underscores the importance of application-specific amino acid similarity matrices in computational biology, since molecular context determines which matrices are more meaningful.
In the context of predicting MHC class I binding peptides, we have shown that SMMPMBEC - which uses PMBEC as a Bayesian prior - has a significantly better prediction performance than SMM. For those alleles with small data sets, performance improvements were more apparent, indicating that the use of PMBEC as a Bayesian prior is likely compensating for inadequate peptide sequence coverage. We have also shown that SMMPMBEC's ability to infer 'true' binding energy contributions of intentionally excluded residues from others present in the binding data is limited. This is probably due to subtle differences of MHC molecule binding specificities that have so far not been appreciated. For comparison, a Bayesian prior based on BLOSUM62 also displays this property.
We have also shown that the average performance difference of SMMPMBEC and NetMHC (the best performing method for peptide:MHC class I binding prediction according to recent benchmarks) is not statistically significant. Our results indicate that, at least for peptide:MHC class I binding predictions, scoring matrices can provide competitive prediction performances. This close prediction performances between a non-linear model and a linear one suggests that the limit of a linear model has been reached. From a practical standpoint, the key advantages of using a scoring matrix are that the model is easy to understand, interpret, and communicate.
Lastly, although the use of PMBEC improves prediction performance, the simplicity of the approach comes with a number of inherent limitations. One notable limitation is that the same set of amino acid similarity rules encoded in PMBEC is used for all peptide positions. Since there are position dependent influences observed for peptide:MHC binding motifs, such information may be used in the near future with more experimental data.
PMBEC is a novel amino acid similarity matrix derived for peptide:MHC class I binding. One prominent feature of the matrix is that it disfavors substitutions of amino acids with opposite charges. This is likely a general feature of peptide:protein interactions. We have also demonstrated the usefulness of PMBEC in the context of peptide:MHC class I binding predictions, by using it as a Bayesian prior in a new prediction method SMMPMBEC. Results from a large-scale benchmark indicate that its prediction performance rivals that of one of the best performing methods in the field.
Positional scanning combinatorial peptide library and peptide synthesis
The combinatorial library was synthesized as previously described . Each pool in the library contains 9-mer peptides with one fixed residue at a single position. With each of the 20 naturally occurring residues represented at each position along the 9-mer backbone, the entire library consisted of 180 peptide mixtures. Peptides utilized in screening studies were synthesized as described elsewhere , or purchased as crude material from Mimotopes (Minneapolis, MN/Clayton, Victoria, Australia), Pepscan Systems B.V. (Lelystad, Netherland) or A and A Labs (San Diego, CA). Peptides synthesized for use as radiolabeled ligands were synthesized by A and A Labs and purified to >95% homogeneity by reverse phase HPLC. Purity of these peptides was determined using analytical reverse-phase HPLC and amino acid analysis, sequencing, and/or mass spectrometry. Peptides were radiolabeled with the chloramine T method . Lyophilized peptides were re-suspended at 4-20 mg/ml in 100% DMSO, then diluted to required concentrations in PBS +0.05% (v/v) nonidet P40 (Fluka Biochemika, Buchs, Switzerland).
Target major histocompatibility complex (MHC) molecules for the generation of binding data using combinatorial peptide mixtures
For each one of the 24 MHC molecules, a scoring matrix was generated. The target MHC molecules came from four organisms: human (A*0201, A*6802, A*3201, A*3001, B*5802, B*5801, B*5401, B*5301, B*5101, B*3501, B*2705, B*1503, B*1501, B*0801, and B*0702); mouse (H-2 Kk, H-2 Kd, H-2 Dd, and H-2 Db); chimpanzee (Patr A*0401 and Patr A*0301); and macaque (Mamu B*08, Mamu B*03, and Mamu B*01). Binding data for human and mouse have been published in ; those for chimpanzee in ; and those for macaque have been submitted.
MHC purification and peptide binding assays
MHC purification and quantitative binding assays based on the inhibition of binding of a high affinity radiolabeled ligand were performed essentially as described elsewhere [18–21]. In competition assays, each mixture or individual peptide was 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 utilized, where the concentration of the labelled ligand is less than that of MHC molecule and IC50 is greater than the concentration of MHC molecule, the measured IC50 values are reasonable approximations of dissociation constant, KD.
Comparing amino acid similarity matrices from different sources
where H = I - J/n, I is an identity matrix, J is a matrix composed of ones, and n is 20. Centering of matrices reduces the influence of data source dependent expected probabilities. Following this, Pearson's correlation coefficients of matrices with respect to PMBEC were calculated. Out of 210 unique elements for each symmetric 20 × 20 matrix, 190 non-diagonal elements were considered for the calculation of correlation since we are interested in relationships involving two different residues.
Application of PMBEC as a Bayesian prior
We have previously established the SMM prediction method , which models the peptide binding specificity for a given MHC allele as a scoring matrix. Each matrix entry corresponds to the predicted binding energy contribution of a residue in a fixed length peptide. The matrix is determined by minimizing the difference between predicted and measured binding affinities in combination with a regularization term, which serves to push any entry in the scoring matrix toward zero for which no clear contribution to binding can be determined from the training data. This approach has been shown to vastly outperform other scoring matrix based predictions such as ARB and Rankpep [6, 7], but it is outperformed by NetMHC, which uses artificial neural networks and amino acid similarity information.
Motivated by this observation, we reasoned that prediction performances of SMM can also be improved by incorporating amino acid similarity information. We reframed SMM from a Bayesian viewpoint, and assigned prior probabilities to scoring matrices by assuming that their matrix entries follow the multivariate normal distribution defined by the covariance values in PMBEC. The minimization of differences between measured and predicted affinities corresponds to maximizing the likelihood of observing the measured binding affinities for a given scoring matrix. It can be shown that the regularization term in the original SMM approach corresponds to the use of a Bayesian prior that assumes a multivariate normal distribution for the scoring matrix but assumes no correlation between different amino acids. In essence, the use of this Bayesian prior now favors scoring matrices that reflect amino acid similarity information encoded in PMBEC. Details of the derivation are shown in the following section.
From here on, SMM refers to the original version of the method, where an identity matrix is used in a prior. Likewise, SMMPMBEC and SMMBLOSUM use PMBEC and BLOSUM62 matrices in their priors, respectively.
Implementation of Bayesian prior into SMM
To build a model of binding specificity of an MHC molecule for 9-mer peptides, a scoring matrix is trained, given N binding affinity measurements, b m (Nx1). The scoring matrix is represented as a vector including an offset variable, yielding a total of 20 × 9 + 1 = 181 rows: w(181x1). Peptide sequences are represented using a sparse encoding scheme, where a binary vector of length 20 is associated with each residue position. Thus, N 9-mer peptides can be represented by a matrix, H(Nx181). Binding predictions are carried out by a matrix multiplication H and w, resulting in a vector of predicted binding affinities for the peptides.
Here, the binding data was split into five parts where one part is labelled 'blind' and the remaining 'training'. The model w was generated based on the 'training' set and was used to make predictions for the 'blind' set. The cross-validated distance uses a sum of squared errors as the norm.
To generate a robust model, the process of splitting the binding data, optimizing λ, and generating w is repeated 10 times, yielding 10 models. The final model returned by the prediction method is an average of these 10 models.
Evaluation of the prediction methods
Five-fold cross validation was used to measure prediction performances of methods. Here, a binding dataset is divided into 5 equal subsets, where one subset is labelled a testing set and the remaining 4/5 subsets a training set. A model was generated based on the training set only and used to make predictions for the testing set. This step was repeated 4 more times by rotating around the testing set label, resulting in blind predictions for the full dataset in the end. Performance was measured by calculating an Area-Under-Curve (AUC) of Receiver-Operating-Characteristic (ROC) curve of the blind predictions with respect to corresponding measured binding affinities. An AUC value has a range from 0.5 to 1.0, where it can be interpreted as the probability of distinguishing binders from non-binders if they were picked randomly. Thus, a value of 1.0 indicates a perfect prediction where as that of 0.5 indicates random.
This work was supported by the National Institutes of Health contract HHSN26620040006C, and PSWRCE grant U54 AI065359 to BP. This is Kyowa Hakko Kirin publication number 1104.
- Henikoff S, Henikoff JG: Amino Acid Substitution Matrices from Protein Blocks. Proc Natl Acad Sci USA 1992, 89(22):10915–10919. 10.1073/pnas.89.22.10915PubMed CentralView ArticlePubMedGoogle Scholar
- Dayhoff MO, Schwartz RM, Orcutt BC: A model of evolutionary change in proteins. Atlas of Protein Sequence and Structure 1978, 345–352.Google Scholar
- Altschul SF, Madden TL, Schaffer AA, Zhang J, Zhang Z, Miller W, Lipman DJ: Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucl Acids Res 1997, 25(17):3389–3402. 10.1093/nar/25.17.3389PubMed CentralView ArticlePubMedGoogle Scholar
- Peters B, Bulik S, Tampe R, van Endert PM, Holzhutter H-G: Identifying MHC Class I Epitopes by Predicting the TAP Transport Efficiency of Epitope Precursors. J Immunol 2003, 171(4):1741–1749.View ArticlePubMedGoogle Scholar
- Burgevin A, Saveanu L, Kim Y, Barilleau E, Kotturi M, Sette A, van Endert P, Peters B: A Detailed Analysis of the Murine TAP Transporter Substrate Specificity. PLoS ONE 2008, 3(6):e2402. 10.1371/journal.pone.0002402PubMed CentralView ArticlePubMedGoogle Scholar
- Lin H, Ray S, Tongchusak S, Reinherz E, Brusic V: Evaluation of MHC class I peptide binding prediction servers: Applications for vaccine research. BMC Immunology 2008, 9(1):8. 10.1186/1471-2172-9-8PubMed CentralView ArticlePubMedGoogle Scholar
- Peters B, Bui H-H, Frankild S, Nielsen M, Lundegaard C, Kostem E, Basch D, Lamberth K, Harndahl M, Fleri W, et al.: A Community Resource Benchmarking Predictions of Peptide Binding to MHC-I Molecules. PLoS Comput Biol 2006, 2(6):e65. 10.1371/journal.pcbi.0020065PubMed CentralView ArticlePubMedGoogle Scholar
- Nielsen M, Lundegaard C, Worning P, Lauemoller SL, Lamberth K, Buus S, Brunak S, Lund O: Reliable prediction of T-cell epitopes using neural networks with novel sequence representations. Protein Sci 2003, 12(5):1007–1017. 10.1110/ps.0239403PubMed CentralView ArticlePubMedGoogle Scholar
- Buus S, Lauemoller SL, Worning P, Kesmir C, Frimurer T, Corbet S, Fomsgaard A, Hilden J, Holm A, Brunak S: Sensitive quantitative predictions of peptide-MHC binding by a 'Query by Committee' artificial neural network approach. Tissue Antigens 2003, 62(5):378–384. 10.1034/j.1399-0039.2003.00112.xView ArticlePubMedGoogle Scholar
- Lundegaard C, Lamberth K, Harndahl M, Buus S, Lund O, Nielsen M: NetMHC-3.0: accurate web accessible predictions of human, mouse and monkey MHC class I affinities for peptides of length 8–11. Nucl Acids Res 2008, 36(suppl_2):W509–512. 10.1093/nar/gkn202PubMed CentralView ArticlePubMedGoogle Scholar
- Lundegaard C, Lund O, Nielsen M: Accurate approximation method for prediction of class I MHC affinities for peptides of length 8, 10 and 11 using prediction tools trained on 9mers. Bioinformatics 2008, 24(11):1397–1398. 10.1093/bioinformatics/btn128View ArticlePubMedGoogle Scholar
- Lundegaard C, Nielsen M, Lund O: The validity of predicted T-cell epitopes. Trends in Biotechnology 2006, 24(12):537–538. 10.1016/j.tibtech.2006.10.001View ArticlePubMedGoogle Scholar
- Nielsen M, Lundegaard C, Worning P, Lauemoller SL, Lamberth K, Buus S, Brunak S, Lund O: Reliable prediction of T-cell epitopes using neural networks with novel sequence representations. Protein Sci 2003, 12(5):1007–1017. 10.1110/ps.0239403PubMed CentralView ArticlePubMedGoogle Scholar
- Kawashima S, Kanehisa M: AAindex: Amino Acid index database. Nucl Acids Res 2000, 28(1):374. 10.1093/nar/28.1.374PubMed CentralView ArticlePubMedGoogle Scholar
- Peters B, Sette A: Generating quantitative models describing the sequence specificity of biological processes with the stabilized matrix method. BMC Bioinformatics 2005, 6(1):132. 10.1186/1471-2105-6-132PubMed CentralView ArticlePubMedGoogle Scholar
- Pinilla C, Appel J, Blanc P, Houghten R: Rapid identification of high affinity peptide ligands using positional scanning synthetic peptide combinatorial libraries. Biotechniques 1992, 13(6):901–905.PubMedGoogle Scholar
- Sidney J, Southwood S, Mann DL, Fernandez-Vina MA, Newman MJ, Sette A: Majority of peptides binding HLA-A*0201 with high affinity crossreact with other A2-supertype molecules. Hum Immunol 2001, 62: 1200–1216. 10.1016/S0198-8859(01)00319-6View ArticlePubMedGoogle Scholar
- Sidney J, Southwood S, Oseroff C, Del Guercio MF, Sette A, Grey H: Measurement of MHC/Peptide Interactions by Gel Filtration. Curr Protocols Immunol 1998, 18.13.11–18.13.19.Google Scholar
- Sidney J, Assarsson E, Moore C, Ngo S, Pinilla C, Sette A, Peters B: Quantitative peptide binding motifs for 19 human and mouse MHC class I molecules derived using positional scanning combinatorial peptide libraries. Immunome Research 2008, 4(1):2. 10.1186/1745-7580-4-2PubMed CentralView ArticlePubMedGoogle Scholar
- Sidney J, Peters B, Moore C, Pencille TJ, Ngo S, Masterman K-A, Asabe S, Pinilla C, Chisari FV, Sette A: Characterization of the peptide-binding specificity of the chimpanzee class I alleles A*0301 and A*0401 using a combinatorial peptide library. Immunogenetics 2007, 59: 745–751. 10.1007/s00251-007-0243-5View ArticlePubMedGoogle Scholar
- Sidney J, Southwood S, Sette A: Classification of A1- and A24- supertype molecules by analysis of their MHC-peptide binding repertoires. Immunogenetics 2005, 57: 393–408. 10.1007/s00251-005-0004-2View ArticlePubMedGoogle Scholar
- Pokarowski P, Kloczkowski A, Nowakowski S, Pokarowska M, Jernigan RL, Kolinski A: Ideal amino acid exchange forms for approximating substitution matrices. Proteins 2007, 69(2):379–393. 10.1002/prot.21509View ArticlePubMedGoogle Scholar
- Kann M, Qian B, Goldstein RA: Optimization of a new score function for the detection of remote homologs. Proteins: Structure, Function, and Genetics 2000, 41(4):498–503. 10.1002/1097-0134(20001201)41:4<498::AID-PROT70>3.0.CO;2-3View ArticleGoogle Scholar
- Johnson MS, Overington JP: A Structural Basis for Sequence Comparisons. An Evaluation of Scoring Methodologies. J Mol Biol 1993, 233: 716–738. 10.1006/jmbi.1993.1548View ArticlePubMedGoogle Scholar
- Koshi JM, Goldstein RA: Context-dependent optimal substitution matrices. Protein Eng 1995, 8: 641–645. 10.1093/peds/8.7.641View ArticlePubMedGoogle Scholar
- Overington J, Donnelly D, Johnson MS, Sali A, Blundell TL: Environment-specific amino acid substitution tables: Tertiary templates and prediction of protein folds. Protein Sci 1992, 1(2):216–226. 10.1002/pro.5560010203PubMed CentralView ArticlePubMedGoogle Scholar
- Bennet SA, Cohen MA, Gonnet GH: Amino acid substitution during functionally constrained divergent evolution of protein sequences. Protein Eng 1994, 7(11):1323–1332. 10.1093/protein/7.11.1323View ArticleGoogle Scholar
- Vogt G, Etzold T, Argos P: An Assessment of Amino Acid Exchange Matrices in Aligning Protein Sequences: The Twilight Zone Revisited. J Mol Biol 1995, 249(4):816–831. 10.1006/jmbi.1995.0340View ArticlePubMedGoogle Scholar
- Gonnet GH, Cohen MA, Bennet SA: Exhaustive matching of the entire protein sequence database. Science 1992, 256: 1443–1445. 10.1126/science.1604319View ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.